The role of frequency in children's learning of morphological constructions
1211
2013
978-3-8233-7840-2
978-3-8233-6840-3
Gunter Narr Verlag
Anne-Kristin Cordes
According to usage-based, constructionist accounts the linguistic input in general and input frequencies in particular play an important role in children's language learning. English-speaking children have been shown to be able to learn entirely novel, invented word order constructions from their input. This book aims to extend this line of research to the area of morphology. Two experimental studies investigate German-speaking and English-speaking children's ability to learn novel morphological constructions from the input. The effects of input frequencies on this learning process are examined in detail. A corpus study provides the morphological background data for the invented constructions and presents additional support for frequency effects from naturalistic language learning. By combining two empirical methods and by exploring morphological learning in two different languages this book provides new insights into the cognitive processes that are assumed to be involved in children's language learning and reveals how these processes are affected by different kinds of input frequencies.
According to usage-based, constructionist accounts the linguistic input in general and input frequencies in particular play an important role in children's language learning. English-speaking children have been shown to be able to learn entirely novel, invented word order constructions from their input. This book aims to extend this line of research to the area of morphology. Two experimental studies investigate German-speaking and English-speaking children's ability to learn novel morphological constructions from the input. The effects of input frequencies on this learning process are examined in detail. A corpus study provides the morphological background data for the invented constructions and presents additional support for frequency effects from naturalistic language learning. By combining two empirical methods and by exploring morphological learning in two different languages this book provides new insights into the cognitive processes that are assumed to be involved in children's language learning and reveals how these processes are affected by different kinds of input frequencies.
<?page no="0"?> Anne-Kristin Cordes The role of frequency in children’s learning of morphological constructions Language in Performance LiP <?page no="1"?> The role of frequency in children’s learning of morphological constructions <?page no="2"?> 48 Edited by Werner Hüllen (†) and Rainer Schulze Advisory Board: Thomas Herbst (Erlangen), Andreas Jucker (Zürich), Manfred Krug (Bamberg), Christian Mair (Freiburg i.Br.), Ute Römer (Atlanta, GA, USA), Andrea Sand (Trier), Hans-Jörg Schmid (München), Josef Schmied (Chemnitz) and Edgar W. Schneider (Regensburg) <?page no="3"?> Anne-Kristin Cordes The role of frequency in children’s learning of morphological constructions <?page no="4"?> Bibliografische Information der Deutschen Nationalbibliothek Die Deutsche Nationalbibliothek verzeichnet diese Publikation in der Deutschen Nationalbibliografie; detaillierte bibliografische Daten sind im Internet über http: / / dnb.dnb.de abrufbar. Diese Arbeit wurde im Jahr 2012 als Dissertation an der Ludwig-Maximilians- Universität München angenommen. Gedruckt mit Unterstützung des Förderungs- und Beihilfefonds Wissenschaft der VG Wort. © 2014 · Narr Francke Attempto Verlag GmbH + Co. KG Dischingerweg 5 · D-72070 Tübingen Das Werk einschließlich aller seiner Teile ist urheberrechtlich geschützt. Jede Verwertung außerhalb der engen Grenzen des Urheberrechtsgesetzes ist ohne Zustimmung des Verlages unzulässig und strafbar. Das gilt insbesondere für Vervielfältigungen, Übersetzungen, Mikroverfilmungen und die Einspeicherung und Verarbeitung in elektronischen Systemen. Gedruckt auf säurefreiem und alterungsbeständigem Werkdruckpapier. Internet: www.narr.de E-Mail: info@narr.de Druck und Bindung: Hubert & Co, Göttingen Printed in Germany ISSN 0939-9399 ISBN 978-3-8233-6840-3 <?page no="5"?> v Contents List of figures ix List of tables xi Acknowledgements xiii 1 Introduction and outline 1 2 Language learning from the input? 8 2.1 Contrasting key issues 8 2.1.1 The issue of human uniqueness 8 2.1.2 Language knowledge in the child and in the adult 11 2.1.3 The role of the input 17 2.2 The poverty-of-the-stimulus argument versus learning from the input 18 2.2.1 Input quality 18 2.2.2 Input quantity 20 2.2.3 Learning from the input 24 2.3 Conclusion 33 3 The learning process 36 3.1 Starting point: Categorization and analogy in language learning 36 3.2 Categorization—The prototype account 38 3.2.1 Concept and process 38 3.2.2 The development of linguistic categories 44 3.2.3 Summary 46 3.3 Analogies 46 3.3.1 Concept and process 46 3.3.2 The development of analogizing abilities 50 3.3.3 The role of similarity 51 3.3.4 Summary 54 3.4 Starting point revisited: Categorization and analogy in language learning 54 3.4.1 Comparison of categorization and analogy 54 3.4.2 Redefinitions of categorization and analogy in construction learning 57 3.4.3 Steps of construction learning 60 3.4.4 Analogy and categorization in construction learning 61 <?page no="6"?> vi 4 Frequency effects in language learning 66 4.1 Frequency in language representation and language learning 66 4.2 Frequency and concrete constructions 69 4.3 Frequency and partially-filled constructions 70 4.3.1 Morphological constructions 71 4.3.2 Other partially-filled constructions 77 4.4 Frequency and abstract constructions 79 4.4.1 Abstract constructions 79 4.4.2 Complex constructions 82 4.4.3 Typological features 83 4.5 Summary of previous research 84 4.6 Introduction of the pattern level 85 4.7 Major research questions 87 5 Corpus study: Frequency of the derivational verb prefix pattern in caregiver and child speech in German and English 92 5.1 Background 93 5.1.1 Motivation 93 5.1.2 Verb prefixes in German and English 95 5.1.3 The development of derivational morphology in children 101 5.1.4 The operationalization of pattern frequency 104 5.1.5 Hypotheses for the corpus study 105 5.2 Corpus analysis 107 5.2.1 Database 107 5.2.2 Procedure and results 108 5.3 Discussion 115 6 Experiment 1: The role of input frequencies at constructional and pattern levels in novel morphology learning 119 6.1 Background and hypotheses 119 6.2 Method 123 6.2.1 Participants 123 6.2.2 Material 124 6.2.3 Procedure 128 6.3 Results 133 6.3.1 Statistical techniques used to analyze the data 133 6.3.2 Learning of a novel verb prefix construction in German and English 135 6.3.3 Learning of a novel reduplication construction in German and English 148 6.3.4 Learning of the novel prefix versus the novel reduplication construction 158 <?page no="7"?> vii 6.3.5 Type frequency effects 160 6.4 Discussion 162 6.4.1 Summary 162 6.4.2 Learning and age 163 6.4.3 Pattern effects 166 6.4.4 Token frequency effects at the constructional level 171 6.4.5 Type frequency effects at the constructional level 172 6.4.6 Conclusion 173 7 Experiment 2: The role of the shape of the input distribution in the learning of a novel partially-filled construction 175 7.1 Background 175 7.2 Method 178 7.2.1 Participants 178 7.2.2 Materials and procedure 178 7.3 Results 181 7.4 Discussion 184 8 Summary and general discussion 189 8.1 Summary of the work 189 8.2 Frequency effects in morphological learning 193 8.2.1 Construction learning based on the input 194 8.2.2 Effects of token frequency, type frequency and typetoken ratio at the constructional level 195 8.2.3 Effects of type and token frequency at the pattern level 198 8.2.4 Summary 202 8.3 Broader theoretical implications 203 8.3.1 Abstractions in language 203 8.3.2 The domain-generality of cognitive processes in language learning 205 8.3.3 The role of salience 207 8.4 Open issues and future directions 209 8.4.1 Inflectional versus derivational morphological constructions 209 8.4.2 Issues regarding the corpus analysis 211 8.4.3 Future research directions 213 8.5 Conclusion 215 References 217 <?page no="8"?> viii APPENDIX I Chapter 5 235 I.1 Excerpt from Leo corpus (4; 5; 3) 235 I.2 Excerpt from Thomas corpus (4; 3; 2) 235 I.3 Lists of manually selected verb prefix constructions and prefix verbs in English and German 236 I.4 Search results for manually selected verb prefix constructions in English and German 245 II Chapter 6 254 II.1 One-sample t-tests assessing novel construction learning 254 II.2 Proportions of correct responses on real items by task, pattern and language 261 II.3 Memorization and generalization of the novel construction by task and pattern 263 II.4 Novel prefix and reduplication learning in Germanspeaking adults 265 <?page no="9"?> ix List of figures Figure 1. A selection of vessels used in Labov’s study (further pictures without handles, with content etc. were also used; taken from Labov 1973: 354). 40 Figure 2. The categorization process. 44 Figure 3. Major components of analogical reasoning (figure taken from Holyoak 2012: 236). 48 Figure 4. The role of supporting object similarity in analogy formation. 53 Figure 5. Illustration of the categorization of a retriever and an Alsatian as dogs. 56 Figure 6. Illustration of the analogy “an atom is like the solar system”. 56 Figure 7. Analogy between a retriever and an Alsatian and categorization of an atom and the solar system as examples of physical models involving orbital structure. 57 Figure 8. Depiction of the pivot schema Let’s [X] as transparencies stacked on top of each other. 62 Figure 9. Depiction of the pivot schema Let’s [X] as a constructional category. 62 Figure 10. Comparisons of stored examples resulting in schematization. 64 Figure 11. The formation of paradigmatic categories. 65 Figure 12. Schematic depiction of the proposed constructional levels. 87 Figure 13. Constructional levels and their interrelations. 126 Figure 14. Experimental token frequencies of verbs used in the novel construction in the different phases of the experiment. 130 Figure 15. Act-out task (German, prefix). Proportions of correct responses by age group. 136 Figure 16. Act-out task (English, prefix). Proportions of correct responses by age group. 138 Figure 17. Forced-choice task (German, prefix). Proportions of correct responses on pretend items by age group. 139 Figure 18. Forced-choice task (English, prefix). Proportions of correct responses on pretend items by age group. 141 Figure 19. Production task (German, prefix). Proportions of correct responses on pretend items by age group. 143 <?page no="10"?> x Figure 20. Production task (English, prefix). Proportions of correct responses on pretend items by age group. 145 Figure 21. Act-out task (German, reduplication). Proportions of correct responses by age group. 149 Figure 22. Act-out task (English, reduplication). Proportions of correct responses by age group. 150 Figure 23. Forced-choice task (German, reduplication). Proportions of correct responses on pretend items. 151 Figure 24. Forced-choice task (English, reduplication). Proportions of correct responses on pretend items. 152 Figure 25. Production task (German, reduplication). Proportions of correct responses on pretend items by age group. 154 Figure 26. Production task (English, reduplication). Proportions of correct responses on pretend items by age group. 155 Figure 27. Proportions of correct responses as a function of condition and construction type. 182 Figure a. Forced-choice task (German, prefix, reduplication). Proportions of correct responses on real items by age group. 261 Figure b. Forced-choice task (English, prefix, reduplication). Proportions of correct responses on real items by age group. 261 Figure c. Production task (German, prefix, reduplication). Proportions of correct responses on real items by age group. 262 Figure d. Production task (English, prefix, reduplication). Proportions of correct responses on real items by age group. 262 <?page no="11"?> xi List of tables Table 1. Classification of idioms proposed by Fillmore and colleagues (1988: 506-510). 27 Table 2. Brian’s mean overgeneralization rates of irregular verbs divided into groups of token frequency (table taken from Maslen et al. 2004: 1325). 75 Table 3. Counts of verbs used by French nursery school children during play (taken from Bybee 1995: 433). 76 Table 4. German verb prefixes, their meanings and examples. 96 Table 5. English verb prefixes, their meanings and examples. 98 Table 6. Word and verb tokens in English and German corpora. 108 Table 7. Verb prefixes, prefix verb types and tokens as coded by transcribers. 109 Table 8. Prefix verbs selected manually. 111 Table 9. Number of verbs in the novel construction by frequency level for each experimental phase. 128 Table 10. Production of previously encountered verbs in the novel construction as a function of token frequency (prefix, all children). 146 Table 11. Production of previously encountered verbs in the novel construction as a function of token frequency (reduplication, all children). 157 Table 12. Act-out, forced-choice and production tasks. Number of children who memorized and generalized 0 to x verbs in the novel construction. 161 Table 13. Verbs used in the training phase and the forcedchoice task. 179 Table 14. Fixed effects in the Condition x Construction model. 183 Table 15. Table 16. Fixed effects in the Condition model. Verb prefix constructions and verb prefixes selected manually. 183 212 Table a. Manually selected verb prefix constructions and prefix verbs for English corpus. 236 Table b. Manually selected verb prefix constructions and prefix verbs for German corpus. 237 Table c. Prefix constructions, prefix verb types and tokens used by Thomas and his caregiver. 245 <?page no="12"?> xii Table d. Prefix constructions, prefix verb types and tokens used by Leo and his caregiver. 246 Table e. Counts of verb types and tokens per prefix construction as produced by Thomas’ caregiver (CG) and Thomas. 252 Table f. Counts of verb types and tokens per prefix construction as produced by Leo’s caregiver (CG) and Leo. 252 Table g. Act-out task. T-test values by age group and frequency level (German, prefix). 254 Table h. Act-out task. T-test values by age group and frequency level (English, prefix). 254 Table i. Forced-choice task. T-test values by age group and frequency level (German, prefix). 255 Table j. Forced-choice task. T-test values by age group and frequency level (English, prefix). 255 Table k. Production task. T-test values by age group and frequency level (German, prefix). 256 Table l. Production task. T-test values by age group and frequency level (English, prefix). 257 Table m. Act-out task. T-test values by age group and frequency level (German, reduplication). 257 Table n. Act-out task. T-test values by age group and frequency level (English, reduplication). 258 Table o. Forced-choice task. T-test values by age group and frequency level (German, reduplication). 258 Table p. Forced-choice task. T-test values by age group and frequency level (English, reduplication). 259 Table q. Production task. T-test values by age group and frequency level (German, reduplication). 259 Table r. Production task. T-test values by age group and frequency level (English, reduplication). 260 Table s. Act-out, forced-choice and production tasks. Number of children who memorized and generalized 0 to x verbs in the novel prefix construction. 263 Table t. Act-out, forced-choice and production tasks. Number of children who memorized and generalized 0 to x verbs in the novel reduplication construction. 264 Table u. Proportions of correct responses and standard errors (SE) for German-speaking adults in forced-choice comprehension task. 265 Table v. Proportions of correct responses and standard errors (SE) for German-speaking adults in production task. 266 <?page no="13"?> xiii Acknowledgements This book is my PhD thesis, which was completed at the Ludwig- Maximilians-Universität Munich, Germany, in 2012. An enormous debt is owed to my supervisor, Hans-Jörg Schmid, who allowed me the freedom to pursue my own line of research and nevertheless let me profit from his guidance and insights. I want to express my thanks to Beate Sodian, who agreed without hesitation to be my second advisor, and to Kathrin Lindner, who willingly accepted my request to serve on the defence committee. My special thanks go to Elena Lieven for giving me the incredible opportunity to spend a very stimulating year at the Max Planck Child Study Centre in Manchester. I appreciated this opportunity so much. Many thanks go to the GAES (Deutscher Akademischer Austauschdienst) for supporting my stay with a 1-year scholarship. I also want to thank all the children that participated in my studies in Germany and England, their parents for letting them participate and the teachers for allowing me to conduct hours and hours of research in their nurseries and schools. Further thanks go to the students at Ludwig-Maximilians-Universität, who participated in the adult study, and in particular to Franziska Günther for volunteering her course. I am also indebted to Anne Bäumler and Thorben Cordes, who let me benefit from their great acting talent for the film stimuli, to Claire Noble, who recorded the English voiceovers, and to Sarah-Katharina Siebenborn, who produced a number of fantastic drawings especially to my order. This thesis benefitted enormously from many stimulating discussions at the Manchester Child Lab (thanks to Elena Lieven, Grzegorz Krajewski, Paul Ibbotson, Eileen Graf and Claire Noble). Additional thanks for their helpful input go to Adele Goldberg, Ron Kuzar, Wolfgang Falkner, Ursula Kania, Anne Bäumler and Carolyn Seybel. I would like to thank Kristina Gerz and Sarah-Katharina Siebenborn for their emotional support at any time of the day and the endless encouragement they gave me during my PhD. My very special thanks go to my parents, Christina and Hans-Werner Siebenborn, for their endless love, unconditional support and encouragement all my life. Thank you for being exactly the way you are! And finally, thanks so much to my husband Thorben Cordes for everything I could wish for, especially for keeping me happy—and nourished—even in the most stressful of times, and to our little son Jakob, who was so incredibly well-behaved in my belly that I could finish and defend my thesis just before he was born. Munich 2012/ Leer 2013, Anne-Kristin Cordes <?page no="14"?> 1 1 Introduction and outline Humans are profoundly sensitive to frequencies of occurrences of events (Ellis 2002; Hasher and Zacks 1984). Encoding and retaining frequency information are automatic and unintentional processes that are largely unaffected by age, cognitive abilities and voluntary goals or strategies (Hasher and Zacks 1984; Hasher, Zacks, Rose and Sanft 1987; Posner and Snyder 1975). Despite the unconscious character of these processes, “[p]eople are surprisingly accurate” when answering questions inquiring general frequency information (Hasher and Zacks 1984: 1372), e.g., Are there more nurses or more linguists? Do more people die of botulism or cancer? The same is true for frequencies in linguistic contexts. People are good at judging the relative frequencies of words, e.g., Which word occurs more frequently in English - bacon or pastrami? (Hasher and Zacks 1984: 1372), the frequencies with which words occur in a previously read list (Hasher and Chromiak 1977) as well as the frequencies of single letters (Attneave 1953) and letter pairs in English (Underwood 1971). In both non-linguistic and linguistic contexts, the frequency of occurrence of events affects human processing in a number of ways: It influences the strength of memory representations, the ability to recognize or recall previous encounters of the same event, the ability to encode new events that are similar to previously encountered ones and the expectations of what follows an event (Arnon and Snider 2010; Bybee 2006; Diessel 2007; Ellis 2002; Gathercole and Hoff 2009; Hasher et al. 1987; Hasher and Zacks 1984; Lieven 2010; Matthews and Bannard 2010; Slobin 1997). Due to these far-reaching effects of frequency on cognitive processing, it is assumed that frequency also plays a crucial role in language learning. This is why the present work explores how input frequencies affect constructional learning in children. The so-called usage-based, constructionist model of language learning 1 provides the theoretical framework for these investigations (Ambridge and Lieven 2011; Behrens 2009; Bybee 1995, 2010; Goldberg 1995, 2006; Tomasello 2000b, 2003). Frequency effects are examined in a corpus study and two experiments in children’s learning of naturalistic and invented (novel) constructions. All constructions belong to the area of derivational morphology, which is as yet largely unexplored with respect to input frequencies. The experimental studies investigate effects of input token frequency, input type frequency and the shape of the input distribution on children’s construction learning abilities. In these analyses, the frequencies of individual constructions are in focus. Additionally, one 1 In future references to this view, I mostly use the abbreviated form the usage-based model. <?page no="15"?> 2 of the experiments and the corpus study explore the role of input frequencies of constructions that are closely related to the construction in question - though at a higher level of constructional abstractness (the so-called pattern level; pattern frequency, see below). All analyses do not only serve the examination of frequency effects. They also help to gain new insights into the cognitive processes that have been proposed to be involved in construction learning. The book is structured as follows. Chapter 2 explores the question whether it seems possible that all language knowledge is learned from the input without the help of language-specific processes or structures. Since frequency is one characteristic of the input, this question is crucial to the present work. The argument for fundamentally input-based learning is developed by contrasting the usage-based view of language learning with the generative point of view. The usage-based position holds that children use their social-cultural and cognitive skills to learn language from their input. Language is thought to develop in joint attentional scenes in which children map the input form they hear to the meaning intended by the speaker. All form-meaning mappings are referred to as constructions (Goldberg 2006: 5). They range from lexemes (e.g., tree) and morphemes 2 (e.g., [verb]ing 3 - cooking ‘ongoing action or process’) to partially-filled (e.g., the [X]er the [Y]er - the bigger the better ‘proportional increase or decrease’) and entirely abstract 4 constructions (e.g., the ditransitive [Subject] [Verb] [Object1] [Object2] - He gave her a book. ‘transfer’). Chapter 3 zooms in on the cognitive processes that children are thought to use in order to build their constructional inventories in usage-based models. Children are assumed to form increasingly abstract constructions based on simpler ones. They first learn words and fixed word strings, e.g., tree, let’s eat, I want it. Subsequently, they create variable slots in fixed strings, starting with a single slot, e.g., let’s [X], and progressing to more and more open slots, e.g., [X] hugs [Y]. Eventually, they form entirely abstract constructions where all positions are variable (Lieven, Behrens, Speares and Tomasello 2003; Lieven, Salomo and Tomasello 2009; Tomasello 2003: 146-161, 2008: 23-24). In doing so, children make use of 2 Morphemes can be found at different levels of abstractness. Individual morphemes can be concrete, e.g., {tree}, but there are also partially-filled and abstract morpheme constructions (e.g., in reduplication). 3 Variable slots in constructions are indicated by [ ] in this work. 4 The term abstract is used to refer to constructions that are schematic, i.e., do not consist of stable linguistic material. Conversely, concrete constructions are those that are item-based and contain stable linguistic material. The schematic—item-based terminology might be preferable because the concrete—abstract dichotomy is also used in the context of nouns (e.g., concrete: table, abstract: peace). The concrete— abstract pair is used here regardless of this fact because it has already become established in the literature on usage-based language learning. <?page no="16"?> 3 categorization, schematization and analogy formation. Previous conceptions of these processes in linguistics and psychology (Goldberg 2006: 70- 75; Langacker 1987a: 371-372; Tomasello 2003: 114-121, 122-126, 161-169, 174) are elaborated and refined in Chapter 3 before they are re-applied to language learning. Essentially, it is proposed that children store two or more examples of an emergent construction, which they then align in an analogical comparison process. This comparison yields a schematization of the commonalities of the examples. This schematization is assumed to be subsequently stored together with the examples as a constructional category. Schema and examples can then be used in generalizations of the construction to new cases. In Chapter 4 these ideas are linked to frequency and to previous research. The processes of analogical comparison, schematization, categorization and generalization are claimed to be susceptible to frequency effects. Token frequency is the “frequency of concrete expressions”; type frequency is “the number of linguistic expressions that instantiate a constructional schema” (Diessel 2004: 29). Token frequency has been linked to the representational strength of types in people’s minds, i.e., high token frequencies strengthen representations (e.g., Bybee 2006, 2010: 19-20). Type frequency is necessary for the creation of variable positions in otherwise fixed constructions and for generalizations of a construction to new cases (Bybee 1995; Bybee 2010: 95-96). The type-token ratio relates type and token frequency. It describes the number of tokens per type and thus reflects the input distribution over several types. Previous research has revealed that input token frequency affects the speed and order of acquisition as well as the number of errors children make when learning constructions of different degrees of abstractness (e.g., Hart 1991; Maratsos 2000; Marchman 1997; Maslen, Theakston, Lieven and Tomasello 2004; Moerk 1978; Naigles and Hoff-Ginsberg 1998; Theakston, Lieven, Pine and Rowland 2004). Input type frequency has been shown to influence productivity including the number of overgeneralizations (e.g., Bybee 1995; Clark 1993: 137-138; Guillaume 1927/ 1973). It was further revealed that not only the frequency of the construction in question, but also frequencies of loosely-related constructions affect learning, e.g., morphological constructions are learned earlier in morphology-rich languages (e.g., Aksu-Koç and Slobin 1985; De Villiers and De Villiers 1985; Dressler 1997; Smoczynska 1985). Based on these findings, it is assumed that the frequencies of constructions that are more closely related to the construction(s) in question will also influence learning. To explore this idea, a more abstract constructional level directly above that of the individual construction(s) explored is introduced. For ease of reference, it is termed pattern level in this book. <?page no="17"?> 4 In the corpus study in Chapter 5 a first step towards the exploration of frequency at the pattern level is taken. The pattern investigated is the derivational verb prefix pattern. It captures the commonalities of all derivational verb prefix constructions, e.g., un[base] verb ‘reverse an action’ and re[base] verb ‘do an action again’, at a more abstract level. It takes the form [derivational prefix][base] verb and the meaning ‘encode a contrast in action’ or ‘encode a contrast to the normal, expected action’. The speech of a German-speaking and an English-speaking child and their caregivers is analyzed in two dense corpora. The aim of the corpus analysis is to examine whether input pattern frequency is reflected in the frequencies in children’s speech. Such an effect is predicted because of previously found relations between input and output distributions of word classes and particle verbs (Behrens 2003, 2006). It is further assessed whether there are differences in pattern frequency between the two languages. Due to the slightly higher number of verb prefixes reported for German (Barz 2006; Engel 2004: 230; Fleischer and Barz 1995: 37, 327; Kühnhold and Wellman 1973: 144-154; Marchand 1960: 139-208; Quirk, Greenbaum, Leech and Svartvik 1984: 981-992; Schmid 2005: 150-161), pattern frequency is also expected to be higher for German than for English in the child speech corpora. Finally, children’s innovative generalizations of verb prefix constructions are analyzed in order to identify further potential differences between German and English and to assess whether they are related to construction type frequencies. Chapter 6 reports Experiment 1, which investigates frequency effects at pattern and constructional levels in children’s learning of novel constructions. It is examined whether the frequency of the pattern underlying a novel construction affects the acquisition of this novel construction. To this end, German-speaking children between 3 and 8 and English-speaking children between 4 and 6 are exposed to a novel verb prefix construction. The underlying pattern frequency was examined in Chapter 5 and the revealed differences are expected to be reflected in this study in novel construction learning. A second group of children is trained on a different novel construction - a verb-initial reduplication construction - that is based on a pattern that exists neither in German nor in English and consequently has a pattern frequency of zero. This is done to assess whether learning of this reduplication construction is equally difficult for German-speaking and English-speaking children, given that neither group has experience with the underlying pattern. Moreover, novel prefix learning and novel reduplication learning can thus be compared. Because of the higher frequency of the underlying pattern, prefix learning is expected to be easier. However, this issue is only assessed tentatively, since additional differences between the novel prefix and novel reduplication might play a role <?page no="18"?> 5 as well (for this reason this issue is not listed in the major research questions). Additionally, the influence of experimental token and type frequencies is examined for novel prefix and novel reduplication learning. Input token frequency rises differentially for types of the novel construction over the course of the experiment. This circumstance allows the assessment of token frequency effects on children’s memory of previously encountered types and on their ability to generalize the new constructions to unfamiliar types. Based on previous research children’s performance is expected to increase for types with higher token frequencies and to be lowest for items requiring generalization. Further, the number of types children store before generalization is examined. On the basis of an earlier study (Wonnacott, Boyd, Thomson and Goldberg 2012) and the requirements of the analogical comparison process in construction learning, the storage of at least two types of the novel construction is thought to precede generalizations. All these research questions are based on the assumption that children are able to learn either of the two novel morphological constructions from the input, though the degree of learning might vary with age. Previous research has so far shown novel construction learning only for abstract word order constructions (Boyd and Goldberg 2011a; Casenhiser and Goldberg 2005; Wonnacott et al. 2012). One of these studies also revealed an age effect, which is also expected in the present study (Boyd and Goldberg 2011a). In Chapter 7 a second training study (Experiment 2) is reported. It investigates the role of input type and token frequency in relation to each other, that is, the role of the input distribution. Previous research has revealed that children profit from skewed input, i.e., input where one type occurs much more frequently in the novel construction than all others, when learning a novel abstract word order (Casenhiser and Goldberg 2005). With respect to partially-filled constructions, on the other hand, it has been proposed that balanced input, i.e., input where all types occur roughly equally frequently, is more conducive (Matthews and Bannard 2010). This latter assumption has not yet been addressed in novel construction learning research. In this book it is suggested that these two positions can be reconciled. It is proposed that the abstractness of the respective construction determines which input distribution is more beneficial, because the tasks children have to perform in learning vary with the abstractness of the construction. A recurring anchor provided by the skewed input is expected to be helpful in abstract construction learning, because it facilitates the comparison of constructional types that otherwise vary in each position. Balanced input is predicted to be more conducive in partiallyfilled construction learning, since the balanced input makes the variable slot become apparent sooner. This second part of this hypothesis is tested <?page no="19"?> 6 in Experiment 2. English-speaking children between 5 and 7 are trained on a novel partially-filled construction. The input distribution is varied between conditions in order to examine whether balanced input is in fact more conducive to the learning of a novel partially-filled construction than skewed input. In the final chapter, Chapter 8, this work and its major results are summarized. The findings are reviewed in relation to each other and with a view to previous research. Three more general issues that surfaced in this book are taken up and discussed in more detail. Specifically, the question of which abstractions people make and store in their minds, the domaingenerality of the cognitive processes involved in language learning and the role of salience in this learning process are debated. Two related, additional issues receive extra attention before future directions for research are proposed and the book is concluded. At its heart, this work aims at contributing to the amelioration of a situation described by Abbot-Smith and Tomasello (2006: 287): “There are very few experimental investigations of the abstraction/ generalization process, and its relation to input and input frequencies of various types”. This situation is further specified by Tomasello (2003: 125) who surmises that “appropriate token and type variation” is presumably necessary for constructional learning, but stresses that very little is known “about the frequencies that might be required in different cases”. The major research questions of the book strive to improve this situation. They are listed in the following. These questions are derived from the discussion of the previous literature in Chapters 2 to 4. Due to this close connection to the theoretical background, they are elaborated on and presented (again) in context at the end of the last theoretical chapter (Chapter 4). Research questions: I Pattern frequency in naturalistic language learning a Are input and output frequencies of the derivational verb prefix pattern related? b Are there differences in the frequency of the derivational verb prefix pattern between German and English in the language to and of children? II Novel construction learning and pattern frequency a Pattern present in native language; frequency varies between languages i Learnability: Can German-speaking and English-speaking children learn a novel construction based on a pattern that is used in their languages and does the success of learning increase with age? ii Frequency: Is learning a novel construction easier for children in whose native language the underlying pattern is more frequent? <?page no="20"?> 7 b Pattern absent from native language i Learnability: Can German-speaking and English-speaking children learn a novel construction based on a pattern that is absent from their languages and does the success of learning increase with age? ii Frequency: Is learning a novel construction equally difficult for all children given that the underlying pattern is absent from both languages? III Token frequency in novel prefix and novel reduplication learning Does token frequency increase novel construction learning, i.e., correct responses? IV Type frequency in novel prefix and novel reduplication learning Do children use at least two previously experienced types of the novel construction correctly before forming correct generalizations? V Type-token ratio in partially-filled construction learning Is balanced input more conducive to learning a partially-filled construction than skewed input? <?page no="21"?> 8 2 Language learning from the input? The present chapter discusses the question whether there is sufficient reason to proceed from the usage-based, constructionist hypothesis that language can be learned from the input with the help of social-cultural skills and domain-general cognitive abilities (Ambridge and Lieven 2011; Behrens 2009; Bybee 1995, 2010; Goldberg 1995, 2006; Tomasello 2000b, 2003). This position is adopted in this book, since the interest in input frequency effects is based on the assumption that input plays a crucial role in language learning. The usage-based view competes with the historically earlier and perhaps still more common generative standpoint that the input is too poor a source for language learning, which is one reason why language acquisition is said to be based on innate language-specific structures (Chomsky 1959, 1965, 1975, 1980a, 1988, 2000; Pinker 1995, 1999). In the first part of this chapter the two theories are contrasted with respect to a few fundamental issues. Due to the focus on the usage-based view in this work, this side of the arguments tends to be presented in more detail. The second part of the chapter is then devoted to a detailed discussion of the input. Evidence against the generative idea that the input cannot serve as the basis for learning all structures of a language is brought forward and usage-based mechanisms of learning from the input in absence of language-specific structures or constraints are presented. 1 2.1 Contrasting key issues 2.1.1 The issue of human uniqueness The recurring and connecting element between many fundamental issues in language learning is the input. Its presumed deficient character is the starting point for the assumption held by many generative accounts that language acquisition must be based on innate language-specific structures (nativism). 2 Since language is present in humans but in no other species, however closely-related evolutionarily, the nativist assumption holds that language is the one feature that is uniquely human (Chomsky 1968/ 2006: 59; Leiss 2009; Pinker 1995; Valian 2009). Arguments brought forward in 1 A detailed account of all relevant theories belonging to each of the two traditions and their development is not attempted, since this would go far beyond the scope and aim of this book. 2 The generative position does not necessarily entail assumptions of linguistic nativism, but most generative accounts are linked to nativist assumptions. <?page no="22"?> 9 favour of this position are the rapidity of children’s language development in absence of formal instruction and despite the deficient input (e.g., slips of the tongue, elliptical structures, errors) and the fact that all members of a language community converge on the same linguistic system (Chomsky 1959: 57; Pinker 1995). 3 Chomsky points out that there is no alternative way of explaining why his granddaughter learns to speak when exposed to language while a rock and a rabbit do not, given the same input, than by invoking innate language-specific knowledge: To say that ‘language is not innate’ is to say that there is no difference between my granddaughter, a rock and a rabbit. In other words, if you take a rock, a rabbit and my granddaughter and put them in a community where people are talking English, they’ll all learn English. If people believe that, then they believe that language is not innate. If they believe that there is a difference between my granddaughter, a rabbit and a rock, then they believe that language is innate. So, people who are proposing that there is something debatable about the assumption that language is innate are just confused. So deeply confused that there is no way of answering their arguments. (Chomsky 2000: 50-51) This somewhat polemic suggestion precludes the possibility put forward in usage-based accounts, namely, that the difference between humans and other non-talking animals (let alone between humans and rocks) may lie elsewhere. Indeed, Goldberg (2006: 69) argues that it is “[s]urely […] premature to give up hope that humans, with our rich cognitive abilities, complex social skills, predilection to imitate, and 100-billion-neuron brains, can learn language from the available input.” It is these skills that are considered to be involved in language learning according to the usage-based model. Instead of assuming the innateness of language-specific knowledge, proponents of this view hold that the social-cultural ability to read intentions sets humans apart from other closely related species. This ability includes basic intention-reading skills which are thought to develop during the first one and a half years of life (Tomasello 1995; 2003: 3) as well as more advanced theory-of-mind abilities that develop a little later in childhood (Wellman and Liu 2004; Wimmer and Perner 1983; see Wellman, Cross and Watson 2001 for a review). Early skills comprise children’s abilities to share attention with other people (joint attention), to follow other people’s pointing gestures, to produce their first own gestures to direct other people’s attention and to begin 3 Whether speakers do in fact converge on “one grammar” is discussed in the literature. More than 20 years ago Langacker (1987a: 376) suggested that linguistic systems of b- "# $ % & & differences between people with different educational attainments. <?page no="23"?> 10 to understand and imitate intentional actions of others (Tomasello 2003: 3). The more advanced theory of mind entails children’s abilities to attribute mental states, such as intentions, goal-directedness and desires, to other people (Premack and Woodruff 1978; Wimmer and Perner 1983). Evidence for children’s intention-reading skills comes from numerous experiments. Preferential looking studies using a violation-of-expectation paradigm showed that young 1-year-olds form expectations about people’s behaviour. They expected people to look for objects in the place where they left them, but not elsewhere (Surian, Caldi and Sperber 2007; Onishi and Baillargeon 2005). A task with an active-helping paradigm further revealed the ability of older 1-year-olds and 2-year-olds to predict other people’s beliefs and intentions (Buttelmann, Carpenter and Tomasello 2009). Children read the experimenter’s intention and inferred his false belief from his unsuccessful attempts to do something. This was evident in children’s behaviour. When witnessing the experimenter’s unsuccessful attempt to open a box where he had left an object, children concluded that he wanted to retrieve this object, which had been moved in his absence, and pointed the correct location out to him or helped him open the correct box. In a more classical format, this change-of-location task involves a story about a protagonist who puts an object in a place, which is removed in his absence. Children have to predict where the protagonist will search for the object. They usually succeed at this task later, between 4 and 6 years of age, not least because the story-telling mode is cognitively more demanding (Wimmer and Perner 1983). There seems to be a continuous development from children’s recently revealed early abilities of intention-reading to their later more sophisticated theory-of-mind skills that were found in more demanding tasks. Evidence for this relation comes from a study showing that children’s understanding of intentional actions (in a habituation task) in their first year of life predicted their theory-of-mind understanding at preschool age (Wellman, Lopez-Duran, LaBounty and Hami 2008). There are several reasons why the proposed social-cultural skills of intention-reading should be unique to humans. Children’s understanding and their production of intentional actions are connected to their imitative abilities (Tomasello 2003: 21-28). Children tend to imitate actions on the reasoning that the model they are imitating had a reason for performing the respective action in a certain way, that is, they assume there is an underlying intentionality behind the way an action is carried out (Gergely, Bekkering and Király 2002). 4 The human tendency to imitate actions is 4 This understanding of imitation contrasts with the behaviorist understanding of imitation, i.e., imitation that requires reinforcement (Skinner 1957: 59-60). It is this <?page no="24"?> 11 contrasted by other primates’ preference to merely emulate, that is, to copy results (Tennie, Call and Tomasello 2006; Tomasello 1999: 29-30). Moreover, contrary to weak claims to the opposite (Hauser, Chomsky and Fitch 2002), there is evidence that non-human primates do not use signals referentially and meaningfully to direct their conspecifics’ attention or affect others’ mental states, which is what humans do (Tomasello 2003: 9-11). Rather, non-human primates aim to affect others’ behaviour or motivational states directly, e.g., to indicate food sources to others or alarm them in case of approaching danger. Further strong support for the uniqueness of human social skills comes from a recent study by Herrmann and colleagues (Hermann, Hernández-Lloreda, Call, Hare and Tomasello 2010), who compared the cognitive abilities of 106 chimpanzees and 105 2-yearold children. Their factor analysis revealed that a separate social cognition factor underlay human children’s but not chimpanzees’ performance. Children further performed significantly better than chimpanzees on tasks relating to this social cognition factor. The authors suggest that their findings imply that a specialized set of social-cognitive skills, “cultural intelligence”, has evolved in humans. To sum up, proponents of nativist, generative approaches see innate language-specific structures as uniquely human. In contrast, supporters of the usage-based model take as a starting point the null hypothesis that language is not innate. Instead they propose that intention-reading skills set humans apart from other species. These skills involve joint-attentional abilities, pointing and other gestures of directing attention as well as the imitation and understanding of intentional behaviour. These abilities have been shown to emerge early in children’s development and to be absent from the behavioural repertoire of primates that are closely-related to humans. 2.1.2 Language knowledge in the child and in the adult Apart from these different assumptions of what children bring with them to the task of language learning, there are also differences as to how children’s and adults’ language knowledge is conceptualized in the generative and the usage-based account. This issue and the related question of how child and adult representations are connected are discussed in the following. Proponents of generative, nativist accounts presume that adult native speakers’ language knowledge (or competence using the appropriate terminology) comprises an innate core grammar (Universal Grammar) and a definition of imitation that generativists have claimed to be unnecessary or insufficient for language learning (Chomsky 1959). <?page no="25"?> 12 learned periphery. The core is often equated with a syntax module, even though it is usually thought to comprise a phonological, a semantic and a morphological module as well. Each module is considered to be autonomous but modules are nevertheless thought to interact with each other and with the lexicon. The lexicon constitutes the periphery and contains a list of lexical items of a language including information about word classes and so-called subcategorization (i.e., which types of arguments and how many arguments a lexical item takes). The primary interest of generative grammar is placed on the description of the core grammar rather than on accounting for word learning. The core grammar allows the acquisition of rules and is thus associated with humans’ ability to generate an infinite number of grammatically correct sentences. 5 Rules are semantically empty and formal and operate on meaningful lexical units, which is the reason why this account is also known as the words-and-rule approach (Pinker 1999). Rules are formulated in the most abstract way possible. This makes the system of grammatical knowledge highly parsimonious and descriptively elegant but at the same time impossible to learn from the input because it is considered to be not immediately obvious in the structure of any physical utterance due to its complexity (Chomsky 1959). The type and scope of the rules proposed in the generative tradition vary considerably between different stages and different positions (Brown 2004). Overall, a shift towards minimizing the rule apparatus can be observed over time from a very high number of proposed rules 6 (Brown 2004; Chomsky 1957: 26-28, 44-46, 111-113; 1965) to principles and parameters (Brown 2004; Chomsky 1981; Eisenbeiß 2009) and recently to the very restricted language faculty with its major component of recursion (Chomksy 1995; Hauser et al. 2002). 7 In place of these different understandings the principle-and-parameter account, which is still popular among many theorists and researchers, is presented briefly. 5 This understanding of language’s generativity goes back to Wilhelm von Humboldt (“von endlichen Mitteln einen unendlichen Gebrauch machen” - ‘to make infinite use of finite means’ (my translation), Humboldt 1836: 221). It was considered so fundamental that it became widely-used as a term to describe generative theories. 6 They comprised, for instance, phrase-structure rules that defined possible constituent arrangements on the so-called deep structure, i.e., the grammatical structure that underlies a related, so-called surface structure. Surface structures were thought to be derived from deep structures by the application of transformational rules. Lexical items were inserted only on the surface structure. 7 Many of the original, descriptive categories (deep and surface structures, government, Xbar theory) were abandoned in an attempt to increase the efficiency and economy of description. However, there remained a computational rule system that executes certain operations (e.g., merge) and interfaces with phonological and semantic systems. <?page no="26"?> 13 According to this position, Universal Grammar contains both principles and parameters. Innate, universal language knowledge takes the format of unlearnable principles. An example is the structure dependency principle, which defines which constituent can be moved where in so-called transformations (e.g., when forming a question from a declarative) in dependence of its structural role in the sentence (Brown 2004). In contrast to principles, parameters need to be set to one of two (or sometimes more) values. This parameter setting requires language input. For instance, the pro-drop parameter determines whether subjects must be expressed overtly, as in English, or not, as in Spanish. Based on the language input this parameter is set to either pro-drop (in Spanish) or non-pro-drop (in English) in each language. Through their flexible settings, parameters are thought to account for cross-linguistic variety. In this view, adults thus command a core grammar consisting of highlyabstract, innate rule-like principles and set parameters as well as a lexicon that comprises all less systematic knowledge. The question is how children’s language systems are organized. Undoubtedly, children’s language use is more restricted and limited than that of adults. It proved, however, impossible for generativists to propose an account of how children progress from their seemingly limited knowledge to the highly complex and abstract language system that they credit adults with. As a consequence, generativists assume continuity (known as the continuity assumption) between child and adult language knowledge (competence). Children are thought to operate with the same abstract categories as adults and to command the same innate principles (Pinker 1984: 7-8), although it is possible that not all parameters are set yet at a given time. Once they are set, children’s systems are identical to those of adults and any errors children make are explained in terms of inadequate performance as are adult errors. Inadequate performance is explained by “such grammatically irrelevant conditions as memory limitations, distractions, shifts of attention and interest, and errors” (Chomsky 1965: 3). Clahsen summarizes this position: [Y]oung children when they begin to produce sentences already have full grammatical competence of the particular language they are exposed to, and […] differences between sentences children produce and adults’ sentences should be attributed to external factors, i.e. to developments in other domains than grammatical competence. (Clahsen 1996: xix) An obvious problem with this position is that the continuity assumption can only be upheld because the inadequate performance explanation is invoked for all differences between children’s and adults’ language that cannot be accounted for by as-yet unset parameters. This line of argumentation renders the continuity hypothesis unfalsifiable and makes Akhtar <?page no="27"?> 14 (2004: 460) wonder “what could possibly count as evidence against continuity”, when performance factors explain any counter-argument away. There is one area, namely children’s developmental progression, which potentially brings forward evidence against the continuity assumption. The generative account predicts that as soon as a particular parameter is set, children will be able to comprehend and produce all utterances with this parameter accurately. Radford illustrates this with an example: Once a child is able to parse an utterance such as ‘Close the door! ’, he will be able to infer from the fact that the verb ‘close’ in English precedes its complement ‘the door,’ that all verbs in English precede their complements (since Universal Grammar excludes the possibility that some verbs may precede and other follow their complements). (Radford 1990: 61) Radford goes on to suggest that such a sentence will further allow children to generalize that other categories, e.g., nouns or adjectives, also precede their complements, based on the fact that the same structure (the so-called X-bar template) underlies all phrase structures. Only very minimal exposure is thus sufficient for triggering the setting of parameters like the word order parameter (Radford 1990: 61). Children’s performance on structures that are determined by a parameter should consequently improve abruptly from random usage, reflecting the unset parameter (e.g., using variable word order), to perfect or at least near-perfect usage, reflecting the set parameter (e.g., using correct word order). There should be leaps in their language acquisition progress. However, this prediction is not borne out in analyses of child language corpora or in experimental studies. Children’s learning has instead been found to be gradual and item-based. For instance, children’s morphological marking has been shown to be restricted to some verbs or certain personnumber combinations before being gradually extended to others (Pizzuto and Caselli 1992; Tomasello 1992: 223-236). The same pattern of gradual learning is found in studies with novel verbs and weird word orders. When trained on a novel verb in one construction, e.g., the intransitive, 2to 3-year-olds were reluctant to use it in a different construction, e.g., the transitive, without having encountered any evidence that this is possible (Akhtar and Tomasello 1997; Tomasello and Brooks 1998). After hearing novel verbs in a weird non-canonical word order, children of the same age stuck to these patterns for the new verbs and only older children corrected such sentences to the canonical English word order that the word order parameter prescribes (Akhtar 1999). In the face of empirical evidence that learning progresses gradually rather than in leaps, there have been generative attempts to account for this gradualness of development by proposing particular procedures of parameter setting. Classical parameter setting in terms of triggering on the <?page no="28"?> 15 basis of minimal input (according to some accounts a single example) proves inadequate to explain the results, unless performance errors are invoked again to account for incorrect productions. Instead it has thus been proposed that more input or unambiguous input is required for parameter setting (Eisenbeiß 2009; see Roeper and deVilliers 1992 or Roeper and Weissenborn 1990 for the unique trigger hypothesis). In both cases, application of the parameter in production would be expected to occur later. The latter account would require an additional mechanism to determine what constitutes unambiguous input. Alternatively, parameters have been suggested to mature like other body systems, e.g., vision (Clahsen 1996: xix; Eisenbeiß 2009). Parameter maturation would also cause competence to increase more slowly, with adult-like competence being exhibited only after complete maturation. Finally, parameter setting has been likened to unconscious hypothesis-testing (Valian 1990; 2009), in which hypotheses about parameters are constrained by their possible values. Assessing the adequacy of different hypotheses is again thought to require more time. While these suggestions explain that correct uses may be delayed and that errors may occur during the setting procedure, they do not necessarily predict that learning will be gradual. More importantly, they do not account for all the available data on children’s gradual language development. Many errors children make were shown to be frequency-based (cf. Chapter 4), i.e., they occur predominantly on low-frequency lexemes (Kirjavainen, Theakston and Lieven 2009; Maratsos 2000; Marchman 1997; Maslen et al. 2004; Theakston, Lieven and Tomasello 2003). Such errors sometimes persist even when parallel correct uses suggest that the respective parameter is already set. Moreover, the child should also be able to apply the parameter to new cases, once it is set. However, in the novel verb learning study with transitive sentences reported earlier (Akhtar and Tomasello 1997, see also Tomasello 2003: 187; Tomasello and Brooks 1998), children who did use the transitive productively (implying the parameter is set) nevertheless failed to apply the correct word order to novel verbs. Even after presumed parameter setting, these errors occurred. Since they occurred selectively on low-frequency (here: novel) verbs, they cannot be explained by performance limitations, which are by definition random rather than systematic. Parameter-based accounts would have to bring forward proposals explaining how this systematicity results from their position. In the present form, different generative accounts of parameter setting have difficulties explaining the systematicity and frequencydependence of children’s errors and therefore face problems when accounting for the continuity assumption between children’s and adults’ language knowledge. <?page no="29"?> 16 The findings that children’s language development progresses gradually and that their errors in this process are frequency-dependent are taken into account and are, in fact, even predicted in the usage-based account of language learning, as will be discussed in more detail in Chapters 3 and 4. Following ideas of construction grammar (Croft 2001; Goldberg 1995, 2006), the usage-based account describes the end-point of language development, or adult language knowledge, as an inventory of constructions, i.e., learned form-meaning or form-function pairings (Goldberg 2006: 5). The crucial difference to the generative account is thus the important role that is attributed to the functional side of the construction and the communicative side of language in general. It is no longer empty formal rules that are applied to lexical units, instead form and meaning are closely intertwined in constructions. Form-meaning pairs vary in schematicity, such that they range from completely concrete or item-based constructions (e.g., lexemes) to partially-filled (e.g., the morpheme [verb]ing ‘ongoing action or process’) and to entirely abstract or schematic ones (e.g., the ditransitive [Subject] [Verb] [Object1] [Object2] - He gave her a book. ‘transfer’). Crucially, more item-based, specific constructions are not expected to be abandoned in favour of more abstract ones. Language knowledge is stored redundantly, which poses no problem in usage-based accounts, since the aim is not parsimony but psychological plausibility. Like adults, children are assumed to have structured inventories of meaningful constructions that differ in schematicity - with more itembased constructions existing alongside more schematic ones (Behrens 2009; Goldberg 2006: 12; Tomasello 2003: 98-99; ). Thus, children’s representations of language are not in principle different from those found in adults. What differs between children and adults is that children are only beginning to build their constructional inventories. Children’s early constructions are less general, more closely input-related, “simpler and more concrete, with fewer abstractions - because they are based on less linguistic experience” (Tomasello 2003: 97). This is reflected in children’s language use, which is also less general and more item-based than that of adults. In the process of language learning, children’s inventories expand because children gradually form increasingly abstract constructions from simpler ones. Compared to children’s language knowledge, adults’ knowledge merely consists of a much more comprehensive and diverse inventory of constructions including more abstract constructions. The relation between children’s and adults’ knowledge is thus “[not] continuity […] of structures […] but […] continuity of processes” involved in its construction (Tomasello 2000b: 237). The processes that are responsible for building the constructional inventories, e.g., categorization and analogy formation (cf. <?page no="30"?> 17 2.2.3; Chapter 3), rather than the language-specific knowledge itself, are the same in all stages 8 of language learning and language use. To sum up, proponents of the generative position hold that language knowledge comprises a highly abstract rule system, part of which is innate and part of which is acquired (e.g., parameter settings), as well as the lexicon, which is learned from the input. The rule system consists of semantically empty, abstract formal rules that are applied to meaningful lexical units. Continuity between children’s and adults’ language knowledge is presupposed. The only temporary difference that the account permits is caused by rules children have not yet acquired at a given point (i.e., parameters that are not set yet). As illustrated, several problems occur when the continuity assumption and parameter setting procedures are used to account for children’s development, among them the difficulty to explain the gradualness of children’s development and the frequency dependence of errors. These aspects present no problem to the usage-based approach, where language knowledge is conceptualized as a rich network of meaningful constructions that are stored redundantly at different levels of abstractness. The network is thought to be learned gradually from the input and learning is expected to be affected by frequency. The difference between children’s and adults’ language knowledge is that children are still building their networks, which are initially more concrete and item-based before more abstract representations are developed. Children’s and adults’ knowledge are related through the continuity of the processes that are involved in construction learning. 2.1.3 The role of the input The fact that the role that is attributed to the input varies immensely between the generative and the usage-based position has already surfaced in the previous sections. The role of the input bears on a number of issues in both accounts. According to the generative view, the input is too poor to serve as the main basis for language acquisition, which warrants the necessity of postulating an innate grammar system. This innate system enables humans to acquire and use the complex language system they are credited with in generative accounts; at the same time innate constraints impose restrictions on the developing language knowledge, thus ensuring that 8 The finding that language learning during the so-called critical period is easier and results in more native-like language abilities than later attempts to learn language is not ascribed to differences in the cognitive processes used in learning. Instead, it is attributed to a gradual decline of learning abilities with age, which in turn goes back to interference and transfer from the highly entrenched first language and competition between the first and the second language (MacWhinney 2005, 2012). <?page no="31"?> 18 speakers of a language community converge on the same language. Proponents of the usage-based account, who do consider the input to be the main source of language learning, need to provide evidence for the adequacy of the input as a source. Moreover, since the existence of innate rules is negated, they need to propose viable alternatives of how language learning is constrained. Because of the prominent role of the input in this work, these issues are elaborated in detail in the subsequent section. In doing so, the burden of explication and justification is put on the usage-based account, since it is put forward in the present book. 2.2 The poverty-of-the-stimulus argument versus learning from the input The usage-based proposal that the entire language knowledge (as conceptualized in their view) might be learned from the input without the support of abstract, innate structures or rules runs counter the generativist claim of the poverty of the stimulus. The argument is presented in terms of input quality and quantity; counter-arguments are brought forward with respect both areas. Following the insight that the input is not as poor as previously assumed, the usage-based proposal about how language might be learned from the input is described. Since language learning in the usage-based model involves generalizations based on the input that are not constrained innately, the final section presents ways of limiting potential generalizations. 2.2.1 Input quality The poverty-of-the-stimulus argument in fact subsumes several arguments. Depending on the aspect in focus, the poverty-of-the-stimulus argument is known as the logical problem of language acquisition, Baker’s Paradox, Gold’s Theorem, Plato’s Problem, Chomsky’s Problem or the problem of no negative evidence. It is the main reason why proponents of generative models of language acquisition hold that language cannot solely be learned from the input. The first part of the argument refers to the quality of the input. It has been pointed out repeatedly by Chomsky (1965: 58, 1975: 22) how “degenerate”, erroneous and elliptical the language input to children is. Even though he offered no evidence from speech corpora or experiments, it has since been claimed that a large percentage of caregivers’ utterances are erroneous and ungrammatical (Sampson 2005: 43), including, for instance, sentences without subjects in English, which have been said to confuse children ((1) and (2)): <?page no="32"?> 19 (1) Speaker A: Can you come and help me for a second? Speaker B: Coming. (2) Mother: Want your lunch now? The latter example (2) is taken from Valian (1990: 120), who claims that it is ungrammatical. Sampson (2005: 44) argues that Valian’s use of ‘grammatical’ means “conforming to the artificial standards required for formal language in high-prestige contexts”. However, as Sampson points out, such a definition of grammaticality might not be particularly insightful, since adults do regularly produce utterances like speaker B in (1) or the mother in (2). And crucially, these ‘elliptical’ utterances are informative and easy to comprehend in the given situations, because they are systematically related to given referents in the respective situations. It is further likely that adults produce these utterances because they learned at some point during their development that such utterances are indeed acceptable in certain contexts. This situation would in turn imply that children in fact need to be exposed to these supposedly ungrammatical utterances in order to learn that subjects can be omitted in spoken language when they are recoverable from the context (even in non-pro-drop languages like English and German). From this perspective, numerous so-called ungrammatical utterances cannot in fact be considered ungrammatical. Newport, Gleitman and Gleitman (1977: 121) cast further doubt on the argument that children’s input is at all error-ridden. In analyses of caregiver speech directed at children between 1 and 3 years they found that “the speech of mothers to children is unswervingly well formed. Only one utterance out of 1500 spoken to children was a disfluency.” In their study the ungrammatical utterances amount to just 0.067% of all utterances produced by the mothers. This proportion is presumably too low to provide support for the poverty-of-thestimulus argument. But Newport and colleagues (1977) go on to point out that in speech among adults, there are about 5% incorrect utterances. 9 Simply overhearing these utterances has been claimed to be enough to create a problem, particularly so because adults do not mark ungrammatical utterances as such (Hornstein and Lightfoot 1981: 11). However, children have been shown to be able to distinguish between child-directed and adultdirected speech, and prefer speech directed at them (Fernald 1985; Pinker 1995). This preference is presumably caused by the special features of child-directed speech, e.g., short sentences, high pitch, repetitions and redundancies, which have been revealed to catch children’s attention and 9 To my knowledge there are no data as to the percentage of grammatically correct utterances children hear from different interlocutors in cultures where caregivers do not use child-directed speech and children learn primarily from siblings or their mothers when they speak for them rather than with them (on Kaluli: Ochs and Schieffelin 2009: 307; Schieffelin 1979: 86). <?page no="33"?> 20 thus be particularly conducive to language learning. Considering these circumstances, the extent of the negative influence of 5% ungrammatical utterances in the speech between adults is at least debatable (Sampson 2005: 43) and a very weak base for the low-quality side of the poverty-ofthe-stimulus argument. 2.2.2 Input quantity The second part of the poverty-of-the-stimulus argument holds that the quantity of the input children receive is too limited as to serve as the basis for language learning, which has several negative consequences. It has been claimed that children only receive “relatively slight exposure” (Chomsky 1975: 4) to language. This assertion raises the question of how much language children actually hear. Hart and Risley (1995: 132) give word token counts and suggest that children (depending on their social class) will have been exposed to 10-30 million words by age three. Cameron-Faulkner, Lieven and Tomasello (2003) analyzed child-directed speech of 12 English-speaking mothers to their 2-year-old children. They found that children at that age hear an estimated 5,000-7,000 utterances per day including about one-quarter to one-third questions. Even though it is difficult to generalize from these counts since younger children might receive less input and older children might receive more input than 2-yearolds, a crude calculation would suggest that by the time an Englishspeaking child turns three, he or she will have heard about 5,475,000- 7,665,000 utterances. Although Chomsky’s term “slight” is very vague and poorly defined, it is questionable that the extent of the reported exposure can be considered as such. The quantity of the input has further been claimed insufficient for forming adequate hypotheses about the target language. Language acquisition or parameter setting is sometimes portrayed as hypothesis-testing process (Pinker 1995; Valian 1990, 2009), in which children form and test hypotheses in order to determine the grammar of their target language. The problem is that however much input children hear (even if all utterances were grammatical), it can only ever constitute a random sample of all theoretically possible sentences. This random sample is thought to underdetermine adult language competence because it is congruent with a number of theoretically possible languages rather than only with the ambient language. Nevertheless, children within a language community all converge on the same language. Generativists consider negative evidence, that is, feedback on every ungrammatical utterance children hear or produce, the only way out of this dilemma. Unless children consistently receive negative evidence to restrict potential alternative generalizations, constraints that are innate in Univer- <?page no="34"?> 21 sal Grammar are concluded to be necessary in order to limit the hypothesis space (Eisenbeiß 2009; Pinker 1995, 2004; Marcus 1993). Even though caregivers have been found to react in different ways to well-formed and illformed utterances (Demetras, Post and Snow 1986; Hirsh-Pasek, Treiman and Schneiderman 1984), negative evidence has been argued to be too infrequent and often ineffective (Marcus 1993: 53) as the example below illustrates (Braine 1971: 160-161, emphasis in the original). (3) Child: Want other one spoon, Daddy Daddy: You mean, you want THE OTHER SPOON Child: Yes, I want other one spoon, please, Daddy Daddy: Can you say “the other spoon”? Child: Other… one… spoon Daddy: Say… “other” Child: Other Daddy: “Spoon” Child: Spoon Daddy: “Other… spoon” Child: Other… spoon. Now give me other one spoon? Negative feedback further frequently refers to the semantic incorrectness or pragmatic inappropriateness of the content of an utterance rather than to its ungrammaticality (Brown and Hanlon 1970; Maratsos 1983: 732). For these reasons, it seems doubtful that children only rely on negative feedback for determining the correct hypotheses about the grammar of their target language. Another source of potential evidence in hypothesis-testing is indirect negative evidence, that is, the lack of certain structures in the input. However, there is an abundance of structures children never hear and are still able to rate as grammatical as adults (Eisenbeiß 2009; Valian 2009), e.g., triply embedded structures. This source thus neither provides sufficient evidence for constraining the hypothesis space, again calling for innate structures to take on this task. The generative assumption of indirect positive or no positive evidence is claimed to provide an additional reason for the necessity of such constraints. According to the argument, certain structures do not occur in the input (i.e., there is no positive evidence of certain structures in the input) and yet no errors are made in their production: “People attain knowledge of the structure of their language for which no evidence is available in the data to which they are exposed as children” (Hornstein and Lightfoot 1981: 9; see also Crain 1991: 598). This phenomenon is often illustrated by yes/ noquestions that are thought to be formed from complex sentences. Chomksy (1980b: 40) himself says about this structure that “[a] person might go through much or all of his life without ever having been exposed to the <?page no="35"?> 22 relevant evidence”. But speakers would nevertheless on the first occasion form the respective question correctly, because they rightly apply one but not the other of two possible principles. Based on simple yes/ no-questions like (4b) (examples taken from Eisenbeiß 2009: 277) in the input children might contemplate the following two hypotheses: an order-dependent principle that the leftmost auxiliary is fronted or a structure-dependent one that the auxiliary of the main clause is fronted. This circumstance amounts to a problem in questions that are formed from complex sentences and thus contain one auxiliary in the main clause and one in the subordinate clause (5a). Because such questions are supposedly not in the input, only the innate structure-dependency principle is thought to be able to provide the correct answer: It is the auxiliary of the main clause that is fronted (Eisenbeiß 2009). In (5b) and (5c) the position of the fronted auxiliary in the corresponding declarative clause (5a) is underlined. (4) a The chicken is running. b Is the chicken running? (5) a The farmer who is running is bald. b *Is the farmer who _ running is bald? c Is the farmer who is running _ bald? According to the no-positive-evidence argument children (and adults) are said to virtually never hear sentences like (5c) in their input, so that the possibility of learning from experience is precluded (Chomksy 1980b). However, there is mounting evidence that the relevant sentences are not as rare in the input as generativists suggest. Pullum and Scholz (2002: 43, emphasis in the original) argue that “it is implausible that one could expect to live one’s whole life as an English speaker, or even reach kindergarten, without running into any sentences of the sort” illustrated in (6) to (8) (taken from Pullum and Scholz 2002: 43). In these examples auxiliaries are in bold print and the position of the fronted auxiliary in the corresponding declarative sentences is underlined again. (6) Can a helicopter that has lost its tail rotor __ still fly? (7) Is the boy who was bothering you __ still here? (8) Could those who are coming __ raise their hands? Sampson (1989) also stresses his ability to think of examples of this structure from a poem by William Blake that almost every English-speaking child is likely to encounter. The line Did He who made the lamb make thee? in Tiger does not contain two auxiliaries, but due to the required do-support it still represents a case where the structure-dependency rule should hold, since the order-related rule would have resulted in *Did He who make the lamb made thee. In further support, Sampson (2005: 46) quotes an example of a yes/ no-question taken from Arthur Mee’s Children’s Encyclopdia, which <?page no="36"?> 23 is designed for 6to 10-year-olds: If a man flew above the air would he be able to hear? Again the order-dependence principle would have predicted the incorrect structure *Did if a man fly above the air he would be able to hear? Additional evidence comes from a child language corpus. Pullum and Scholz (2002) explored a single, random file of the Suppes Corpus in the CHILDES database (Suppes 1974). They found several examples of whquestions that also exemplify the structure-dependency principle. During the one-hour recording (N INA 05. CHA ) the caregiver used utterances (9a)- (11a) in her speech to the two-year-old child: (9) a Where’s the little blue crib that was in the house before? (10) a Where’s the other dolly that was in there? (11) a Where’s the other doll that goes in there? An order-dependency rule would have predicted the ungrammatical sentences (9)b-(11)b. (9) b *Where was the little blue crib that __ in the house before is? (10) b *Where was the other dolly that __ in there is? (11) b *Where did the other doll that __ go in there is? Pullum and Scholz (2002) go on to provide further evidence from the 40million-word corpus of the Wall Street Journal. They found that 12% of the yes/ no-questions in the corpus are evidence of the structure discussed and can therefore serve to disprove the potential alternative hypothesis of order dependency. Even though this corpus contains more sophisticated adult language, it still refutes the supposition that people go through “much or all” their lives without any evidence for this construction. The data presented challenge the accurateness of the no-positiveevidence argument for a frequently used example. Additional evidence against the argument comes from connectionist models (Morris, Cottrell and Elman 2000; Reali and Christiansen 2005). Connectionist models are models that explore what can be learned from which type of input with minimal requirements. A simple recurrent network (Morris et al. 2000) and a simple neural network (Reali and Christiansen 2005) were shown to be capable of producing correct complex questions without any positive evidence, solely based on exposure to simpler structures (e.g., simple declaratives with information on agreement, subcategorization and selectional restrictions, simple questions, complex declarative sentences and noun phrase constituency). This finding suggests that even if children did not hear any examples of complex question (which they do, as shown above), they might be able to produce them correctly based on the exposure to simpler structures alone. Finally, Tomasello (2005) also points out that children do not even need examples of complex questions in their input if they are credited with a <?page no="37"?> 24 functional understanding of language. Such functional understanding entails that children understand a phrase, e.g., a noun phrase plus a relative clause, as one single act of reference. If children understand a phrase like the other dolly that was in there in example (10) as a single act of reference to a particular object (dolly), there is no danger that they are tempted to extract an auxiliary from this phrase. A functional understanding of language would thus also allow children to form questions involving such phrases correctly, without the need for an innate structure-dependency principle or previous exposure to such questions. A considerable amount of evidence has been accumulated to counter different aspects of the poverty-of-the-stimulus argument: language to children was found to be very well-formed rather than defective, exposure to language was by no means slight, and the existence of a no-positiveevidence problem seems highly unlikely given the rich input children were shown to receive and potentially even unnecessary if children are credited with insights into the functional nature of language. On the other hand the over-availability of indirect negative evidence and insufficient availability of negative evidence still call for innate restrictions (Universal Grammar or similar) or an alternative way of solving the problem of how the hypothesis space in language learning is constrained. 2.2.3 Learning from the input In order to present itself as a viable alternative to the generative account, the usage-based position must describe how language is learned in absence of an innate Universal Grammar (or similar structure) and the related abstract, formal rules that guide children to acquire the same language as their linguistic environment and impose the necessary restrictions on the acquisition process. The subsequent sections set out to provide such an alternative account by describing how language can be learned from the input and how the abstractions and generalizations that are formed in this process are otherwise constrained. Forming generalizations According to the usage-based account, children’s intention-reading abilities are responsible for their understanding of the functional, pragmatic side of language. Children learn symbols, i.e., form-meaning pairs, in scenes of joint attention by reading their interlocutor’s communicative intentions and attaching meanings to the linguistic forms the interlocutor uses. Because their intention-reading skills combine with the powerful cognitive abilities of pattern-finding, more abstract form-meaning constructions can be formed as well. Pattern-finding comprises the abilities to segment the in- <?page no="38"?> 25 put, to form schemas, categories and analogies, and to perform distributional analyses over the input (Tomasello 2003: 4, 2005). More details about these abilities are given in what follows. In order to segment the input into meaningful units, it is vital to detect recurrent patterns. Humans, including young infants, have been found to be particularly good at that. Saffran, Aslin and Newport (1996) showed that 8-month-old infants can identify word-like units in a continuous speechlike stream of sounds. The stream consisted of nonsense syllables. Certain syllable combinations were thought of as ‘words’, even though they did not carry any meaning. Children were able to identify these recurrent syllable combinations by relying exclusively on transitional probabilities, i.e, the probability of one syllable following another. These probabilities are higher within than between words. 10 In real, meaningful language, children can additionally make use of stress patterns, phonotactic information and, crucially, the meaning or function in order to individuate words in the speech stream (Jusczyk and Aslin 1995; Jusczyk, Hohne and Bauman 1998). Based on meaningful words or word strings that are individuated in this manner first abstractions are formed. Children’s emerging abstraction skills are evident in the development of linguistic as well as non-linguistic schemas. Early in their first year of life, infants begin to form sensorymotor schemas based on simple reflexes like sucking or prehension (Piaget 1935/ 1970: 120-121, 139). When executing such an action repeatedly and on different objects, e.g., grasping a baby’s rattle, grasping a ball, grasping someone’s thumb, a prehension schema is formed and new objects can be assimilated to it. The fact that children can apply the schema to different objects demonstrates the slot-and-filler structure of the schema, e.g., grasp X. The sensory grasping schema is similar to pivot schemas in language, which children extract from their input. If children hear Throw the ball, Throw the pillow, and Throw the box repeatedly in their input and see the respective actions at the same time, the more abstract slot-and-filler schema Throw the [X] can be formed (Tomasello 2003: 174). New fillers can then be put in the slot and subsequently extend the schema. In order to learn more abstract constructions from the input and from previously formed simpler schemas children make use of analogies. Analogies are understood as mappings of relational structures (Gentner 1988; Gentner 2003; Gentner and Markman 1997). So-called item-based schemas (e.g, [X] hugs [Y]) are built by aligning two or more simpler structures and abstracting over them (e.g., John hugs [Y], [X] hugs Susan.); abstract, fully schematic schemas (e.g., [Agent] [Verb Phrase] [Patient]) are 10 The authors give the real language example of pretty baby. The probability that pre is followed by ty is higher than the likelihood that ty is followed by ba. <?page no="39"?> 26 formed by structurally aligning several item-based schemas (e.g., [X] hugs [Y], [A] kicked [B]). In these fully schematic schemas, only the relational structure between the respective elements remains the same in different instances of the construction, while all the concrete lexical material varies with each instantiation. Distributional analyses are performed on the language input simultaneously in order to form paradigmatic categories (Tomasello 2003: 172). The emergence of this ability is closely linked to conceptual development: Linguistically and non-linguistically, children group together things that have similar functions and behave in similar ways (Mandler 2000). According to Nelson (1974) things are what they do, e.g., different dogs bark, bite and play, they like to be patted and so forth. In the same manner, different nouns serve similar communicative functions (e.g., they refer to things that can be thrown in Throw the [X], or more generally to objects or routines) and are very frequently found in similar linguistic co-texts (e.g., the [N]s, a [N], some [N]s, etc.). For instance, children might start forming a verb category by grouping items like hit, kick and eat together because all of them refer to actions and can appear in the slots of frames like I’m [X]ing it, He’s gonna [X] it, etc. (Ambridge and Lieven 2011: 85), before later extending the category to include verbs that are not transitive (e.g., sleep) or do not refer to (visible) actions (e.g., think) because they share their filler quality with familiar verbs with respect to other slots. Two studies by Lieven and colleagues underpin the idea of the formation of progressively abstract constructions and paradigmatic categories with empirical data from naturalistic language learning (Lieven et al. 2003; Lieven et al. 2009). In these studies corpus data from five English-speaking children aged 2; 0-2; 1 were used. For six weeks children were recorded five times a week for one hour each. All the multi-word utterances they produced in the last hour of recording were traced back to earlier productions of the same child. Lieven and colleagues did that in order to analyze how abstract children’s constructions were and how much they relied on previously used word strings. They found that 20-40% of children’s utterances at 2; 1 were verbatim repetitions of previous utterances. Another 40-50% of their utterances differed from previous ones by only one substitution. This means that roughly half the utterances children used could be described as pivot schemas with one variable slot. The remaining 10-40% contained more than one slot. With respect to slot fillers, Lieven and colleagues revealed that the majority of changes to a previously used schema involved the substitution of the referent (usually a noun or noun phrase). With increasing mean length of utterance children used fewer exact repetitions of previous material, formed a higher number of slots and were more flexible <?page no="40"?> 27 with slot-fillers. Schemas thus increased in their abstractness and distributional slot-filler categories increased in size, variability and number used. There is an advantage to conceptualizing language as constructions of varying degrees of abstractness with different numbers of variable slots and viewing learning as forming the relevant abstractions over the input. It allows the accommodation of so-called schematic idioms, which poses a problem in the generative account. Fillmore, Kay and O’Connor (1988) distinguish between substantive idioms, which are completely lexically filled (e.g., kith and kin ‘family and friends’, all of a sudden ‘suddenly’, tickle the ivories ‘play the piano’), and schematic idioms, which are only partially lexically filled (e.g., the [X]er the [Y]er ‘proportional increase or decrease’, [N]th cousin [M] times removed ‘cousin of the Nth degree M generations apart’, pull [X]’s leg ‘tease X’). They further categorize idioms with respect to their lexical, syntactic and semantic characteristics, resulting in the classification given in Table 1. Table 1. Classification of idioms proposed by Fillmore and colleagues (1988: 506- 510). Category Lexical, syntactic and semantic characteristics Substantive example Schematic example unfamiliar pieces unfamiliarly arranged lexically, syntactically and semantically irregular kith and kin the X-er the Yer familiar pieces unfamiliarly arranged lexically regular, but syntactically and semantically irregular all of a sudden Nth cousin M times removed familiar pieces familiarly arranged lexically and syntactically regular, but semantically irregular tickle the ivories pull X’s leg For the usage-based account neither substantive nor schematic idioms pose any difficulties. Substantive idioms are learned as fixed strings of form and function. Schematic idioms are also learned as units of form and function. The slots are formed through the experience of variation in the relevant positions and can then be filled with new material. Idioms are thus treated in the same way as any other units of language - as form-meaning pairings with or without variable slots. In the generative account the distinction between grammar and lexicon must be taken into account. Substantive idioms can be accommodated by the lexicon, since they are entirely entirely fixed, provided that multi-word units are allowed in the lexicon. Schematic <?page no="41"?> 28 idioms, on the other hand, do not fit into the grammar system because most of them have their own irregular syntax that runs counter to other grammatical rules. But neither do they rightfully belong to the lexicon, since they contain variable positions that cannot be filled based on regular rules. Nevertheless, because the idiosyncrasies of a language are usually part of the lexicon, placing schematic idioms there seems to be the more likely choice, in particular because they must be rote-learned. If this placement entails that learning partially lexically-filled schematic idioms with idiosyncratic meanings is possible in generative accounts, the question of whether all language knowledge might not be learnable in this way arises. This in turn is precisely what the usage-based position argues (Tomasello 2003: 102-105). If children’s social-cultural intention-reading abilities are thus combined with their cognitive skills, children are able to learn more and less abstract constructions including different types of idioms from the input. However, there must be ways in which children’s potential generalizations are constrained, since more generalizations are theoretically possible than in fact occur in the language. In the generative account this problem is solved by postulating innate constraints. The usage-based solution is brought forward in the following. Constraining generalizations According to the usage-based account, it is not hypotheses about the target language, but generalizations that call for constraints. Goldberg (2006: 93- 104) and others (Bybee 2010: 95; Bybee and Scheibman 1999; Bybee and Slobin 1982; Tomasello 2003: 175-181, 2005) argue that entrenchment, statistical preemption and the openness of constructions conspire to limit potential generalizations. Entrenchment is understood as the independent storage of word forms or constructions as a result of high frequency (Croft and Cruse 2004: 292, 308-310). Statistical preemption refers to repeatedly witnessing a competing construction (Goldberg 2006: 93, 95) in a situation in which a particular other construction is expected. Restrictions, e.g., semantic or phonological restrictions, on the items that can occur in a construction define its degree of openness (Goldberg 2006: 93). Entrenchment Entrenchment involves the independent storage of a construction resulting from high frequencies of occurrence (Langacker 1987a: 59-60). The idea is <?page no="42"?> 29 that every time a token or type 11 of a construction occurs, it causes the activation of the constructional schema in the mind. Frequent encounters lead to higher degrees of representational strength and cause the storage in the form of a unit (Croft and Cruse 2004: 292), which entails that the construction or type in a construction becomes “progressively entrenched” with “repeated use” (Langacker 1987a: 59; cf. Chapter 4 for a more detailed account). The following two examples illustrate the effect of entrenchment for frequent types in a construction. If a lexeme frequently occurs in a particular construction leading to its entrenchment in this construction, it becomes unlikely that its use is overgeneralized incorrectly to a different construction, in which it has not been experienced. Evidence for this relation comes from a study with high-frequency and low-frequency fixed transitivity verbs, i.e., verbs that are used either exclusively transitively or exclusively intransitively (Brooks, Tomasello, Dodson and Lewis 1999). From 3 years on, children were more likely to use low-frequency verbs, such as vanish, in incorrect constructions than high-frequency verbs, such as disappear, e.g., they were more likely to produce *He vanished it than *He disappeared it. The presumed reason is that high-frequency lexemes were more deeply entrenched in the correct construction and thus resisted generalization to an incorrect construction, in which the lexeme had not been heard. A second example comes from inflectional morphology. Several studies revealed that children are more likely to overgeneralize the English regular past tense to low-frequency irregulars than to high-frequency irregulars (Bybee and Slobin 1982; Maratsos 2000; Marchman 1997; Maslen et al. 2004; cf. Chapter 4). These two examples reveal that entrenchment induced by high frequency contributes to constraining potential generalizations at different levels of constructional abstractness by limiting children’s willingness to use a frequent lexeme in a construction that is has not been witnessed in. Entrenchment and the related issue of representational strength are further discussed in Chapter 4, where frequency effects are in focus. Preemption Statistical preemption is a particular type of indirect negative evidence. In cases where two very similar constructions can be used (e.g., the doubleobject and the prepositional-object ditransitive) preemption may cause learners to start avoiding one of them: “In a situation in which construction A might have been expected to be uttered, the learner can infer that construction A is not after all appropriate if, consistently, construction B is 11 Token frequency causes the entrenchment of types of concrete and more abstract constructions. Type frequency is necessary for the entrenchment of more abstract constructions with variable positions (cf. Chapter 4). <?page no="43"?> 30 heard instead” (Goldberg 2006: 96; see also Boyd and Goldberg 2011b; Braine and Brooks 1995: 367 on “unique argument structure preference”; Brooks and Tomasello 1999; Goldberg 2011; Tomasello 2003: 178). The repeated and consistent occurrence of construction B thus leads to B’s entrenchment and A’s consistent failure to occur in the same situation (thus providing indirect negative evidence) to A’s preemption. Preemption is, for instance, found in morphological constructions, where this process is known as blocking. The use of *goed is preempted by went because we hear it consistently in situations were *goed should have been at least as appropriate. Similarly, children preempts *childs, men and women preempt *mans and *womans (Goldberg 1995: 30). In the case of argument structure constructions, the situation is slightly more complicated because different constructions usually have slightly different meanings (Goldberg 2006: 95). However, as long as a context licences two different constructions, preemption can apply. Children’s sensitivity to preemptive information has been shown in several studies (Braine and Brooks 1995; Brooks and Tomasello 1999; Theakston 2004) 12 and has further been modelled computationally (Alishahi and Stevenson 2005; Regier 1996: 59-80). Brooks and Tomasello (1999) demonstrated that preschool children are able to use preemptive information of novel verbs appearing in certain argument structure constructions. Children were exposed to a novel fixed-transitivity verb. Children in the preemption condition heard the novel verb in a preemptive construction, e.g., they heard a novel intransitive verb in the construction The doll is helping the car tam (‘The doll is helping the car move up the hill.’) rather than the transitive. Children in the no-preemption condition heard the novel verb only in a non-informative construction, e.g., the intransitive construction The car is tamming. In response to questions that called for the agent as the subject, children in the preemptive condition were significantly less likely than children in the no-preemption condition to produce an intransitive novel verb in the non-licensed transitive construction *The doll is tamming the car. The exposure to the alternative, intransitive help [verb] [X] construction, thus preempted the use of the theoretically “equally” possible transitive construction. Additionally, entrenchment might have been at work at the same time, since the alternative help-construction might have become a little entrenched during training at least in comparison to the transitive construction, in which the novel verb was never heard. Children in the nopreemption condition on the other hand discounted their information as uninformative, which is why their generalizations were unconstrained. 12 Adults have also been shown to be sensitive to preemptive information and to discount non-preemptive information (Boyd and Goldberg 2011b). <?page no="44"?> 31 In the previously reported study by Brooks and colleagues (1999), in which frequency-induced entrenchment shielded children from incorrect overgeneralizations of argument structure constructions, preemption is likely to have played a role as well. The reason is that children were more likely to have experienced high-frequency rather than low-frequency verbs in preemptive constructions, e.g., they are more likely to have been exposed at some point in their development to sentences such as He made the bunny disappear, whose occurrence is likely in a magician context, than to sentences such as He made it vanish, because vanish is generally much less frequent, in particular in the speech to children. These examples illustrate that the two mechanisms of entrenchment and preemption often work together to prevent incorrect generalizations. Evidence of preemption from experimental studies might soon be complemented by corpus findings. Goldberg (2011) demonstrated how such evidence might be retrieved from corpora. She (2011: 135) suggests that preemption in corpora is determined by the conditional probability of a verb i appearing in construction B (CxB), given a context where construction A (CxA) is at least as appropriate (12): (12) P(CxB| a discourse context at least as suitable for CxA and verb i ) (13) P(prepositional transitive 13 | (double-object transitive or prepositional transitive) and verb explain ) For instance, the probability for explain to appear in the prepositional transitive in contexts where the double-object transitive is theoretically equally appropriate (13) is approaching 1, since explain is used almost exclusively in this pattern. Token frequency must also be taken into account, since it determines how certain it is for an item to be used in a particular construction (statistical preemption). Witnessing explain once in the prepositional transitive and never in the double-object transitive should not lead to the preemption of the double-object transitive yet; after 100 instances, however, it would be reasonable to preempt this alternative. While the probability equalled 1 in both scenarios, so-called confidence increased. Goldberg (2011) suggests that the natural log function captures the confidence growth. A combination of probability and confidence then provides ample evidence for speakers to constrain their generalizations appropriately. In the future, analyses of child language corpora following Goldberg’s suggestions might serve to reveal relations between preemptive evidence provided by the caregiver and productions by the child. 13 Goldberg (2011) uses the term dative instead of prepositional transitive. The term prepositional transitive is preferred in this work, because it is seen as the more fitting term for English, where case is not overtly marked. <?page no="45"?> 32 Finally, a non-linguistic study supports the findings of children’s abilities to use preemptive information in learning and shows that children are further able to discount irrelevant information (Gergeley et al. 2002; see also Meltzoff 1988). One-year-old infants witnessed an adult turn on a light-box by leaning forwards and touching the switch on its top with her forehead. In one condition the experimenter’s hands were free, in the other condition her hands were occupied (pretending to be cold she wrapped a blanket around herself and held on to it with both hands). After watching the experimenter, the infants were asked to turn on the light. In the handsfree condition 69% turned it on with their heads; in the hands-occupied condition 21% used their head. The fact that the adult used her head when her hands were free, that is, when the (more usual) alternative of using the hands was equally available, preempted the use of their hands in infants. However, when the hands were occupied, children assumed this to be the reason for the head action and discounted the use of the head on this basis. They thus only used the right kind of information for preemption. These findings suggest that children’s preemptive abilities are not exclusively linguistic but of a more general social-cognitive nature. Pattern openness and coverage The third factor that has been proposed to limit possible generalizations in language learning is pattern openness. Pattern openness restricts the application of a pattern, for instance to certain phonological or semantic contexts, and therewith affects its productivity (Bybee 1995; Prasada and Pinker 1993). An example of a phonologically restricted pattern is the English irregular past tense schema [C (C) (C) velar/ nasal] past (clung, sung, stung), which applies to many base forms that take the form [C (C) (C) velar/ nasal] present (cling, sing, sting; Bybee 1995; Croft and Cruse 2004: 297-298). Such restrictions tend to be related to lower type frequencies, which have generally been revealed to make generalizations less likely. Constructions with higher type frequency on the other hand are more likely to be extended to new types because their extension has already been witnessed repeatedly (Bybee 1995; Clausner and Croft 1997; Goldberg 1995: 134). Goldberg (2006: 99-101) proposes that a particular aspect of pattern openness, a pattern’s coverage, restricts generalizations with respect to the semantics in cases where entrenchment and preemption do not apply. Coverage relates to the semantic similarity between lexemes usually occurring in a construction and a new lexeme that is to be inserted in the relevant position. Coverage is “the degree to which previously attested instances fill a semantic space that includes the potential target” (Goldberg 2006: 100). It determines whether the new previously unattested combina- <?page no="46"?> 33 tion of construction and lexeme are comprehensible in communication. In example (14) (Goldberg 2006: 100) the occurrence of sneeze in the causedmotion construction is possible, because its meaning can be construed as semantically similar to the meanings of verbs that regularly occur in the construction. Sneeze is similar to blow in that both involve air, it is similar to knock since both are non-volitional. The categorization of sneeze with these verbs explains why it might be used in this construction and why native speakers consider it fully acceptable. (14) She sneezed the foam off the cappuccino. Goldberg argues that the use of sneeze in the caused-motion construction cannot be explained in terms of entrenchment or preemption. It occurs with such a low frequency that its entrenchment would be impossible and it is semantically and pragmatically too specialized for there to be another construction to preempt it (Goldberg 2006: 98). Coverage is also responsible for speakers’ expectations as to the potential contexts in which a novel, previously unfamiliar, lexeme might be used, when entrenchment and preemption cannot apply (Goldberg 2006: 103). Summary The usage-based account suggests that several mechanisms collaborate in order to constrain potential generalizations. Entrenchment and preemption guide children towards the right uses and to the appropriate extensions when more than one alternative is available. Pattern openness and coverage lead children to appropriate categorizations in cases where there are no competing patterns or when pattern frequency is too low for the regular occurrence of an alternative. Altogether, the input provides the learner with an abundance of information and cues for constraining generalizations. 2.3 Conclusion This chapter has pursued the question of whether there is the possibility that children learn language from the input without being guided by innate language-specific structures or constraints. The idea of learning from the input forms a central part of the usage-based view of language learning and is contrasted by the more classical, generative account of language acquisition. According to the generative view it is impossible to learn the entire language system on the basis of the input. The input is seen as too deficient as to allow language learning. This is why humans are thought to be endowed with innate, uniquely-human, language-specific structures. These contain the innate core grammar (Universal Grammar, or, in more <?page no="47"?> 34 recent models, the language faculty (e.g., Hauser et al. 2002), which allows humans to form abstract rules. The innate structures and the rules based on them are thought to be necessary in order to constrain hypotheses about the target language so that children’s grammars do not become overgeneral. Apart from the core grammar, language knowledge comprises the lexicon, which is assumed to be learned. Children and adults are credited with the same language knowledge. Differences between children’s and adults’ language use are explained in term of as-yet unset rules (parameters) or performance limitations. The usage-based theory, on the other hand, holds that humans bring with them the unique social-cultural ability to read intentions. This skill is thought to lay the foundation for communication and the use of symbols. Together with domain-general cognitive processes, such as categorization or analogy formation, children are thought to learn all language structures from the input. Language knowledge is understood as an inventory of constructions with varying degrees of abstractness. Children’s inventories are different from adults’ in that they are less comprehensive, more itembased and concrete and contain fewer abstractions. The reason is that children are still in the process of building their inventories. The cognitive processes they rely on are thought to be the same ones that adults use. Generalizations of more abstract constructions are constrained by other means than innate structures. The processes of entrenchment and preemption as well as restrictions as to the openness of a pattern are assumed to work together in preventing inappropriate generalizations. To conclude this chapter, it is important to emphasize one point. All models of language learning agree that children must learn all the idiosyncratic aspects of their language (e.g., lexemes, phonology, presumably idioms) and that this learning is input-based. The differences between the generative and the usage-based position lie in the extent of learning from the input, the mechanisms involved and the form of language knowledge. Proponents of generative approaches consider it impossible that all language knowledge is learned from the input, in particular the highlyabstract, formal and semantically empty rule system that they assume to be part of people’s language knowledge. Because of the gap between the input and this complex system, innate structures are presumed and different learning mechanisms are assumed to be necessary for acquiring aspects of this abstract core (e.g., set parameter) and the periphery. The usage-based position on the other hand starts from the null hypothesis that socialcultural and domain-general cognitive processes are sufficient to learn all language knowledge - conceptualized as form-meaning pairs of different degrees of abstractness - from the input. Given the evidence presented in the present chapter, the usage-based position is taken as a starting point in <?page no="48"?> 35 this work. 14 Following from the idea of learning from the input is the assumption that the input characteristic frequency will affect the cognitive processes in all areas of language learning. Before this proposal is explored theoretically in Chapter 4 and empirically in Chapters 5-7, the cognitive processes involved in learning come into focus in Chapter 3. 14 This procedure does not exclude the possibility that additional assumptions about innate structure may turn out to be necessary. <?page no="49"?> 36 3 The learning process This chapter explores the cognitive processes that have been proposed to be involved in language learning from the input. Their role in the language learning process is briefly introduced first; then the two main processes of categorization and analogy formation are presented in more detail (Tomasello 2003: 4, 174). To this end more general background is provided for each of them. This has the added advantage that similarities of these processes in linguistic and non-linguistic processing might surface and shed light on their domain-generality. Subsequently, categorization and analogy formation are compared. This comparison is followed by redefinitions of these two and related terms with regard to the language learning process. A number of steps that children might take in the construction learning process are then proposed. Finally, the redefinitions and steps are applied to the learning of constructions and paradigmatic categories. 3.1 Starting point: Categorization and analogy in language learning Tomasello labels the cognitive abilities 1 involved in language learning “categorization” or “pattern-finding” (2003: 4). They are understood to involve - “the ability to form perceptual and conceptual categories of ‘similar’ objects and events”, - “the ability to form sensory-motor schemas from recurrent patterns of perception and action”, - “the ability to perform statistically based distributional analyses on various kinds of perceptual and behavioural sequences” and - “the ability to create analogies (structure mappings) across two or more complex wholes, based on the similar functional roles of some elements in these different wholes” (Tomasello 2003: 4). The formation of perceptual and conceptual categories of similar objects and events is, for instance, necessary in word learning. Only if children are able to group together situations that are similar in a certain respect will they be capable of learning and using a linguistic label for an entire group of referents. In Tomasello’s description of this ability, the detection of similarity between objects or events is fundamental. Even though similarity has 1 The term ability is used interchangeably with the term (cognitive) process, since both are meant to refer to ‘the ability to perform a cognitive process’. <?page no="50"?> 37 been claimed to be too vague and fuzzy a concept to be used in scholarly theories and have any explanatory power (Goodman 1972) because the exact dimensions of similarity are often difficult to specify, it seems to be a powerful factor in cognitive processing that cannot be left unacknowledged. Indeed Quine noted that there is “nothing more basic to thought and language than our sense of similarity” (1969: 116). The present chapter supports this view in revealing the relevance of similarity in categorization and analogy formation. The ability to form sensory-motor schemas, which is relevant for learning behavioural patterns and constructional pivot schemas, has already been mentioned briefly (cf. 2.2.3; Tomasello 2003: 123-124). In this skill the likeness of linguistic and non-linguistic learning becomes evident. Twoyear-olds’ ability to learn a behavioural sequence was shown in a study by Brown and Kane (1988). After being taught how to pull one object, children were able to transfer ‘pulling’ to other objects and thus learned the schema ‘pull something’. Tomasello proposes that the same cognitive ability is required for learning the equivalent language pattern, i.e., the pivot schema Pull [X] (2003: 14). In both cases learning requires appropriate 2 input of similar examples 3 of pulling behaviour, which needs to be paired with the Pull [X] construction in pivot schema learning. Certain type variation and token frequency in the slot position are thought to allow the schematization of the variable slot. Functionally-based statistical distributional analyses of behavioural and perceptual sequences are thought to be used to fill variable slots in constructions, e.g., in pivot schemas, with “the right kind of word” (cf. 2.2.3; Tomasello 2003: 169-173). Members of paradigmatic categories like nouns or verbs tend to carry similar communicative functions and to co-occur with particular linguistic material. Nouns are often associated with concrete objects, they occur in similar constructions (e.g., Look, a [X]), and are frequently preceded by determiners, such as a or the, and followed by the plural morpheme {-s}. Verbs are often related to actions, they occur in other patterns (e.g., Let’s [X]) and are frequently followed by the morphemes {-s}, {-ed}, {-ing} and sometimes preceded by auxiliaries, such as am, is, are, was, were etc. Keeping track of frequencies of functionally-similar cooccurrences allows children to form paradigmatic categories. 2 Tomasello (2003: 124-124) points out that there are no studies that address “the simple question of exactly what kinds of linguistic experience children must have in order to form a productive slot in a pivot schema”. So, the ‘appropriate’ input has as yet to be determined. 3 Example, exemplar and instantiation are used interchangeably to refer to one realization of a construction with certain linguistic material, i.e., a token exemplifying a type. The term member is also used to refer to a token instantiating a type and highlights the fact that the realized type is a member of a constructional category. <?page no="51"?> 38 The ability to form analogies across complex wholes is thought to be necessary for more abstract construction learning (Tomasello 2003: 4). Tomasello points out that this process “is very like the process of schematization” in categorization; “it is just that analogies are more abstract” (2003: 164). Just as in pivot schema formation input with appropriate type variation is needed. However, in contrast to pivot schemas, item-based schemas have only one stable linguistic item (e.g., [X] hugs [Y]) and abstract schemas do not contain any recurring linguistic material (e.g., [Agent] [Verb Phrase] [Patient]). It becomes obvious that the processes suggested for construction learning at different levels of abstractness are very similar. Tomasello states that “[o]ne special form of schematization is analogy - or, alternatively, […] one special form of analogy is schematization. Both exemplify the process by which children try to categorize […] linguistic constructions” (2003: 298). Categorization thus seems to serve as a superordinate term for schematization and analogy. On the other hand, categorization and analogy formation are often understood to potentially result in schematization (e.g., Holyoak 2005; Langacker 2000b: 4). In order to clarify the relation between these processes, categorization and analogy formation are analyzed in greater detail in what follows, not least with a view to schematization. Both concepts were originally borrowed from psychology and the related processes have been heavily researched in both psychology and linguistics. The concept of analogies perhaps still features more prominently in psychological accounts, whereas the idea of categorization was taken up by many (mostly cognitive) linguists and refined in various linguistic accounts. The reported literature 4 was selected to reflect this circumstance and to provide a basis for the discussion of the role of two processes in usage-based language learning, which lies within the area of overlap between linguistics and psychology. 3.2 Categorization—The prototype account 3.2.1 Concept and process Development of prototype theory Categorization profoundly influences how humans make sense of the world. The automaticity and naturalness of this process becomes apparent when somebody lacks the ability to categorize. Borges’ Funes is a fictional example of such a person (1964: 68). 4 Please note that a comprehensive review of categorization and analogy in the fields of psychology and linguistics is not attempted. Instead the reported literature serves the subsequent explication of the two processes in language learning (cf. 3.4). <?page no="52"?> 39 Funes remembered not only every leaf of every tree of every wood, but also every one of the times he had perceived or imagined it. […] He was, let us not forget, almost incapable of ideas of a general, Platonic sort. Not only was it difficult for him to comprehend that the generic symbol dog embraces so many unlike individuals of diverse size and form; it bothered him that the dog at three fourteen (seen from the side) should have the same name as the dog at three fifteen (seen from the front). His own face in the mirror, his own hands, surprised him every time he saw them. This example illustrates what categorization entails. It is the implicit process of grouping together elements that are “judged equivalent for some purpose” (Langacker 2000b: 17). It allows humans to draw inferences about new situations on the basis of known, familiar ones. In this way, categorization contributes to the acquisition of knowledge similarly to analogies. But categories further provide a means of retaining and representing previously learned knowledge (Kruschke 2005). For a long time, mental categories of how we perceive the extralinguistic world were assumed to be arbitrary and determined by the language of the categorizer. This view followed Whorf’s linguistic relativity assumption that “all observers are not led by the same physical evidence to the same picture of the universe, unless their linguistic backgrounds are similar” (1956: 214). Research into colour naming in different languages (Brown and Lenneberg 1954; Lenneberg 1967) supported Whorf’s idea, since colour terms and the number of colour terms vary considerably between languages and were thus seen as arbitrary. However, a comprehensive investigation with speakers of 20 different languages called this assumption into question (Berlin and Kay 1969). Participants were shown 329 colour chips that varied in saturation, hue and brightness. They were asked to name the basic colours of their language and point out the focal point and the outer boundary for each colour their language had a name for (Berlin and Kay 1969: 5). Again the number of colour terms varied considerably between languages. However, despite this fact, speakers of all languages converged on the same reference points for each colour they had a name for. These reference points were subsequently called focal colours by Berlin and Kay (1969: 7-10; see also Kay and McDaniel 1978). The universal preference for focal colours suggests that categorization is rooted in human perception. Berlin and Kay’s insights further challenged the classical Aristotelian theory of categories. The fact that people consistently chose certain colour shades as most typical necessarily entails that others were seen as less typical. This circumstance runs counter the classical idea of categories, according to which categories are abstract containers with distinct boundaries. Category membership is determined by the presence or absence of clearlydelimited necessary and sufficient features that are common to all mem- <?page no="53"?> 40 bers. Membership is thus an all-or-nothing characteristic and all category members are seen as equally representative. Since this classical theory proved unsuccessful in accounting for the colour research results, alternative explanations were required. Rosch extended Berlin and Kay’s findings and developed the so-called prototype theory of categories (Rosch 1973, 1975; Rosch and Mervis 1975; Rosch, Mervis, Gray, Johnson and Boyes-Braem 1976). She demonstrated that many categories are organized around a perceptually salient prototype or a best example that is more accurately remembered, more rapidly produced and acquired earlier than other category members. In contrast to classical categories, prototype categories allow and indeed predict membership to be graded: from best examples, to less typical ones and to those whose membership is debatable. Rosch and Mervis (1975) showed this gradience of typicality in their goodness-of-example experiments, where participants were asked to judge how good an example of a certain category several (potential) members were (e.g., How good an example of the category FRUIT is apple/ banana/ date? ). Labov (1973) provided further evidence for graded category structure as well as for the vagueness of category boundaries. He asked participants to name drinking vessels, such as the ones depicted in Figure 1. The finding that certain vessels were easy to categorize, whereas there was little consensus for others, illustrates graded membership. Participants’ decreasing consensus on borderline cases demonstrates a certain degree of fuzziness at the edges of the category. Classical theory was unable to account for this phenomenon. Figure 1. A selection of vessels used in Labov’s study (further pictures without handles, with content etc. were also used; taken from Labov 1973: 354 5 ). 5 Reprinted with permission from “The boundaries of words and their meaning” (p. 354) by W. Labov from New Ways of Analyzing Variation in English edited by Charles-J. Bailey and Roger W. Shuy, 340-373. © 1973 by Georgetown University Press. www. georgetown.edu. All rights reserved. <?page no="54"?> 41 With graded category structure and fuzzy boundaries new means of determining category coherence were required, which were different from classical necessary and sufficient features. Wittgenstein (1958: §66) discerned how categories cohere in his discussion of GAMES . He revealed for the category GAMES that its members share different attributes with some and other attributes with other members. Consider for example the proceedings that we call ‘games’. I mean boardgames, card-games, ball-games, Olympic games, and so on. What is common to them all? […] Are they all ‘amusing’? Compare chess with noughts and crosses. Or is there always winning and losing, or competition between players? Think of patience. In ball-games there is winning and losing; but when a child throws his ball at the wall and catches it again, this feature has disappeared. Look at the parts played by skill and luck; and at the difference between skill in chess and skill in tennis. Think now of games like ring-a-ring-a-roses; here is the element of amusement, but how many other characteristic features have disappeared! And we can go through the many, many other groups of games in the same way; we see how similarities crop up and disappear. There is, however, not a single attribute that all members share. Category members are held together by a “network of overlapping similarities” (Ungerer and Schmid 2006: 29). Rosch and Mervis (1975) showed how these overlapping attributes again support the graded category structure found in prototype theory. They asked participants to list attributes of members of a category, e.g., BIRD . Good examples shared the highest number of attributes with other category members; bad examples shared only a few with others. Since sharing several, but not all, attributes suffices for category membership, damaged examples can be accounted for as well. A bird that has only one wing left after an accident can still be categorized as a bird even though it might no longer be able to fly. Such examples posed a problem to the classical account, because a necessary feature, i.e., the ability to fly, was violated. Prototype effects, category representation and categorization of new members Within prototype theory, there are different conceptions of what a prototype is, what a prototype represents and whether the category is represented solely by the prototype. These issues are explored in the following. Prototype effects The prototype may be understood as the most typical (Langacker 1987a: 371) or salient member that comes to mind most readily, or as the member that is most frequent or unites the most frequent attributes. Alternatively, the prototype may be seen as a central tendency, the average or as an ideal- <?page no="55"?> 42 ized caricature, which is as distinctive from adjacent categories as possible (Kruschke 2005). Or it might be a more abstract representation of what is common to all members. Such an integrated structure that consists of “the commonality that emerges from distinct structures when one abstracts away from their points of difference by portraying them with lesser precision and specificity” is sometimes also referred to as a schema (Langacker 2000b: 4). A prototype in this sense includes more abstract knowledge about a category. Another position is proposed by many cognitive linguists (Aitchison 2003: 70; Croft and Cruse 2004: 95-105; Langacker 1987a: 147; Taylor 2009: 87-91; Ungerer and Schmid 2006: 47-59). They see prototypes as representatives of internal theories people hold. 6 These theories comprise patterns of cultural and common knowledge, observations, experience, beliefs and assumptions about the world (Aitchison 2003: 70; Taylor 2009: 91). The prototype might thus be an actual category member, an idealized member, a schematization of commonalities or a representative of internal theories people hold. Basic level categories typically have prototypes as well as subordinate categories. Nevertheless, some categories, usually superordinate (e.g., FURNITURE ) and functional (e.g., THINGS TO TAKE TO A DE- SERT ISLAND ) categories, do not possess perceptually stable prototypes. They may, however, have a prototype in terms of a schema based on the common functions of their members. Category representation The discussion of what constitutes a prototype already introduced the question of how category knowledge is represented. If the prototype is an actual example, the question is whether any additional information apart from the members themselves is stored. ‘Pure’ exemplar theory assumes the mere learning of instances (see Chandler 2002, for a review). It is thought that there are no abstractions based on the category’s exemplars, neither in terms of necessary and sufficient features nor in the form of a more abstract prototype. Due to different frequencies and interrelations between members it is expected that members differ in centrality. However, “the exemplar view seems to take away the ‘categoriness’ of categories” (Ross and Makin 1999: 215). It is questionable that exemplars can be perceived as a category at all, if only exemplars are stored. It seems difficult to determine category membership of new stimuli or to form category la- 6 Internal theories have also been labelled mental models, frames, scripts, internalized cognitive models (ICMs) or cognitive domains (Aitchison 2003: 72). When referring to the temporal sequencing of events, the term script is preferred (Taylor 2009: 91). While internal theories are mostly seen as relatively stable, Croft and Cruse (2004: 98-99) imagine each situation to result in a particular construal of an internal theory. They argue that this understanding allows interpretations that fit the context. <?page no="56"?> 43 bels without any more abstract representations. In fact, Ross, Perkins and Tenpenny (1990) showed that people learn more general aspects about a category, even if the experimental design forces them to resort exclusively to exemplar-based comparisons. Certain schematizations thus seem to be the inevitable consequence of comparison processes in categorization. Based on such evidence Abbot-Smith and Tomasello (2006) and Ross and Makin (1999) argue that it is impossible to store only exemplars. Instead, they propose that exemplars and abstractions are represented in the mind (exemplar-abstraction model) and that both the exemplars and the schematized abstraction might be involved in categorizing new stimuli. According to this account there is no reason why exemplars and abstractions over several exemplars should not be stored together. Categorization of new members The question of how new potential members are categorized is closely related to the question of how a category is represented. Langacker assumes that new “elements are assimilated to the category on the basis of their perceived resemblance” to the prototype, which is a typical instance in his terminology, or to the schema (1987a: 371). He envisages this process as “capture by an attractor” (2000b: 7). The stimulus that is to be categorized activates a number of established categories. The activated categories are thought to be represented by a prototypical exemplar or a schema, which constitute the potential standards of comparison. The new stimulus is compared to the standards of the activated categories. The standard that is most fully activated is selected to categorize the stimulus (see connectionist modelling for very elaborate accounts of learning by activation spreading, e.g., Elman, Bates, Johnson, Karmiloff-Smith, Parisi and Plunkett 1996; McClelland, Rumelhart and the PDP Research Group 1986; Rumelhart, McClelland and the PDP Research Group 1986). If the stimulus fulfils all specifications of the standard of the most highly activated category, Langacker refers to the process as instantiation or elaboration. In this case he claims the standard of comparison to be a schema. The only difference between the new member and the schema is that the new member is usually more specific and detailed. If the discrepancies between the standard of comparison and the stimulus are more considerable but below a certain threshold, Langacker labels the process extension. This is because the category is broadened through the acceptance of the new member (Langacker 1987a: 371-372). In this case Langacker claims that the standard of comparison is an exemplar rather than a schema. How the adequate standard of comparison (schema, prototypical exemplar) is determined and how the two suggested modes of categorization are differentiated in prac- <?page no="57"?> 44 tice 7 remains open. Langacker’s account is one example of a similaritybased approach. 8 The comparison process involved in categorization might be depicted 9 as illustrated in Figure 2: Category membership of a new stimulus (the object in the circle in (a)) involves comparing it to a highly similar exemplar (b), a highly frequent or particularly salient exemplar or prototype (c) or a more abstract representation of category commonalities (d) of the most highly activated category ( BIRD ). Figure 2. The categorization process. (Pictures courtesy of Sarah-K. Siebenborn). 3.2.2 The development of linguistic categories When a 1-year-old calls a sheep in the field “sheep” after having learned this word in reference to a different sheep at a different time, he or she performs an act of linguistic categorization. In order to use language adequately children need to learn to use a linguistic form for the appropriate 7 Langacker (1987a: 372) admits this himself: “[T]he two modes are sometimes difficult to distinguish in practice”. He does not, however, consider this missing discriminability a “matter of concern” (1987a: 372). 8 Similarity-based views of categorization are common in psychology as well, where similarity is often modelled in terms of spatial distance between exemplars (Nosofsky 1986; Shepard 1962a, 1962b; Tversky 1977; Tversky and Gati 1978). 9 This figure is solely for illustration. Due to its graphic character it is limited to the expression of visual attributes. This is not to imply that all or even the majority of attributes are of visual nature. <?page no="58"?> 45 meanings or, conversely, to figure out how the world is partitioned in their language at varying levels of constructional abstractness (see packaging task in Aitchison 2003: 191-196). Children’s learning of the meaning scope of a linguistic form is gradual and piecemeal and there is much evidence that it depends on prototypicality. In lexical, morphological and syntactic learning children begin with more prototypical meanings and proceed to learn less typical ones. Aitchison (2003: 194-195), for instance, reports that a child, Eva, started out using the meaning of ‘kicking a ball’ as a prototype for kick presumably based on certain attributes, e.g., a waving limb, sudden sharp contact between body and object, forward propulsion of the object, and progressed to less prototypical uses (Bowerman 1978; Griffiths 1986). In morphological learning, English-speaking children have been shown to acquire the central sense of the past tense morpheme or the progressive morpheme first. They tend to first associate {-ing} with ongoing actions, e.g., bathing, and {-ed} with a change of state, e.g., broke, and only later extend the progressive to stative verbs or the past tense to activity verbs (Clark 1996; Li and Shirai 2000: 149- 184; Shirai and Anderson 1995). Goldberg (2006: 76, see also Goldberg, Casenhiser and Sethuraman 2004) showed a similar progression for abstract constructions. She reanalyzed data from the Bates, Bretherton and Snyder corpus (1988) and showed for three different constructions ([Subject] [Verb] [Adverbial], [Subject] [Verb] [Object] [Adverbial], [Subject] [Verb] [Object 1 ] [Object 2 ]) that the verb used most frequently in the construction by caregivers was the one that is most prototypical of the meaning of the entire construction (go, put, give respectively). As a consequence, children used each construction predominantly with verbs expressing the prototypical meaning initially and extended the category to less typical meanings only later. 10 Children’s initial underextension of the constructional categories as well as their progression from more to less prototypical meanings is in line with the prototype theory and with the usage-based understanding of language learning (Taylor 2009: 268-269). The reason is that children will often lack sufficient background (e.g., internal theories/ cognitive domains) to use a linguistic form appropriately in all possible contexts (Taylor 2009: 276). Consequently categories crystallize around the representation of the prototype initially and new members are added based on similarity to the prototype. Children’s initial uses of a linguistic form, i.e., the category label, thus tend to be narrower than those of adults. It is possible that children’s and adults’ semantic prototypes of a linguistic category differ initially, but children’s prototypes gradually move towards the adult ones. 10 But see Ninio (1999) for an alternative account. <?page no="59"?> 46 In contrast, classical categorization theory requires children to acquire the relevant features of each category. If children lack one or more of these features, overgeneralization errors occur. With every feature that is added correctly or removed after incorrect attribution, children are expected to perform a leap towards the correct use of the construction. This account is not in keeping with children’s developmental progression, since neither leaps nor frequent overextensions are characteristic of the process (Taylor 2009: 268). Conversely, the use of linguistic labels has been shown to be helpful in the process of forming categories of similar objects, even if children were not previously familiar with the labels or nonce (invented) labels were used. Labels were revealed to facilitate categorization in children, presumably because they highlight commonalities between objects or invite children to look for such commonalities (Kemler Nelson 1995; Sloutsky and Fisher 2004; Waxmann 1999; Waxman and Booth 2003; Waxman and Markow 1995). 3.2.3 Summary Categorization and categories serve as a means for humans to organize and represent the world. The classical view of categories which sees them as being defined by necessary and sufficient features fails to account for prototype effects shown in many studies. Prototype theory on the other hand predicts these effects. Members are linked by shared attributes, but not all members share all attributes with all other members. As a consequence, category membership is graded and fuzzy towards the boundaries. Category knowledge is represented in category members, but according to exemplar-abstraction models also in the form of more abstract schematizations (prototype; schema). Comparisons between a potential new member and familiar members or more abstract representations allow the categorization of new members. Prototype effects have also been shown in children’s learning of linguistic categories (i.e., constructions at different levels of abstractness). They progress from more prototypical to less prototypical meanings. Categorization is facilitated by linguistic labels, presumably because they highlight similarity between examples. 3.3 Analogies 3.3.1 Concept and process The second cognitive process that is proposed to be at work in constructional learning is analogy. A very well-known example of analogy used in numerous textbooks of cognitive psychology is the fictive radiation prob- <?page no="60"?> 47 lem. In a first step of their study, Gick and Holyoak (1980: 307-308) confronted participants with the so-called tumour problem: Suppose you are a doctor faced with a patient who has a malignant tumor in his stomach. It is impossible to operate on the patient, but unless the tumor is destroyed the patient will die. There is a kind of ray that can be used to destroy the tumor. If the rays reach the tumor all at once at a sufficiently high intensity, the tumour will be destroyed. Unfortunately, at this intensity the healthy tissue that the rays pass through on the way to the tumor will also be destroyed. At lower intensities the rays are harmless to the healthy tissue, but they will not affect the tumor either. What type of procedure might be used to destroy the tumor with the rays, and at the same time avoid destroying the healthy tissue? Most participants found it very difficult to find an appropriate solution. Only 10% of the participants spontaneously proposed a successful, feasible solution. However, telling participants the fortress story changed the situation (Gick and Holyoak 1980: 352). A small country fell under the iron rule of a dictator. The dictator ruled the country from a strong fortress. The fortress was situated in the middle of the country, surrounded by farms and villages. Many roads radiated outward from the fortress like spokes on a wheel. A great general arose who raised a large army at the border and vowed to capture the fortress and free the country of the dictator. The general knew that if his entire army could attack the fortress at once it could be captured. His troops were poised at the head of one of the roads leading to the fortress, ready to attack. However, a spy brought the general a disturbing report. The ruthless dictator had planted mines on each of the roads. The mines were set so that small bodies of men could pass over them safely, since the dictator needed to be able to move troops and workers to and from the fortress. However, any large force would detonate the mines. Not only would this blow up the road and render it impassable, but the dictator would then destroy many villages in retaliation. A full-scale direct attack on the fortress therefore appeared impossible. The general, however, knew what to do. He divided his army up into small groups and dispatched each group to the head of a different road. When all was ready he gave the signal, and each group marched down a different road. Each group continued down its road to the fortress, so that the entire army finally arrived together at the fortress at the same time. In this way, the general was able to capture the fortress, and thus overthrow the dictator. After hearing this story and being encouraged to use it to solve the tumour problem, 75% of participants suggested that the doctor should administer low-intensity rays from different angles at the same time, so that they all meet on the tumour resulting in high-intensity treatment of the tumour <?page no="61"?> 48 without harming healthy tissue due to the low intensity of individual rays (Holyoak and Thagard 1995: 112-114). Finding this solution involved a comparison between the two stories that linked the fortress to the tumour, the army to the rays and the general to the doctor. The tumour and the fortress with the dictator constituted the respective danger or problem. The general was unable to send his entire army down one single route because the mines would have exploded, just as the doctor was unable to send all rays down the same way for danger of damaging healthy tissue. The analogy is incomplete at first, because there are different routes to the point of danger only in the fortress story. However, if the comparison increases awareness of the similar relations between the other components, the missing part of the analogy can be filled in and solves the tumour problem: Sending the rays down different routes avoids damage to the healthy tissue and allows the successful treatment of the malignant tumour. The tumour problem exemplifies a definition of analogies given by Holyoak (2012: 234), which conceives the compared entities in terms of source and target. The latter refers to a representation of a new situation; the former to the representation of a well-known situation. Figure 3 illustrates the definition and steps involved. Figure 3. Major components of analogical reasoning (figure taken from Holyoak 2012: 236 11 ). 11 Reprinted with permission from Chapter 13 “Analogy and Rational Reasoning” (p. 236, Figure 13.1) by Keith J. Holyoak from The Oxford Handbook of Thinking and Rea- <?page no="62"?> 49 Two situations are analogous if they share a common pattern of relationships among their constituent elements even though the elements themselves differ across the two situations. Typically […] the source or base is more familiar or better understood than the […] target. […] This asymmetry in initial knowledge provides the basis for analogical transfer—using the source to generate inferences about the target. Upon encountering the target, the source is usually retrieved from memory (in the tumour example the source was in fact only introduced after the presentation of the target). Objects and relations of the source and the target situations are aligned and mapped to each other. This structuremapping process is crucial to the analogy, because commonalities and differences between the two situations become apparent when the best “structurally consistent match” is determined (Gentner 2003: 201). This matching process is guided by three principles: 1) one-to-one mappings: each element in one situation corresponds to one element in the second situation, 2) parallel connectivity: when two elements correspond, the elements they govern correspond as well, e.g., the doctor corresponds to the general, the general commands the army, the doctor commands the rays, 3) systematicity: there is a preference for the deepest and richest match over more superficial ones. A successful mapping allows inferences about the target based on more comprehensive knowledge about the source, which in the case of the tumour problem results in its solution. Inferences can also extend or restructure the original source or result in the abstraction of a schema from source and target. In the present example, for instance, a schematization of ‘a central danger can be controlled by applying moderate forces that converge on the danger from different directions’ may be formed. Analogy is thus a powerful mechanism for acquiring new knowledge, but it may also lead to schematizations of commonalities between source and target that are retained for future recall (Gentner 2003: 203- 204). 12 soning edited by Keith J. Holyoak and Robert G. Morrison, 234-259. © 2013 Oxford University Press. www.oup.com. All rights reserved. 12 The capacity of analogy to acquire new knowledge is more controversial than might be expected. Like many psychologists investigating analogy, Gentner does not assume that the tertium comparationis (i.e., the point of comparison) and the potential abstractions are known before the analogy is formed (2003: 203). Under the label the “problem of inductive reasoning” it has, however, been claimed that it is impossible to recognize similarities between situations that lead to abstractions of any sort without knowing the dimension of similarity and having the abstractions a priori (Chomsky 1988: 147). This problem is not testable and impossible to refute on logical grounds (Tomasello 2000a: 241). For this reason it will not be considered any further. <?page no="63"?> 50 3.3.2 The development of analogizing abilities There has been disagreement as to how early children are able to form analogies. Piaget (Piaget, Montangero and Billeter 1977) claimed that analogical reasoning emerges only in adolescence. However, this assumption turned out to be based on results from studies where analogizing abilities were confounded with other factors. For instance, analogy tasks were used that required the application of mathematical operations (e.g., cubing numbers) that children of the relevant age group did not know yet. In marked contrast to Piaget’s position, Gentner (2003: 200-201) assumes that children form analogies from infancy on. She proposes that experiencing temporally close repetitions of the same action highlights very small differences between them and results in ever so slightly more abstract representations of the relevant action. Progressively larger differences are tolerated and lead to increasingly abstract representations. For children from at least three years of age, there is experimental evidence for their ability to form systematic analogies. Goswami and Brown (1989) used a picture version of a is to b as c is to d (a: b: : c: d) analogies where d was missing. Children between 3 and 6 years were tested on relations that suited their age, e.g., melting or cutting. Following Goswami (1991) the choice of appropriate relations is particularly important, because knowledge about a certain domain (e.g., the fact that certain things can melt or be cut) is a prerequisite for the formation of analogies in the respective area. In Goswami and Brown’s study children were asked to choose one out of five options to fill in the missing part of the picture analogy. Performance increased from 37% to 91% correct from the youngest to the oldest children, but given the number of response choices even the youngest group were above chance level. Goswami and Brown assume that it is in fact children’s understanding of domains rather than their analogizing abilities per se that develops with age and results in a higher number of successful analogies. They thus oppose the Piagetian position that analogical reasoning itself is late-developing. Gentner (2003) proposes that the ability to form relational analogies receives an extra boost in humans (compared to other species) because of their use of relational language. Relational language includes but is not limited to spatial, causal and social-communicative expression such as in, on, buy, sell, target or parent. Initially, children tend to interpret such expressions non-relationally, e.g., preschool children describe uncles as “friends” that are “usually male” and “in a certain age range” (Keil and Batterman 1984: 224), before understanding their relational character, i.e., grasping that uncles are the brothers of one’s mother or father. Nevertheless, even at a very young age children profit from the use of familiar relational expressions in analogizing tasks. For instance, the number of children’s correct analogies increased in a matching task where objects in two scenes were to <?page no="64"?> 51 be mapped to each other based on size if the familiar relational terms baby/ mommy/ daddy were used to refer to small/ medium/ large objects respectively (Rattermann and Gentner 1998). The presumed reason for the facilitative effect of language is that it preserves the aligned relations linguistically and thus makes them more salient and easier to identify in new situations (Gentner 2003; Gentner and Loewenstein 2002; Gentner and Rattermann 1991; Loewenstein and Gentner 1998, 2005). 3.3.3 The role of similarity Several factors that influence analogy formations (i.e., structure alignment and mapping) in children and adults have been proposed, most importantly similarity. It has been shown that mapping of source and target is easiest and the most straightforward in cases of literal similarity (i.e., sameness: when source and target are identical) or high concrete similarity between source and target (Gentner 2003; Gentner and Kurtz 2006: 636). Similarity of rich matches, e.g., two identical dachshunds, is perceived as higher than that of sparse matches, e.g., two identical circles (Gentner 2003: 200). The role of relational similarity increases developmentally. Gentner and colleagues have shown in a series of studies that children initially compare two situations rather holistically. This phase is followed by special attention to particular objects that are identical or similar across the compared situations. Finally, children shift their attentional focus towards the relations that are similar in the two situations. This shift away from object similarity towards relations has been named relational shift hypothesis (Gentner 1988, 2003; Gentner and Rattermann 1991; Kotovsky and Gentner 1996; Markman and Gentner 1993; Rattermann and Gentner 1998). The relational shift is often accompanied by an attentional move from perceptual to functional and conceptual commonalities. The developmental shift from object to relational similarity is mirrored in children’s and adults’ information processing: Both children and adults process object similarities before relational ones. The relational shift allows for the formation of analogies in absence of supporting object similarity. Evidence for the relational shift hypothesis comes from numerous studies in several areas. One study explored the shift in children’s and adults’ interpretations of metaphors (Gentner 1988). Attributional and relational metaphors were investigated. Attributional metaphors require responses based on perceptual object similarity; relational metaphors call for relational interpretations. 13 The number of attributional responses to attribu- 13 A snake is like a hose is an example of an attributional metaphor, e.g., both share attributes in that they are long and wiggly. A cloud is like a sponge exemplifies a relational metaphor, e.g., both share similar functional relations - they hold and give off water. <?page no="65"?> 52 tional metaphors was stable across preschoolers, primary school children and adults, whereas relational responses to relational metaphors increased with age. There was a shift from a sole focus on perceptual-configurational similarity to functional and conceptual commonalities. Crucially, this shift towards relational similarity was shown not to be a move away from object similarity but rather an additional, more in-depth consideration of relations (Kotovsky and Gentner 1996). Object and relational similarity might also work together. Object similarity facilitates the recognition of relational similarity in situations where object similarity is in line with relational commonalities (DeLoache 1990; Gentner and Markman 1997; Gentner and Medina 1998; Gentner and Toupin 1986; Holyoak and Thagard 1995: 83-84; Keane 1987). While younger children often need object similarity to guide them to relational commonalities, older children are able to rely solely on relational similarity. Nevertheless, even adults form analogies more readily if they are supported by object similarity, i.e., if objects belong to the same domain (Keane 1987). Moreover, the repeated exposure to examples with concrete similarity facilitates the subsequent recognition of higher-order relational similarity (progressive alignment hypothesis; Kotovsky and Gentner 1996). In cases where object and relational similarity are pitted against each other there are certain factors that determine which match - the object match or the analogy - prevails. This question has been explored in crossmapping studies. Crucially, cross-mapping tasks permit both an object match and a relational match. They usually involve pictures of two scenes expressing the same relation. One object re-occurs in both scenes. For instance, a woman receives food from a delivery man in one scene and feeds a squirrel hazelnuts in the other. The non-analogical object match would match the woman in both scenes, whereas the relational match would link the two givers of food (delivery man and woman), the two receivers of food (woman and squirrel) and the different types of food (Markman and Gentner 1993). When more than a single match has to be made or participants are asked to judge the similarity between the scenes before making the match, analogical relational matching prevails over object matching. The likely reason is that similarity rating brings object and relational similarities to attention. Since relational matches are perceived as deeper, richer and more powerful than object matches (see systematicity principle), analogical relational matching is preferred. This preference of relational over object similarity has also been shown in studies where people are asked to rate the similarity of two situations: People tend to judge situations as more similar when they are based on relational rather than object similarity (Goldstone, Medin and Gentner 1991; Markman and Gentner 1993). There are, however, certain characteristics that can tip the scale towards object matches. A move toward object matches is prompted, when the coherence <?page no="66"?> 53 of the potential relational match is decreased or object attributes are enriched, i.e., when the salience of relations is minimized or the salience of objects is maximized (Markman and Gentner 1993). To sum up, relational similarity is by definition a prerequisite for analogy formation. The structure alignment and mapping process ultimately “promotes relational commonalities over common object properties” and thus analogy formation over alternative mappings (Gentner 2003: 201). Supporting object similarity is not necessary but it facilitates the process and presumably sets it in motion in young children. Figure 4. The role of supporting object similarity in analogy formation. Object similarity decreases from left (a) to right (d), whereas relational similarity is given in all four cases. Figure 4 illustrates the facilitative effect of object similarity, which decreases from left to right in the depiction. In the leftmost situation (a) literal similarity (i.e., sameness) makes the structure alignment and mapping process very straightforward. Relational similarity is necessarily a given as well, but reliance on object similarity alone allows for the comparison. The schema that might potentially be abstracted is no more abstract than the two examples (source and target) it is based on. In (b) object similarity is high and might thus guide the analogizer into seeing that relations are alignable. The potential schema is slightly more abstract than the examples. In (c) object similarity is lower but still present. Nevertheless the reliance on relational similarity becomes increasingly vital to align source and target. The potential schema takes a more abstract form as well. Example (d), depicted on the right, involves no object similarity and is based solely on similar relations. Analogizing over source and target might thus result in a fully abstract schema. Relational similarity is given in all four cases but its significance increases as the supportive object similarity decreases. The <?page no="67"?> 54 degree of object similarity is related to the degree of abstractness of the potential schema. The figure does not show cases of cross-mapping where object and relational similarity conflict. But as reported earlier, it is assumed that generally analogical relational matches prevail unless object features are highly salient or relational similarity is incoherent. 3.3.4 Summary Analogies serve the acquisition of new knowledge through a comparison of a new target and a better known source and by drawing inferences about the target based on the source. Both situations are aligned structurally, such that relations are mapped to each other. In this mapping each element in the source must correspond to one element in the target; elements that are governed by corresponding elements must correspond as well; the deepest, richest match is selected. As a result of analogies a schema that represents the commonalities of the two situations may be formed and retained in memory. Children shift their attention from similarity in objects towards similar relations. The development of this preference of relational over object similarity eventually allows them to form analogies even if object similarity suggests a less appropriate (possibly non-analogical) alternative or if only relational similarity is given. Relational similarity might be supported by congruent object similarity. Relational language facilitates the focus on relations and with it the alignment and mapping of relations. 3.4 Starting point revisited: Categorization and analogy in language learning 3.4.1 Comparison of categorization and analogy After separate discussions of categorization and analogy formation the two processes are compared in this section. Categorization is the process of grouping together similar items, usually under a common label. Initial category formation as well as the categorization of potential new members entails a comparison process. In the latter case the standard against which the potential exemplar is compared has been the topic of much debate. Regardless of the standard of comparison, the driving force behind the process is the similarity between the standard and the potential member. Similarity does not bear on perceptual aspects alone, but also involves the similarity of attributes that other members, the prototype or schema share with the potential member. For initial category formation, there is much less debate as to the standard of comparison, presumably because it is logically required that it is two examples that are compared. Categorization is seen as a process that is based on people’s knowledge of the world. Cate- <?page no="68"?> 55 gories provide a tool to impose order on how we perceive the world and to represent this system. Analogies also involve a comparison. The structures of a better-known source and a new target are aligned and mapped so that relations between the two situations match. Potential results are the extension of knowledge about the target based on what is known about the source and the abstraction of a schema that represents source and target and may be but is not necessarily stored in memory and labelled linguistically. A successful alignment process presupposes relational similarity between the source and the target and may be facilitated by object similarity if it points towards aligning the two situations in the same way. Young children tend to focus on object similarity and might thus fail to form analogies in crossmapping situations. Their developing attention to relations is manifested in the relational shift, which results in the tendency to favour relational over object matches because they are perceived as deeper, richer and more powerful. What categorization and analogy share is thus a comparison whose outcome is determined by similarity. A potential difference between the comparison involved in the categorization and the comparison in analogy formation is the respective standard of comparison. According to Langacker (1987a: 371-372) a category member (typical member) or the schema (here: commonalities between members) may serve as the standard in the categorization of new potential members. In analogy formation, the standard is usually thought to be an example, i.e., an instance at the same level of abstractness as the target rather than a schema (Langacker 2000b). 14 The problem of thus attempting to distinguish categorization and analogy is that it is usually impossible to determine the standard since only the result of the comparison can be measured and the standard itself is not accessible to investigation. Moreover, initial category formation is very similar to analogy formation because both involve a comparison of exemplars. Given these similarities it is not surprising that categorization and analogy formation are not usually distinguished by the standard of comparison. In fact, there are a number of authors who use the terms categorization and analogy interchangeably at times (e.g., Gentner and Namy 1999; Kotovsky and Gentner 1996; Tomasello 2003: 4). Whether the two terms are indeed interchangeable is explored by a simple test in the following section. Categorizations are usually phrased as an A is a C (as is B), e.g., a retriever is a dog (as is an Alsatian). Figure 5 illustrates this example. The linguistic term DOG is the hyperonym and label of the category. It stands for the commonalities between exemplars 14 Analogies are formed directly “on the model of others [other exemplars], not on the basis of stored abstracted patterns” (Langacker 2000b: 59). <?page no="69"?> 56 that are already established or for internal theories about dogs. Retriever and Alsatian are two examples of dogs. DOG schema (category label, hyperonym) exemplar A exemplar B retriever Alsatian (hyponym) (hyponym) Figure 5. Illustration of the categorization of a retriever and an Alsatian as DOGS . (Pictures courtesy of Sarah-K. Siebenborn). Analogies on the other hand are usually phrased as A is like B in way C, e.g., an atom is like the solar system with respect to its configuration. 15 This analogy is illustrated in Figure 6. The nucleus is mapped to the sun; nucleus and sun are the centre of the respective systems. The electrons are mapped to the planets. The relation between nucleus and electron is mapped to the relation between sun and planets. Figure 6. Illustration of the analogy “an atom is like the solar system”. (Picure of atom courtesy of Thorben Cordes, picture of solar system courtesy of Sarah-K. Siebenborn). If the two terms categorization and analogy formation are indeed interchangeable, then it should be perfectly possible to describe the category DOG as an analogy and the atom-solar system analogy as a category. Retrievers are like Alsatians in that their builds, attributes (e.g., quadrupeds, tails), characteristics (e.g., barking) and so forth can be mapped (see Figure 7). Conversely, the atom and the solar system might both been seen as 15 At least, early 20 th century models of the atom, which have since been refuted, relied on the structural similarity expressed in this analogy (Bohr 1913a, b, c). nucleus : sun electrons : planets nucleus : electrons : : sun : planets : <?page no="70"?> 57 exemplars of the category PHYSICAL MODELS INVOLVING ORBITAL STRUCTURE , which, of course, does not reflect a hypostatized concept for most people. Figure 7. Analogy between a retriever and an Alsatian and categorization of an atom and the solar system as examples of PHYSICAL MODELS INVOLVING ORBITAL STRUCTURE . However, there is one problem with this view. The change in perspective does not seem to work for all types of categories. It works well with basic level and subordinate categories, but problems arise with superordinate or functional categories. Different members of the superordinate category FURNITURE , e.g., a chair and a cupboard, cannot be structurally aligned and be compared successfully in an analogy. The same is true for functional categories, such as THINGS TO TAKE TO A DESERT ISLAND . The question is which factor determines whether analogy and categorization can be used interchangeably or not. The determining factor proposed here is that comparisons in categorization need not be based on the structural alignment of parallel relations. This is true for both comparisons in the categorization of new members as well as for comparisons in initial category formation. In contrast, analogies necessarily involve the alignment of parallel relations. The question is whether and, if so, how this difference manifests in construction learning. 3.4.2 Redefinitions of categorization and analogy in construction learning If the major difference between analogy and categorization in fact lies in the relational alignment requirement for analogies, it follows that the two terms can be used interchangeably in situations where the relational requirement is fulfilled. Moreover, the term analogy is used to zero in on an initial comparison that may lead to schematization. The term category formation also focuses on the initial comparison, but presumes more strongly that its consequence is a schematization. The term categorization on the PHYSICAL MODELS INVOLVING ORBITAL STRUCTURE : <?page no="71"?> 58 other hand focuses on the admission of new potential members to an existing category. These differences suggest that it might be useful to distinguish between an initial comparison process and a comparison process to categorize new members. These points are now elaborated with respect to linguistic constructions. In order to form clear-cut definitions of the processes involved in language learning I refer to the reported literature as well as to the reflections in the previous sections. The exemplars and the schema of a linguistic construction do share the same relational structure (Tomasello 2003: 326). The nature of constructions therefore entails that comparisons necessarily involve the structural alignment of parallel relations. This suggests that the terms analogy and categorization can indeed be used interchangeably in the present case. Because of the distinction between category formation and categorization made above, the use of the terms in construction learning is specified further in the following. The initial comparison process allows the abstraction of a schema; subsequent comparisons mainly serve the categorization of new members, but may also refine the schema. I understand these comparison processes to be of uniform nature, seeing that relational alignments are involved in all cases. Because of this relational mapping, I also use the term analogy to refer to any of these comparisons. The term initial comparisons can be used interchangeably with category formation. The only difference between the two terms is their focus - initial comparison highlights the comparison itself, whereas category formation highlights the consequences of the comparison process, that is, the formation of a schema and/ or a category. Schemas are understood as more abstract or more general representations of the commonalities of the compared instances (Langacker 2000b). I use the term schema to refer to these commonalities regardless of whether they are just momentary occurrences during the comparison process or achieve long-term representation, so-called well-entrenched representation (schemas versus schemas with unit status in Langacker’s terminology, Langacker 2000b). My use of the term is further not restricted to a certain level of abstractness or granularity, but includes simple schemas as well as more general and more abstract structures that developed from simpler ones (Langacker 2000b). Initial comparisons, in which schemas can be formed, are thought to be based on two or more exemplars, because at this point there are no more abstract schematizations available yet. With respect to the comparison involved in the categorization of new members, it has become clear that it is as yet impossible to determine the standard of comparison in individual cases in practice (Langacker 1987a: 371-372, 2000b). Consequently, my use of the term comparison in the categorization of new potential members and the term categorization that I <?page no="72"?> 59 also use in this case are free of any presuppositions as to the hypothesized standard in individual cases, but with one exception. I assume the principle that “[m]ore specific knowledge [usually] takes precedence over more general knowledge” holds (Lakoff 1987: 147 about Wilensky’s Law; Wilensky 1983: 25, 145). This principle is certainly true in everyday situations. If asked about a friend’s dog’s behaviour, I will first consult my knowledge about dogs of the same breed that I know, then about dogs in general and finally about animals in general. A reason for such behaviour is given by Langacker (2000b: 16). He proposes that the standard of comparison 16 is selected on the basis of the “degree of stimulation”, which is “roughly proportional to the number of features shared”. Therefore, “lower level schemas” and possibly individual exemplars “have a built-in advantage in the competition with respect to higher-level schemas”, because they are more finely grained and hence tend to share a higher number of features with the target than more abstract schemas, where individual differences are levelled out. Following this logic, similar exemplars should generally be the preferred standards of comparison, followed by low-level schemas and then by increasingly general higher-level schemas in all types of comparisons. In individual cases, there are, however, presumably additional factors that influence the selection of the standard. My understanding of categorization is further very closely related to generalization. Generalization is highly relevant to construction learning research, since evidence for the abstraction of schemas (of fleeting or enduring character) is difficult to provide. My definition of the term generalization diverges from that of others who use it synonymously with abstraction to imply a more abstract level of generality (e.g., Langacker 1987a: 437). Rather, I see generalization as a potential consequence of the successful formation of a constructional category. It refers to the extension of the category by the categorization of a new member in comprehension or production. In comprehension, a new utterance is understood in terms of a known construction. In production, a new member is actively formed on the basis of familiar exemplars or the schema. 17 In this manner categorization of a new member and generalization can be used interchangeably in order to refer to the extension of a category. One final point is important with respect to constructional categories, schemas and generalizations. As mentioned above, it is impossible to de- 16 Or, the potential categorizing structure in his terminology. 17 In the language acquisition literature, generalization is often used interchangeably with productivity. Sometimes productivity is taken to refer to all uses of a particular structure or pattern of a child; sometimes it is used to refer exclusively to extensions beyond the child’s input. In the latter case, productivity is similar to my definition of generalization, but it tends to concern the number of a child’s generalizations rather than generalizations per se. <?page no="73"?> 60 termine whether a schema has been abstracted and stored. However, according to the exemplar-abstraction view it is impossible to perceive a category as such without any more abstract representation of the commonalities of its members (Abbot-Smith and Tomasello 2006; Ross and Makin 1999). 18 This position is adopted here as well. As soon as examples are stored together as a category, a more abstract schema is thought to have been formed and to be stored as well. While this point is very difficult to investigate experimentally, generalizations are easier to reveal. The idea that generalizations are indicative of constructional schema abstraction is also implicitly put forward by Tomasello (2003: 314-320). Within this work, generalizations are thought to require an underlying category and are thus taken as evidence for the preceding formation of a constructional category and its representation in the form of exemplars plus schema. Since the issue of which abstractions humans form and store is a controversial issue in constructionist views of language and language learning, it receives additional attention in the General Discussion (8.3.1). 3.4.3 Steps of construction learning Based on these redefinitions and the reviewed literature, in particular Tomasello’s (2003) understanding of constructional learning and generalization, and the exemplar-abstraction models by Abbot-Smith and Tomasello (2006) and Ross and Makin (1999), several steps of constructional learning are proposed in the following. As a first step, examples of an emergent construction must be retained in memory, e.g., John hugs Mary and Peter hugs Susan. At this point, they are not yet stored as types of the emergent construction. They may, however, be connected to other previously formed constructions, for instance, the constructions of the respective lexemes, the morphological [verb]s construction and so forth. As soon as two or more examples are stored, a comparison can be formed between them on the basis of relational similarity and with potential support of object similarity in part of the construction. In the comparison two or more types are aligned structurally and their relations are mapped to each other (i.e., an analogy is formed). This and potentially further similar comparisons yield a more abstract schema of the construction at hand, e.g., [X] hugs [Y] or [Agent] hugs [Patient]. Together with the exemplars the schema is then stored as a constructional category. As mentioned in the previous section, categories are thus seen as exemplar-abstraction categories. Cate- 18 Ross and Makin assume that not only one general abstraction is formed for each category, but that local abstractions are formed as well and are later used in the categorization of new members. I also expect humans to form local abstractions alongside more general ones. It is merely for ease of the description and reference that I tend to refer to the constructional schema. <?page no="74"?> 61 gory formation allows generalizations. Generalizations in comprehension are categorizations of new exemplars as category members, when the relevant constructional category is the most highly activated one (Langacker’s “capture by an attractor”, 2000b: 7). Whether an exemplar or the schema serve as the standard of the analogical comparison in this categorization remains open. Generalizations in production involve the formation of a new exemplar on the basis of a comparison (analogy) to a familiar example or the constructional schema. The proposed steps are in keeping with previous studies in the usage-based tradition. Some of these studies are discussed in Chapter 4. This book aims to provide further evidence for the proposed steps. 3.4.4 Analogy and categorization in construction learning This chapter closes with a description of construction learning at different levels of constructional abstractness. It takes into account the reflections on and re-descriptions of the cognitive processes involved as well as the proposed steps of constructional learning. Learning is described at the levels of a) concrete lexemes and chunks, b) pivot schemas, c) item-based constructions and d) abstract constructions. Further, learning of e) paradigmatic categories is presented. a) In order to learn small, concrete units of language, children compare similar-sounding instances with highly similar communicative functions (meanings), which they have previously stored from their input. In each realization of a lexeme or chunk, children experience slight differences due to factors that vary, such as different speakers, speed of speech, context, and so forth. Comparing these minimally distinct instantiations resembles stacking highly similar, but not identical, overhead transparencies on top of each other. Tomasello (2003: 124) borrowed this transparency metaphor from Langacker (2000a: 214-215). It highlights that the learning process involves analogical comparisons, i.e., relational alignment and mapping. In lexeme or chunk learning, the comparison allows children to form a slightly more abstract representation (constructional schema) of the word or chunk reflecting commonalities between different instantiations but neglecting minimal differences. A very similar process has been proposed by Gentner (2003) for the formation of non-linguistic conceptual schemas that are only minimally more abstract than the individual instances to begin with. In the present case, the constructional schema of the lexeme or chunk is thought to be subsequently stored together with the different realizations. b) Pivot schemas are more abstract or more general because they are chunks with a variable slot, e.g., Let’s [X] (Tomasello 2003: 123-124). Stacking overhead transparencies on top of each other illustrates which part of the construction remains stable and where the slot is abstracted. This is <?page no="75"?> 62 illustrated in Figure 8. Through the alignment and mapping of several stored examples in a comparison, e.g., Let’s play, Let’s eat, Let’s go, children identify the systematic variability in the slot position and then schematize it at a more abstract level, e.g., Let’s [verb], Let’s [action], Let’s [X]. Precisely how abstract the slot is and how much semantic information it includes is presently unresolved. 19 In the constructional schema this position is consequently more abstract. Figure 9 gives an example of the constructional schema and several exemplars that are thought to be stored together with it in the constructional category. The exemplars illustrate how the slot position can be filled. Figure 8. Depiction of the pivot schema Let’s [X] as transparencies stacked on top of each other. Figure 9. Depiction of the pivot schema Let’s [X] as a constructional category. c) Patterns known as item-based schemas (e.g., [X] kicks [Y]) centre on a concrete lexical item and allow more than one variable slot (Tomasello 2003: 125). They are also formed by structural alignment and mapping of several previously stored exemplars. The schema resulting from the comparison is more general and reveals where variability is possible, e.g., [X] 19 Presently, there is no generally accepted system as to how abstract slots in partiallyfilled constructions or abstract constructions are described. Some authors include semantic information, such as agent, others describe the variable positions merely formally, e.g., noun phrase. Strictly speaking, the choice necessitates investigations as to how children and adults actually represent the individual constructions at different points in time. For the present theoretical discussion, this issue can remain open. Let’s go Let’s play Let’s jump Let’s eat Let’s go Let’s play Let’s jump Let’s eat Let’s [X] Let’s play Let’s jump Let’s eat Let’s go <?page no="76"?> 63 kicks [Y], [X] gives [Y] [Z]. Each category member is a realization of this schema and provides potential slot fillers, e.g., Joseph kicks Jack, Abigail kicks Sam, He kicks her. In the case of item-based schemas, there is less recurrent linguistic material than in lexemes, chunks or pivot schemas, i.e., concrete object similarity is reduced. Concrete similarity is now restricted to a single position. Nevertheless, it might serve to guide the structural alignment of relations in the comparison of different examples. d) In each of the cases discussed so far, at least some of the lexical material was fixed. This concrete object similarity in a part of the construction presumably facilitates structural alignment and mapping. The reason is that concrete similarity of objects involved in the comparison makes the detection of relational similarity easier (DeLoache 1990; Gentner and Markman 1997; Gentner and Medina 1998; Gentner and Toupin 1986; Holyoak and Thagard 1995: 83-84; Keane 1987). Structural alignment is more difficult if only the relations, but none of the concrete linguistic material, match. This is the situation for so-called abstract constructions. Laying transparencies on top of each other results in a completely blurry picture (Tomasello 2003: 164). This is why children have to rely solely on the functional interrelationships of the previously stored examples that are “superimposed” or compared. The resulting constructional categories have schemas that are entirely abstract (e.g., [Agent] [Verb Phrase] [Patient] or [Noun Phrase] [Verb Phrase] [Noun Phrase]). They are stored together with category members that exemplify how the slots can be filled (e.g., The man kicked the woman, A child hugs his mother, The boy hit the girl). The schemas of all the constructions described so far varied in their generality, from least to most general. Despite the fact that it is these patterns that were labelled by Tomasello and selected for more detailed description here, it is important to note that they are points on a scale ranging from schemas that are only slightly more general than their instances to entirely abstract, very general schemas that consist solely of variable slots. Moreover, as has been illustrated in this section, the relationship between constructional abstractness and literal object similarity is inverse. The more abstract a construction the less stable linguistic material is present and the more children are forced to rely on relational similarity exclusively. This is presumably one of the main reasons why children start out with more concrete constructions and progressively build more abstract ones from them. Figure 10 illustrates a)-d). <?page no="77"?> 64 Figure 10. Comparisons of stored examples resulting in schematization. x) Perfect literal similarity (sameness) is virtually never given, since pronunciation and context always vary to some degree. a) Very high object similarity between realizations of a concrete construction. b) High object similarity between instances of a pivot schema. c) Lower object similarity between instances of an item-based schema. d) No object similarity between instances of an abstract construction. e) Parallel to schematizing increasingly abstract constructions, children learn which kind of material can go into the slots. These paradigmatic categories are formed by so-called distributional analyses (Tomasello 2003: 169- 173). Children group together previously stored linguistic items that “play similar communicative roles in the utterances they hear” (Tomasello 2003: 170). For instance, the noun category may develop based on the function nouns usually fulfil: They tend to refer to objects or, as Langacker (1987b) put it, invite the construal of “bounded entities”. Verbs tend to refer to actions and processes and thus invite the construal of experiences as such. These communicative functions are the reason why linguistic items of a paradigmatic category tend to coincide with certain other linguistic elements. Nouns regularly co-occur with determiners or with plural morphemes, which allow the localization of one or more referents in space. Verbs co-occur with tense markers allowing localization of processes and actions in time. Early on, paradigmatic categories are built around prototypical items, e.g., concrete nouns or concrete activity verbs, and are then gradually extended (cf. 3.2.2). Initial category formation as well as category extension is supported by the alignment of structures with similar functions co-occurring with certain linguistic material. Figure 11 illustrates the comparisons involved in the formation of distributional categories based on the functional similarity of elements. <?page no="78"?> 65 Figure 11. The formation of paradigmatic categories. Lexemes with similar communicative functions, which recur in similar positions (in bold green print) and with similar linguistic material (in bold blue print), are grouped together in paradigmatic categories. The likeness of non-linguistic categorization and analogy formation and these processes in constructional learning has become apparent throughout this chapter. It supports the idea of the domain-generality of these processes (cf. 8.3.2). Further support for this assumption comes from factors that were suggested to influence analogical comparisons in both areas. Object similarity facilitates the alignment and mapping of parallel relations in non-linguistic analogies, if it is in line with the relational mapping (DeLoache 1990; Gentner and Markman 1997; Gentner and Medina 1998; Gentner and Toupin 1986; Holyoak and Thagard 1995: 83-84; Keane 1987). The same supportive effect of object similarity was proposed for constructional learning. Repetition was further suggested to be beneficial to nonlinguistic analogy formation, because it causes the entrenchment of concrete similarity which subsequently facilitates the recognition of higherorder similarity (Kotovsky and Gentner 1996). In constructional learning, repetition might likewise strengthen or entrench representations and thus support both the learning of examples and the comparison and abstraction processes. Repetition is realized by frequency in constructional learning. Its role is explored in detail in the next chapter (Chapter 4). To sum up, children are assumed to first form and later extend both syntagmatic and paradigmatic categories by comparisons of functionallysimilar units in a process of aligning and mapping similar relations. The resulting categories are thought to form a highly interconnected network of constructional knowledge. When building this network, children are assumed to proceed from concrete, item-based to more abstract, schematic constructions. Such a progression allows children to perform initial structure mappings with the help of (partial) literal object similarity (e.g., in pivot or item-based schemas) before building more abstract constructions based on functional and relational similarity alone. Repetition (frequency) may further facilitate constructional learning just as it did with nonlinguistic analogies. Give me a book Give me a toy Give me a hug Give me a break Give me some cars Give me some books Give me some toys Give me some grapes Give me the doggie Give me the book Give me the baby Give me the toy <?page no="79"?> 66 4 Frequency effects in language learning The important role of frequency in human cognitive processing was already pointed out in the introduction. Humans encode and store nonlinguistic and linguistic frequency information automatically and unconsciously (Hasher and Zacks 1984; Hasher et al. 1987; Posner and Snyder 1975). Frequencies of occurrences of events thus result in advantages in future processing of similar events (Hasher et al. 1987; Hasher and Zacks 1984). These advantages are evident in learning in general and language learning in particular (Arnon and Snider 2010; Bybee 2006; Diessel 2007; Ellis 2002; Gathercole and Hoff 2009; Lieven 2010; Matthews and Bannard 2010; Slobin 1997). The present chapter starts out by describing how different types of input frequency are thought to affect language representations and language learning. The subsequent, major part of this chapter is then devoted to the discussion of previous research on frequency effects in language learning. This report serves to establish the current state of knowledge and to expose areas where frequency research is still very limited. The review of frequency effects is divided into abstraction levels, from concrete to partiallyfilled and to abstract constructions. Before the major research questions of this book are derived from the gaps in the literature, special attention is drawn to a particular level of constructional abstractness that is proposed to be relevant in constructional learning and plays in important role in the research of this book. 4.1 Frequency in language representation and language learning According to the usage-based, constructionist model language knowledge is represented in interrelated networks of constructions with different degrees of abstractness (Behrens 2009; Bybee 2006; Goldberg 2006: 12; Tomasello 2003: 98-99). Bybee’s network model 1 provides insights into the role of frequencies in such a network (2006, 2010). The idea is that represen- 1 Bybee (2006, 2010: 19-20, 79, 109 and more) refers to her model as an “exemplar model” because she assumes that exemplars are stored in the mind. Since she explicitly expects abstractions to be formed (e.g., Bybee 2006, 2010: 174 and others), her model is by no means restricted to exemplars only. It is thus similar to the exemplarabstraction model proposed by Abbot-Smith and Tomasello (2006) and Ross and Makin (1999, see Chapter 3). <?page no="80"?> 67 tations of exemplars 2 , i.e., types of a construction, and constructional schemas are probabilistic and graded. Evidence of graded representations comes from an abundance of psycholinguistic research, including studies showing that humans are able to recognize a previously encountered event without being able to recall it (Anderson and Bower 1972; MacDougall 1904) and studies of tip-of-the-tongue experiences, where retrieval fails but is felt to be imminent (Brown 1991). Representations are thus not either present or absent, but they vary in representational strength based on frequency. According to Bybee (2006, 2010: 19-20) every token of experience is represented in the constructional network. If a token is identical to an existing exemplar, it is mapped onto it, thereby strengthening it. If a token is not identical to an existing exemplar, it is stored as a new exemplar near similar ones, metaphorically speaking. Every time humans process a token of an exemplar, the processing time is reduced marking a practice increment, the perceptual and motor systems become more finely tuned to the comprehension and production of the type, the speed of access for subsequent recognition and recall of tokens of the same type decreases and the future activation of the same type is facilitated (Bybee 1985: 119, 2010: 24; Ellis 2002; Lieven 2010). High input token frequency thus strengthens memory representations and causes entrenchment in the individual’s mind. Representational strength and entrenchment are often used interchangeably, but it is possible to tease their meanings apart. Representational strength refers to the depth or autonomy of a type’s storage (Bybee 1995), whereas entrenchment concerns the ease of activation and use of a type, i.e., “the degree to which the formation and activation of a cognitive unit is routinized and automated” (Schmid 2007: 119; see also Langacker 1987a: 59-60). Representational strength and entrenchment are related, which is evident in the fact that types with higher representational strength are more easily and more highly activated by the same input than types with lower strength. Strongly represented types are, for instance, more likely than weakly-represented types to serve in the categorization of new structures, i.e., in generalization (Bybee 2010: 78-80; Langacker 2000b). While high token frequency strengthens a concrete, lexically-fixed type of a construction, high type frequency is important for strengthening more abstract schemas. In partially-filled constructions, repeated exposure to different types strengthens the representation of stable positions and at the 2 According to Bybee (2010: 19) “an exemplar is built up from a set of tokens that are considered by the organism to be the same on some dimension”. This definition implies that a certain number of tokens is necessary for the existence of an exemplar. However, Bybee also states that tokens that do not match an existing exemplar are stored as new exemplars. This use of the term is in line with the definition of an exemplar as a type, which is used in the present work. <?page no="81"?> 68 same times makes slot positions apparent. Type frequency thus bolsters the representation of the more abstract schema of the partially-filled construction (Bybee 1995). Variability in the slot positions further allows the formation of categories of items that can fill these slots (Tomasello 2003: 124-125). High type frequency is linked to productivity. The higher the number of different types that have been experienced in a slot (and the fewer restrictions on how the slot can be filled) the more likely new types are to occur in the slot, i.e., the more productive the slot is (Bybee 1995; Bybee 2010: 95). Types with extremely high token frequencies on the other hand are stored deeply and highly independently so that they do not usually contribute to slot formation and productivity (Bybee 2010: 95-96). Type frequency in entirely abstract constructions allows the formation of slots in each position of the construction. Each position thus becomes associated with a category of potential slot fillers. Each type of the construction relates the slots to each other by exemplifying the abstract schema. Type frequency thus provides the variability that is necessary for forming and strengthening the more abstract schemas of partially-filled and abstract constructions. Token frequency ensures the representational strength of the individual types and is important as well. When more abstract constructions are recruited in the categorization of new structures (i.e., generalization), representational strength as determined by type and token frequency is consequently relevant to the degree of activation of the potential categorizing constructions. If constructions fail to be reinforced by types and tokens over a longer period of time, their representations are thought to weaken. The described effects of frequency on representation are relevant to language learning (Tomasello 2003: 106-107). The steps of constructional learning proposed in Chapter 3 can be specified with respect to the role of frequency. Memory of exemplars is more adequately referred to as their representational strength, since it is not an all-or-nothing phenomenon. Token frequency affects the representational strength and entrenchment of exemplars, which is important for several reasons. It is via representational strength that input token frequency is thought to affect the speed and order of acquisition of types. It is further expected that a certain token frequency ensuring relatively stable representation is a prerequisite for a type to participate in a comparison process resulting in schema formation. This comparison process also requires type variation, i.e., at least two types are necessary. Token and type frequency are also important in generalization, because new potential members are captured more easily by highfrequency types and schemas. Type frequency is particularly relevant for the generalization of more abstract constructions, because it determines their productivity. <?page no="82"?> 69 The present section has explored the ways in which frequencies affect language representation and learning. The studies that are presented in the following report empirical evidence of the effects of token frequency, type frequency and type-token ratio on constructional learning at different levels of abstractness. It is assumed that the described effects come about because frequency affects representational strength of types and constructional schemas and the productivity of constructions. 4.2 Frequency and concrete constructions Concrete constructions are words and fixed word strings. Research on frequency in this area so far concentrates on words. For words 3 token frequency plays the major role, because there is no type variation in the usual sense. Of course, tokens exhibit some variation on a smaller scale, because there are often several different morphological forms of a word and because different realizations of a word differ from each other based on speaker, context and additional factors. Input token frequency affects the speed and the order of acquisition of words. Hart (1991) showed this in a corpus study with 11to 17-month-old children. Her analyses revealed that children’s first words were highly frequent in their input. The converse was, however, not accurate. The words with the highest frequencies in the input were not the first ones children learned. The highest-frequency words are usually function words, such as the or of (Gathercole and Hoff 2009: 114). This fact already highlights that frequency is not the only factor that plays a role in word learning, “otherwise we would never get beyond the definite article in our speech” (Ellis 2002: 178). Nevertheless, further studies about the order of verb emergence support the notion that the relation between input token frequency and children’s productions is strong. Naigles and Hoff-Ginsberg (1998) assessed the order of acquisition of 25 verbs in 57 children using speech samples. They found that the frequency of verbs in mothers’ speech was significantly correlated with the order in which children began producing these verbs. Theakston and colleagues (2004) explored the effect of frequency and semantic generality of verbs (e.g., general vs specific verb meaning) on the age of acquisition. They showed that the verbs’ relative input frequency strongly affected children’s verb learning, whereas semantic generality only played a minor role. That word semantics are nevertheless not negligible was revealed in a study by Harris, Barrett, Jones and Brookes (1988). They found a strong relationship between children’s initial word uses and the most frequent meaning in their input: Children tended to use that meaning of a word first that their mothers used most frequently. 3 In this context a word is understood as a pairing of a form and a meaning. <?page no="83"?> 70 Moreover, overall input token frequency affects early vocabulary growth in children. Huttenlocher, Haight, Bryk, Seltzer and Lyons (1991) found that the overall amount of parental speech to their children was proportionate to children’s vocabulary size. Differences in input frequencies of different words were again related to the speed of acquisition, with higher-frequency words being learned faster than lower-frequency ones. A study by Rice, Oetting, Marquis, Bode and Pae (1994) explored the number of tokens per word that normally-developing children and children with specific language impairment required before their successful performance in a comprehension task. It was revealed that younger children and children suffering from specific language impairment required higher token frequencies than older and normally-developing children in order to reach similar levels of lexical comprehension. Goldinger (1998) explored token frequency with respect to the abstractness of word representations. Since each token of the same word is slightly different from the others due to a number of contextual factors, the representation of a lexeme also eventually becomes more abstract than the individual tokens. Goldinger showed that representations of words with lower input frequency were closer to the actual input than those of highfrequency words. As a consequence, pronunciation variations caused by different speakers or distortions of the speech signal were more likely to be tolerated for high-frequency words than for lower-frequency ones. To sum up, high token frequency of lexemes in the input was linked to earlier and faster emergence of these words in child speech as well as to more comprehensive vocabularies in children. The number of tokens necessary for relatively stable representations (as evidenced by correct comprehension) decreased with children’s age. High input token frequencies were further associated with more abstract representations in the child’s mind and higher tolerance of speech signal deviations. 4.3 Frequency and partially-filled constructions The term partially-filled constructions subsumes all constructions that contain both lexically-fixed positions and one or more variable slots (pivot schemas and item-based constructions in Tomasello’s terminology, 2003: 114-121). The development of partially-filled constructions involves the abstraction of the slot positions. Children presumably start with the storage of unanalyzed chunks of language (Arnon 2010: 6; Lieven et al. 2003, Lieven et al. 2009; Tomasello 2003: 114-121). They subsequently segment them and begin to develop slots in variable positions. Token frequency is important for the storage of chunks; type frequency provides variability, which contributes to slot formation. <?page no="84"?> 71 Many morphological constructions are partially-filled, for instance English [verb]ing ‘ongoing action or process’, where the verb position constitutes the variable slot and the morpheme {-ing} is stable. There are also entirely abstract morphological constructions such as reduplications. 4 Despite this fact, the literature on frequency in morphological learning presently focuses on partially-filled morphological constructions. For this reason, morphological constructions are presented in this section. They receive special attention in the present chapter because the empirical studies in this book explore frequency effects in morphology. The part on morphological constructions is followed by research concerning other partially-filled constructions. 4.3.1 Morphological constructions Morphology is usually divided into derivation and inflection. Derivation is traditionally said to form new lexemes from a bound lexical and a free lexical morpheme, e.g., {un}+{do}, while inflection is thought to produce merely new word forms by adding a bound grammatical morpheme to a free lexical one, e.g., {play}+{-s} (Haspelmath and Sims 2010: 18). With respect to frequency effects in language learning, derivational morphology is not very well-researched at all. For this reason the part reporting the relevant literature is relatively brief and illustrates that there is room for further research. Children’s learning of inflectional morphology and frequency effects in this development are much better researched. Derivational morphology There is one study exploring overall input frequency effects on derivational morphology learning in children (Behrens 2003). A few more studies investigated children’s creative overgeneralizations, usually referred to as coinages or innovations when they concern word formation, in relation to the productivity of derivational morphemes (Clark 1993: 137-138; Clark and Hecht 1982; Clark and Cohen 1984). Behrens (2003) explored a special type of prefix verbs in German, i.e., particle verbs. Particle verbs are separable in certain grammatical contexts and their meaning is usually compositional, e.g., weggehen ‘away.go’ - ‘go away’ (cf. 5.1.2 for more details on particle verbs). Because of their separability, both morphemes would more adequately be described as free lexical ones, e.g., {weg}{gehen}. Despite this peculiarity they are usually classified as (derivational) prefix verbs and consequently discussed here. In her study of these verbs, Behrens used a highly-dense corpus of one German- 4 One example is the Pangasinan [-consonant+vowel-] [noun] ‘plural’ as in amigo ‘friend’—amimigo ‘friends’ (Rubino 2001: 540). <?page no="85"?> 72 speaking child, Leo, who was recorded between 1; 11 and 4; 11. She analyzed Leo’s speech and a size-matched sample of his caregiver’s speech with respect to the proportional frequency of particle verbs. Leo’s input was not adjusted over time to Leo’s increasing language abilities. Instead the proportion of particle verbs remained relatively stable throughout the recording interval. Leo’s productions of particle verbs, on the other hand, became increasingly attuned to the distributional properties in the input, i.e., the proportion of particle verbs in Leo’s speech approached those in his input with increasing age. Frequencies in the child’s productions thus mirrored frequency distributions in the input more and more (see Behrens 2006, where the same relationship is demonstrated for different word classes). Crucially, this study only concerns the overall distributions and does not explore how frequency affects the way in which children analyze and learn the separable prefix construction, i.e., whether token frequency strengthens memory representations or type frequency influences schema formation. With respect to more classical cases of derivational morphemes, the effects of productivity on innovations have been explored. Productivity is associated with high type frequency, since it invites further applications of the respective constructional schema (Clark 1993: 138). It was shown that the most productive of several near-synonymous morpheme constructions is learned earliest and used innovatively most frequently, e.g., the English agentive -er in contrast to -ian and -ist (Clark 1993: 120; Clark and Hecht 1982). In a study with novel nouns that all ended in -er, all ended in -ian or all ended in -ist, children were also found to recall the most productive -er best and use it most frequently to substitute the other two morphemes (Clark and Cohen 1984). Children’s preferences for innovative uses were further shown to shift to other constructions if those became more productive in adults’ language (Clark 2003: 288). To sum up, a corpus study showed that the proportions of particle verbs, a sub-group of prefix verbs, in a child’s speech developed to reflect the proportions in his input. High productivity of a morpheme construction, which is accompanied by high type frequency, was further revealed to cause its speedy acquisition and high numbers of innovative uses in children’s productions. Inflectional morphology Research on frequency effects in inflectional morphology focuses on the English past tense and is largely concerned with the order of morpheme acquisition and with children’s errors. These studies provide valuable insights into frequency effects on children’s progression and are reported subsequently. <?page no="86"?> 73 Order of acquisition Brown (1973: 277) and de Villiers and de Villiers (1973) were the first to analyze children’s speech in early child speech corpora in order to assess children’s provision of grammatical morphemes in obligatory contexts. Brown’s corpus analysis resulted in a well-known rank order of grammatical morpheme acquisition in English, which was supported by de Villiers and de Villiers’ analyses. What remained neglected in both studies was whether the rank orders were related to children’s input. Moerk (1978) reanalyzed Brown’s data with a view to the input and revealed that the morpheme that was most frequent in a child’s parental input was the first to reach the 90% correct criterion (provision of the respective morpheme in 90% of obligatory contexts) in that child’s uses. Further, input frequency was shown to be related not only to the order but also to the speed of morpheme acquisition. For each child, there was a correlation between input frequency and age of acquisition. Conversely, morpheme constructions that were not modelled in the input were not present in children’s productions. Morpheme acquisition in second language learning was also revealed to be a function of the frequency of occurrence in the input (Goldschneider and DeKeyser 2001). Errors When learning a particular morpheme, children use frequent inflected forms before less frequent ones and make fewer errors with them (e.g., Rubino and Pine 1998). The presence or absence of several types of errors in children’s inflectional morpheme learning can thus shed some light on how children’s productions are related to input frequency. The following discussion explores how far children’s errors are affected by token frequency and type frequency. Token frequency. Overgeneralizations 5 are one type of errors children produce in development. Overgeneralization involves the application of an incorrect morphological schema (constructional schema). 6 Frequently it is a 5 A frequently-used alternative term is overregularization. This term is misleading because (as will be shown) overgeneralization is not restricted to the regularization of irregulars. Instead irregulars can also be overgeneralized to regulars, i.e., ‘overirregularization’. 6 The idea of different schemas serving in inflection is based on the single-route model of inflectional morphology (Bybee 1995, 2006, 2010; Bybee and Moder 1983; Bybee and Slobin 1982). Children are thought to form all these schemas based on the input. This process is not different for what is traditionally referred to as regular and irregular inflection. Type frequency simply tends to be highest for the regular inflection because it is the most frequent schema, which is also the reason why it is most frequently overgeneralized. The single-route view of inflectional morphology learning is a part of usage-based modelling and therefore presupposed here. It stands in con- <?page no="87"?> 74 schema with higher type frequency that is applied incorrectly to a lexeme whose correct inflection follows a schema with lower type frequency. For the English past tense the regular [verb]ed is the schema with the highest type frequency and the schema that is overgeneralized most frequently. The role of token frequency is examined in this section, the role of type frequency, which was already mentioned here is revisited in the subsequent section. Early research examined only overall rates of overgeneralization of irregulars. Marcus and colleagues (1992), for instance, averaged corpus data of 25 children and found an overgeneralization rate of 4.2%. However, their analysis neglected the potential influence of token frequency. Maratsos (2000) reanalyzed the same data 7 and revealed that the types with the lowest token frequency in children’s productions exhibited much higher rates of overgeneralization than high-frequency types, e.g., the overregularization rate for verbs with token frequencies between 1 and 9 in children’s speech was 54% rather than 4.2%. A corpus study by Maslen and colleagues (2004) supports this finding and additionally relates children’s overgeneralizations not only to token frequency in their own speech but also in the input. The authors used a high-density corpus that covered 8-10% of the productions of an Englishspeaking child, Brian, between ages 2; 0 and 3; 11. Again, overgeneralization rates varied with token frequency in Brian’s own speech. In addition, there was a significant negative correlation between token frequency Brian’s input and his overgeneralizations (r = -0.62). Table 2 gives the mean overgeneralization rates with respect to token frequencies in Brian’s and his caregiver’s speech. The mean overgeneralization rate increased when token frequency in Brian’s and his caregiver’s speech decreased. It was highest, at 23.1%, for verbs Brian produced between 1 and 9 times and his caregiver used about 50 times on average and lowest, at 1.85%, for verbs Brian produced over 100 times and his caregiver used nearly 1000 times on average. trast to the dual-route model, according to which innate rules are responsible for regular inflection while irregulars are learned from the input and stored in the lexicon (Pinker and Prince 1988; Pinker and Ullman 2002; Marcus, Pinker, Ullman, Hollander, Rosen and Xu 1992). The two views are not discussed in this book due to space restrictions. 7 Marcus et al. excluded overgeneralized verbs that occurred fewer than ten times on the reasoning that including them would “not yield reliable estimates of the overregularization rates” (1992: 29). This is a very questionable procedure because the corpora they used covered only about 1-2% of children’s productions, which means that even forms that appeared only a few times in the corpus could potentially be 50 to 100 times as frequent in a child’s speech altogether (Maratsos 2000). The exclusion of rare overgeneralized forms hence neglected a considerable portion of the available data. In Maratsos’ reanalysis (2000) all the data was taken into account. <?page no="88"?> 75 Table 2. Brian’s mean overgeneralization rates of irregular verbs divided into groups of token frequency (table taken from Maslen et al. 2004: 1325 8 ). Token frequency Number of verbs M tokens (child) M tokens (input) M overgeneralization rate 1-9 29 2.97 52.24 23.1 % 10-49 16 25.06 157.19 13.75 % 50-99 3 69 736.3 14.27 % > 100 4 146.25 960.5 1.85 % A past-tense elicitation study by Marchman (1997) provides further support for these findings. The frequency of test items varied and was estimated based on adult speech samples. Children between 3 and 13 years were more likely to overgeneralize low-frequency than high-frequency irregular verbs (see also Bybee and Slobin 1982). To sum up, the rates of children’s overgeneralizations as well as their correct productions depend on the token frequencies of the respective verbs in the input and in children’s own speech. Zero-marking errors are a second type of errors that has also been related to input token frequency (Marchman 1997; Theakston et al. 2003). Zero-marking describes the impermissible omission of an inflectional morpheme (e.g., *He play instead of He played in a past-tense context). In Marchman’s past tense elicitation study high-frequency verbs were less vulnerable to zero-marking than low-frequency verbs. Theakston and colleagues (2003) explored the {-s} morpheme that is used to mark the third person singular present tense. They experimentally varied the frequency with which 2½and 3-year-old children heard this morpheme on novel verbs by using novel verbs with different constructions. Children in one condition heard constructions without {-s}, e.g., Will it tam? Children in the other condition heard the same verbs in constructions with {-s}, e.g. It tams. When children were subsequently prompted to use the novel verbs themselves, children in the first but not the second condition frequently produced zero-marking errors in contexts that required the inflected form. High input frequency of the morpheme with the novel verbs thus averted zero-marking errors, whereas the non-occurrence of this morpheme fostered them. To sum up, all types of inflectional errors occur predominantly on types that are low in token frequency in children’s input as well as in their own 8 Reprinted with permission from "A dense corpus study of past tense and plural overregularization in English" (p. 1325, Table 2) by Robert J. Maslen, Anna L. Theakston, Elena V. Lieven and Michael Tomasello, Journal of Speech, Language, and Hearing Research, 47, 1319-1333. © 2004. American-Speech-Language-Hearing Association. All rights reserved. <?page no="89"?> 76 productions. High-frequency types, on the other hand, are relatively resistant against such errors. The fact that low-frequency types are particularly prone to these errors is in line with the idea that the correct forms of these types have low representational strength in children’s minds. Type frequency. As I claimed initially, overgeneralization involves the application of an incorrect schema, usually a schema that is higher in type frequency than the correct schema. 9 The presumed reason is that high type frequency is conducive to the productivity and generalizability of a schema or construction. Experiencing a high number of different types in the slot position partially-filled schema invites the extension of the schema to new cases, including incorrect overgeneralizations. The frequent overgeneralization of the regular past tense inflection in English, which has the highest type frequency, provides first support for this idea. Bybee (1995) reports a study by Guillaume (1927/ 1973) who showed the same effect. Guillaume analyzed verb forms produced by French-speaking preschool children. Table 3 (Bybee 1995: 433, Table 1, based on Guillaume 1927/ 1973: 240-251) gives the token and type frequencies of verbs in children’s spontaneous speech. Children most frequently overgeneralized the first conjugation class that has the highest type frequency (76%) rather than the third class verbs, where token frequency was highest. To sum up, constructions with high type frequencies are usually applied in overgeneralizations by children. Table 3. Counts of verbs used by French nursery school children during play (taken from Bybee 1995: 433 10 ). Conjugation class Number of Uses (Tokens) Number of Verbs (Types) First (chanter ‘sing’) 1,060 (36.2%) 124 (76.0%) Second (finir ‘finish’) 173 (6.0%) 10 (6.1%) Third (vendre ‘sell’) 1,706 (57.8%) 29 (17.9%) Type frequency also affects overgeneralizations and zero-marking errors in another way: The type frequency of so-called phonologically similar friends and enemies influences the choice of the inflectional schema. Friends form the inflected form in the same manner as the lexeme in question (e.g., stay-stayed would be a friend of play-played), whereas enemies 9 Other factors, such as phonological similarity (Bybee 1995), also play a role and are mentioned at the end of this section. 10 Reprinted with permission from "Regular morphology and the lexicon" (p. 433, Table 1) by Joan Bybee, Language and Cognitive Processes, 10, 425-455 © 1995. Psychology Press. All rights reserved. <?page no="90"?> 77 form the inflected form differently than the lexeme in question (e.g., saysaid would be an enemy of play-played). Friends and enemies can be both regular and irregular. In her past-tense elicitation study Marchman (1997) revealed that verbs with higher numbers of enemies (high type frequency of enemies) were more prone to overgeneralization and zero-marking, whereas a high number of friends resulted in fewer overgeneralizations and higher resistance against zero-marking. It is thus not only the type frequency of different inflectional schemas, but also the type frequency of phonologically similar and dissimilar lexemes in the different schemas that affect children’s errors. Summary. Numerous frequency effects were shown in research on inflectional morphology learning. Tokens that are highly frequent in the input develop high representational strength in children’s minds and are produced developmentally early by children (Moerk 1978). As a consequence, overgeneralization and zero-marking errors are inversely related to a type’s token frequency (Maratsos 2000; Marchman 1997; Maslen and collegues 2004; Theakston et al. 2003). The schema that is selected for overgeneralization is usually high in type frequency, since high type frequency is linked to high productivity (Guillaume 1927/ 1973). The type frequency of phonologically similar friends and enemies further hinders or supports the occurrence of overgeneralization and zero-marking errors (Marchman 1977). 4.3.2 Other partially-filled constructions Research on partially-filled constructions is not limited to morphological constructions. Input token frequency also affects learning of other partiallyfilled constructions. Type frequency is relevant to the abstraction of variable slots in non-morphological constructions as well, and the relation of types and tokens, i.e., the type-token ratio, was further proposed to play a role in this process. Several studies showed the relation between the order of acquisition of partially-filled constructions and their token frequency in the input. A corpus analysis by Lieven (2008) found correlations between the frequency of low-scope, auxiliary frames in child-directed speech and the order of their emergence in children. Rowland and colleagues (Rowland and Pine 2000; Rowland, Pine, Lieven and Theakston 2003) investigated wh-question frames in 2to 4-year-olds. They revealed that the order of the acquisition of particular [wh-word][auxiliary] or [wh-word][verb] frames was predicted by their token frequencies in the input. Further investigations showed that errors, i.e., inversion errors, were more frequent with less frequent frames (Ambridge and Rowland 2009; Rowland 2007), which mirrors insights about errors in inflectional construction learning. <?page no="91"?> 78 Effects of type frequency were shown in a word-string repetition task. Matthews and Bannard (2010) found that 2and 3-year-old children were more likely to have formed a slot at the end of a four-word sequence if the last word was difficult to predict. Difficulty in predictability increased with the number of different types occurring in the variable position and thus benefitted slot abstraction. Presumably not only type frequency but also the relation of tokens to types is important in slot formation, i.e. the type-token ratio. Matthews and Bannard (2010) propose that predictability of all positions is high if a word string is fixed in the input most of the time, e.g., if children hear Throw the bottle 118 times and Throw the ball and Throw the teddy once each. In this case children are unlikely to form a productive slot in the final position (i.e., Throw the [X]). In contrast predictability is much lower if children hear each of the three sentences 40 times. Variability is more obvious in this case and children are much more likely to abstract the slot in the final position. The predictability or probability of seeing a particular word in an emergent slot position is thus possibly determined by the type-token ratio of several types, i.e., by the input distribution. A distribution where the slot filler is highly predictable is skewed towards this type, e.g., 120 tokens and 3 types distributed as follows: 118-1-1. Conversely, a distribution where the slot filler is unpredictable is more balanced, e.g., 120 tokens and 3 types distributed as follows: 40-40-40. 11 Input token frequency of partially-filled constructions affects the order of acquisition and the number of errors in children’s uses. High type frequency facilitates slot formation. Matthews and Bannard’s reflections reveal that it is probably not mere type frequency, but the relation of tokens to types that plays a role in slot formation. However, their example is purely hypothetical and research has not yet followed up on this issue. There are now a few new studies on the effects of type-token ratio (the shape of the input distribution) in the area of abstract constructions (cf. 4.4), but with respect to partially-filled construction learning these effects are unexplored. 11 The measure of so-called entropy is sometimes used to indicate the shape of the input distribution (see Shannon and Weaver 1949: 6 for a formula to calculate this measure). High entropy refers to a high degree of uncertainty, that is, a low predictability of how the slot position is filled. It can thus be used to describe a balanced input distribution. Conversely, low entropy refers to a high degree of certainty, i.e., to high predictability of how the slot position is filled, and can thus be used to describe a skewed input distribution. High entropy is consequently more conducive to slot formation than low entropy. <?page no="92"?> 79 4.4 Frequency and abstract constructions The term abstract construction subsumes all constructions whose form side is entirely abstract, i.e., does not contain any lexically-fixed material. Abstract constructions nevertheless vary in the degree of generality and complexity. For example, the form side of a construction might be described as [prefix][base] verb or as [morpheme][morpheme]. The latter is more general, since it also includes other combinations of morphemes, for instance, in compounds. The present section discusses different types of abstract constructions. First, word order constructions, which are most commonly referred to as abstract constructions, are presented. The second group reported are so-called complex constructions that describe complex sentence structures. Finally, the highly general level of so-called typological features is briefly mentioned, even though it has to be noted that it is probably not a constructional level where a form is paired with a common meaning or function. 4.4.1 Abstract constructions Children’s use of abstract constructions was shown to be affected not only by the overall frequency of such constructions in the input but also by the token frequency of particular words in these constructions and the type variability in certain positions of the abstract construction. The latter two issues are discussed first, followed by the effects of overall frequency of abstract constructions on learning. Frequency of lexemes in abstract constructions There are several studies that explored how the input token frequency of particular words in abstract constructions affects children’s behaviour. A previously reported study explored children’s willingness to overgeneralize high-frequency and low-frequency fixed-transitivity verbs to incorrect constructions (e.g., an intransitive verb to a transitive; Brooks et al. 1999; cf. 2.2.3). Children were less willing to overgeneralize high-frequency verbs than low-frequency ones, e.g., they were less likely to produce *He disappeared it than *He vanished it. Theakston (2004) and Ambridge and colleagues (Ambridge, Pine, Rowland and Young 2008) investigated grammaticality judgements of children, aged 5 and 7, and adults. Children and adults judged argument structure overgeneralizations as more ungrammatical if they contained highly frequent verbs than when they contained lower-frequency verbs. In weird word order studies, where scenes are described using non-canonical word orders (e.g., SOV in English), 2-yearold English-speaking and French-speaking children were more likely to correct sentences with high-frequency verbs to canonical word orders than <?page no="93"?> 80 sentences with medium-frequency or low-frequency verbs. Three-year-olds corrected all sentences regardless of verb token frequency, suggesting that the additional year of language input provided children with the experience they needed to generalize correct, canonical word order to all sentences (Matthews, Lieven, Theakston and Tomasello 2005, 2007). While children are able to generalize canonical word orders to sentences with verbs of all frequencies at age 3, even older children are more reluctant to generalize a novel word order after minimal input. Boyd and Goldberg (2011a) trained 5and 7-year-olds and adults on a novel, invented argument structure construction with the form [Noun Phrase1] [Noun Phrase2] [Verb Phrase novel ] and the meaning ‘approach’, e.g., The doctor the construction worker vakoes. ‘The doctor hops to the construction worker.’ As in the example, the verbs that were used were novel as well. At test, 5year-olds were above chance level only on sentences they had been exposed to during training and on sentences with noun phrases that were familiar from training. The higher frequencies of these sentences or at least of the noun phrases in these sentences compared to entirely new sentences were presumably responsible for children’s above-chance performance. At 5 years old, children were, however, unable to generalize the novel construction to completely new sentences. Seven-year-olds and adults were able to do so, after receiving the same input. Young children’s knowledge was thus restricted to familiar examples of the novel construction or examples with familiar elements, whereas older children and adults were able to generalize the novel construction to entirely new examples. Casenhiser and Goldberg (2005) examined the role of the type-token ratios of the verbs in novel construction learning. Therefore, they did not assess whether children performed more successfully on a particular sentence if they had heard the respective verb before in the construction. Rather Casenhiser and Goldberg were interested in the question of whether the input distribution of verbs affected children’s ability to generalize the novel construction. Based on the finding that naturalistic input is skewed towards one verb for several constructions, e.g., go occurs in 39% of all [Subject][Verb][Location] constructions (Goldberg et al.’s 2004 reanalysis of the Bates et al. corpus 1988), Goldberg and Casenhiser expected skewed input to be conducive to novel construction learning. They trained 5to 7year-old children on a novel argument structure construction with the form [Noun Phrase theme ] [Noun Phrase location ] [Verb Phrase novel ] and the meaning ‘appearance’, e.g., The sailor the pond neebod meant ‘The sailor sailed onto the pond from out of sight.’ For children in the skewed condition one verb occurred disproportionately frequently in the novel construction (4-1-1-1-1). Children in the balanced condition heard each verb approximately equally frequently (2-2-2-1-1). The overall number of types and tokens was the same in both conditions. Results revealed that children in both groups <?page no="94"?> 81 performed better than a control group who did not receive any auditory input of the novel construction in training. Importantly, children in the skewed input condition were significantly better than their peers in the balanced condition. Skewing thus facilitated generalizations of the novel abstract construction. The highly frequent verb might have served as anchor for the entire construction. An anchor is particularly useful in abstract construction learning because it indicates which structures can be successfully aligned and mapped in the analogical comparison, in a situation where all other lexical material varies in each instantiation. In a training study Childers and Tomasello (2001) found a similar advantage for 2½-year-olds on the familiar transitive construction. Training sentences in one condition took pronouns and full noun phrases as subjects and objects, whereas only full noun phrases were used in all sentences in the other condition. Children in the former condition profited more from the training, presumably because the frequently reoccurring pronouns also served as anchors for the otherwise entirely abstract construction. The result that skewed input facilitates abstract constructional learning stands in contrast to the predictions for partially-filled construction learning. For partially-stable constructions, balanced input has been suggested to facilitate the abstraction of slots (cf. 4.3.2, example by Matthews and Bannard 2010). Given the present findings, it is thus possible that the ideal input distribution, or type-token ratio, is dependent on the level of constructional abstractness, i.e., balanced input for more item-based constructions, skewed input for abstract constructions. This assumption has not been subjected to test. Finally, a study by Wonnacott and colleagues (2012) explored the effect of verb type frequency on novel construction learning. Wonnacott and colleagues trained 5-year-olds on the same novel argument structure construction as Boyd and Goldberg (2011a). Type frequency of verbs was varied between groups: there was only one verb type for children in the 1-type condition and four verb types for children in the 4-types condition. Children in the 1-type and 4-types conditions performed equally when familiar verbs from the training were used at test. On unfamiliar verbs requiring the generalization of the novel construction, however, children who had been exposed to four types were significantly better than children whose exposure had been limited to one type. This was not the case for adults, who were tested in the same design and performed equally well regardless of the number of types in their input. In order to generalize the novel word order, children thus required exposure to more than one verb type in their input. To sum up, token frequency of particular lexemes in abstract constructions affected children’s overgeneralization errors, their grammaticality judgements and their ability to correct non-canonical word orders. Young <?page no="95"?> 82 children’s comprehension of a novel abstract construction was shown to be limited to familiar types and types with elements they had previously experienced in the novel construction. Children’s ability to generalize a novel abstract construction was facilitated by a skewed input distribution. Generalizations further required children’s exposure to examples with more than one different verb type. Frequency of abstract constructions The effect of token frequency on novel construction learning was also examined by Wonnacott and colleagues (2012). They exposed children to tokens of the novel abstract construction on three successive days and found that children’s comprehension of the novel construction increased from day 1 to day 3. Further research explored input frequency in naturalistic construction learning and revealed effects on the speed of learning. For instance, the dynamic passive voice, e.g., The doll was torn by the girl, is not very common and thus very infrequent in the input (Gathercole and Hoff 2009). For this reason, English children learn this structure relatively late. If the frequency of the passive in children’s input was increased experimentally, children were shown to learn it sooner (de Villiers 1980). The same was revealed to be true for other languages, where the passive is more common and thus more frequent in children’s input (Demuth 1989). In another study, Gathercole (1986) found effects of input frequency not only on the speed of acquisition but also on the order of sub-meaning learning. Gathercole compared the use of the English present perfect in monolingual children from the US and from Scotland. While it is possible to substitute the perfect by the simple past in several contexts in American English (e.g., Did you eat yet? instead of Have you eaten yet? ), this replacement is not legitimate in British English. Gathercole (1986) showed that, consequently, there were large differences in the frequencies of the present perfect in the input to American and Scottish children. Due to the higher frequency in the input, Scottish children used the present perfect long before and much more frequently than American children did. The input frequency also affected the order in which different sub-uses of the perfect were learned. At the level of abstract constructions, input frequencies thus influenced the degree of learning, the speed of learning and the order, in which different meanings of a construction were learned. 4.4.2 Complex constructions Complex constructions are utterances consisting of at least two clauses, e.g., a main clause and a subordinate clause or two main clauses. Frequency effects in this area have been explored for all complex construction <?page no="96"?> 83 types together, i.e., counts summed up over all kinds of complex constructions were used. Huttenlocher, Vasilyeva, Cymerman and Levine (2002) assessed the input-output relationship of complex structures in this manner. They related the overall frequency of complex sentences in the speech of 4to 5-year-olds to the frequency of complex sentences in the input of children’s parents and preschool teachers. They discovered that the frequencies of children’s uses of more complex syntactic structures were directly related to the frequencies of such structures in the speech of their parents and teachers. Moreover, Huttenlocher and colleagues’ analyses revealed that there was more syntactic growth in children’s competence over one year in classes where teachers used syntactically more complex speech compared to classes that received less complex input from their teachers. The finding that the frequency of complex sentences in the input is related to the speed of development of syntactic complexity in children’s utterances receives support from further studies. Hoff-Ginsberg (1998) explored the language of children between 1; 6 and 2; 5 years and their parents. Her analysis revealed that first-born children were faster in their syntactic development than later-born children. According to Hoff-Ginsberg the reason for this advantage lay in the greater frequency of parental speech to first-born children. This finding is in line with results from a study by Barnes, Gutfreund, Satterly and Wells (1983), who also investigated the language to and of 2year-olds. They showed that children with early syntactic skills received more parental input than children with later-developing syntax skills. The frequency of complex constructions in parents’ and teachers’ input thus affects children’s productions of complex sentences as well as the speed of their development. 4.4.3 Typological features Research in the previous section revealed that it is not only the frequency of a particular construction itself that affects its learning. Instead, research on complex constructions has shown that the overall frequency at a more general level is also relevant to learning, e.g., the learning of complex constructions was shown to be related to the frequency of syntactically complex constructions in the input. For this reason, the even more general level of typological features is explored in the following. A typological feature is a “structural property of language that describes one aspect of crosslinguistic diversity” (WALS 12 ), e.g., word order or inflectional morphology. Dressler and others analyzed how the same feature develops in children cross-linguistically (e.g., Aksu-Koç and Slobin 1985; De Villiers and De Villiers 1985; Dressler 2005; Slobin 1973; Smoczynska 1985). Their analyses 12 Source: Word Atlas of Language Structure <http: / / wals.info/ feature>01.02.12. <?page no="97"?> 84 revealed that children become sensitive to the typological structures in their language and that they detect and store frequent structures earlier in comparison to less frequent ones and, crucially, in comparison to children with a different native language in which such structures are less frequent. In this manner, inflectional morphological categories were shown to be acquired earlier by children learning morphology-rich languages 13 , such as Slavic languages or Turkish, than by children whose language have poorer morphological systems, e.g., English 14 (Aksu-Koç and Slobin 1985; De Villiers and De Villiers 1985; Dressler 1997; Smoczynska 1985). Children acquiring morphology-rich languages thus become “more tuned to morphology” (Dressler 2005: 10) than children in whose language the same relations and functions are expressed differently, e.g., through word order. The frequency of typological features or structures consequently affects children’s development of the respective structure. Just as in the case of complex structures, typological features involved counting numerous different constructions, e.g., all inflectional morphological constructions. Examples from these two areas show that frequency is at work at very general levels as well. The general levels at which frequency was assessed, however, make it difficult to determine the levels of cause and effect. It is, for instance, possible that the frequencies at a specific level of generality are particularly relevant to the development of a certain construction. This issue is linked to the general question of which is the most appropriate level of description and which level is psychological real in that humans operate with abstractions at this level (Fischer and Stefanowitsch 2008: 12-13; cf. discussion in 8.3.2). 4.5 Summary of previous research Frequency affects language development in a myriad of ways. Effects of input token frequency, input type frequency and the type-token ratio in the input were examined. Input token frequency influences the speed, order and accuracy of acquisition of concrete, partially-filled and abstract constructions. As discussed initially, it is assumed that the effects are produced via children’s representations (Bybee 2010: 19-21). The idea is that input token frequency increases the representational strength of a structure in children’s minds, which is not measured directly but in turn causes the reported effects, e.g., leads to a higher speed of acquisition. Type frequency was revealed to be necessary for the development of slots in partially-filled 13 Morphological richness is understood as the number of productive inflectional morphological categories (Dressler 2003: 47). 14 This is despite the fact that morphology is usually easier in morphologically-poor languages. <?page no="98"?> 85 and entirely abstract constructions. Type frequency entails variability and thus makes the developing abstract position apparent. Type variability is linked to productivity including overgeneralizations and creative uses (Bybee 1995). The shape of the input distribution, i.e., the type-token ratio of several types, was further shown to affect learning. Skewed input facilitated abstract construction learning, presumably because the skewed lexemes anchored the entire construction (Casenhiser and Goldberg 2005; Childers and Tomasello 2001). For partially-filled constructions, balanced input was hypothesized to be more conducive, the idea being that variability in the slot position becomes apparent sooner (Matthews and Bannard 2010). Despite the numerous insights that the reported studies revealed with respect to frequency in language learning, several gaps also became apparent. Specifically, research on token and type frequency was very limited in the area of derivational morphological constructions. Moreover, the question of whether the role of the input distribution (in terms of the type-token ratio) differentially affects learning at different levels of abstractness awaits examination with a partially-filled construction. Finally, research on abstract constructions revealed that frequency can be counted at different levels of generality, i.e., only the construction in question or also including related structures. The highly abstract levels are, however, not necessarily constructional levels anymore, i.e., connected by a common form and function. To assess the influence of the frequency of closely related constructions it thus necessary to define a certain abstract level and examine effects at this level. 4.6 Introduction of the pattern level One aim of the present work is the exploration of frequency at such a more general abstraction level, which is - in contrast to previous studies - defined explicitly and is still considered as a construction (i.e., a formmeaning pairing). Given the previously reported gaps in the research of effects of token frequency, type frequency and type-token ratio in derivational morphology, a derivational verb prefix construction was used as a starting point for determining the more abstract level of analysis. Derivational verb prefix constructions are intuitively described by partially-filled constructions because the prefix remains stable in different instantiations of the construction whereas the other position is an open slot that can be filled with variable material, e.g., un[base] verb ‘reverse an action’, re[base] verb ‘do again’. These and other derivational verb prefix constructions, can be represented at a more abstract and more general level by the [derivational prefix][base] verb construction. For this level to be rightfully called a construction, a meaning side is required in complementation of the pro- <?page no="99"?> 86 posed form side. Schmid (2011: 160-162) proposes that the cognitive function of ‘encoding contrasts’ underlies prefixation in general. This suggested contrast is most apparent in negative prefixes, such as un- ‘not X’ (happy - unhappy), but also present in re- ‘do again’ (build - rebuild), where a finished action is taken up again contrary to what would be expected, or dis- ‘reverse an action’ (appear - disappear), which involves the reversal of a previously fulfilled action. According to Schmid, further prefixes express deviations from the norm in space, time, quantity or attitude (e.g., pre-war as opposed to ‘during or after the war’, ultra-right ‘excessively right from the speaker’s point of view’). The prefix construction presently proposed for further investigation is limited to verbs. Since verbs tend to express actions, the meaning ‘encode a contrast in action’ or ‘encode a contrast to the normal, expected action’ is suggested for the [derivational prefix][base] verb construction. This abstract construction is exemplified by different prefix constructions (e.g., un[base] verb , re[base] verb ), each of which is realized by a certain number of verb types (e.g., undo, untie, unload). Each of these types is in turn realized by a certain number of tokens in actual speech. For ease of reference, these levels need to be termed unambiguously. One possibility would be to use the classification of macro-constructions, meso-constructions, micro-constructions and constructs proposed by Traugott (2008: 236). She describes constructions at the lowest level of “empirically attested tokens” as constructs. They correspond to verb tokens in the present situation. Micro-constructions are “individual construction types” according to Traugott and refer to verb types in the present case (e.g., undo, untie). Meso-constructions are defined as “sets of similarly-behaving specific constructions”. They correspond to construction types (e.g., un[verb], re[verb]). Macro-constructions are “meaning-form pairings that are defined by structure and function” and refer to the newly-introduced abstract level in the present situation ([derivational prefix][base] verb ). Traugott points out that the hierarchy is not limited to these four levels. This is also apparent in the present case, since there are further more general and more abstract constructions, e.g., the [prefix][base] construction that is open as to the resultant word class or the even more abstract [morpheme][morpheme] construction. The circumstance that there is more than one abstract level calls for a more precise term than abstract construction for the [derivational prefix][base] verb level. The term macro-construction could be used in accord with Traugott. However, this use would imply a generalizability of any findings with respect to the proposed macro-construction to other macroconstructions. Moreover, for other constructions, e.g., word order constructions, the macro-level might involve a very different degree of abstractness. One reason for this difference is that word order constructions are per se more general than prefix constructions, but a second more general reason is that there is no generally-accepted system for the description of construc- <?page no="100"?> 87 tion hierarchies. The use of the term macro-constructions would thus imply the generalizability of the findings to constructions that are not clearly defined at all. In order to steer clear of such problems the new term pattern is introduced for the [derivational prefix][base] verb construction and used throughout the book. It describes the commonalities of all verb prefix constructions and is precisely one step more abstract than the verb prefix constructions themselves. The use of the term pattern is not related to any assumptions as to the pattern’s or its schema’s psychological reality or representational strength in people’s minds. For the time being, it only serves as a precise descriptive tool for one particular level at which frequency can be explored. The pattern level is more abstract and more general than the prefix constructions themselves, but at the same time it is itself a construction and thus more precise than, for instance, typological features. The generalizability of any results at the pattern level is expected to require separate discussion. The use of the pattern term and the terms used for the other, lower constructional levels in this book are given in Figure 12. Figure 12. Schematic depiction of the proposed constructional levels. The pattern level describes the commonalities of the different derivational prefix construction types, which in turn exemplify it. Each derivational prefix construction is exemplified by a number of verb types. Each verb type is realized by its tokens. 4.7 Major research questions The usage-based account holds that children learn language from their input based on social-cultural skills and with the help of their cognitive categorization and analogizing abilities. It follows from these assumptions that characteristics of the input will affect the cognitive learning processes. One of these input characteristics is frequency. The review of previous research on frequency effects in language learning revealed the important role of frequency but at the same time made open questions apparent: the <?page no="101"?> 88 effects of type and token ratio are largely unexplored in the area of derivational morphology, it is not clear whether effects of the shape of the input distribution (type-token ratio) differ between partially-filled and abstract constructions, and little is known about the effects of frequencies of constructions that are closely related to the construction in question (i.e., other construction types exemplifying the same pattern). Moreover, most previous research focussed on English, while other languages were mainly neglected. In order to fill the reported gaps, derivational verb constructions were used to examine different frequency effects in two languages, German and English. Effects in naturalistic language learning were explored by a corpus study; two experiments served to investigate frequency effects in novel construction learning. A novel construction is a form-meaning pairing that is unfamiliar to children because it is not present in their native language, i.e., the meaning is usually expressed in a different way and the form does not exist in the language. Using a novel construction has the advantage that it is possible to precisely control children’s exposure to this new construction. If constructions that are present in naturalistic language are used, exposure before the experiment cannot be controlled and may vary considerably between children. Furthermore, novel constructions are a useful means of assessing the validity of the idea that children learn from the input. The corpus study examined the frequency of the derivational prefix pattern in the speech of a German-speaking child and his caregiver, and an English-speaking child and his caregiver. Pattern frequency was assessed by counts of derivational prefix construction types, prefix verb types and prefix verb tokens. The corpus study had two main aims. The first one was based on research showing that construction learning is affected by input frequency at the level of individual constructions (e.g., Gathercole 1986; Gathercole and Hoff 2009; Hart 1991; Huttenlocher et al. 1991; Moerk 1978; Naigles and Hoff-Ginsberg 1998) as well as at the very general typological level (e.g., Aksu-Koç and Slobin 1985; De Villiers and De Villiers 1985; Dressler 2005; Slobin 1973, Smoczynska 1985) and on findings that related constructions additionally facilitate construction learning (Abbot-Smith and Behrens 2006). The first aim was thus to explore whether input frequency at the pattern level also affects children’s output frequencies (I a). The second aim was to assess whether there are differences between German and English with respect to frequencies at the pattern level (I b). This question served a practical end. The idea was to identify a pattern whose frequency varied between German and English, because this would allow predictions for the experimental study (see below). An additional minor aim was the analysis of children’s creative uses of derivational prefix constructions with respect to type frequency. Previous research had suggested <?page no="102"?> 89 that constructions with high type frequency are very productive and prone to overapplications (Bybee 1995; Clark 1993: 137-138; Clark and Cohen 1984; Clark and Hecht 1982; Guillaume 1927/ 1973). I Pattern frequency in naturalistic language learning a Are input and output frequencies of the derivational verb prefix pattern related? b Are there differences in the frequency of the derivational verb prefix pattern between German and English in the language to and of children? In the first experiment frequency effects on novel construction learning were explored for German-speaking and English-speaking children. Half the children were exposed to a novel prefix construction that exemplified the derivational verb prefix pattern. The other half of the children were trained on a construction that was based on a pattern that was notably absent from both German and English thus yielding a pattern frequency of zero in both languages. This construction was a verb-initial reduplication. Previous research had shown that novel word order constructions can be learned from the input and that the success of learning increases with age (Boyd and Goldberg 2011a; Casenhiser and Goldberg 2005; Wonnacott et al. 2012). The present study explored whether this finding extends to a partially-filled novel prefix construction and an abstract novel reduplication construction (II a i, II b i). Earlier studies at the general typological feature level had revealed that learning is affected by frequency of a feature cross-linguistically (Aksu-Koç and Slobin 1985; De Villiers and De Villiers 1985; Dressler 1997, 2005; Smoczynska 1985). In the present experiment it was assessed whether frequency at the pattern level has similar effects (II a ii, II b ii). On the assumption that frequent constructions are learned earlier (Gathercole and Hoff 2009), it was further explored whether the construction based on a frequent pattern is easier to learn than a construction based on an infrequent, i.e., absent, pattern. This investigation only had preliminary character, since it was not only the underlying pattern frequency that varied between the compared constructions (cf. 6.4.3). This is why this issue is not listed in the main research questions. II Novel construction learning and pattern frequency a Pattern present in native language; frequency varies between languages i Learnability and age: Can German-speaking and Englishspeaking children learn a novel construction based on a pattern that is used in their languages and does the success of learning increase with age? ii Frequency: Is learning a novel construction easier for children in whose native language the underlying pattern is more frequent? <?page no="103"?> 90 b Pattern absent from native language i Learnability and age: Can German-speaking and Englishspeaking children learn a novel construction based on a pattern that is absent from their languages and does the success of learning increase with age? ii Frequency: Is learning a novel construction equally difficult for all children given that the underlying pattern is absent from both languages? Effects of input token frequency were also examined in this first experiment. In earlier studies token frequency had been related to the speed of acquisition of constructions other than derivational morphological constructions (Ambridge et al. 2008; Hart 1991; Matthews et al. 2005; Matthews et al. 2007; Theakston 2004; Theakston et al. 2004). In inflectional morphology learning, the rate of errors was further inversely related to token frequency (Marchman 1997; Maratsos 2000; Maslen et al. 2004). One novel construction learning study additionally revealed that token frequency increased children’s comprehension of the novel construction (Wonnacott et al. 2012). The present experiment aimed at exploring the role of input token frequency in novel derivational morphology learning by assessing children’s performance on instances whose experimental input token frequency varied. III Token frequency in novel prefix and novel reduplication learning Does token frequency increase novel construction learning, i.e., correct responses? The importance of type frequency at the level of partially-filled and abstract constructions has been pointed out repeatedly, since it is necessary for slot formation and subsequent generalization (Clark 1993: 137-138; Clark and Hecht 1982; Guillaume 1927/ 1973; Marchman 1997; Matthews and Bannard 2010). Previous research suggested that a single type in the input is insufficient for children to generalize abstract constructions (Wonnacott et al. 2012). This finding is in line with the account of constructional learning followed in this work, since comparison and schema abstraction that are thought to logically precede generalization require the involvement of at least two exemplars. The present study thus explored whether children indeed store two or more types of a novel morphological construction before generalizing it to new instances. IV Type frequency in novel prefix and novel reduplication learning Do children use at least two previously experienced types of the novel construction correctly before forming correct generalizations? The second experiment investigated the effect of the input distribution on English-speaking children’s learning of a partially-filled novel construction. Previous research has shown that children and adults profit from <?page no="104"?> 91 skewed input in novel argument structure construction learning (Casenhiser and Goldberg 2005; Goldberg et al. 2004). The likely reason is that the highly frequent lexeme serves as an anchor for the entire novel construction, facilitating analogies over several examples and with it schema abstraction and generalization. Hypothetical considerations by Matthews and Bannard (2010) on the other hand suggested an advantage for balanced over skewed input for more item-based constructions. The reason is that balanced input is presumed to be more conducive to slot formation in an otherwise fixed string, because the variability in the slot position becomes apparent sooner than in the case of skewed input. This assumption had not been tested experimentally. Its exploration was thus the main aim of the second study. V Type-token ratio in partially-filled construction learning Is balanced input more conducive to learning a partially-filled construction than skewed input? The research questions presented here served as the basis of the more specific hypotheses that are given for each empirical study in the respective chapter and ensure that the studies cover all relevant issues. The empirical studies in this book serve to explore different effects of frequency at the constructional and the more general pattern level. The aim of the studies is thus to gain new insights into the role of frequency in construction learning and more generally to extend the present understanding of the constructional learning process in children. <?page no="105"?> 92 5 Corpus study: Frequency of the derivational verb prefix pattern in caregiver and child speech in German and English The present chapter presents the study of one morphological pattern in a German and an English child speech corpus. The pattern was the derivational verb prefix pattern [derivational prefix][base] verb . 1 The study served two main aims and a minor goal. The main aims concerned the exploration of pattern frequency. The first one was to examine whether input frequency at the newly-introduced pattern level influences children’s own productions in a naturalistic context. The second major aim was to explore whether pattern frequency was different for German and English. The identification of a pattern with different cross-linguistic frequencies was necessary so that one of the novel constructions in the experimental study in Chapter 6 could be based on this pattern. Any differences in pattern frequency between German and English revealed in this chapter were expected to become reflected in German-speaking versus English-speaking children’s performance in novel construction learning. An additional goal of the corpus study was to analyze children’s innovative uses in the light of the particular constructions that were overgeneralized and with a view to their type frequency. The present chapter starts with a brief motivation of the study. Subsequently, previous insights as to verb prefixes in German and English are presented and previous research on children’s learning of derivational morphology is briefly reported. Then, it is described how pattern frequency is operationalized and the hypotheses are formed. The corpus study itself includes two analyses as well as the examination of innovative overgeneralizations. Finally, the results are discussed with respect to the hypotheses and their relevance to the experimental study in Chapter 6. 1 The use of ‘verb’ in the terms ‘derivational verb prefix pattern’ and ‘derivational verb prefix construction’ is not to imply that the prefix only attaches to verbal bases, but instead it is to express that the result of the prefixation is a verb. This is also marked by the fact that the general term ‘base’ is used in the description of the pattern or construction and by adding ‘verb’ in subscript to the description of the form side, e.g., [derivational prefix][base] verb . <?page no="106"?> 93 5.1 Background 5.1.1 Motivation Pattern frequency Previous research has explored effects of input frequency on language learning at different levels of generality (cf. Chapter 4). The literature on construction learning revealed that the input token frequency of particular constructions affects how easily (in terms of errors), how quickly and in which order children learn them. This was shown for concrete, partiallyfilled and abstract constructions (Hart 1991; Maratsos 2000; Marchman 1997; Maslen et al. 2004; Moerk 1978; Naigles and Hoff-Ginsberg 1998; Theakston et al. 2004). Type variation in the input as well as the shape of the input distribution was shown to be important for the learning of partially-filled and abstract constructions (Casenhiser and Goldberg 2005; Clark and Cohen 1984; Clark and Hecht 1982; Guillaume 1927/ 1973; Matthews and Bannard 2010; Wonnacott et al. 2012). In all these cases frequency of one particular construction was considered. Research on complex construction learning, on the other hand, summed up input frequencies over all kinds of complex constructions and assessed whether they influenced children’s productivity with complex constructions in general. Higher overall input frequencies were shown to lead to higher and earlier productivity in children (Barnes et al. 1983; Hoff-Ginsberg 1998; Huttenlocher et al. 2002). It is thus not only the case that the frequency of a particular construction affects the learning of this construction, but also that the frequencies of a group of similar constructions influence the learning of all these constructions. A similar finding comes from typological research. If a language is rich with respect to a particular feature, e.g. inflectional morphology, it means that there is a high number of productive categories involving this feature. There are consequently numerous constructions with this feature, making it very frequent. Research revealed that children learning a morphology-rich language, such as Slavic languages or Turkish, acquire morphological structures earlier than children whose language has a poorer morphological system, e.g., English (Aksu-Koç and Slobin 1985; De Villiers and De Villiers 1985; Dressler 1997; Smoczynska 1985). Here, counts were formed at a very general level, but frequency affected learning presumably because all involved constructions were related in that they were morphological. Further research revealed that constructions that are related in form may also influence learning, both positively and negatively. Constructions that are supported by previously learned ones, e.g., the German sein-passive (sein ‘to be’) is supported by the copula sein, are acquired earlier and faster than unsupported ones, e.g., the German werden-passive (werden ‘to become’; Abbot-Smith and Behrens 2006). If a construction is only seemingly supported by another frequent <?page no="107"?> 94 one, errors are more frequent and learning can be delayed (e.g., frequent Let me do it in the input led to me-for-I errors, e.g., children produced *Me do it instead of I do it; Kirjavainen et al. 2009). Similar results come from psychological research revealing that a learning task is easier to solve if it shares common elements with a previously learned task (Singley and Anderson 1989: 2; Thorndike 1906: 243). These results support the idea that frequencies of related constructions affect constructional learning. Pattern level frequency was explored in the this work for several reasons: The pattern level, [derivational prefix][base] verb 2 , precisely determines and summarizes the commonalities of closely-related constructions, such as [un][base] verb and [dis][base] verb , which are expected to influence each other’s acquisition, given previous research at both more and less general levels. At the same time, being able to trace back potential effects to the pattern level might allow more precise claims as to cause and effect 3 than was possible in the studies of complex constructions and typological features reported earlier. Following the procedure in typological research, frequencies at the pattern level were explored cross-linguistically, i.e., for German and English. Expectations as to potential differences were based on the literature review that is presented subsequently in 5.1.2. Input and output frequencies were compared because previous research revealed a similar relation for other categories (word classes, particle verbs; Behrens 2003, 2006). Type frequency and innovative overgeneralizations Type frequency is highly relevant in the development of partially-filled and abstract constructions, which are formed from more concrete constructions. The emergent slot positions become apparent through type variation (Lieven et al. 2003; Lieven et al. 2009; Matthews and Bannard 2010). Slot formation (i.e., schema abstraction) is a prerequisite for subsequent generalization. If type frequency in the input is high (and the schema has a certain degree of openness), children are likely to extend it productively, i.e., to use it with numerous types themselves (Bybee 1995; Clark 1993: 137-138; Clark and Cohen 1984; Clark and Hecht 1982; Guillaume 1927/ 1973). Due to the fact that corpus data only provides partial insight into children’s input, it is impossible to assess children’s productivity on this basis. That is, all forms children produce that are not attested in the input might have been part of their non-recorded input. Children’s innovative overgeneralizations on the other hand are unlikely to have appeared in the input at any point, because they are non-canonical. Based on previous findings, 2 For brevity, only the form of a construction or pattern is sometimes given in reference to the entire form-meaning pair. 3 This tracing back is possible in experimental studies, e.g., Experiment 1. <?page no="108"?> 95 constructions that are high in type frequency are expected to be particularly prone to such innovative extensions. 5.1.2 Verb prefixes in German and English Before the hypotheses are given and the corpus study is presented, verb prefixes in German and English are examined. This description is necessary in order to form predictions as to potential differences in frequency between the two languages. The characteristic feature of the prefix construction is that morphological information is located before the base, which is most frequently, but not necessarily, a free lexical morpheme (Marchand 1960: 129). In German the term ‘prefix verb’ is traditionally used for derivational prefix verbs (e.g., bemalen ‘paint on’) and separable particle verbs (aufmachen ‘open’). It should be added that there are also inflectional circumfixes in German, where part of the morphological information is placed verb-initially and part of it is placed verb-finally (ge[stem]t, e.g., gemalt ‘drawn PAST.PARTICIPLE ’, ‘painted PAST.PARTICIPLE ’ and ge[stem]en, e.g. gelaufen ‘run PAST.PARTICIPLE ’, ). In English only derivational prefix verbs (e.g., undo) place an additional morpheme verb-initially. Particles always follow the verb (e.g., look up) and there are no circumfixes. Derivational verb prefixes In the formation of derivational prefix verbs a bound lexical morpheme usually combines with a free lexical morpheme. Prefix and verb base stand in the semantic relation of modification and determination with the head typically determining the word class of the complex lexeme (Schmid 2005: 148). The cognitive function of derivational prefixes has been described as ‘profiling a contrast’, i.e., the concept of ‘different from’ (Schmid 2005: 162- 165). This function becomes particularly apparent in negative prefixes un-, inor reversative prefixes deand dis-. Derivational prefixes form verbs from verbal, nominal, adjectival or particle bases 4 (e.g., befahren ‘drive/ travel on’ based on fahren verb ‘drive’, zerfleischen ‘mangle’ based on 4 There are also atypical cases, where the bases do not exist as free morphemes in English or German, but seemingly different prefixes are attachable to the would-be base (e.g., *ceive as in perceive, conceive, deceive). Such cases are either analysed as monomorphemic due to their synchronic opacity or they are split into prefix and socalled bound root. The difficulty for synchronic analysis and the impression that there is a regular pattern behind such cases are results of historical developments. These patterns were usually productive in other languages before several examples of the pattern came into the English (or German) language, e.g., Lat. percipere ‘perceive’, Lat. concipere ‘conceive’, Lat. decipere ‘deceive’. The status of such cases as prefix verbs is controversial. <?page no="109"?> 96 Fleisch noun ‘meat’, verlängern ‘lengthen’ based on lang adjective ‘long’, verneinen ‘negate’ based on nein particle ‘no’; Barz 2006: 699; undo, empower; Quirk et al. 1984: 982-992). Verbal bases are the most common in English and German, which coincides with the fact that prefixes tend to preserve word class (Schmid 2005: 148). German Derivational prefixes cannot be separated from their bases. In German the infinitival marker zu ‘to’ precedes the prefix verb (1) and the past participle formation goes without the initial (-)geof the circumfix ge[stem]t/ en. Instead the prefix is maintained (2). (1) Er plant, seinen Onkel zu begrüßen. ‘He is planning to greet his uncle.’ (2) Er hat seinen Onkel begrüßt. ‘He greeted his uncle.’ Productivity of prefix types varies. Verbal bases are most productive and combine with the highest number of prefix types; denominal derivations most frequently attach to the prefixes be-, entand zer-; and adjectival bases prefer the prefix ver- (Barz 2006: 700-701). The overall most frequent prefix type is ver- (Fleischer and Barz 1995: 325). The prefix types proposed vary considerably in the literature. Table 4 is a compilation of prefix types given in Barz (2006: 699-701), Engel (2004: 230), Fleischer and Barz (1995: 37, 327), Kühnhold and Wellman (1973: 144- 154). 5 Sample meanings and examples are provided, whenever given in the literature. Table 4. German verb prefixes, their meanings and examples. Prefix Meaning Example be- ‘causing’, ‘covering’ begrenzen ‘delimit’ bemalen ‘paint on’ de-/ des- ‘opposite’, removal’ destabilisieren ‘destabilize’, desillusionieren ‘disillusion’ dis- ‘opposite’, ‘separation’ disqualifizieren ‘disqualify’ durch- ‘from beginning to end’, ‘pervading’ durchwandern ‘hike through’ ent- ‘privative’, ‘distance’, ‘beginning’ entflammen ‘kindle’ er- ‘beginning’, ‘result’ erblassen ‘turn pale’ exexkommunizieren ‘excommunicate’ ge- ‘opposite of miss-’ gefallen ‘like’ 5 Kühnhold and Wellman (1973: 144-154) do not differentiate between inseparable prefix verbs and separable particle verbs in their listing. Please note that only prefix types with examples of inseparable prefix verbs were taken into account here. <?page no="110"?> 97 hinter- ‘negative’ hinterziehen ‘evade (tax)’ in- ‘intruding’ injizieren ‘inject’ interinteragieren ‘interact’ ko-/ kol-/ kom-/ kon-/ kor- ‘together’, ‘cooperation’ konzentrieren ‘concentrate’ miss- ‘negative’ misslingen ‘fail’ obobwalten ‘prevail’ perperzipieren ‘perceive’ präpräkludieren ‘preclude’ proproponieren ‘propose’ re- ‘form back’ reprivatisieren ‘reprivatize’ subsubstituieren ‘substitute’ transtransponieren ‘transpose’ über- ‘spatially higher’, ‘exceeding a norm’ überspringen ‘leap/ skip (over)’ um- ‘encompassing’ umgeben ‘surround’ unter- ‘spatially lower’, ‘preventive’ unterbinden ‘prevent/ stop’ ver- ‘resultative’, ‘erroneous’, ‘decorative’ vergolden ‘gild’ vollvollbringen ‘accomplish’ wider- ‘opposite’ widersprechen ‘contradict’ wieder- ‘iterative’ wiederholen ‘repeat’ zer- ‘moving apart’ zertrümmern ‘smash’ For a number of cases it is difficult to resolve whether they are in fact cases of prefixation. They are mostly morphemes of Latin origin that attach predominantly to bound roots: ab- (e.g., abbreviieren ‘abbreviate’, Scholze- Stubenrecht, Eickhoff and Mang 2001: 17), ad- (e.g., addizieren ‘allot’, ‘adjudge’, 25), ante- (e.g., antedatieren ‘predate’, ‘antedate’, 70), en- (e.g., enkodieren ‘encode’, 269), intro- (‘into’, e.g., introduzieren ‘introduce’, 457) and post- (‘behind’, ‘after’, e.g. postponieren, ‘put behind’, 793). These prefixes were not listed above for three reasons: They do not usually combine with free morphemes, their character is exclusively foreign and the reference works were very divided over their prefix status. If considered as prefixes, they are certainly peripheral members of this category. English The most productive English derivational prefixes attach to verbal and occasionally to nominal bases. Schmid (2005: 153-161) analyzed English prefixation patterns in a 41,000-word corpus (BUMC). He did not limit his analysis to prefix verbs, but considered other word classes as well. He reports un- (adjectives and verbs) to be the most frequent prefix in terms of types and tokens, followed by in- (adjectives) and re- (verbs). The compilation in Table 5 compiles the verb prefixes reported by Quirk and colleagues <?page no="111"?> 98 (1984: 981-992, “living prefixes”), Marchand (1960: 139-208) and Schmid (2005: 150-161). Table 5. English verb prefixes, their meanings and examples. Prefix Meaning Example be- ‘affect’, ‘provide’, ‘surround’, ‘cover with’ bedazzle circum- ‘move around’ circumambulate, circumbind co- ‘accompanying’, ‘with’, ‘joint’ cooperate counter- ‘against’, ‘in opposition to’ counteract de- ‘to reverse an action’, ‘to get rid of’ decode dis- ‘the opposite of’, ‘not’, ‘to reverse the action’ dislike, disown en-/ em- ‘to make/ put/ get into’ endanger, empower fore- ‘before’ foretell inter- ‘between’, ‘among’ interweave mal- ‘badly’, ‘bad’ malfunction mis- ‘wrongly’, ‘astray’ miscalculate out- ‘(to do something) better, faster, longer etc. than…’ outlive over- ‘too much’ overdo post- ‘after’ postpone pre- ‘before’ presuppose, precook re- ‘again’, ‘back’ rebuild sub- ‘under’, ‘beneath’, ‘lesser in rank’ sublet supra- ‘over’ supra-saturate super- ‘over’ super-accumulate trans- ‘across’, ‘from one place to another’ transship un- ‘to reverse the action’ with verbal base, ‘deprive of’ or ‘release from’ with nominal base undo, unhorse under- ‘too little’ undercook Additional, more peripheral members of the prefix category that only combine with bound roots were again not listed in the table due to their debatable status as prefixes: ab- (‘remove’, ‘away’, e.g., abduct), contra- (‘opposite’, e.g., contradict), ex- (‘out of’, e.g., export), in- (‘in’, ‘into’, e.g., insert) and per- (‘through, completely, wrongly, exceedingly’, e.g., permeate). Given the two lists, German and English seem not entirely different with regard to the number of derivational verb prefix types. Nevertheless, their number is slightly higher in German. It should be added that English has a particularly high number of foreign prefixes, which are unlikely to occur in the speech to and of children (see Table 5, e.g., circum-, co-, counter-, mal-, and so forth). In order to fully determine the prevalence of the deriva- <?page no="112"?> 99 tional verb prefix pattern in each language, the number of prefix types, or more precisely the number of prefix construction types, would have to be explored in comparable corpora of German and English. In fact, not only the number of prefix types, but also the number of types per prefix (prefix verb types) and the number of tokens would call for consideration. It is thus obvious that research into frequencies of this area in German and English is as yet very limited. Verb particles In German a second class of morphemes is frequently referred to as prefixes, because they are regularly preposed to verbs. They are also known as separable prefixes (Fleischer and Barz 1995: 208), Halbpräfixe ‘semi prefixes’ or trennbare Verbzusätze ‘separable verb adjuncts’ (Wellmann 1998: 439; 454- 455), postponierbare Präverben ‘postponable pre- @ \ ^ _ $``~ "% Nachverben ‘after-verbs’ (Weinrich 1993: 32), prefixoids (Donalies 2005: 28) or verb particles (Eichinger 1989: 92; Fehlisch 1998: 150). This last term has become prevalent in the literature (Fehlisch 1998: 150) and is used here. German particles are separable from the verb in certain syntactic contexts. They are separated morphologically from the base in the participle (3) and the zu-infinitive ‘to-infinitive’ (4). Syntactic separations of base and particle occur in main clauses, where the verb occurs in verb-second position and the particle is found in final position (5), as well as in imperatives (6) (Behrens 1998: 682-683; Eisenberg 2006: 264). (3) Sie hat ihren Mann angerufen. ‘She has called her husband.’ (4) Sie plant, ihren Mann anzurufen. ‘She is planning to call her husband.’ (5) Sie ruft ihren Mann an. ‘She is calling her husband.’ (6) Ruf(t) ihn an! ‘Call SINGULAR(PLURAL) him! ’ Base and particle are not separated in subordinate clauses following subordinating conjunctions (7), in uses with the bare infinitive (8), with modal auxiliaries (9) or the future (10). (7) Er freut sich, wenn seine Frau anruft. ‘He is glad when his wife calls.’ (8) jemanden anrufen ‘to call somebody’ (9) Sie muss/ soll/ kann/ darf ihn anrufen. ‘She must/ should/ can/ may call him.’ (10) Sie wird ihn anrufen. ‘She’ll call him.’ In the majority of cases particle verbs take verbal bases (e.g., abholen ‘pick up’), but there is also a limited number of nominal (e.g., Leine ‘leash’ > anleinen ‘to leash’) and adjectival bases (e.g., dünn ‘thin’ > ausdünnen ‘thin out’). The most common bases are simplex verbs (e.g., ausreisen ‘leave (the country)’, umsteigen ‘change (trains)’). Complex bases with prefixes or suffixes can also take particles (e.g., vorbesprechen ‘discuss preliminarily’, vor- <?page no="113"?> 100 beimarschieren ‘march by’), whereas particle verbs cannot take a second particle (e.g., *zurückweggehen ‘go away back’; Barz 2006: 705). Verb particles are usually homonymous to free morphemes. Semantically, they tend to imply the time or direction of the movement that may be made explicit in an adverbial. Metaphorical extensions of those meanings are also frequent (Barz 2006: 711-712). Depending on their word class the following particles are distinguished: prepositional particles (11), adverbial particles (12), adjectival particles (13) and nominal particles (14) (Barz 2006: 706; all examples taken from Barz 2006: 706, translations added). Please note that these are merely examples rather than an exhaustive list: (11) ab- ‘from’/ ‘off’/ ‘away’, an- ‘at’/ ‘on’/ ‘to’, auf- ‘on’, aus- ‘from’/ ‘off’, bei- ‘by’/ ‘at’/ ‘for’, durch- ‘through’, hinter- ‘behind’/ ‘after’, mit- ‘with’/ ‘by’, nach- ‘after’/ ‘to(wards)’, über- ‘over’/ ‘above’, um- ‘around’, unter- ‘under-’/ ‘below’/ ‘among’, vor- ‘before’/ ‘in front of’, wider- ‘against’, zu- ‘to(wards)’, ein- ‘in’/ ‘into’ (corresponds to preposition in ‘in’/ ‘into’) (12) her- ‘here’/ ‘from’/ ‘to’, hin- ‘here’/ ‘to’, herunter- ‘down’, hinunter- ‘down’, dahin- ‘there’/ ‘to’ (13) fest- ‘firm’/ ‘tight’, frei- ‘free’, hoch- ‘high’/ ‘up’ (14) stand- ‘position’/ ‘stand’ as in standhalten ‘withstand’/ ‘defy’, preis- ‘price’ as in preisgeben ‘reveal’/ ‘divulge’ Nearly every verb can be combined with a number of particles (15), resulting in extremely high productivity (Barz 2006: 707; Weinrich 1993: 1033; example taken from Barz 2006: 707, translations added). (15) nehmen ‘take’: abnehmen ‘decrease’/ ‘lose weight’, annehmen ‘accept’, aufnehmen ‘incorporate’/ ‘include’/ ‘accept’, ausnehmen ‘fleece sb’/ ‘except’, durchnehmen ‘deal with’, einnehmen ‘ingest’/ ‘occupy’, mitnehmen ‘take along’, übernehmen ‘take over’/ ‘take on’, unternehmen ‘undertake’, vornehmen ‘plan’/ ‘undertake’, wegnehmen ‘take away’/ ‘remove’, zunehmen ‘increase’/ ‘gain weight’, hernehmen ‘get’/ ‘take from’, herunternehmen ‘take down’, zurücknehmen ‘take back’ German particle verbs correspond to English phrasal verbs in many cases (see examples in (15), e.g., zurücknehmen ‘take back’). In English, particles always follow the verb (e.g., get over, show up, switch off, talk into). 6 They either follow the verb directly (e.g., He took back what he had said earlier.) or they are separated syntactically (e.g., He took it back.), depending on the weight of the intervening noun phrase (Olsen 1997: 52-53). Since the present chapter is concerned with morphological material that precedes rather than follows verb bases, English phrasal verbs are not discussed in any more detail here (cf. Greenbaum and Quirk 1990: 336-343; Lamprecht 1986: 36-38). 6 A rare exception are inverted sentences, e.g., It was supposed to blow up at noon and up it blew. But even here the particle never directly precedes the verb because of the intervening subject. <?page no="114"?> 101 Circumfixes In German, but not English, there is also a circumfix: the past participle ge[stem]t and ge[stem]en, e.g., spielen—gespielt 7 ‘play’--‘played PAST.PARTICIPLE ’, laufen—gelaufen ‘run’--‘run PAST.PARTICIPLE ’. The past participle is used with different auxiliaries to form the present perfect, the past perfect, the future perfect and the passive (all tenses), e.g., ich habe gespielt ‘I have played’, ich hatte gespielt ‘I had played’, ich werde gespielt haben ‘I will have played’, es wird gespielt ‘it is played’, ich bin gelaufen ‘I have run’, ich war gelaufen ‘I had run’, ich werde gelaufen sein ‘I will have run’, es wird gelaufen ‘it is run’. In addition, the past participle is used attributively with nouns, e.g., der gespielte Ball ‘the played ball’/ ‘the ball that was played’, das gelaufene Rennen ‘the run race’/ ‘the race that was run’, and predicatively to express passive meanings, e.g., Sie servierte das Fleisch lecker gewürzt ‘She served the meat deliciously seasoned’ (Engel 2004: 225-226). Additionally, it is possible to form a related noun from the participle, e.g., sie wird gesucht—die Gesuchte ‘she is searched for’--‘the searched-for (person)’ (Engel 2004: 226-227). With particle verbs -geoccurs in mid-verb position between the particle and the stem, e.g., abholen—abgeholt ‘pick up’--‘picked up PAST.PARTICIPLE ’, der abgeholte Kuchen ‘the picked-up cake’/ ‘the cake that was picked up’. The first part of the circumfix, (-)ge-, is used with most verbs except for derivational prefix verbs, e.g., beenden—beendet (*gebeendet or *begeendet) ‘finish’—‘finished PAST.PARTICIPLE ’, verbs ending in -ieren, e.g., massieren— massiert (*gemassiert) ‘massage’—‘massaged PAST.PARTICIPLE ’ and verbs with a derivational prefix and a particle, e.g., anvertrauen—anvertraut (*geanvertraut, *angevertraut or *anvergetraut) ‘confide’—‘confided PAST.PARTI- CIPLE ’ (Engel 2004: 225-226). Altogether, the circumfix occurs in numerous grammatical contexts and is particularly frequent in spoken discourse, where the present perfect is preferred over the simple past in German (Lindgren 1957: 117-118; Schwitalla 1997: 138). When considering all morphemes (or parts of morphemes) that are placed before the verb base, German might have an advantage over English. The reason is that there are slightly more derivational prefix verb types in German, there is a high number of particle verbs in German and there is the first part of the highly frequent circumfix in German. 5.1.3 The development of derivational morphology in children The question how children learn derivational morphology constitutes the final part of the content-related background necessary for the present corpus study. Much of the knowledge about children’s learning of deriva- 7 The final part of the circumfix is determined by whether the verb belongs to the strong (-en) or the weak (-(e)t) conjugation (Engel 2004: 225). <?page no="115"?> 102 tional morphology is derived from their innovative uses. These overgeneralizations are cases, in which children extend a derivational morphological construction creatively but non-canonically. Such uses are different for each child. They depend on a child’s vocabulary (e.g., on whether there is a canonical form that can preempt the innovation) as well as more general factors such as the transparency of the morphological construction, the difficulty of constructing the innovation, the productivity and type frequency of the construction (cf. 4.3.1; Clark 2003: 281). Factors such as growing vocabularies help children recover from using their innovations permanently, since conventional forms eventually preempt their uses. Innovations in derivational morphology arise around the second birthday when children use nouns as verbs (denominal conversion/ zeroderivation), for instance *to brush (the floor) or *to broom meaning ‘to sweep using a brush/ broom’ or *to band-aid meaning ‘to put a band-aid on’ (Clark and Clark 1979). At around 2½ years of age, children start using the first affixes innovatively. The diminutive suffix -ie and agentive suffix -er are usually the first ones. Clark (2003: 292) provides examples (16-17) from unpublished diary data: (16) 2; 5; 26 (reaching over to put some meat in a bowl at the counter, helping the father cook): I reached right over. I’m a big *reacher. (17) 2; 5; 13 (clapping his hands): I a *clapper. Later in the third and in the fourth years of life, children also start using -er with instrumental meaning, -y to form adjectives and unon verbs (Clark 1993: 118; Clark and Hecht 1982; Clark, Carpenter and Deutsch 1995). A child whose language was examined by Clark (1993: 118) used -y first on familiar adjectives that do not take -y (e.g., *dark-y) and shortly afterwards used it innovatively to form adjectives from nouns (e.g., *crumb-y). In an elicitation game where children had to contradict the experimenter, Clark and colleagues (1995) explored the development of the prefix un-. Before age 3 children expressed reversal by particles (e.g., on/ off) and general purpose verbs. They also used undo, but the authors presumed that children had not yet analyzed its components at this time an. From around 3 years old, children started using a number of different bases with unand produced several innovative overgeneralizations (1995: 645, 646), such as the ones shown in the following examples. (18) 3; 1; 5 (child needs to put blocks in a bag to take them elsewhere): First I *unbuild it, okay? (19) 3; 10 (child trying to pull loose): *uncapture me Strikingly, children rarely used unwith irreversible actions, e.g., burn or hit. It was concluded that children developed the meaning of unfrom the general-purpose verb undo and their productive knowledge from several examples when they grew to understand their compositionality. <?page no="116"?> 103 A corpus study by Behrens (1998) revealed further details as to the developmental progression of prefix verb learning in German and English. 8 In their first three years of life, children used very few prefix verbs. More prefix verbs were produced in the fourth and fifth years of life including creative overgeneralizations evidencing increasing productivity, but overall prefix verb tokens accounted only for around 1% of children’s verb tokens (examples taken from Behrens 1998: 706). (20) Simone 3; 5; 15 (German): Die sind alle verloren *vergangen. - ‘They have all gone lost.’ (21) Naomi 3; 8; 19 (English): *unspilled on the floor Behrens further analyzed particle verbs, which she revealed were more frequent in the input reaching 22-30% of verb tokens in German and 11% in English. 9 Productivity was high for both [particle][verb] and [prefix][verb] combinations. German-speaking children used an average of 67% and English-speaking children an average of 32% of verb bases with at least one prefix or particle (1998: 708). Unfortunately, these numbers were not reported separately for prefix verbs and particle verbs. Clark (2003: 280) proposes that children store a repertoire of words in memory and use them as a template for interpreting unfamiliar ones and producing novel ones. Her idea is similar to the proposed steps of constructional learning. According to these steps children store at least two examples of an emergent construction, form a comparison over them that allows them to abstract their commonalities in a schema. This schema and the exemplars are subsequently stored as a constructional category and can be resorted to in the comprehension of unfamiliar examples and the production of new ones. Schema abstraction or, in the case of derivational morphemes, slot formation and productive generalization have been shown to be facilitated by type frequency (Clark 1993: 137-138, 146-149; Clark and Cohen 1984; Clark and Hecht 1982). Semantic transparency has further been suggested to be beneficial. Insights into the separability of prefix and base may be further facilitated by the high number of separable particle verbs in German. The fact that particle and base carry separate meanings might provide children 8 Behrens (1998) also analyzed corpus data of Dutch-speaking children. Due to the scope of this chapter, these data are not reported here. 9 Incidentally, Behrens did not find a lag of German children (particle either preposed to verb or separated) behind their English peers (particle always follows the verb) in particle verb learning. These results suggest that prefixed information is not in general disadvantaged in comparison to suffixed information (see also Weist 1983; Weist and Konienczna 1985; but see Hupp, Sloutsky and Culicover 2009 for an alternative point of view). Instead, “children seem to be perceptive to where which kind of information is encoded in the target language, provided the language offers sufficiently regular perceptual cues” (Behrens 1998: 707). <?page no="117"?> 104 with a way into understanding the compositionality of prefix verbs as well and thus give German-speaking children a head start on prefix learning. The high number of particle verbs and the inflectional circumfix (-)gemight further guide German children’s attention to the beginning of words more and make them more prepared to look for meanings of ‘word parts’ in this position. 10 5.1.4 The operationalization of pattern frequency In the discussion of the motivation of the present study, the pattern level was described as summarizing the commonalities of the closely-related verb prefix constructions. In a parallel manner, pattern frequency is operationalized in terms of cumulated frequencies. Frequencies are counted at the three levels that can be used in the exemplification of the pattern (cf. introduction of the pattern level in 4.6). The first level are verb prefix construction types (e.g., un[base] verb ‘to reverse the action expressed by the verb’, re[base] verb ‘to perform an action again’), the second level are prefix verb types (e.g., undo, untie, rename, redress) and the third level are prefix verb tokens. Counts are always summed up over all participating constructions so as to represent the frequency of the pattern. It is true that the fact that a lexeme in the corpus can be categorized as type of a particular construction does not entail that the speaker has categorized the lexeme as such. It is thus important to note that counting these lexemes as types of a construction or counting the constructions as types of the pattern is not to imply such a categorization in the respective speaker’s mind. In contrast, it is thought that speakers do not classify types as exemplars of a more abstract construction before they have abstracted the respective constructional schema and formed a category. 11 Even after schema abstraction and category formation a type that is experienced is not necessarily categorized as a member of a particular more abstract construction. The exact point of schematization or category formation is very difficult, if not impossible, to determine on the basis of canonical forms in the corpus data. However, creative overgeneralizations of constructions can provide insights into constructions that have previously been abstracted and are represented in the speaker’s mind. 10 Interestingly, German children tend to place the particle verb-initially in their early uses, e.g., Simone 1; 11; 27: Bilderbuch anguck ‘picture book at.look’ - ‘look at picture book’ (Behrens 1998: 702), even if this non-separation is not licensed by the grammatical context. This finding suggests that children in fact prefer the initial position for the particle early in the learning process. 11 It is possible and likely for one lexeme, word string or longer utterance to be a member of more than one constructional category. They can further be re-classified by speakers as members of different constructional categories and so forth. <?page no="118"?> 105 At this point it is important to address one issue. The previously reported literature on verb prefixes in German and English was naturally concerned with prefixes. The present study on the other hand is interested in prefix constructions. The difference is that prefixes can carry more than one meaning, whereas prefix constructions are pairings of a form and a single meaning. The counts of prefix types in 5.1.2 (based on Tables 4 and 5) thus only provides very limited insights into the frequencies of prefix constructions. 12 Regardless of this limitation, the higher number of prefix types in German (and the considerable number of foreign prefixes in English) served as a basis for the prediction that there would be more prefix constructions in German as well. 5.1.5 Hypotheses for the corpus study The subsequent corpus analyses set out to explore pattern frequency. Frequency at the pattern level was expected to be relevant to language learning because frequency both at higher and lower levels of generality had been reported to influence construction learning (e.g., Dressler 2005: 10; Gathercole 1986, 2002a, 2002b, 2002c; Hoff-Ginsberg 1998; Huttenlocher et al. 2002; Lieven 2008; Moerk 1978; Naigles and Ginsberg 1998; Theakston et al. 2004) and because frequencies of related constructions had also been shown to affect learning (Abbot-Smith and Behrens 2006; Kirjavainen et al. 2009). The frequency of the derivational verb prefix pattern was assessed in terms of verb prefix construction types, prefix verb types and prefix verb tokens. In absence of corpus studies in this area, expectations as to potential differences in pattern frequency between German and English relied mainly on the count of prefix types in the two languages and on the insight that there was a particularly high number of foreign prefixes in English that are unlikely found in the language to and of children. Hypothesis I predicted that the input frequency and thus availability of the verb prefix pattern is higher in German than in English. The second hypothesis was additionally based on the idea that German-speaking children might profit from their experience with the highly frequent and transparent particle pattern. Because of the separability and the separate meanings of particle and verb they might be more inclined than English-speaking children to analyse prefix verbs in the same way (Behrens 1998). Hypothesis II predicted the frequency of the prefix pattern to be higher in the Germanspeaking child’s productions than in those of his English-speaking peer. I Input: German versus English There are more verb prefix construction types, prefix verb types and prefix verb tokens in the input in German than in English. 12 Prefixes with more than one meaning were not subdivided into several prefix constructions with separate meanings at this point. <?page no="119"?> 106 II Output: German versus English There are more verb prefix construction types, prefix verb types and prefix verb tokens in German child speech than in English child speech. Previous research at the constructional level linked input and output distributions at different levels of abstractness (Behrens 2003, 2006; Huttenlocher et al. 1991). Hypothesis III deals with the relation between input and output frequencies in derivational morphology learning. Higher input frequencies were thought to be reflected in higher production rates in child speech. III Input-output relationship Higher frequencies of verb prefix construction types, prefix verb types and prefix verb tokens in the input are reflected in higher frequencies in child speech. Innovations, or overgeneralizations, received special attention. For canonical forms children produced, it was generally impossible to determine whether they were generalizations of an abstracted schema or previously encountered instances retrieved from memory. Even if a form was not present in caregiver speech in the corpus, there was the possibility that the child had experienced the form in his or her non-recorded input. Canonical forms thus do not reveal insights as to children’s ability to generalize productively. In contrast, overgeneralizations reveal a certain degree of analysis during the learning process, even though they are a transitional phenomenon. They imply that children have separated prefix and base, attributed meanings to each morpheme and formed a more abstract representation (schema/ construction) that can be generalized. These steps are necessary for constructional learning according to the account presented in Chapter 3. For this process, high type frequency in the slot position is particularly important, as studies on morphological and other partially-filled constructions showed (Clark and Cohen 1984; Clark and Hecht 1982; Guillaume 1927/ 1973; Lieven et al. 2003; Lieven et al. 2009; Matthews and Bannard 2010). Although innovative overgeneralizations decrease with age and vocabulary growth, in the explored age range a higher number of overgeneralizations was considered to be related to more advanced development. In line with this assumption and the higher input and output frequencies postulated in the previous hypotheses, the last hypothesis (Hypothesis IV) predicted a higher number of overgeneralizations for the German-speaking than for English-speaking child. More overgeneralizations were expected for prefixes with high prefix verb type frequency. IV Innovations in child speech There are more innovative overgeneralizations in German child speech than in English child speech in terms of overgeneralized verb <?page no="120"?> 107 prefix construction types, prefix verb types (different overgeneralizations using a prefix type) and prefix verb tokens (total number of overgeneralized tokens). Overgeneralizations are more common for prefix types with higher prefix verb type frequencies. 5.2 Corpus analysis 5.2.1 Database Two corpora were selected: One contained the speech of and to a Germanspeaking child, Leo (Behrens 2006); the other comprised comparable data for an English-speaking child, Thomas (Lieven et al. 2009). Both children came from middle-class backgrounds. These two sources were used, because they were comparable in terms of corpus size, length, age range covered, number of word tokens and density. The data taken into account here covered ages 2; 0 to 4; 11. Between 2; 0 and 2; 11 Leo was recorded five times per week for one hour each. From 3; 0 up to 4; 11 he was recorded five times per week for one week each month. For Thomas there were five one-hour recordings per week between 2; 0 and 3; 2. Between 3; 3 and 4; 11 there were five hour-long recordings for one week each month. The density of the data, particularly in the first year of recordings, was 5 to 10 times higher than usual. The abundance of data is the reason why only one child per language was selected for this analysis. The Thomas corpus was available on CHILDES (Child Language Data Exchange System; MacWhinney 1995) <http: / / childes.psy.cmu.edu/ data/ Eng-UK/ >. It included an additional morphology-line that gave word classes and further morphological information for each utterance produced by child and caregiver. The Leo corpus was also downloadable from CHILDES <http: / / childes.psy.cmu. edu/ data/ Germanic/ >. However, this version of Leo’s data did not include the morphology-line. This extra line was crucial for searches, since it allowed combined searches for all inflectional forms of a verb. The more inclusive version of the Leo corpus was made available by courtesy of the Max Planck Institute for Evolutionary Anthropology, Leipzig. The morphology-line was missing for the caregiver speech in a number of files of the Leo corpus. For this reason all such files from the German corpus and the files of corresponding ages of the English corpus were omitted from searches. Consequently, an equal amount of speech data from the following ages was included in the analysis of the two languages: 2; 0, 2; 1, 2; 7, 2; 8, 3; 2, 3; 3, 3; 9, 3; 10, 4; 4, 4; 5, 4; 6, 4; 7, 4; 8, 4; 9, 4; 10, 4; 11. CLAN software was used to perform the corpus searches. This programme was also available on CHILDES <http: / / childes.psy.cmu. edu/ clan/ >. <?page no="121"?> 108 5.2.2 Procedure and results Corpus size The subsequent analyses were based on corpora with the token counts given in Table 6. 13 The number of word tokens produced by Leo and Thomas was comparable, and so was the number of word tokens produced by the caregivers. The same was true for the numbers of verb tokens produced by children and caregivers. If anything, there was a slight trend towards higher totals for Thomas and his caregiver than for Leo and his caregiver, except for the number of children’s verb tokens. Only token counts are given in Table 6 because CLAN counts each different form of a lexeme as a type. Consequently, adopting the type counts would have overestimated German counts since the number of different inflections is higher than in English. 14 In the analyses of caregiver speech the entire corpora (as described in the table below) were used instead of sub-corpora that were sizematched with children’s corpora. The reason for this procedure was that the present study was interested in overall input frequencies of the verb prefix pattern children were exposed to. Table 6. Word and verb tokens in English and German corpora. Speaker Total word tokens Total verb tokens Thomas 383,662 28,136 Caregiver Thomas 1,386,834 116,100 Leo 341,404 28,663 Caregiver Leo 1,011,081 100,873 Counts based on transcribers’ coding In a first survey analysis the original coding of prefixes as per the transcribers served to identify prefix verbs. In the German corpus derivational prefixes were separated from the base by ‘+’ on the morphology-line (e.g., 13 Verb token counts were not checked manually for miscodings (e.g., lexemes mistakenly tagged as verbs and verbs mistakenly not tagged as such), since this was impossible due to their size. 14 For instance, there are ten possible forms of German gehen ‘go’ and only five for English go, when the different auxiliaries are not considered. German gehen ‘go’: gehe ‘go’ 1 st person singular present tense, gehst ‘go’ 2 nd person singular present tense, geht ‘go’ 3 rd person singular present tense and 2 nd person plural present tense, gehen ‘go’ infinitive, 1 st and 3 rd person plural present tense, ging ‘went’ 1 st and 3 rd person singular past tense, gingst ‘went’ 2 nd person singular past tense, gingen ‘went’ 1 st and 3 rd person plural past tense, gingt ‘went’ 2 nd person plural past tense, gegangen ‘gone’ past participle, gehend ‘going’ present participle; English go: go present tense all persons except 3 rd singular and infinitive, goes 3 rd person singular present tense, went past tense, going present participle, gone past participle. <?page no="122"?> 109 be+streuen ‘sprinkle on’), in the English corpus ‘#’ was used (e.g., un#do). 15 Verb prefixes were counted manually from the lists the search returned. Prefix verb types were also counted manually, so that frequencies were not distorted based on the number of inflections in each language. Thus, one type included all inflected forms of a verb. Children’s overgeneralizations were included in the analysis. In these preliminary counts, verb prefixes rather than verb prefix constructions were counted, because prefix meanings could not be ascribed in all cases due to transcription errors. Counts are summarized in Table 7. They revealed that Thomas’ caregiver used much fewer verb prefixes, prefix verb types and prefix verb tokens than Leo’s caregiver. The same discrepancy was found between Thomas and Leo. The relation between caregiver input and children’s output in both languages were such that the higher overall verb prefix, prefix verb type and prefix verb token frequencies were in the input, the higher they were in children’s uses too. The number of overgeneralizations was not analyzed at this stage. Table 7. Verb prefixes, prefix verb types and tokens as coded by transcribers. Speaker Total verb prefixes Total prefix verb types Total prefix verb tokens Thomas 1 4 8 Caregiver Thomas 5 18 80 Leo 21 218 874 Caregiver Leo 30 463 3203 Note: These numbers of prefix verb types were checked for repeated symbol use, e.g., even if un#do and un##do were displayed separately, they were counted as the same type. The output was not edited in any other way. Problems with transcribers’ coding and coping strategies A series of problems surfaced and called for corrective measures before more informative counts were made. One problem was that not all prefixes were marked as such by the operators ‘+’ and ‘#’. Instead, numerous cases of prefixes were not marked at all or marked erroneously with the operator used in the other language (e.g., ‘+’ instead of ‘#’). To remedy this and further shortcomings prefix verbs were subsequently selected on the basis of complete lists of verbs of both corpora (for prefix selection criteria see below). In the ensuing search, all selected prefix verbs were searched using the operators ‘+’ and ‘#’ (each one in each language), using repeated operators (‘++’ and ‘##’) and using no operators at all. The manual prefix selec- 15 See Appendix I.1 and I.2 for brief excerpts from each corpus. <?page no="123"?> 110 tion had the added advantage of simultaneously allowing the exclusion of miscoded cases, such as V| out#side, which is not a verb, or emp+fehlen ‘recommend’, where empis not a prefix. It had the additional advantage that erroneous codings of particle verbs as prefix verbs were of no consequence, since only verbs included in the prefix verb list were searched for regardless of their original coding. In the case of homographic particle and prefix verbs, e.g., umsegeln, which means ‘hit while sailing’ as a particle verb and ‘sail around’ as a prefix verb, the context was consulted to resolve whether the verb was used as a prefix verb or not. The problem that CLAN counts each different form of a verb as a type was bypassed again by manual verb prefix and prefix verb counts. The manual selection of prefix verbs based on a list of all verbs was also necessary because the definition of what transcribers considered a prefix was not entirely transparent and did not seem theory-driven given the examples. For the manual prefix verb selection, the following criteria were applied. Prefixes had to carry an identifiable meaning relatively independently of the base 16 they were attached to, that is, prefixes had to be relatively transparent. This is not a necessary part of the traditional definition of prefixes. But it was thought logically impossible for children to generalize opaque structures, since children were expected to process opaque prefix verbs in the same way as simplex verbs. As a result of this measure extremely highly frequent types of several prefix patterns were excluded, for instance ver-stehen ‘under-stand’. High frequency has been reported to go hand in hand with processing as chunk and loss of transparency (Bybee 2006). In inflectional morphology learning, types with the highest token frequencies have been shown to contribute very little - if anything - to schema formation (Bybee 1995), which justifies the exclusions. The approach applied resulted in a relatively high degree of compositionality of prefix verbs. This was intentional 17 and did not preclude the possibility that the composite meaning of [derivational prefix]+[base] verb was more comprehensive than those of prefix and base combined (Schmid 2005: 165). Even though the corpus constituted only a part of the language children heard, each verb prefix construction had to occur with at least two different verb types in the entire corpus in order to ensure it could potentially be perceived as a construction. This was done because Wonnacott and colleagues (2012) found that children did not generalize a new argument 16 Please note that bases were mainly verbal, but other bases were accepted also. This criterion proved to be appropriate in the experiments (see Chapter 6) were several children attached a novel verb prefix to nominal instead of verbal bases in a production task. 17 The criterion of relatively high compositionality was applied in the corpus study, because the newly invented construction that was used in the experiment in Chapter 6 was also necessarily compositional in nature. <?page no="124"?> 111 structure construction after exposure to a single type. All the criteria applied caused many more exclusions in German than in English. This was accepted, since it resulted in a more conservative test of the hypotheses. Prefix meanings were ascribed on the basis of the literature. Each pair of a prefix and a meaning was considered a prefix construction (type). It was attempted to use as many distinct meanings as necessary, but not to make particularly fine-grained subdivisions that children would be unlikely to make. Several German prefixes had more than one meaning. Further details of this procedure including difficulties are discussed in 8.4.2. Comprehensive lists of English and German prefix constructions are given in Appendix I.3 (Tables a and b) including the sources used for the classification of meanings. Whether multiple prefix meanings were cases of prefix polysemy or homonymy was considered irrelevant to the matter investigated. Counts based on manual selection of prefix constructions Table 8 gives the results of the counts of verb prefix construction types, prefix verb types and prefix verb tokens based on the manual selection of prefix verbs. Appendix I.4 provides more comprehensive tables (Tables c and d) including frequencies of all prefix verbs used by the two children and their caregivers. Despite the more conservative criteria applied in the second corpus search, the overall picture remained the same. Leo was exposed to nearly nine times as many verb prefix construction types as Thomas ( 2 (1) = 18.241, p < .001), eight times as many prefix verb types ( 2 (1) = 136.111, p < .001) and more than six times as many prefix verb tokens ( 2 (1) = 726.450, p < .001). All chi-square tests were significant. These findings support Hypothesis I that input frequencies of the verb prefix pattern are higher in German than in English. Table 8. Prefix verbs selected manually. Speaker Total verb prefix constructions Total prefix verb types Total prefix verb tokens Thomas 2 7 25 Caregiver Thomas 3 25 177 Leo 22 106 412 Caregiver Leo 26 200 1164 Children’s productivity pointed into the same direction - if anything, even more so: Leo produced 11 times as many verb prefix construction types as Thomas ( 2 (1) = 16.667, p < .001); he used more than 15 times as many prefix verb types ( 2 (1) = 86.735, p < .001) and over 16 times as many prefix verb tokens ( 2 (1) = 342.721, p < .001). Again all chi-square tests were <?page no="125"?> 112 significant. These counts support Hypothesis II that the verb prefix pattern is more frequent in the speech of the German-speaking child than in that of his English-speaking peer. The input-output relation was examined overall using the same table. Higher input frequencies were demonstrably related to higher frequencies in children’s productions, both for Leo and Thomas, and with respect to the frequency of verb prefix constructions (3-2, 26-22), prefix verb types (25-7, 200-106) and prefix verb tokens (177-25, 1164-412). These findings provide support for Hypothesis III. A breakdown of tokens and types by construction types implied that input and output frequencies are also related at the constructional level (see Appendix I.4, Tables e and f). For English the un[base] verb construction was most frequent in terms of types and tokens in Thomas’ input and his own uses. Re[base] verb was lower in frequencies in both input and output and over[base] verb was lowest and only appeared in caregiver speech. The frequency rank orders for child and caregiver were thus the same. Due to the low number of construction types, no further statistical analyses were performed. In the German corpus the prefix constructions with highest and the lowest numbers of verb types and tokens were the same for caregiver and child, but rank orders did not correspond one to one for all constructions. Due to the high number of construction types in German it was possible to correlate input and output type and token frequencies. Correlations between input and output were significant for both token frequencies (r(24) = 0.813, p < .001) and type frequencies (r(24) = 0.892, p < .001). These findings show the input-output relations for the constructions types summarized by the pattern and thus provide further support to Hypothesis III. Subsequently, forms children produced that were not present in the input, in particular their innovative uses, were analyzed. In Thomas’ speech there was one canonical form that was not present in the input; Leo produced 24 such forms involving 10 different prefix constructions. Due to the limited nature of corpus data it was indeterminable whether children had heard these forms in their non-recorded input or generalized them on the basis of an abstraction. Such forms therefore provide only very weak support for the suggestion that children abstracted and generalized the respective constructional schemas. More reliable evidence for schema abstraction and generalization comes from innovative overgeneralizations children produced. In Thomas’ speech, there was one such form: unopen. 18 In this noncanonical form the most frequent prefix construction un[base] verb was com- 18 Searches for overgeneralizations in Thomas’ and Leo’s speech were based on the entire corpus, that is, all files between 2; 0 and 4; 11, since the morphology-line was given throughout for the two children. The morphology-line was missing only for the German caregiver, which did not affect the present counts. <?page no="126"?> 113 bined with the free lexical morpheme open. Thomas produced it at 4; 7; 10 when he and his caregiver discussed how bottles are opened. The context does not make Thomas’ intended meaning of unopen entirely clear. It remains open whether he intended it to mean ‘close’, i.e., ‘undo the opening action’, or the opposite ‘open’ in vague analogy to uncork. Consequently, this form might be an overgeneralization of the un[base] verb prefix construction, but it is by no means unambiguously so. At best, there was thus probably one overgeneralization in the entire corpus for Thomas. Leo produced 11 different non-canonical forms (11 types), some of which he used repeatedly (16 tokens). Their analysis is presented chronologically, starting with the earliest form. At 2; 6; 14 Leo used verschneiden ‘ver+cut’. 19 This form might be an example of the construction ver 1 [base] verb 20 ‘to perform an erroneous action; unintended event; to act in the wrong way’, e.g., ‘to cut by mistake or in the wrong way’ or ‘to cut something off unintentionally’. The transcriber’s comment however suggested a different interpretation: a blend of verletzen ‘harm’ and schneiden ‘cut’, presumably meaning ‘getting harmed when cutting’. This first instance is thus not an innovative use of a prefix construction. At 2; 7; 6 Leo produced verschlängeln ‘ver+snake’ in connection to a Schlange ‘snake’. The intended meaning could not be inferred unambiguously from the context. It might have referred to an unsuccessful attempt of drawing a zigzag line or to a snake that took the wrong way or got entangled. In this case it would be an example of the ver 1 [base] verb ‘to perform an erroneous action; unintended event; to act in the wrong way’ construction. But it remained open whether this use was a genuine overgeneralization. The next case was characterized by an over-application of morphemes: durchzerschneiden ‘through+zer+cut’ (2; 7; 18). The child used the particle construction durch[base] verb ‘through/ in pieces’ and the prefix construction zer 1 [base] verb ‘action or event resulting in the subdivision into smaller parts’. The same meaning was thus expressed by durchand zer-. And while the use of either one construction would have been appropriate (both zerschneiden and durchschneiden appeared in the corpus), the combination was not. Nevertheless, Leo’s over-application suggests that he was able to generalize each of the two constructions. At 2; 8; 15 Leo produced his first entirely unambiguous overgeneralization. He and his caregiver talked about Christmas cookies they were baking and Leo called sprinkling powdered sugar on top of the cookies beschneien ‘be+snow’, i.e., ‘cover with what looks like snow’. The following day Leo (2; 8; 16) referred to snow-covered cars with beschneien. Both these 19 In the English translation only the base is translated. The German prefix is given in German rather than being translated, since the meaning of the prefix is subsequently discussed. 20 The subscript numbers indicate the prefix meaning as listed in Appendix I.3. <?page no="127"?> 114 cases were overgeneralizations of the construction be 1 [base] verb ‘apply something; equip/ cover something with something; attach something to something; can be driven by a purpose’. The first usage additionally involved an extension of the meaning ‘snow’. The two tokens of beschneien were thus even used in different contexts. Leo used zerstürzen ‘zer+tumble’ to refer to a damaged train track aged 2; 9; 17. This is again an unclear case. In this situation zerstören ‘destroy’ or the particle verb einstürzen ‘tumble down’, if the track is on a bridge, might be used. Leo’s use of the prefix construction zer 2 [base] verb ‘performance of an intentionally or unintentionally damaging action’ implied the destructive character of the event. The base verb stürzen alone means ‘fall, topple’. The context made it impossible to determine whether Leo’s form was a mispronunciation of zerstören or a genuine overgeneralization of the zerconstruction to the base stürzen with the meaning of ‘damaging something by falling into it’. At ages 2; 10; 20, 2; 10; 23, 2; 11; 4 and 2; 11; 27 Leo used vertarnen ‘ver+disguise’ to refer to an animal (a snake) that wanted to hide and go unnoticed. He overgeneralized ver 3b [base] verb ‘action or event resulting in more obscure/ different state’, a constructional category that also includes verhüllen ‘veil, disguise’, verkleiden ‘dress up, disguise’, verdecken ‘cover’. He did not realize that in the case of vertarnen, the meaning of ‘disguise’ was already incorporated in the verb tarnen that he used as a base. This overgeneralization occurred repeatedly over a limited period of time. Subsequently, at 3; 4; 10 Leo used bewinken ‘be+wave’ twice in connection to a train track or journey. His caregiver asked him what he meant, but he did not give conclusive indications, so the meaning remained entirely unclear. Aged 4; 1; 14 Leo produced befärben ‘be+colour’ to refer to his fingers that were covered in stamp ink. Ich hab mir ein bisschen die Finger mit Stempelfarbe befärbt. ‘I have me a little bit the fingers with stamp ink be+coloured.’— ‘I covered my fingers with a little bit of stamp ink.’ Here, Leo overgeneralized the prefix construction be 1 [base] verb ‘apply something; equip/ cover something with something; attach something to something; can be driven by a purpose’, in order to express ‘cover in colour/ ink’. It is a clear case of overgeneralization. Leo produced verschwierigen ‘ver+harder’ twice at 4; 6; 15 when playing a card game. His caregiver asked him why he did something complicated and Leo answered um es mir zu verschwierigen ‘to it me ver+harder’ - ‘to make it harder for me’. He used verdickern ‘ver+thicker’ the following day aged 4; 6; 16 when discussing a bump on the head. Leo inquired whether the bump was a result of the veins getting thicker. Both these cases were overgeneralizations of the same prefix construction ver 3a [base] verb ‘to make or become (more) [adjective]’. More precisely the construction involves <?page no="128"?> 115 ver[(comparative of) adjective]+(e)n. Verschwierigen has the antonymic analog vereinfachen ‘make easier’, which is based on this schema, and verdickern closely resembles semantically related verlängern ‘make longer’. At 4; 10; 5 Leo used verkrickelkrackeln when he and his caregiver tried to write. Leo complained that the paper was nearly all used up, since his caregiver had verkrickelkrackelt ‘ver+chicken-scratched’ all the paper. This instance was likely an overgeneralization of construction ver 1 [base] verb ‘to perform an erroneous action; unintended event; to act in the wrong way’, assuming that Leo thought it was wrong to use up all the paper by chickenscratching. Leo’s innovative productions made use of three to five prefix constructions. For three constructions Leo’s uses were clearly defined as overgeneralizations. Two constructions, he generalized to two different types. All five constructions Leo potentially used innovatively were among the seven most frequent constructions in terms of type frequency both in his input and his own productions. High type frequency thus seemed to foster innovative overgeneralizations. Nevertheless, not all prefix constructions with high type frequencies were overgeneralized. One reason might be that not all overgeneralized forms Leo produced were recorded, since corpus data is limited. Nevertheless, this fact suggests that other factors were involved as well. Innovative uses are not solely dependent on constructional type frequency, but also further affected by factors such as transparency of meanings, frequency of preempting forms, productivity of the morpheme, ease of formation of the novel form, the size of the child’s vocabulary, topics of conversation and potential restrictions on the bases a prefix can attach to (Clark 2003: 281, 283). Leo is eventually expected to retreat from using the creative generalizations he formed due to the mechanisms of entrenchment and preemption discussed in Chapter 2 (2.2.2). Alternative forms will be more frequent in the input than the creative overgeneralizations Leo formed and will thus become progressively entrenched and at the same time preempt the persistence of Leo’s creative forms. As predicted by Hypothesis IV Leo overgeneralized more than Thomas. Leo produced overgeneralizations of three to five prefix construction types compared to none or a single one in Thomas’ speech. Leo formed more than one novel prefix verbs type for at least two prefix constructions; and the overall number of tokens of Leo’s overgeneralizations was higher than Thomas’, even if ambiguous cases were disregarded. Leo further overgeneralized those prefix constructions that had very high type frequencies. 5.3 Discussion In summary, the data supported the hypotheses. The availability of the verb prefix pattern in terms of verb prefix construction types, prefix verb types and tokens in the input was significantly higher for the German- <?page no="129"?> 116 speaking child than for his English-speaking peer (I). Leo also produced higher numbers of verb prefix construction types, prefix verb types and prefix verb tokens in this own speech than did Thomas (II). Comparing input and output of the pattern showed that overall, higher numbers of verb prefix construction types, prefix verb types and tokens in the input were related to higher frequencies in child speech. Inspecting the inputoutput relationship of the individual prefix construction types revealed the same tendency, i.e., the children tended to produce more types and tokens of prefix types or meanings the more frequent these were in the input (III). The numbers of innovative overgeneralizations in terms of prefix construction types, prefix verb types and tokens were higher in Leo’s than in Thomas’ speech (IV), providing support for the underlying supposition that the German child had already abstracted the schemas of several prefix construction types. The fact that Leo selectively overgeneralized high type frequency constructions further supports the idea that type frequency contributed to schematization and thus generalization. There was hardly any previous research on the frequencies of derivational verb prefix constructions and the related pattern in German or English, least so for spoken language to and of children. It was merely the number of verb prefix types according to grammars of German and English that suggested that the frequency of the verb prefix pattern in caregiver speech might be higher for German than for English. The expectation that German-speaking children might be more productive with the prefix pattern was further supported by the idea that the high frequency of separable particle verbs in German might help children separate prefix and base. The present corpus study is the first to provide empirical evidence for the higher frequency of the verb prefix pattern in German than in English (i.e., more different prefix constructions, prefix verb types and tokens). Incidentally, the most frequent prefix constructions in child and caregiver speech in German and English coincided with the prefixes reported to be most frequent in the literature, despite the differences in genre and modality (Fleischer and Barz 1995: 325; Schmid 2005: 161). The idea that frequencies added up at the pattern level should be relevant to construction learning is based on previous research at lower and higher levels of generality. Research summing up all constructions that use a certain typological feature in order to explore children’s acquisition of that feature and studies adding up frequencies of all complex constructions in order to investigate children’s complex construction learning had revealed very general effects of input frequency (Aksu-Koç and Slobin 1985; De Villiers and De Villiers 1985; Dressler 2005: 10; Hoff-Ginsberg 1998; Huttenlocher et al. 2002; Smoczynska 1985). Moreover, studies of individual constructions, which constitute a lower level than the pattern level, had shown effects of input frequency on the speed of acquisition, the order of <?page no="130"?> 117 acquisition, the number of errors and so forth both within a language and cross-linguistically (e.g., Gathercole 1986, 2002a, 2002b, 2002c, 2009: 114; Lieven 2008; Marchman 1997; Maratsos 2000; Maslen et al. 2004; Moerk 1978; Naigles and Ginsberg 1998; Theakston et al. 2004). Additional research had suggested that constructional learning is not only affected by the frequency of the construction itself in the input, but also be the frequency of related structures (Abbot-Smith and Behrens 2006; Kirjavainen et al. 2009). The present study provides first support to the idea that input frequencies summed up over different construction types, i.e., pattern frequency, are related to output frequencies. Further research is needed to explore the effects of pattern frequency on construction learning in more detail. The finding that innovative uses of prefix constructions were more common for Leo provides indications that higher pattern frequency might facilitate the learning of individual prefix constructions. Moreover, the result that Leo tended to use construction types with very high type frequencies in his overgeneralizations supports previous research. High type frequency is thought to be conducive to the abstraction of variable slots in partially-filled constructions, which in turn invite generalization to new cases (Bybee 1995; Bybee 2010: 95-96). Guillaume (1927/ 1973) showed this relation for inflectional morphology in French. In his study, children overgeneralized the verb conjugation class with the highest type frequency rather than the one with the highest token frequency. Clark and Cohen (1984) and Clark and Hecht (1982) showed the influence of type frequency on innovative uses of derivational morphology on nouns. The present study extended these findings to several derivational verb prefix constructions and explicitly related input type frequencies and innovative uses of the same child. As proposed earlier, it is possible that the advantage for Germanspeaking Leo was not singularly based on the higher availability of the prefix pattern in the input, but might have been supported by the existence of particle verbs. Leo’s experience with separable particle verbs might have made him more inclined than Thomas to attach separate meanings to verbinitial inseparable prefixes as well (Behrens 1998). Exposure to particle verbs and verbs using the circumfix ge[stem]t/ ge[stem]en, (where at least part of the form is positioned verb-initially) might further have caused the German child to become increasingly inclined to attend to the beginnings of verbs. A similarly strong bias is unlikely in English, because only prefix verbs occur in this position and they are less frequent than in German. Incidentally, it should be noted that frequency is not understood to be the only factor in the learning process. Factors such as attention, recency, context, semantic transparency, perceptual salience, linguistic complexity, frequency of preempting forms, productivity of the morpheme, ease of <?page no="131"?> 118 formation of the novel form, the size of the child’s vocabulary and potential restrictions on the bases a prefix can attach to are expected to moderate the relationship between input and output (Clark 2003: 281, 283; Diessel 2004: 30; Gathercole and Hoff 2009; Ebbinghaus 1913: 90-123; Kline and Demuth 2010; Narasimhan and Gullberg 2011). A comprehensive model of the learning process will have to include all these and possibly more influencing factors. Naturally, the generalizability of the findings is limited, since only the speech of two children and their caregivers was analyzed. Future research will have to explore the effects of pattern frequency on construction learning further. A first step towards this aim was taken in the experimental studies in this book. Pattern frequency was manipulated in one of the experiments as well as input token frequency, type frequency and type-token ratio in order to gain insights into the role of frequencies in morphological construction learning. The findings of the present study were used in this endeavour. Since the derivational prefix pattern had been revealed to differ in frequency between German and English (at least in the speech to and of children), it could subsequently serve as a basis for the formation of one of the novel constructions and was further the source of the prediction that this novel construction would be easier to learn for German-speaking children than for English-speaking children. <?page no="132"?> 119 6 Experiment 1: The role of input frequencies at constructional and pattern levels in novel morphology learning Usage-based, constructionist approaches hold that children learn language from the input they hear (cf. Chapter 2) using cognitive abilities of categorization and analogy as well as their social-cultural skills (cf. Chapter 3; Goldberg 1995, 2006; Tomasello 2003). Among other things, this assumption entails two implications: Firstly, for this account to model language learning adequately evidence is required showing that children are indeed able to learn form-meaning pairings from what they hear with the help of domain-general cognitive processes. And secondly, there should be evidence that input characteristics, e.g., input frequencies, affect learning. The present study aims at providing such evidence. Children were exposed to one of two novel constructions and then their ability to act-out, comprehend and produce the novel structure was assessed. The novelty of the construction allowed the exploration of learning from the input. Effects of input frequency were examined at a number of levels. The impact of the frequency of the pattern that the novel constructions were based on was explored as well as the effect of experimental token frequency on the representational strength and the entrenchment of types and the role of type frequency in generalization. A relatively wide age range of children was examined so as to follow the developmental progression and to assess differences in the degree of learning. In the following, previous research (as presented in Chapters 2-5) is summarized briefly in order to provide a background to the hypotheses of the present study. Subsequently, the method is described and the results are presented. Finally, the results are discussed with a view to the hypotheses and the broader research context. 6.1 Background and hypotheses Children’s ability to learn novel constructions was investigated in the area of derivational morphology. Novel constructions were used in order to control children’s previous exposure. Morphological constructions were selected so as to extend previous research on novel construction learning, which had remained limited to abstract word order constructions (Boyd and Goldberg 2011a; Casenhiser and Goldberg 2005; Wonnacott et al. 2012). Children with different native languages - German-speaking children and English-speaking children - were tested. This approach had the advantage <?page no="133"?> 120 that effects of the underlying pattern frequency that varies between the two languages could be assessed. Moreover, using two languages made it possible to explore whether the same learning mechanisms are at work in different languages. Additionally, previous studies with novel constructions had been restricted to English (Boyd and Goldberg 2011a; Casenhiser and Goldberg 2005; Wonnacott et al. 2012), so that any investigation of a different language potentially broadens the knowledge about novel construction learning. Frequency was examined at a very general level that had not been under investigation in novel construction learning before, the pattern level. As discussed in Chapter 4, a pattern captures the commonalities of several constructions at a more abstract level, e.g., the [derivational prefix][base] verb pattern captures the commonalities of un[base] verb , dis[base] verb and other similar constructions. Frequencies added up at the pattern level (pattern frequency) were expected to affect novel construction learning because of several previous findings. The input frequency of individual constructions had been shown to affect the speed, order and errors in children’s learning of the respective construction within a language and cross-linguistically (e.g., Gathercole 1986, 2002a, 2002b, 2002c, 2009: 114; Lieven 2008; Marchman 1997; Maratsos 2000; Maslen et al. 2004; Moerk 1978; Naigles and Ginsberg 1998; Theakston et al. 2004, cf. Chapter 4). Further research had revealed that not only the frequency of a construction itself but also the frequency of related structures influences construction learning (Abbot- Smith and Behrens 2006; Kirjavainen et al. 2009). The frequency of related structures was also taken into account in studies exploring the effects of summed-up frequencies over all constructions that share a typological feature on children’s acquisition of such constructions (Aksu-Koç and Slobin 1985; De Villiers and De Villiers 1985; Dressler 1997; Smoczynska 1985; cf. 4.4.3) and research on complex construction learning investigating the effects of summed-up frequencies over all complex constructions (Barnes et al. 1983; Hoff-Ginsberg 1998; Huttenlocher et al. 2002, cf. 4.4.2). Finally, the corpus study (cf. Chapter 5) provided first evidence of pattern frequency effects in naturalistic language. The present study is the first to explore effects of pattern frequency on novel construction learning. The first novel construction was based on the derivational verb prefix pattern that had been shown in Chapter 5 to be more frequent in German than in English in the speech to and of children. As a consequence, learning the novel verb prefix construction was expected to be easier for Germanspeaking children than for their English-speaking peers (Hypothesis I b). At the same time, it was predicted that children would be able to learn the construction to some degree (Hypothesis I a). The second construction was selected following the insight that none of the previous novel construction learning studies (including the novel prefix <?page no="134"?> 121 learning) had used a genuinely novel construction, that is, a construction based on a pattern that does not exist in the participants’ native language. Earlier studies had made use of novel word orders in English (Boyd and Goldberg 2011a; Casenhiser and Goldberg 2005; Wonnacott et al. 2012) and the present study used a novel verb prefix in German and English. While the particular constructions were indeed novel in all cases (e.g., the word order [Noun Phrase 1 ] [Noun Phrase 2 ] [Verb Phrase] was not known to express ‘appearance’ or ‘approach’), word order and prefix constructions in general were known to carry meaning in the respective languages. For this reason, the second novel construction was based on a reduplication pattern that is completely absent from both German and English. More specifically, a novel verb-initial reduplication construction was used. Since underlying pattern frequency was zero for both languages, no differences were expected between languages for the learning of the novel reduplication construction (Hypothesis II b). It was, however, predicted that even the genuinely novel construction would be learned by the children to some degree (Hypothesis II a). Seeing that underlying pattern frequency was different for the novel prefix compared to the novel reduplication construction, the comparison between the two suggested itself. Since the learning of a genuinely novel construction had never been tested before and the pattern level was as yet unexplored, no such comparison had been reported in the literature. It was performed in the present study with the prediction that the prefix construction based on the familiar pattern would be easier to learn than the entirely unfamiliar reduplication construction (Hypothesis III). However, due to further potential differences between the two constructions (e.g., level of abstractness), the comparison and its results only have very tentative character. Learning was followed over an age range from 3 to 8 years in the present study. The reason was that learning was expected to be incomplete after the brief experimental exposure and that the degree of learning would vary. The assessment of children of different ages was thought to allow conclusions as to children’s developmental progression in the learning process. A previous novel construction learning studie had revealed age effects (Boyd and Goldberg 2011a). Performance was thus expected to increase with age in the present study both for prefix and for reduplication learning (Hypotheses I c and II c). Input token frequency was examined in the present study because previous research had implied that it affects the representational strength of exemplars in children’s minds and with them language learning. In particular, higher input token frequencies had been revealed to lead to speedier and earlier acquisition of constructions (lexemes: Hart 1991; Huttenlocher et al. 1991; Naigles and Hoff-Ginsberg 1998; Theakston et al. 2004; <?page no="135"?> 122 inflectional morphemes: Moerk 1978; partially-filled constructions: Rowland and Pine 2000; Rowland et al. 2003; abstract constructions: Ambridge et al. 2008; Gathercole and Hoff 2009; Matthews et al. 2005, 2007) and to fewer errors (inflectional morphology: Maratsos 2000; Marchman 1997; Maslen et al. 2004; Theakston, et al. 2003; abstract constructions: Theakston 2004). While none of these studies had explored token frequency effects on novel construction learning in the realm of morphology, similar effects were expected in the present study. Input token frequency was increased gradually throughout the experiment, i.e., different verb types were used more frequently in the novel construction than others. Children’s performance on types of the novel construction was predicted to reflect the respective token frequency, i.e., to increase concomitantly with token frequency (Hypotheses I d and II d). New types with a previous token frequency of zero were expected to be the most difficult, since they required generalization. A certain number of types were thought to be necessary for children to form a more abstract schema of the construction and subsequently generalize it to new verbs. This assumption was based on the understanding of the construction learning process developed in Chapters 3 and 4. It was proposed that a schema is abstracted following a comparison process where two or more types of a construction are relationally aligned and mapped. The logical prerequisite is thus that at least two types of a construction have reached a degree of representational strength that allows this comparison. While comparison and schema abstraction are unamenable to testing, its implications - children’s generalizations - are testable. Previous research supports the assumption that a single type is insufficient for comparison and subsequent generalization in children. Wonnacott and colleagues (2012) found that 5-year-olds were unable to generalize a novel word order construction after exposure to a single type, even if its token frequency was so high as to ensure storage. Rather than examining the number of types children were exposed to, the present study assessed the number of types children had reliably stored (as measured by their ability to act them out, comprehend or produce them) and related it to the number of generalizations the children formed. Based on the evidence from Wonnacott and colleagues and the logical requirements of the comparison process, it was expected that children would show storage of at least two types before generalizing the novel construction (Hypothesis IV). No differences were expected between prefix and reduplication learning, since there was no reason to assume that the minimum number of types needed should differ between the partially-filled prefix construction and the more abstract reduplication construction. The present study thus aimed at exploring effects of pattern frequency as well as input token and type frequency on novel construction learning. <?page no="136"?> 123 Higher pattern frequency was predicted to facilitate learning of a novel construction. Due to its relation to representational strength and entrenchment, higher input token frequencies were thought to facilitate children’s use of the respective types. Two types were expected to be necessary for generalizations of the novel construction. It was further assumed that learning would improve with age. Exploring two languages was thought to provide insights as to the equivalence and the generality of the learning processes involved. Novel morphological constructions were used to extend knowledge on novel construction learning and morphological learning. The hypotheses are summarized below. The results section roughly follows this structure. I Learning of a novel prefix construction is affected by the pattern frequency in the native language, i.e., the frequency of the pattern that the novel construction exemplifies. Higher pattern frequency results in more successful learning. a Children are able to learn a novel prefix construction from the input. b German-speaking children do better than English-speaking children in novel prefix construction learning. c Learning increases with age. d Performance increases concomitantly with token frequency. II Learning a novel reduplication construction exemplifying a pattern that is not present in the learners’ native language (i.e., whose pattern frequency equals zero) is equally difficult for speakers of different native languages. a Children are able to learn a novel reduplication construction from the input. b German-speaking and English-speaking children learn the novel reduplication construction equally well. c Learning increases with age. d Performance increases concomitantly with token frequency. III Learning a novel construction exemplifying a familiar pattern is easier than learning a novel construction exemplifying an entirely unfamiliar pattern (i.e., learning the novel prefix construction is easier than learning the novel reduplication construction). IV At least two types (i.e., different verbs in the novel construction) are stored before the novel construction is generalized to new cases. 6.2 Method 6.2.1 Participants One hundred sixty-eight German-speaking children aged 3 (M=3; 6), 4 (M=4; 5), 5 (M=5; 5), 6 (M=6; 5) and 8 (M=8; 9) were recruited from seven <?page no="137"?> 124 nurseries and one primary school in Munich, Landshut and Wolfsburg, Germany. The data of eight children were excluded when it emerged after testing that they suffered from specific language impairment. Ninety-six English-speaking children aged 4 (M=4; 5), 5 (M=5; 5) and 6 (M=6; 6) were recruited from four nurseries and primary schools in Greater Manchester and Cheshire, England. The age range was greater in German to allow tracking the entire developmental progression. Since the development between 4 and 6 emerged to be the most relevant to the present study, the age range was restricted to 4 to 6 years for the English sample. In each age group there were 32 children (16 boys and 16 girls) in each language. Children were randomly assigned to one of two novel construction learning conditions (pattern). Gender was balanced within each group (8 boys and 8 girls). Children received stickers as a small reward for their participation and participating nurseries and schools received gifts after the completion of the study. 6.2.2 Material The first novel construction was based on the derivational verb prefix pattern. The novel prefix construction used in German was ta[verb]. Tacannot stand alone in German and it does not mean anything on its own. In the experiment, taappeared only before verbal bases. The meaning assigned to it was ‘pretend to perform an action’. This meaning is not expressed by a morpheme in German; instead it is usually expressed by the phrase so tun als ob ‘do as if’. The novel prefix construction in English was va[verb]. One reason for this difference was that two English-speaking children confused the novel prefix tawith English to (tabuild versus to build) during piloting. The second reason was that ta can be used as a short form of thanks in the Manchester region, where many children were tested. Even though this change to va[verb] resulted in slightly different constructions in German and English, it was done as a precautionary measure to ensure that the starting point for English children was comparable to that of German ones, i.e., that the novel verb prefix cannot stand alone, does not mean anything on its own and that there is no phonologically similar form that does. In order to keep the form as similar as possible only the first phoneme was replaced. There is little reason to assume that the two different phonemes (the plosive / t/ and the fricative / v/ ) should cause differences in the difficulty of the novel prefix construction. Vawas also placed before verbal bases and expressed the meaning ‘pretend to perform an action’. This meaning is not usually expressed by a single morpheme in English. Instead the pretend to [verb] construction is normally used. One reason why this meaning was selected was that it can be expressed by a prefix construction in other languages. In Arabic, for instance, adding <?page no="138"?> 125 the prefix tato verb stems of a certain group of verbs (pattern I stative verbs) serves to express ‘pretence’, e.g., hamiq ‘be stupid’ > ‘pretend to be stupid’ (Cuvalay-Haak 1959: 104; Wright 1859: 36). The second construction was based on a pattern that does not occur in either German or English: verb-initial reduplication. Reduplication is used in the majority of the world’s languages, but it is markedly absent from Western European languages (Rubino 2005, 2008, but see Wang 2005 on the role of reduplication and repetition in English). It is defined as the “repetition of phonological material within a word for semantic or grammatical purposes” and can thus be derivational or inflectional (Rubino 2008: 114). It takes many forms, from the repetition of an entire word, word stem, or root to simple consonant gemination, vowel lengthening or incomplete copy of the base (Rubino 2008). Reduplication expresses a myriad of meanings including number, argument distribution, tense, aspect, pretence, transitivity, intensity, reciprocity, and lack of control/ carelessness on verbs, as well as number, case, distributivity, indefiniteness, size, reciprocity and associated qualities on nouns (Rubino 2008). While the meaning of ‘repetitiveness’ may be the most transparent and intuitive to speakers of Western European languages, possibly due to its iconicity, it is by no means the only meaning expressed in this way. In the present study, the novel verb-initial reduplication was formed by repeating the beginning of the verb up to and including the first vowel, e.g., tritrinken in German or dridrink in English. Any consonants intervening between the first consonant and the first vowel were also repeated, resulting in the form C(CC)V-. The reduplicated material was pronounced as similarly to how the respective part of the verb is usually pronounced as possible. The meaning of the novel reduplication construction was also ‘pretend to perform an action’. As pointed out initially, reduplication is a pattern that German-speaking and Englishspeaking children are not familiar with. There are few exceptions, such as Wauwau ‘dog’ or Töfftöff ‘motorbike’ in German and choo-choo or bow-wow in English. These forms occur predominantly in child talk, they characteristically comprise onomatopoeic elements and do not constitute a productive pattern. They tend to refer to objects (rather than actions) and usually the entire lexeme is repeated, sometimes with a change in the initial sound, which makes them considerably different from the forms used in the experiment. The second reason why the meaning ‘pretence’ was selected was that there is a language, Ilocano (spoken on the Philippines), that uses so-called automatic reduplication to express ‘pretence’. Automatic reduplication involves a reduplication combined with an additional obligatory affix. Both are necessary to express the meaning, i.e., they are monomorphemic. In order to express pretence the form aginCVis used, e.g., singpet ‘behave’ > <?page no="139"?> 126 agin-si-singpet ‘pretend to behave’ (Rubino 2008: 114). In the present study a simpler form of reduplication, without the additional affix, was preferred. The third reason why ‘pretence’ was chosen as the meaning of the novel constructions was that children between 3 and 8 were expected to be familiar with this meaning through pretend play. Individual pretend play has been reported to emerge at around 12 to 13 months of age (Fein 1981). From 2 years on, children also engage in pretend play with others (Rakoczy, Tomasello and Striano 2004). At this age they show a robust understanding of pretence, e.g., they are not confused about the non-literal treatment of objects in pretence and readily assign different pretend meanings to the same object in subsequent play scenes, e.g., a ball can be an apple and then later an orange (Harris and Kavanaugh 1993; Leslie 1987). Finally, the selected meaning is in line with the meaning proposed for the more abstract prefix pattern [derivational prefix][base] verb ‘encode a contrast in action’/ ‘encode a contrast to the normal, expected action’. In several examples of this pattern in Chapter 5, the deviation from the normal expected action was a deviation in manner. For instance, über 5 [base] and ver 1 [base] express ‘performing an action unsuccessfully or unintentionally’. It is also the manner that contrasts with expectation in the novel prefix construction: contrary to expectation an action is not actually performed in the normal, usual manner but only pretended (ta[verb]/ va[verb] ‘perform a pretend action’). The same meaning ‘encode a contrast in action’/ ‘encode a contrast to the normal, expected action’ is ascribed to the invented pattern level of reduplication; as is the meaning of ‘pretend to perform an action’ to the novel reduplication construction. Both novel constructions are formed with verbal bases only. Generally, the prefix pattern allows other bases as well. However, in the experiment, pretend actions were sometimes contrasted with real action, which made verbal bases necessary. Figure 13 illustrates how patterns and constructions are thought to be related. Figure 13. Constructional levels and their interrelations. The novel constructions are marked in blue. <?page no="140"?> 127 Sixteen verbs were used in each language. German verbs were bauen ‘build’, kehren ‘sweep’, trinken ‘drink’, puzzeln ‘do a jigsaw’, malen ‘draw’, schieben ‘push’, basteln ‘fold’ (e.g., a paperplane), binden ‘tie (a shoelace)’, kleben ‘glue’, nehmen ‘take’, stricken ‘knit’, gießen ‘water (a plant)’, kneten ‘form something with Play-Doh®’, waschen ‘wash’, wischen ‘wipe’ and machen (Foto) ‘take (a picture)’. English verbs were build, sweep, write, drink, cut, push, fold, hammer, read, peel, shake, drill, wash, knit, mix and wipe. Verbs were selected on the basis of the following criteria: It had to be possible to act out distinct real and pretend versions of each verb (for verbs like sleep or think it would be impossible to discern a difference between real and pretend actions). All verbs were transitive. They were either verbs of accomplishment or verbs that take an affected patient. This criterion served to limit the number of different semantic roles. The verbs were mostly two syllables long in the infinitive in German and one in English. As a consequence, most of them were one syllable long in the third person singular, which was used mostly in the experiment. Four verbs in each language were two syllables long in the third person singular (bastelt, bindet, puzzelt, knetet in German and pushes, hammers, washes, mixes in English). Verbs that started with a vowel were excluded, since the novel forms of such verbs would have resulted in unusual German and English phonotactics and might thus have been more difficult to learn, e.g., essen ‘eat’ would have become ta-essen or e-essen in German or va-eat and ea-eat in English. Verbs were selected such that children of all ages could be assumed to have passive and productive knowledge of them. Knowledge was estimated on the basis of informal nursery teacher interviews and specifications in the literature (Dale and Fenson 1996; Masterson and Druks 1998; Masterson, Druks and Gallienne 2008; Stadthagen-Gonzalez and Davis 2006; Szekely et al. 2004). The selection criteria were one reason why verbs in German and English did not overlap entirely. The number of verbs fulfilling the criteria in each language was limited and their suitability for nurseryto schoolage children imposed further restrictions. There were two additional reasons why translation equivalents of German verbs could not be used in English. In some cases English equivalents were monosyllabic verbs ending in a vowel in English, e.g., glue. This is a potential problem because in such cases the entire word would have had to be repeated in the reduplication. This might have made it more difficult for English children to see that the novel construction involves a reduplication of the beginning of the word up to the first vowel rather than a repetition of the entire word. In other cases translation equivalents would have been phrasal verbs in English (e.g., pick up). They were excluded because of the extra particle, which might have posed additional demands in learning. Presenting all pretend actions without any objects might have led children to think the novel meaning was ‘miming’ or ‘invisible acting’ rather <?page no="141"?> 128 than ‘pretending’. This is why two strategies were used to make actions pretend: actions were either demonstrated without any objects, e.g., pretend pushing was done without an object to actually push, or actions were demonstrated using a toy object that cannot be used to perform the real action alongside an object that is needed for the real action, e.g., pretend sweeping was done using a dustpan, a toy block and no dirt. An equal number of pretend-with-objects and pretend-without-objects verbs were assigned to each experimental task. 6.2.3 Procedure Participants were tested individually in a quiet room in their nursery or school. Testing took between 25 and 30 minutes per child. The experiment consisted of four parts: a training film, an act-out training task, a forcedchoice task 1 and a production task. The order of tasks was fixed (see below). Verbs were grouped in four frequency levels. Table 9. Number of verbs in the novel construction by frequency level for each experimental phase. Task Frequency level Level a: new in training film Level b: new in actout task Level c: new in forcedchoice task Level d: new in production task Training film 4 verb types Act-out task 4 verb types 4 verb types Forced-choice task 2 verb types 2 verb types 2 verb types Production task 2 verb types 2 verb types 2 verb types 2 verb types In each task, verbs from a new level were introduced and verbs that had been used in previous tasks reoccurred so as to check effects of verb token frequency on generalization and memory. In the forced-choice and the production task, only half of the verbs of each frequency level were used in the novel construction; the others appeared in ‘normal’ German denoting real situations. This was done to assess not only children’s comprehension and production abilities of the novel construction but also to test their abilities to differentiate it from a normal, familiar construction. Table 9 summa- 1 The use of a training film and a forced-choice comprehension task was based on Casenhiser and Goldberg (2005). <?page no="142"?> 129 rizes the number of verbs in the novel construction by frequency level and task. The frequency level reflected the experimental token frequency of verbs in the novel construction. It did not refer to the frequency of the verb in every-day speech or child-speech corpora. Verbs introduced in the novel construction in the training film were referred to as frequency level a; those used for the first time in the act-out task as frequency level b; items introduced in the forced-choice task as frequency level c; and verbs used for the first time in the production task as frequency level d. Over the course of the experiment, verbs from frequency level a acquired the highest token frequency, followed by verbs from levels b, c and d. Experimental token frequency thus decreased from levels a through to d. Nevertheless, at the point of its first introduction, each verb had a token frequency of 0 in the novel construction. Figure 14 reflects the development of the token frequency of verbs belonging to different frequency levels over the course of the experiment. Verb frequency did not increase in a linear manner. The assignment of verbs to tasks was fixed, because not all actions were suitable for the act-out task, which involved performing several actions with the children in the nurseries and schools. 2 In English verbs were used in the simple present, even though the sentences were usually played along videos of actions in progress, which might have made the progressive the slightly more natural choice. The reason why the simple form was preferred was that the prefix or reduplication already increased verb length. The progressive would have added another syllable resulting in unusually long verbs (e.g., vadrinking or vahammering). Other studies have shown that children readily accept the simple aspect for similar scenarios (Boyd and Goldberg 2011a; Casenhiser and Goldberg 2005; Wonnacott et al. 2012). Additionally, native speakers did not consider the use of the simple present as unusual, presumably because it is used in some similar contexts (e.g., sports commentaries). Using the simple present also caused the number of syllables of verbs in the novel construction to be roughly equivalent in German and English. The added advantage was that the English construction resembled its German equivalent structurally more closely this way (e.g., no auxiliary was used as would have been with the progressive). 2 The fact that the assignment was fixed is not ideal, because other item characteristics, e.g., age of acquisition, difficulty, frequency in real language, could not be controlled for by counterbalancing the assignment between participants. However, apart from the restriction that verbs in the act-out task had to be suitable for act-out with children, the assignment of verbs to task was randomized to minimize potential problems. <?page no="143"?> 130 All German and English sentences were recorded by female native speakers of the respective language with standard accents. Both speakers used child-friendly but not extremely exaggerated intonation. Figure 14. Experimental token frequencies of verbs used in the novel construction in the different phases of the experiment. Note: Read figure as follows. In the training film, previously unheard verbs from level a were introduced in the novel construction. In the act-out task, verbs from level a had already acquired a token frequency of 4 due to previous exposure in the film. These verbs reoccurred in the act-out task (to check children’s memory) and new verbs (level b) with a frequency of 0 at this point in time were introduced. Old and new verbs were used 5 times each in this phase. Before the forced-choice task the training film was replayed, so that token frequency for verbs from the training film increased by 4. In the forced-choice task children were exposed to verbs from levels a and b, which by then had already acquired token frequencies of 13 and 5 respectively, as well as to new verbs (level c). Children heard verbs from each frequency level twice in the novel construction in this task. In the production task, children were exposed to verbs from levels a, b and c, which they had by then already heard 15, 7 and 2 times, as well as to new verbs (level d). Training film All film clips were prepared using MAGIX video software and were presented using Microsoft Powerpoint. Children were seated approximately 35 to 40cm in front of a 14’’ screen. The training film contained eight clips. An actor and an actress performed the same four pretend actions each. The <?page no="144"?> 131 clips were paired with audio descriptions of the scenes in the novel construction. Participants heard the novel construction twice in each clip, e.g., Der Mann tabaut/ baubaut. Der 3 tabaut/ baubaut was in German or the equivalent The man vabuilds/ buibuilds. He vabuilds/ buibuilds something in English. The order of clips was pseudo-randomized such that the same action was never presented twice in a row and the same actor or actress never occured more than twice in a row. Act-out task In the act-out task participant and experimenter moved away from the computer and acted out eight pretend and the corresponding real actions. All materials necessary to perform real and pretend versions of all verbs were positioned within easy reach of the child. The verbs were blocked by frequency level (level b-level a-level b-level a), so that generalization was tested first with new items, followed by old items to test memory and allow further learning, followed again by a block of new and then old items. The following script was used. It is presented here with the English prefix construction. The German equivalents are given in brackets. The same wording was used with the reduplication construction, only the novel verb form was different then. To start with, the experimenter told the child that they wanted to vaverb now. She then asked the child whether they knew how to vaverb, e.g., how to vabuild (Do you know how to vabuild? or Weißt du, wie man tabaut? ) and the child was expected to perform the corresponding pretend action, e.g., pretend to build. If the child did not know how to do it, the experimenter showed the child and asked them to give it a try too. The experimenter then asked the child if they could say the new form as well (Can you also say ‘vabuild’? or Kannst du auch ‘tabauen’ sagen? ). If the child did not say the word, the experimenter repeated it, so that all children heard an equal number of tokens. Child and experimenter then went on to perform the corresponding real action, e.g., actually build. To this end, the experimenter asked the child if they knew how to verb, e.g., how to build (And how do you build? How is that done? or Und wie baut man? Wie macht man das? ). In instances where children repeated the pretend action or did not perform the real action properly, the experimenter showed them the real action. Finally, the experimenter briefly repeated real and pretend actions saying, So, this is building. And this is vabuilding. (Also, das ist Bauen. Und das ist Tabauen.). After the act-out task children were shown the training film for a second time to increase the input of the novel construction. 3 German der is normally used as a definite article. It can, however, also stand on its own similar to a personal pronoun but with a slightly stronger demonstrative force, which is why it was used here. <?page no="145"?> 132 Forced-choice comprehension task In the forced-choice comprehension task children were presented with two films simultaneously. In one of the films the actor or actress performed a pretend action and in the other one the same actor or actress performed the corresponding real action. While the films were shown, the children were asked to point to one of them, e.g., pretend item: Show me the man vabuilds/ buibuilds. He vabuilds/ buibuilds something. (Zeig mir, der Mann tabaut/ baubaut. Der tabaut/ baubaut was.); real item: Show me the man builds. He builds something. (Zeig mir, der Mann baut. Der baut was). If pointing was ambiguous, children were asked which one was better. Twelve verbs were used, half of them occurred in the novel construction denoting the pretend action, half occurred in the normal form denoting the real action. The side on which the target appeared, the gender of the actor and the order were counterbalanced. For half the children the task began with a pretend action, for half it began with a real action. Items were blocked by frequency level (levels c-a-b-c-a-b). One item in each block was used in the novel construction, one in the real construction. Practice trials in which children had to point to one of two films showing known real actions that were not used in the rest of the study were run at the beginning of the experiment so as to familiarize the children with the pointing procedure and to ensure answers would be mostly unambiguous. Production task The production task involved children watching clips in which the actor or actress performed pretend or real actions. Children were asked to tell what the people in the films did. Sixteen verbs were targeted. Items were blocked by frequency level (levels d-a-b-c-d-a-b-c). Each block contained one pretend and one real action. For half the children the task began with a pretend action, for half with a real action. The same actor or actress did not appear more than twice in a row. Scoring In the act-out task, children were awarded a point when they were able to perform both the pretend and the real action of a verb correctly and no point when they failed to do so (e.g., performed the real or pretend action for both, or did not know how to perform one or both actions). In the forced-choice task, children received a point for pointing to the correct film and none for pointing to the incorrect one. In the production task children were awarded a point for producing the target form and none for failing to do so. <?page no="146"?> 133 Order of tasks The order of tasks was fixed. This resulted from the necessity to show the training film first. The act-out task followed so as to give children a break from watching films before the forced-choice task. The production task came last because it was assumed that children would be reluctant to produce the novel construction earlier in the experiment after even less exposure. The variety of tasks served to investigate act-out abilities, as well as comprehension and production skills after minimal training and to assess reliably whether the construction was indeed learned. Due to the fixed task order, the different response formats of the tasks, i.e., act-out, pointing and production, and the fact that children received corrective feedback in the act-out task but not in the other two tasks, there were no assumptions as to how performance in the different tasks would differ and the data was not analyzed in this respect. 6.3 Results 6.3.1 Statistical techniques used to analyze the data In a first step it was tested at which age a significant learning effect occurred, since testing the effects of independent factors (age, frequency level, construction, pattern, language) does not directly inform about the significance of the learning feat. Following the example of previous novel construction learning studies (Casenhiser and Goldberg 2005; Wonnacott et al. 2012), one-sample t-tests were computed to this end. Since it is controversial whether only generalization to new cases constitutes learning or memory of previously experienced exemplars is to be considered part of the learning process, t-tests were performed for each level of frequency. Responses for all tasks were binary, i.e., correct versus incorrect, which means that the mean referred to the mean proportion (of correct responses). Binary outcomes are binomially distributed. As a consequence the variance of the sample proportions are highest for the probability p = 0.5 and decrease symmetrically towards 0 and 1 (Jaeger 2008). When the data is arcsine-square-root transformed, these changes in probability are captured more appropriately because they are more similar to the more adequate logit-transformed data (Jaeger 2008). For this reason, the data were arcsine-square-root transformed before the t-tests. As soon as several conditions are compared within a test (e.g., twosample t-test; analysis of variance ANOVA), an additional problem arises with respect to the variances of different conditions. They are assumed to be homogeneous, because this is a precondition for the application of the respective statistical test. Variances in the different conditions are, however, only homogenous if they happen to be equally distant from p = 0.5. <?page no="147"?> 134 Still, this is usually not a given in practice (Jaeger 2008). Fortunately, there are modern techniques that avert this and further problems of ANOVAs. One alternative technique - mixed-effects logistic regressions (Jaeger 2008) - is now described. It was used in a second step in the analysis of Experiment 1 in order to supplement the more traditional t-tests. It served the more comprehensive analysis of the data, in particular the exploration of effects of independent factors. In logistic regressions (Jaeger 2008; Manning 2003: Ch. 5.7) the logarithm of odds of success log(odds(Y = success)) 4 is used instead of proportions or means of correct responses (Jaeger 2008; Krajewski 2011). The odds capture the characteristics of binomially distributed data accurately. They are greater than 1 when success is more likely than failure and lower than 1 when failure is more likely. The logit transformation results in positive values when more than half of the responses are correct and in negative values when more than half of the responses are incorrect. Regression analyses 5 aim to determine the formula that best predicts the dependent variable (Y) based on the independent variables (X, Z; Jaccard 2001: 3; Krajewski 2011; see [1], taken from Krajewski 2011): [1] Y = a + b*X + c*Z In other words, the regression serves to find the intercept a and coefficients b and c. Testing the effect of an independent variable (or factor) amounts to testing whether its coefficient is significantly different from 0. Using mixed-effects regression models brings with it several advantages. Mixed-effects models do not make the mostly inadequate assumption of homogeneity of variances at all (Jaeger 2008). Moreover, apart from fixed effects both participants and items can be included as random factors in a single model (Baayen, Davidson and Bates 2008; Jaeger 2008; Raaijmakers 2003). In more classical procedures (e.g., ANOVAs) only one random factor can be included at a time. Sometimes only participants are included causing the problem that the set of stimuli theoretically has to comprise the entire population of items. This is usually impossible, especially if language stimuli are used, since they are typically taken from an infinite set of possible utterances. In order to ameliorate this issue, two separate analyses are frequently calculated (i.e., one including participants, one including items), so that generalizations over the sample tested and the items used can be formed. However, these separate analyses neglect interactions. Attempts to remedy this shortcoming involve the calculation of combined measures (e.g., the quasi F-ratio or the min-F’). Nevertheless, the original problem of having two separate analyses is only ameliorated but not resolved. Mixed-effects models are thus a useful alternative allowing a 4 odds(Y = success) = P(Y = success) / P(Y = failure) (Krajewski 2011). 5 Note that ANOVAs can also be viewed as regressions (e.g., Jaeger 2008). <?page no="148"?> 135 single model where both random factors as well as potential interactions between them are simultaneously controlled for. In the analyses of the present experiment, series of mixed-effects logistic regression models were fitted to the data of each task, starting with the most comprehensive model (i.e., including all possible interactions between factors). This backward approach allowed testing the significance of each term hierarchically. When coefficients were returned as nonsignificant, subsequently a model without each such interaction or factor was fitted in order to assess whether or not it contributed significantly to the goodness-of-fit. If the model fit was not significantly worse than the 2 ), the interaction or factor was removed from the model. It was taken into account that factors cannot be removed on lower levels if they are significant in a higher-level term (e.g., a non-significant effect of factor cannot be removed if it is involved in a significant interaction). Categorical independent variables cannot be entered into a regression formula unless they are coded as numerical variables. In this study the most widespread coding convention was followed, i.e., 'dummy coding', in which one level of a factor is assigned a reference status and model coefficients represent differences between corresponding levels and this reference level. This is why (in contrast to ANOVAs) several pairwise comparisons were automatically obtained when fitting a model. If necessary (i.e., if a factor had more than two levels) the same model was re-fitted once or twice with re-ordered levels of the respective factor so as to test additional comparisons. Using both more traditional and modern techniques in the analysis aimed at doing justice to the data. Results are presented in the following for the learning of the novel prefix construction and the novel reduplication construction in German and English; comparisons between the two languages are reported as well. 6.3.2 Learning of a novel verb prefix construction in German and English Results from German-speaking and English-speaking children’s learning of the novel prefix construction as well as a comparison of the two groups are presented here. Results are ordered chronologically by task starting with the act-out, followed by the forced-choice comprehension and the production task. <?page no="149"?> 136 Act-out task German results Figure 15 gives an overview over the proportions of correct responses of German-speaking children learning the novel prefix construction by age in years with respect to frequency level (level a = familiar from training film, level b = new in act-out task) in the act-out task. The figure suggests that performance increased with age and was higher for higher experimental token frequency (frequency level a > b). Figure 15. Act-out task (German, prefix). Proportions of correct responses by age group. Note: The proportions were calculated for illustration. T-tests were based on these proportions, but regression analyses were performed on individual responses (correct/ incorrect) as the dependent variable, with age in months as a continuous predictor. If illustrations had been based on the regression models, it would have been necessary to select random ages for display because of the continuity of the factor age, which is why the representation by age in years was preferred. This is true for all such bar charts and the respective analyses. The error bars in this and all successive bar charts display the standard error of the mean. If no error bars are displayed, there was no variation because participants were either at ceiling or at floor. One-sample t-tests were calculated to assess whether children’s performance (= learning) was significantly different from non-learning. Cor- 3 yrs 4 yrs 5 yrs 6 yrs 8 yrs Age <?page no="150"?> 137 rect responses in this task required children to act out pretend and real actions of a verb adequately on a linguistic cue (e.g., Weißt Du, wie man tabaut? / Und wie baut man? or Do you know how to vabuild? / And how do you build? ). In acting out correct pretend and real actions, children revealed that they were able to select the respective actions from an infinitely large set of possible actions. Non-learning on the other hand was expected to be equivalent to incorrect performance on all items. T-tests thus tested the proportion of correct responses against 0. On previously experienced items (level a) German-speaking children from age 3 performed significantly differently from 0 (t(15) = 2.92, p < .011); on new items requiring generalization (level b) children from 4 years onwards did so (t(15) = 2.83, p < .013). Tables including values for each age group and frequency level for this task as well as all subsequently presented ones (by language and condition) are given in Appendix II.1 (Tables g-r). In the following step mixed effects logistic regression models using Laplace approximation were fitted to the data (Baayen et al. 2008). Correct versus incorrect response served as the dependent variable, children’s age (in month) and frequency (level a/ b) were included as fixed factors. Children and verbs served as random factors. The Age x Frequency interaction was removed, since it did not improve the model fit significantly. The results confirmed the intuitions based on inspecting Figure 15. Performance increased significantly with age (B = 0.13, SE = 0.01, z = 9.52, p < .001) and was higher for verbs children had been exposed to before (level a > b: B = -4.34, SE = 1.38, z = -3.15, p < .002). English results Figure 16 gives an overview over the proportions of correct responses of English-speaking children learning the novel prefix construction by age in years with respect to frequency level in the act-out task. The figure suggests that performance increased with age and was higher for familiar than for new verb types in the novel construction (level a > b). One-sample t-tests testing the proportion of children’s correct responses against 0 were calculated for both frequency levels and all three age groups. They revealed that English-speaking children from age 4 on performed significantly differently from 0 on previously heard items (level a: t(15) = 4.50, p < .001) as well as on new items requiring generalization (level b: t(15) = 5.58, p < .001). Subsequently, mixed-effects logistic regression analyses with the same variables as in the German analyses were performed on the English act-out data. The Age x Frequency interaction was removed for English-speaking children as well, since it did not improve the model fit significantly. The results showed a significant increase with age (B = 0.09, SE = 0.02, z = 3.87, <?page no="151"?> 138 p < .001). The effect of token frequency approached significance (level a > b: B = -1.91, SE = 1.06, z = -1.81, p < .071). 6 Figure 16. Act-out task (English, prefix). Proportions of correct responses by age group. Comparison T-tests had revealed that both German-speaking and English-speaking children from 4 years onwards performed significantly differently from 0 on new items requiring the generalization of the novel construction. On items that were familiar from the training film, children from all age groups speaking either language did so. In order to explore whether there were any differences between languages, mixed-effects logistic regression models with response (correct/ incorrect) as the dependent variable, children’s age (in month), frequency (level a/ b) and language (German/ English) as fixed factors were fitted to the data. Children and verbs were included as random factors. Language did not significantly improve the model fit. Additional effects have already been discussed for each language separately and thus are not reported again. 6 It is possible that the lower number of children tested was responsible for the only marginally significant effect compared to German. 4 yrs 5 yrs 6 yrs Age <?page no="152"?> 139 Forced-choice comprehension task German results Figure 17 gives an overview over the proportions of correct responses by age in years with respect to frequency level (level a = familiar from training film, level b = familiar from act-out task, level c = new in forced-choice task) for German-speaking children learning the novel prefix construction in the forced-choice task. Only pretend items are displayed, because only these required the comprehension of the novel construction. 7 The figure suggests that performance increased with age and tended to be higher for higher experimental token frequencies. Figure 17. Forced-choice task (German, prefix). Proportions of correct responses on pretend items by age group. One-sample t-tests were calculated to test whether children’s performance on pretend items (= learning) was significantly different from nonlearning. Previous studies with similar paradigms (e.g., Casenhiser and Goldberg 2005) compared performance to chance levels, since the forced- 7 Children at all ages did well on real items, which was expected since they were familiar with them. It is nevertheless an important finding because it attests that children did not invariably point to the perhaps more unusual new action without being able to discriminate it from other actions. Such behaviour would have resulted in high performance for pretend but not real items. The high performance on real items was found in the forced-choice task for all groups throughout the experiment. Germanspeaking children’s performance on real items in the forced-choice task is displayed in Appendix II.2, Figure a, for prefix and reduplication learning. 3 yrs 4 yrs 5 yrs 6 yrs 8 yrs Age <?page no="153"?> 140 choice task gave children two response options (chance = 0.5). However, comparison to chance level might be too conservative. A study of the effects of input distributions on novel construction learning (Experiment 2, Chapter 7) used a very similar forced-choice task and included a control group that did not receive any training on the novel construction. Performance on pretend items did not quite reach a proportion of 0.1 correct responses in this control group. Instead children in this group pointed to real items 90% of the time, regardless of the linguistic form they heard. For this reason, children’s performance in Experiment 1 was also tested against the more lax 0.1, even though it has to be borne in mind that this baseline stems from a different study and children’s variance was neglected (a more detailed rationale for the use of this baseline is given in the discussion in 6.4.2). 8 On items familiar from the training film (level a) children from 4 years on were significantly above chance (t 0.5 (15) = 3.91, p 0.5 < .002) and 0.1 (t 0.1 (15) = 7.27, p 0.1 < .001). On items familiar from the act-out task (level b), 4year-olds and older children were significantly above chance (t 0.5 (15) = 2.76, p 0.5 < .015) and 3-year-olds and older children were significantly above 0.1 (t 0.1 (15) = 2.15, p 0.1 < .049). On new items (level c) children from age 5 were significantly better than chance (t 0.5 (15) = 3.23, p 0.5 < .006) and children from 4 years old were significantly better than 0.1 (t 0.1 (15) = 3.87, p 0.1 < .002). In order to analyze further effects, a series of mixed effects logistic regression models with age (in months), frequency (levels a/ b/ c) and construction (pretend/ real) as fixed factors, response (correct/ incorrect) as dependent variable, and children and verbs as random effects was fitted to the data. Interactions that did not improve the fit of the model significantly were excluded one at a time, starting with higher-order interactions. The final model with the best fit was re-fitted once with re-ordered levels of frequency in order to compare frequency levels b and c with each other. There was a significant increase with age (B = 0.10, SE = 0.01, z = 7.38, p < .001). Children did significantly better on real than on pretend items (B = 1.01, SE = 0.35, z = 2.90, p < .004), which was predicted since real items required comprehension of ‘normal’ sentences whereas pretend items required comprehension of the novel construction. This difference decreased with age as evidenced by the Age x Construction interaction (B = -0.05, SE = 0.02, z = -3.12, p < .002), most likely because children performed highly on real items at all ages and did increasingly well on pretend items with age. There was also a significant effect of frequency between levels b and c (B = -1.34, SE = 0.42, z = -3.22, p < .002), suggesting that correct responses 8 Due to differences between the two studies, e.g., different sample sizes, groups were not compared using independent sample t-test. The comparison to the baseline only serves as an additional indicator of children’s abilities, but has to be treated carefully. <?page no="154"?> 141 were less likely on new items than on items that were familiar from either the training film or the act-out task. English results Figure 18 gives an overview over the proportions of correct responses by age in years with respect to frequency level for English-speaking children learning the novel prefix construction in the forced-choice task. Again only pretend items are displayed (for real items, see Appendix II.2, Figure b). The figure suggests that performance increased with age and was higher for higher experimental token frequency. Figure 18. Forced-choice task (English, prefix). Proportions of correct responses on pretend items by age group. First, one-sample t-tests were calculated to test whether children’s performance on pretend items (= learning) was significantly different from non-learning. Parallel to the German data, English results were tested against 0.5 (chance) as well as 0.1 (control group in Experiment 2, Chapter 7). On items familiar from the training film (level a), 6-year-old children performed significantly above chance (t 0.5 (15) = 3.48, p 0.5 < .004) and children from 5 years on were significantly better than 0.1 (t 0.1 (15) = 4.03, p 0.1 < .002). The same was true for items from the act-out task (level b): Six-yearolds were significantly better than chance (t 0.5 (15) = 2.76, p 0.5 < .015) and children from 5 years on performed significantly above 0.1 (t 0.1 (15) = 3.82, p 0.1 < .002). On new items (level c) children aged 5 and older were signifi- 4 yrs 5 yrs 6 yrs Age <?page no="155"?> 142 cantly better than 0.1 (t 0.1 (15) = 3.06, p 0.1 < .008). Compared to chance even 6-year-olds were only marginally better (t 0.5 (15) = 1.78, p 0.5 < .100). In the following step, a series of mixed effects logistic regression models with the same variables that had been used for the analysis of the German data was fitted to the English forced-choice comprehension data. Frequency did not significantly improve the model fit and was subsequently removed. There were significant effects of age (B = 0.09, SE = 0.02, z = 3.87, p < .001) and construction (B = 1.83, SE = 0.24, z = 7.78, p < .001) as well as a significant interaction between the two (B = -0.10, SE = 0.02, z = -4.23, p < .001). Performance of English-speaking children increased with age. It was better on real than on pretend items, but this difference decreased as children grew older. Comparison T-tests had revealed that children did better if items were already familiar than when items were new and required generalization. They further suggested that there might be a difference between German-speaking and English-speaking children, since German-speaking children performed significantly above chance levels on items they had never heard before in the novel construction, whereas English-speaking children did not. Mixed-effects logistic regression models explored these potential language effects further. Response (correct/ incorrect) served as the dependent variable, children’s age (in month), frequency (level a/ b/ c), construction (pretend/ real) and language (German/ English) as fixed factors. Children and verbs were included as random factors. Effects not involving language are not reported, since they have already been discussed. There was a significant effect of language, suggesting that German-speaking children did better than English-speaking children on this task (B = 1.96, SE = 0.54, z = 3.60, p < .001). The difference between frequency levels b and c was significant only for German children (B = -1.25, SE = 0.63, z = -1.98, p < .048), which corresponds to the individual analyses by language group (see above). Production task German results Figure 19 gives an overview over the proportions of correct responses by age in years with respect to frequency (level a = familiar from training film, level b = familiar from act-out task, level c = familiar from forced-choice task, level d = new in production task) for German-speaking children learning the novel prefix construction in the production task. Only pretend <?page no="156"?> 143 items are displayed, because real items 9 did not require the use of the novel construction. The figure suggests that performance increased with age and higher experimental token frequencies (levels a > b > c > d). Figure 19. Production task (German, prefix). Proportions of correct responses on pretend items by age group. One-sample t-tests were calculated to test whether children’s performance on pretend items (= learning) was significantly different from nonlearning. Since correct responses required children to produce the novel construction based on a non-linguistic cue (i.e., a video clip showing the respective action) from an infinite set of possible responses, it was assumed that any correct productions would imply a certain degree of learning. Ttests thus tested the proportion of correct responses against 0. Children from 3 years of age performed significantly above 0 on items familiar from the training film (level a: t(15) = 3.09, p < .008) and items introduced in the act-out task (level b: t(15) = 2.24, p < .041). Children aged 4 years or older were significantly above 0 on items familiar from the forced-choice task 9 Children at all ages did well on real items, which was expected since they were familiar with them. It is nevertheless an important finding because it attests that children did not invariably produce the ‘funny new form’ for any action. Such behaviour would have resulted in high performance for pretend but not real items. The high performance on real items in the production task was found for all groups throughout the experiment. German-speaking children’s performance on real items in the production choice task is displayed in Appendix II.2, Figure c, for prefix and reduplication learning. 3 yrs 4 yrs 5 yrs 6 yrs 8 yrs Age <?page no="157"?> 144 (level c: t(15) = 3.47, p < .004). And children at age 5 or older did significantly better than 0 on new items (level d: t(15) = 3.50, p < .004). Additionally, a series of mixed effects logistic regression models with age (in months), frequency (levels a/ b/ c/ d) and construction (pretend/ real) as fixed factors, response (correct/ incorrect) as dependent variable, and children and verbs as random effects was fitted to the data. Interactions that did not improve the fit of the model significantly were excluded one at a time, starting with higher-order interactions. The final model with the best fit was re-fitted twice with re-ordered levels of frequency in order to calculate all pairwise comparisons. There were significant effects of age (B = 0.09, SE = 0.01, z = 9.22, p < .001) and construction (B = 1.55, SE = 0.43, z = 3.59, p < .001) as well as a significant interaction between the two (B = -0.07, SE = 0.01, z = -7.40, p < .001). This result suggest that the difference between real items, on which children of all ages performed well, and pretend items decreased with age, as children did increasingly well on pretend items as well. Frequency was significant for all comparisons except for that between levels a and b (levels b vs c: B = -1.09, SE = 0.46, z = -2.34, p < .020; levels c vs d: B = -1.28, SE = 0.47, z = -2.74, p < .007), which were the two groups with the most frequent verbs (from training film and act-out task). The significant interaction between frequency and construction (levels c vs d: B = 1.74, SE = 0.49, z = 3.53, p < .001) implies that the discrepancy between frequency levels referred to pretend items. English results Figure 20 gives an overview over the proportions of correct responses by age in years with respect to frequency for English-speaking children learning the novel prefix construction in the production task. Only pretend items are displayed, because real items did not require the use of the novel construction (for real items, see Appendix II.2, Figure d). The figure suggests that performance increased with age. The potential effect of frequency seems not entirely clear. One-sample t-tests were calculated to test whether children’s performance on pretend items (= learning) was significantly different from nonlearning. Parallel to the German analysis, t-tests compared the proportion of correct responses against 0. On items familiar from the training film (level a), children from age 5 performed significantly better than 0 (t(15) = 3.00, p < .009), for 4-year-olds this effect was only marginally significant (t(15) = 1.86, p < .083). Children from 4 years on were significantly above 0 on items introduced in the act-out task (level b: t(15) = 2.45, p < .028) and the forced-choice task (level c: t(15) = 2.15, p < .049). On new items requiring generalization, only 6-year-olds performed significantly above 0 (level d: t(15) = 3.15, p < .007); 5-year-olds did marginally better than 0 (t(15) = 2.08, p < .056). <?page no="158"?> 145 Figure 20. Production task (English, prefix). Proportions of correct responses on pretend items by age group. A series of mixed effects logistic regression models with the same variables that had been used in the German analysis was then fitted to the English production data and served to further explore effects of age, frequency and construction. The regression model revealed that there were significant effects of age (B = 0.10, SE = 0.03, z = 3.51, p < .001) and construction (B = 4.73, SE = 0.63, z = 7.54, p < .001), that is, performance of English-speaking children increased with age and was better on real than on pretend items. Additionally, the difference between performance at frequency levels c and d was significant (B = -1.16, SE = 0.61, z = -1.89, p < .011), revealing that production of familiar items was easier than generalization to new items. For these frequency levels there was also a marginally significant interaction with construction, suggesting that the effect referred to pretend items (levels c vs d: B = 1.47, SE = 0.85, z = 1.74, p < .083). Comparison T-tests had revealed that there might be a difference between languages. German-speaking children performed significantly above 0 on new items requiring generalization at 5-years of age, whereas only 6-year-old Englishspeaking children did so. Mixed-effects logistic regression models with response (correct/ incorrect) as the dependent variable, children’s age (in month), frequency (level a/ b/ c/ d), construction (pretend/ real) and language (German/ English) 4 yrs 5 yrs 6 yrs Age <?page no="159"?> 146 as fixed factors were fitted to the data in order to explore language effects. Children and verbs were included as random factors. Only effects related to language are reported. There was a significant advantage of German over English (B = 1.64, SE = 0.40, z = 4.07, p < .001). Token frequency The repeatedly-shown effect of the frequency level suggests that performance increased with the token frequency of a verb type in the novel construction. Table 10 gives an exact breakdown of the frequency levels of familiar verbs German-speaking and English-speaking children produced, based on the total number of familiar verbs they produced. If children remembered only a single verb in the novel construction, it was most likely a high-frequency one. If they remembered two verbs, the majority tended to be of higher frequency. As children remembered more and more verbs, the number of verbs with lower token frequencies increased, but items with highest token frequency were usually remembered best, unless children remembered all verbs, which necessarily entailed equal performance on familiar verbs of all token frequencies. Table 10. Production of previously encountered verbs in the novel construction as a function of token frequency (prefix, all children). Token frequencies of familiar verbs in experiment Number of correct responses on verbs in novel construction Number of children High (level a) Medium (level b) Low (level c) 0 38 1 8 5 3 2 9 7 7 4 3 8 9 11 4 4 11 21 17 6 5 14 28 25 17 6 40 80 80 80 Note: N German = 80; N English = 48. Read table as follows: Eight children produced one verb correctly in the novel construction, for 5 children this verb was a high-frequency verb (i.e., a verb that had been introduced in the very first phase of the experiment and was repeated throughout) and for 3 it was a medium-frequency verb (i.e., a verb that was introduced in the second phase of the experiment and repeated from then onwards in subsequent tasks). <?page no="160"?> 147 Summary Children were able to learn the novel prefix construction to different degrees depending on their age. T-tests revealed that German-speaking and English-speaking children from 4 years onwards showed significant novel construction learning in the act-out task on new items requiring generalization. In the forced-choice task the age of successful generalization of the novel construction depended on the standard of comparison. Germanspeaking children from 5 years onwards and English-speaking children from 6 years on (marginal effect) performed better than chance. Even younger children - 4-year-old German-speaking and English-speaking children - showed learning when compared to the control group of a very similar study (see Chapter 7). In the production task, there is evidence that German-speaking children from 5 years onwards and English-speaking children from 6 years on generalized the novel construction to new cases. To sum up, German-speaking children between 4 and 5 as well as Englishspeaking children between 4 and 6 revealed their ability to generalize the novel construction to new cases. In terms of memory of previously experienced items, which might also be considered part of the learning process since it is assumed to precede generalizations, learning began earlier, e.g., even 3-year-old German-speaking children were able to act out and produce previously experienced items. In all tasks and for German and English children alike, there was a significant increase of performance with age (overall in the act-out task, and on pretend items in the two subsequent tasks). This effect was suggested by bar charts and t-tests and consolidated by the significant age effects in the regression analyses of all tasks. For German-speaking and English-speaking children there was a significant effect of frequency in the act-out task (for English the effect approached significance). In the forced-choice task, there was a significant effect of frequency only for German-speaking children between previously heard verbs and new ones. In the production task, there were significant effects of frequency between levels b and c and levels c and d in the analysis of German-speaking children and between levels c and d for Englishspeaking children. This result suggests that differences in token frequency between the most frequently heard ones (levels a and b) did not make a difference to children anymore at this stage, whereas lower frequencies made learning increasingly difficult. In the act-out task there was no effect for language (German versus English). A reason might be that only the act-out task included feedback. Namely, the experimenter showed the child how to perform the correct action that corresponded to the respective verb in the novel construction, if the child failed to do so. This behaviour might have allowed for within-task learning effects, which might have improved children’s performance re- <?page no="161"?> 148 gardless of language. In the other two tasks, no corrective was provided. In the forced-choice comprehension and the production tasks, significant effects of language were found, such that German-speaking children outperformed their English-speaking peers. This finding was supported by the t-tests that revealed that German-speaking children of younger ages used and generalized the novel construction successfully. Novel construction learning thus progressed faster for children speaking German. 6.3.3 Learning of a novel reduplication construction in German and English The results of reduplication construction learning are arranged in the same fashion as those for prefix construction learning. They are presented by task. Results of the German-speaking group are always presented before the results of the English-speaking group. The comparison between German and English groups completes each section. Act-out task German results Figure 21 gives an overview over the proportions of correct responses by age in years with respect to frequency (level a = familiar from training film, level b = new in act-out task) for German-speaking children learning the novel reduplication construction in the act-out task. The figure suggests that performance increased with age and was higher for higher experimental token frequency (level a > b). Again one-sample t-tests were calculated to assess whether children’s performance (= learning) was significantly different from non-learning. Since correct responses required children to perform the pretend and real action of a verb adequately, non-learning equalled incorrect performance. T-tests thus tested the proportion of correct responses against 0. On familiar items from the training film, children from age 3 were significantly different from 0 (level a: t(15) = 2.42, p < .029). On new items requiring generalization 4-year-olds and older children performed significantly better than 0 (level b: t(15) = 2.18, p < .046). Parallel to the analyses of the prefix learning data, a series of mixed effects logistic regression models using Laplace approximation was fitted to the data to explore the effects of different factors (Baayen et al. 2008). Correct versus incorrect response served as the dependent variable, children’s age (in month) and frequency (levels a/ b) were included as fixed factors. Children and verbs served as random factors. The Age x Frequency interaction was removed, since it did not improve the model fit significantly. The results confirmed the intuitions based on inspecting Figure 21. There were significant increases with age (B = 0.14, SE = 0.01, z = 10.15, p < .001) and <?page no="162"?> 149 for higher experimental token frequency (level a > b: B = -3.88, SE = 1.09, z = -3.56, p < .001). Figure 21. Act-out task (German, reduplication). Proportions of correct responses by age group. English results Figure 22 gives an overview over the proportions of correct responses by age in years with respect to frequency for English-speaking children learning the novel reduplication construction in the act-out task. The figure suggests that performance increased minimally with age and more considerably with experimental token frequency (level a > b). One-sample t-tests assessed whether children’s performance was significantly different from 0. On previously heard items (level a: t(15) = 4.46, p < .001) as well as on new items requiring generalization (level b: t(15) = 2.18, p < .046), children from 4 years of age performed significantly differently from 0. Regression analyses with the same variables used in the German analyses were performed on the English act-out data. The Age x Frequency interaction was removed here as well, since it did not improve the model fit significantly. There was a significant effect of token frequency, i.e., performance increased with higher experimental token frequency (level a > b: B = -3.16, SE = 1.10, z = -2.87, p < .005). The effect of age approached significance (B = 0.07, SE = 0.04, z = 1.74, p < .082). 3 yrs 4 yrs 5 yrs 6 yrs 8 yrs Age <?page no="163"?> 150 Figure 22. Act-out task (English, reduplication). Proportions of correct responses by age group. Comparison T-tests had shown that children from the youngest age tested in the respective language performed significantly differently from 0 on both familiar (level a) and unfamiliar items (level b). Mixed-effects logistic regression models explored whether there was indeed no effect of language in the act-out task. Response (correct/ incorrect) served as the dependent variable, children’s age (in month), frequency (level a/ b) and language (German/ English) were used as fixed factors. Children and verbs were included as random factors. Language did not significantly improve the model fit. Forced-choice comprehension task German results Figure 23 gives an overview over the proportions of correct responses by age in years with respect to frequency (level a = familiar from training film, level b = familiar from act-out task, level c = new in forced-choice task) for German-speaking children learning the novel reduplication construction in the forced-choice task. Only pretend items are displayed, because only these items required the comprehension of the novel construction (for children’s performance on real items, see Appendix II.2, Figure a). Figure 23 4 yrs 5 yrs 6 yrs Age <?page no="164"?> 151 suggests that performance increased with age and decreased for lower experimental token frequencies (level a > b > c). Figure 23. Forced-choice task (German, reduplication). Proportions of correct responses on pretend items. One-sample t-tests were calculated to test whether children’s performance on pretend items (= learning) was significantly different from nonlearning. Parallel to the analysis of the prefix learning data, t-tests tested whether the proportion of correct responses was significantly different from chance (0.5) and a control group in a similar study (0.1, see Chapter 7). Children from age 6 on were significantly above chance on items introduced in the training film (level a: children were at ceiling, i.e., perfect, no t-test possible) and in the act-out task (level b: t(15) = 10.25, p < .001) as well as on new items requiring generalization (level c: t(15) = 2.78, p < .014). In comparison to 0.1, children aged 4 and older were significantly better on items familiar from the training film (level a: t(15) = 3.61, p < .003) and the act-out task (level b: t(15) = 3.55, p < .003), and children aged 5 or above were significantly better than 0.1 on new items (level c: t(15) = 2.74, p < .016). Subsequently, a series of mixed effects logistic regression models with age (in months), frequency (levels a/ b/ c) and construction (pretend/ real) as fixed factors, response (correct/ incorrect) as dependent variable, and children and verbs as random effects was fitted to the data. Interactions that did not improve the fit of the model significantly were excluded one at a time, starting with higher-order interactions. The final model with the best fit was re-fitted once with re-ordered levels of frequency in order to 3 yrs 4 yrs 5 yrs 6 yrs 8 yrs Age <?page no="165"?> 152 compare frequency levels b and c with each other. There were significant effects of age (B = 0.09, SE = 0.01, z = 8.75, p < .001) and construction (B = 2.96, SE = 0.26, z = 11.54, p < .001). Children’s performance increased with age and was better on real items requiring normal language than on pretend items requiring the novel construction. There was also a significant effect of frequency between levels b and c (B = -0.91, SE = 0.42, z = -2.18, p < .030), suggesting that performance was markedly lower on new items than on items familiar from either the training film or the act-out task. English results Figure 24 gives an overview over the proportions of correct responses by age in years with respect to frequency for English-speaking children learning the novel reduplication construction in the forced-choice task. Only pretend items are displayed (for real items, see Appendix II.2, Figure b). Figure 24 suggests that performance increased with age and decreased from frequency level a to levels b and c. Figure 24. Forced-choice task (English, reduplication). Proportions of correct responses on pretend items. T-tests were used to assess whether the proportion of correct responses was significantly different from chance (0.5) and a control group in a similar study (0.1, see Chapter 7). Only 6-year-olds were significantly above 0.1 on items introduced in the training film (level a: t(15) = 2.99, p < .010) and the act-out task (level b: t(15) = 2.44, p < .028) as well as on new items (level 4 yrs 5 yrs 6 yrs Age <?page no="166"?> 153 c: t(15) = 2.64, p < .019). No age group was significantly above chance on any of the items. Subsequently, regression analyses with the same variables as for German children were fitted to the English forced-choice data. Frequency was excluded from the model, since it did not significantly improve model fit. The same was true for the Age x Construction interaction. There were significant effects of age (B = 0.08, SE = 0.02, z = 3.87, p < .001) and construction (B = 4.28, SE = 0.33, z = 13.01, p < .001), such that performance increased with age and was better on real than on pretend items. The intuition that the differences in frequency between levels a and b might affect learning did not emerge as a significant effect. Comparison T-tests had revealed that there might be a difference between languages. On new items requiring the generalization of the novel construction (level c), German-speaking children from 5 years on performed better than 0.1 and children from 6 years of age were additionally significantly above chance. For English-speaking children, only 6-year-olds performed significantly above 0.1 (they were not above chance, though). On items familiar from the training film (level a) and the act-out task (level b), there seemed to be a similar advantage for German. Mixed-effects logistic regression models served to explore a potential effect of language further. Response (correct/ incorrect) served as the dependent variable, children’s age (in month), frequency (level a/ b/ c), construction (pretend/ real) and language (German/ English) were used as fixed factors. Children and verbs were included as random factors. Interactions that did not significantly improve the model fit were removed oneby-one. Only effects relating to language are discussed. There was a significant effect of language (B = 1.84, SE = 0.37, z = 4.98, p < .001), with Germanspeaking children performing significantly better than their Englishspeaking peers. There was also a significant interaction between language and construction (B = -1.44, SE = 0.47, z = -3.06, p < .003), limiting the advantage of German-speaking participants to pretend items. Production task German results Figure 25 gives an overview over the proportions of correct responses by age in years with respect to frequency (level a = familiar from training film, level b = familiar from act-out task, level c = familiar from forced-choice task, level d = new in production task) for German-speaking children learning the novel reduplication construction in the production task. Only pretend items are displayed, because only these items required the use of the novel construction (for real items, see Appendix II.2, Figure c). The fig- <?page no="167"?> 154 ure suggests that performance increased with age and decreased with lower experimental token frequencies (a > b > c > d). Figure 25. Production task (German, reduplication). Proportions of correct responses on pretend items by age group. One-sample t-tests were calculated to test whether children’s performance on pretend items (= learning) was significantly different from 0 (= non-learning). Children aged 4 years or older performed significantly above 0 on familiar items introduced in the training film (level a: t(15) = 3.66, p < .003), the act-out task (level b: t(15) = 2.74, p < .016) and the forcedchoice task (level c: t(15) = 2.24, p < .041). On new items requiring generalization, 6-year-olds and older children were significantly above 0 (level d: t(15) = 4.47, p < .001). Subsequently, a series of mixed effects logistic regression models with age (in months), frequency (levels a/ b/ c/ d) and construction (pretend/ real) as fixed factors, response (correct/ incorrect) as dependent variable, and children and verbs as random effects was fitted to the data. Interactions that did not improve the fit of the model significantly were excluded one at a time, starting with higher-order interactions. The final model with the best fit was re-fitted twice with re-ordered levels of frequency in order to include all pairwise comparisons. There were significant effects of age (B = 0.11, SE = 0.01, z = 10.91, p < .001) and construction (B = 1.91, SE = 0.38, z = 5.01, p < .001) as well as a significant interaction between the two (B = -0.10, SE = 0.01, z = -9.95, p < .001), suggesting that the difference between real and pretend items decreased with age. Frequency was significant for all comparisons except for that between levels c and d (level a > b: B = 3 yrs 4 yrs 5 yrs 6 yrs 8 yrs Age <?page no="168"?> 155 -0.97, SE = 0.47, z = -2.05, p < .041; level b > c: B = -1.11, SE = 0.48, z = -2.30, p < .022). This finding suggests that items children had heard only in the previous task in the novel form and items they had never heard were equally difficult to produce. English results Figure 26 gives an overview over the proportions of correct responses by age in years with respect to frequency for English-speaking children learning the novel reduplication construction in the production task. Only pretend items are displayed (for real items, see Appendix II.2, Table d). The figure suggests that performance increased with age and decreased somewhat for lower experimental token frequency. Figure 26. Production task (English, reduplication). Proportions of correct responses on pretend items by age group. Parallel to the German data analyses, one-sample t-tests were calculated to test whether English children’s performance on pretend items (= learning) was significantly different from 0 (= non-learning). Only 6-year-olds’ performance was significantly above 0 on items of all frequency levels (level a: t(15) = 3.42, p < .004; level b: t(15) = 3.31, p < .005; level c: t(15) = 3.31, p < .005; level d: t(15) = 2.78, p < .014). On the most familiar items, 5year-olds’ performance also approached significance (level a: t(15) = 2.08, p < .056). 4 yrs 5 yrs 6 yrs Age <?page no="169"?> 156 Regression analyses with the same variables as in the German analyses were fitted to the English production data. Frequency did not improve the model fit significantly and was thus removed. There were significant effects of age (B = 0.14, SE = 0.03, z = 4.35, p < .001) and construction (B = 6.90, SE = 0.49, z = 14.20, p < .001), such that performance increased with age and was better on real than on pretend items. Comparison T-tests had shown that performance on new items was similar for both language groups, with only children from 6 years onwards performing significantly different from chance. On familiar items, however, Germanspeaking children from 4 years of age on were significantly different from 0 compared to only 6-year-old English-speaking children. As a consequence, a difference between German-speaking and Englishspeaking children was expected. It was tested using a series of mixedeffects logistic regression models with response (correct/ incorrect) as the dependent variable; children’s age (in month), frequency (levels a/ b/ c/ d), construction (pretend/ real) and language (German/ English) served as fixed factors in the analysis. Children and verbs were included as random factors. Interactions that did not significantly improve the model fit were removed one by one. Only effects relating to language are discussed. The analysis revealed a significant effect of language (B = 2.51, SE = 0.61, z = 4.11, p < .001), with German-speaking children performing significantly better than their English-speaking peers. There was a significant interaction between language and construction (B = -3.21, SE = 0.50, z = -6.40, p < .001) that limited the advantage of German-speaking participants to pretend items. There was further a significant three-way interaction between age, language and construction (B = -0.11, SE = 0.49, z = -2.16, p < .031), suggesting that the advantage of German-speaking children on pretend items decreased with age. Token frequency Performance on previously encountered verbs in the novel construction increased with their token frequency. Table 11 gives an exact breakdown of the frequency levels of previously heard verbs German-speaking and English-speaking children produced in the novel construction in the production task in relation to the total number of such verbs they produced. If children remembered only a single verb in the novel construction, it was always a high-frequency one. If they remembered two verbs, at least one of them tended to be of highest frequency. If children remembered more and more verbs, the number of verbs with lower token frequencies increased, but nevertheless items with the highest token frequency were usually remembered best, unless children remembered all verbs they had <?page no="170"?> 157 previously been exposed to, which necessarily entailed equal performance on verbs of all token frequencies. Table 11. Production of previously encountered verbs in the novel construction as a function of token frequency (reduplication, all children). Token frequencies of familiar verbs in experiment Number of correct responses on verbs in novel construction Number of children High (level a) Medium (level b) Low (level c) 0 59 1 8 8 2 8 9 6 1 3 7 11 7 3 4 9 16 13 7 5 15 30 27 18 6 22 44 44 44 Note: N German = 80; N English = 48. Read table as follows: Eight children produced one verb correctly in the novel construction, for all of them this verb was a high-frequency verb (i.e., a verb that was introduced in the very first phase of the experiment and repeated throughout the experiment). Summary Children were able to learn the novel reduplication construction. The degree of learning depended on children’s age and the task. In the act-out task, German-speaking and English-speaking children from 4 years on showed significant generalizations of the novel construction to new verbs. In the forced-choice task, generalizations were significant for 5-year-old German speakers when compared to the control group (tested in the input distribution study, Chapter 7) and for 6-year-olds when compared to chance. English-speaking children were significantly better than controls only from 6 years on. In the production task, children of both language groups generalized the novel construction productively to new cases from 6 years on. German-speaking and English-speaking children between 4 and 6 were thus able to generalize the novel construction to some extent. If memory of previously experienced items is considered as well, which is justified on the assumption that it necessarily precedes generalizations of a schema, even younger children showed evidence of novel construction learning in several tasks (and more so in German than in English). As evidenced by the t-tests and regression analyses there was a significant effect of age for all three tasks and for German-speaking and English- <?page no="171"?> 158 speaking children alike (in the act-out task this effect was approaching significance for English-speaking children when analyzed separately). Children’s performance with the novel construction thus increased with age. For German-speaking and English-speaking children there was a significant effect of frequency in the act-out task. In the forced-choice task, there was a significant effect of frequency only for German-speaking children between previously heard verbs in the novel construction and unfamiliar ones. In the production task, there were significant effects of frequency between pretend items of levels a and b and of levels b and c for German-speaking children. This result suggests that the differences in token frequency of familiar items were so big as to affect learning, whereas the difference between the lowest level of frequency in previously heard verbs (level c) and entirely new verbs did not affect children’s performance. In the act-out task there was no effect for language (German versus English). It is possible that this finding is a result of the different task format, which included the provision of corrective feedback. Feedback was not given in the other two tasks, which merely tested children’s knowledge of the novel construction. In the forced-choice comprehension and the production tasks significant effects of language on pretend items were found, such that German-speaking children outperformed their English-speaking peers. This result implies that learning might progress faster for children speaking German. In the production task, the advantage decreased slightly with age, suggesting that English-speaking children may be catching up. 6.3.4 Learning of the novel prefix versus the novel reduplication construction After the separate analyses of prefixation and reduplication construction learning, this section deals with the comparison of the two. Given that the prefixation pattern exists in both German and English, whereas the reduplication pattern does not, it is expected that children do better on the novel prefix than on the novel reduplication construction. A series of mixed-effects logistic regression analyses were fitted to the complete dataset for each task. Response (correct/ incorrect) served as the dependent variable, age (in months), frequency (the number of levels varied in each task), construction (pretend/ real in forced-choice and production tasks), language (German/ English) and pattern (prefix/ reduplication) were included as fixed factors. Participants and verbs served as random factors in each analysis. Non-significant interactions and factors were removed one by one if they did not improve the model fit significantly. Only significant terms related to the factor pattern are reported, <?page no="172"?> 159 since the other effects were already discussed elsewhere in this chapter (in 6.3.2 and 6.3.3). There were significant effects of pattern in all three tasks, with children performing better on the novel prefix construction than on the novel reduplication construction (act-out task: B = -1.05, SE = 0.30, z = -3.52, p < .001; forced-choice task: B = -1.46, SE = 0.24, z = 6.05, p < .001; production task: B = -1.69, SE = 0.45, z = -3.74, p < .001). In the forced-choice (Pattern x Construction: B = 2.20, SE = 0.30, z = 7.34, p < .001) and the production tasks (Pattern x Construction: B = 1.94, SE = 0.50, z = 3.87, p < .001) there were interactions with construction, limiting this effect referred to pretend items. In the production task the advantage of prefix construction learning decreased significantly with age on pretend items (Pattern x Construction x Age: B = -0.09, SE = 0.03, z = -2.99, p < .003). The advantage of Germanspeaking children was smaller for pretend items in reduplication than for those in prefixation learning (Pattern x Construction x Language: B = -1.27, SE = 0.60, z = -2.10, p < .036). At levels a vs b there was an interaction between frequency and pattern, suggesting that the difference between these frequency levels was significantly larger for reduplication than for prefix construction learning (B = -0.76, SE = 0.35, z = -2.74, p < .030). Additional token counts (similar to those shown in Tables 10 and 11, but performed separately for each pattern) further revealed that children in the reduplication condition were more reluctant to produce verbs in the novel construction, even if they had heard them before in this form. Twenty-six of 128 children in the reduplication condition never produced previously experienced instances of the novel construction compared to only 13 of 128 in the prefix condition. Conversely, 16 children learning the reduplication produced all previously heard verbs in the novel construction compared to 28 children learning the novel prefix. The majority of children in the reduplication condition produced 0-3 of 6 previously heard verbs, whereas the majority of children learning the prefix produced 4-6 such verbs. With respect to generalizations, children learning the prefix were also bolder. As further counts revealed, 79 children learning the novel prefix generalized in the act-out task, 99 did so in the forced-choice task, and 51 did it in the production task. Generalizations in the reduplication condition amounted to 58 in the act-out task, 74 in the forced-choice task and 38 in the production task. Additional support for the difference between the two novel constructions comes from the effect of token frequency for German-speaking children. The levels that did and did not differ significantly in the production task varied across the two constructions. The only frequency levels that were not significantly different from each other were levels a and b in prefix learning and levels c and d in reduplication learning. In prefix learning, token frequencies of verbs from levels a and b were thus both high enough <?page no="173"?> 160 to ensure a certain level of performance. In reduplication learning, on the other hand, no difference was made between previous exposure to two tokens and no previous exposure. These findings suggest that reduplication learning was more difficult and potentially required more tokens to ensure the level of representational strength necessary for productions. 6.3.5 Type frequency effects The focus of this section is the question of how type frequency affected generalizations. The effect of token frequency suggested that act-out, comprehension and production of the novel construction were the easier the more frequently a verb had been heard in the novel construction, with new verbs being the most difficult. What remained open is the relation between stored exemplars and generalization of the novel construction, in particular whether there is a particular number of types that children base their generalizations on. In order to explore these issues, a breakdown of memorized and generalized verbs in the novel construction per task is given in Table 12. German-speaking and English-speaking children, prefix and reduplication learning are displayed together. 10 The terms memory and storage are used to refer to the degree of representational strength necessary for the retrieval of a form, as evidenced by the performance on the respective type in the act-out, comprehension or production task. In the act-out task, 9 of 256 children who revealed storage of no (5 children) or one verb type (4 children) generalized the novel construction once. More than a single generalization was found as soon as children showed memory of two or more verbs in the novel construction. Children who remembered three or four verb types were more likely to generalize the novel construction to some extent than not to generalize at all. In the forced-choice task, 12 children generalized the novel construction once without showing any memory of stored instances of the novel construction. Apart from this exception, the number of generalizations increased gradually as the number of memorized verb types did. Children showing evidence of one or two stored verb types showed tentative generalizations; memory of three or four examples of the novel construction made generalization increasingly likely. Seeing that the number of generalizations increased slowly from one memorized exemplar onwards, it might be argued that the 12 (of 256) cases of single generalizations without evidence of memorized exemplars can be explained by the task format. 10 Individual tables for prefix learning and reduplication learning are located in Appendix II.3 (Tables s and t). Since general differences between the novel constructions with respect to the minimum number of types required for generalization were neither expected and nor revealed, the table including both constructions is displayed here. <?page no="174"?> 161 Children were asked to point at the one of the two films that matched the audio description they heard. In contrast to task demands in the act-out and the production tasks, correct pointing was not impossible without profound knowledge of the novel construction. Table 12. Act-out, forced-choice and production tasks. Number of children who memorized and generalized 0 to x verbs in the novel construction. Act-out task Number of memorized verbs Number of generalizations 0 1 2 3 4 0 51 5 0 0 0 1 25 4 0 0 0 2 20 11 1 2 1 3 14 17 14 3 0 4 9 14 38 19 8 Forced-choice comprehension task Number of memorized verbs Number of generalizations 0 1 2 0 46 12 0 1 21 1 1 2 6 5 3 3 3 17 8 4 7 44 82 Production task Number of memorized verbs Number of generalizations 0 1 2 0 97 0 0 1 14 2 0 2 14 3 0 3 13 1 1 4 13 7 0 5 9 13 7 6 7 24 31 Note: N=256 (consisting of N German = 160; N English = 96). The number of memorized verbs refers to correct reponses on verbs children had previously heard in the novel construction; the number of generalized verbs refers to correct responses on verbs that children had not previously heard in the novel construction and which thus required the generalization of the construction. <?page no="175"?> 162 In the production task, there were no generalizations before children showed evidence of at least one memorized exemplar in the novel construction. Two children with memory of one verb type generalized the novel construction once. Even children who remembered two, three or four types in the novel construction did not generalize rampantly. Nevertheless, generalizations increased with the number of stored instances and children with recollection of five or six verb types were in fact more likely to generalize to some degree than not to generalize at all. To conclude, a few children started generalizing without revealing evidence of having memorized types of the novel construction and a few children who had stored a single example generalized once. Broader generalizations, i.e., to more than one new type, awaited storage of two or more verb types in the novel construction. It did not seem to be a prerequisite for children to show correct act-out, comprehension or production of a particular number of exemplars in the novel construction before generalizing. However, it did appear that successful storage of a single type cleared the way for a first generalization, while storage of two or more verb types preceded generalizations to more than one new case (except for one child in the forced-choice task). Overall, the data thus revealed that children’s willingness to generalize the novel construction to new cases increased with the number of exemplars they had stored reliably. 6.4 Discussion 6.4.1 Summary The present study revealed a number new of insights into novel construction learning. a) It was demonstrated that German-speaking children between 3 and 8 and English-speaking children between 4 and 6 were able to learn the novel prefix construction to different degrees depending on their age and the task, when given very limited input. Learning was more successful for German-speaking children than for their English-speaking peers. b) It was further revealed that German-speaking children between 3 and 8 and English-speaking children between 4 and 6 were able to learn the novel reduplication construction to different degrees depending on their age and the task, based on very limited input. Again learning was easier for German-speaking children than for their English-speaking peers. c) The tentative comparison of novel prefix and reduplication construction learning showed that learning of the novel prefix construction was easier. d) There was a significant effect of age in all tasks showing that novel construction learning became easier with age, regardless of the language children spoke and the novel construction they learned. e) Learning was shown to be affected by experimental input token frequency. Children <?page no="176"?> 163 generally performed better on verbs in the novel construction the more frequently they had heard them previously in the experiment, with new items requiring the generalization of the novel construction being the most difficult ones. Frequency effects were stronger for German-speaking than for English-speaking children. For the latter group they were somewhat more pronounced in prefix construction learning. f) Type frequency was also revealed to affect learning to some degree. The more types of the novel construction the children had stored, the more rampantly they generalized the novel construction to new cases. This behaviour was regardless of the language and of the construction that was being learnt. It was, however, difficult to determine the exact number of types that children had stored before starting to generalize. 6.4.2 Learning and age Previous research has shown that English-speaking children are able to learn novel argument structure constructions (Boyd and Goldberg 2011a; Casenhiser and Goldberg 2005; Wonnacott et al. 2012). Boyd and Goldberg (2011a) further found an increase in performance between the ages of 5 and 7. The present study was the first to show that German-speaking and English-speaking children were able to learn novel morphological constructions from the input. The degree of learning, as evidenced by children’s performance, increased with age. Depending on the task (task demands), children from 4 to 6 years of age on were able to generalize the novel constructions to new cases. Memory of familiar exemplars of the novel constructions was found at even earlier ages. The overall level of performance was fairly low for the youngest children, but moved towards ceiling for the oldest children. That performance continues to increase and in fact reaches ceiling for previously experienced verbs in the novel construction as well as for verbs requiring generalization was shown in an additional study with adults, who learned the novel prefix or the novel reduplication construction. Results are reported in more detail in Appendix II.4. This progression suggests that children’s grasp on the novel constructions was emergent. These findings are in keeping with Hypotheses I a, I c, II a, II c. I a Children are able to learn a novel prefix construction from the input. II a Children are able to learn a novel reduplication construction from the input. I/ II c Learning increases with age. One reason why younger children showed lower performance in learning studies might be that they required a higher number of tokens per type in order to ensure the degree of representational strength needed for <?page no="177"?> 164 schema abstraction and generalization. The circumstance that younger children need higher token frequencies to reach similar levels of comprehension as older children has also been shown for lexeme learning (Rice et al. 1994). Younger children’s more limited memory might thus be responsible for their inability to learn the novel construction faster and generalize it sooner. Boyd and Goldberg (2011a) on the other hand propose that it is not storage difficulties that are responsible for later generalizations, but rather young children’s unwillingness to generalize on the basis of the same amount of evidence as older learners. The reason they provide is that younger children have less language experience in general. In holding back on generalizations longer than older learners they minimize the risk of incorrect generalizations. A third reason for differences between younger and older children in learning concerns the relational shift proposed in analogy formation (Gentner 1988, 2003; Gentner and Rattermann 1991; Markman and Gentner 1993; Rattermann and Gentner 1998). According to Gentner and her colleagues younger children (around age 3) attend predominantly to object similarity and tend to recognize relational similarity, which is necessary to form analogical comparisons in construction learning, only if it is supported by object similarity. For older children (from around age 6 onwards) this additional support is unnecessary and analogies over exemplars are more easily formed. The present data is in line with the first and the third proposal. The finding that the age effect was not restricted to new exemplars requiring generalization but was also present for familiar items entails that there were differences in the memory representation of tokens between younger and older children. Younger children might therefore require a higher number of tokens to achieve the necessary degree of representational strength (see Rice et al. 1994), but it might also be more difficult for them to successfully align exemplars in the comparison, in particular in reduplication construction learning where the construction is abstract and supportive object similarity is not present. Differences in children’s performance in the different tasks were expected but not subjected to analysis. The reasons were the different task and response formats, i.e., act-out, pointing and production, the circumstance that the different formats called for different reference levels in the ttests, and the fact that children received corrective feedback in the act-out task but in none of the other two tasks. There is one issue that deserves attention with respect to children’s ability to learn the novel constructions. Learning was assessed in different ways, including t-tests. In the act-out and the production tasks, t-test compared performance against a baseline of 0 on the assumption that children without knowledge of the novel construction would be unable to act-out the appropriate pretend action on a linguistic cue or to produce the novel form in response to a video showing a pretend action. In the forced-choice <?page no="178"?> 165 task, children were given a choice between two options, which suggests that chance performance (i.e., random pointing) might indicate participants’ ignorance (= non-learning), whereas above-chance performance implies learning. Several aspects of children’s performance, however, propose that this measure underestimates children’s knowledge of the novel construction. The youngest age groups were not at chance, but below chance on certain groups of items (e.g., verbs requiring generalization), suggesting that non-learning did not cause random pointing but preference for real actions. Children’s performance further increased steadily with age (see age effects in all tasks). Moreover, the other tasks revealed learning in cases, where children did not do significantly better than chance in the forced-choice task. For instance, 6-year-old English-speaking children learning the reduplication were not above chance in the forced-choice task, but there was evidence that they were able to use and generalize the novel construction in the act-out task and even in the production task, which was presumably more difficult than the forced-choice task (since it required production rather than points). Additionally, in Experiment 2 a control group that did not receive any training pointed to the real films in 90% of the trials, regardless of whether the novel or the familiar construction was used in the linguistic cues. Non-learning was thus mirrored in a baseline of correct performance of 10% (rather than 50%). The presumed reason is that children were experienced language users and thus brought certain language knowledge with them. Rather than learning the novel construction by mutual exclusivity (i.e., concluding that the unusual words must refer to pretend actions based on the fact that the real sentences refer to real actions), children tended to point to real action even if they heard the novel construction, presumably because the novel construction included the real words (e.g., vabuild, buibuild). For these reasons, it was concluded that the test against chance would indeed most likely underestimate children’s knowledge of the novel construction. As a consequence children’s performance in the forced-choice task in Experiment 1 was tested not only against chance but also against the baseline of the control group in Experiment 2. The insights based on this test alone are necessarily limited, because there are several differences between the two studies. Children in the control group were English-speaking, aged between 5 and 7, the verbs used overlapped with but were not entirely identical to those in Experiment 1 and only the prefix construction was used. These circumstances make the comparison to the English-speaking group learning the prefix in the present study most appropriate. However, since reduplication learning was more difficult, it is not to be expected that a control group hearing the reduplication construction would have performed higher. The comparison to German is more difficult, in particular because the German-speaking sample included older age groups. Nevertheless, older German-speaking chil- <?page no="179"?> 166 dren in particular performed very well and were in fact even above chance from age 5 on in prefix learning and from age 6 onwards in reduplication learning. Any conclusions derived from the test against the baseline of the controls in Experiment 2 alone would have to remain preliminary. But they are supplemented by findings from the other two tasks and extended by the more conservative tests against chance, so that conclusions as to children’s abilities to learn the novel constructions are possible. The test against the control group was thus considered acceptable in the present context. Further studies with separate control groups 11 might provide additional insights into the baseline performance of children learning novel constructions at different degrees of abstractness. 6.4.3 Pattern effects Previous research has revealed that not only the input frequency of individual constructions affects children’s learning of the respective constructions (cf.4.2-4.4), but that input frequencies of related structures play a role as well. Related structures include previously learned similar ones with structural overlaps (e.g., learning of the sein-passive (sein ‘to be’) is supported by the previously learned copula sein; Abbot-Smith and Behrens 2006), as well as structures that share a typological feature (e.g., morphological inflections; Aksu-Koç and Slobin 1985; De Villiers and De Villiers 1985; Dressler 1997; Smoczynska 1985) or complex sentence structure (Barnes et al. 1983; Hoff-Ginsberg 1998; Huttenlocher et al. 2002). Finally, the corpus study (Chapter 5) has revealed that higher input frequencies at the pattern level co-occur with higher output frequencies. The present experiment explored whether pattern frequency, i.e., the frequencies of constructions that share the same pattern with the construction in question at a more abstract level, affects novel construction learning. Such a difference was expected to be reflected in an effect of language in novel prefix construction learning because the underlying pattern frequency was higher for German than for English (cf. Chapter 5). Indeed, there was a significant effect of language in the forced-choice and the production tasks. German-speaking children were at least one year ahead of English-speaking children in generalizing the novel prefix construction to new cases and were also better at reproducing familiar examples of the novel construction. These findings were thus in support of Hypothesis I b. They suggest that pattern frequency affects novel construction learning. More precisely, the summed up frequencies over other verb prefix constructions influenced children’s learning of a new exemplar of such 11 Please note that there was no control group for Experiment 1 because the test of a control group would have doubled the number of participants. <?page no="180"?> 167 constructions. This study is the first to show that frequencies of related constructions in terms of pattern frequency affect learning. There was, however, no advantage for German-speaking children in the act-out task; both language groups performed equally well. This first task of the experiment was different from the other two in that the children were given feedback. If children failed to act out a pretend or real action adequately in response to the invitation Show me how to ta[verb]/ [verb]., the experimenter showed how the respective action was done. Following feedback, children might have generalized their understanding only locally to the following action(s). Similar windows for learning were not provided in the forced-choice and the production task, where correct and incorrect responses were acknowledged by somewhat neutral praise (e.g., Well done for pointing.). While the feedback seemed to have benefitted both language groups equally within the act-out task, generalizations of the knowledge in subsequent tasks were more successful for German-speaking children. Altogether, the results thus provide partial support for Hypothesis I b. I Learning of a novel prefix construction is affected by the pattern frequency in the native language, i.e., the frequency of the pattern that the novel construction exemplifies. Higher pattern frequency results in more successful learning. b German-speaking children do better than English-speaking children in novel prefix construction learning. One point regarding the comparison of German-speaking and Englishspeaking children is important to discuss. All comparisons were performed regardless of one difference between the two groups, which was already mentioned in 6.2.2. German-speaking children learned the form ta[verb], whereas English-speaking children learned va[verb]. The reason for this variation was that a few English-speaking children in the pilot study seemed to confuse tawith to, and ta is also used as a short form of thanks in the region where many of the English-speaking children were tested. One reason why ta[verb] might be easier to learn than va[verb] is that the first plosives are frequently acquired before the first fricatives (Jackson and Stockwell 2011: 138). However, when children are 4 years old they usually command a comprehensive phonetic inventory. In order to assess whether there were any differences between ta[verb] and va[verb], an additional group of 16 English-speaking 4-year-olds was tested on the novel ta[verb] construction. Both English-speaking groups were then compared. There were no significant differences between 4-year-olds learning taand vafor any of the three tasks, except for a significant interaction with age (B = -0.28; SE = 0.12; z = -2.26; p < .024) in the forced-choice task. The interaction suggested that young 4-year-olds found ta[verb] more difficult than va[verb], which means that, if anything, va[verb] was easier for them rather than harder. For this reason the comparison of the two language groups <?page no="181"?> 168 was considered acceptable. While it cannot be excluded that the English disadvantage goes back to the different prefix, this explanation does not seem very well-founded. The second novel construction was a verb-initial reduplication. In contrast to all previous novel construction learning studies (Boyd and Goldberg 2011a; Casenhiser and Goldberg 2005; Wonnacott et al. 2012; prefix learning in the present study), this construction was genuinely novel in that the underlying pattern was completely absent from the children’s native languages (German and English). Due to the seemingly similar situation, no difference was expected between German-speaking and English-speaking children in reduplication construction learning. Contrary to this expectation, there was a significant advantage for German-speaking children in the forced-choice and the production tasks. German-speaking children were earlier in their generalizations of the novel construction, they generalized more and revealed memory of more examples. In the act-out task, there was no difference between the children of the two language groups. But since the corrective feedback might have been at least partly responsible for this finding, the two other tasks are given more weight. Consequently, the results are not in keeping with Hypothesis II b. II Learning a novel reduplication construction exemplifying a pattern that is not present in the learners’ native language (i.e., whose pattern frequency equals zero) is equally difficult for speakers of different native languages. b German-speaking and English-speaking children learn the novel reduplication construction equally well. In some ways, learning of the novel reduplication construction was particularly similar to early naturalistic language learning, since form and function were entirely new to children. In other ways, it was not the same as naturalistic learning, not only because it was an experimental situation, but because the children in the experiment were already fairly experienced in their native language. Even the youngest participants had had at least 3 years of exposure to their native language. Aspects of their native language other than the absent reduplication pattern might thus have affected reduplication construction learning. In novel prefix construction learning, German-speaking children profited from the high number of verb prefixes (and their types and tokens) in their language. The question is how this effect arose. It is possible that the verb prefixes children were familiar with led children to look for new information in the position of prefixes and thus made the understanding of the novel prefix construction as one example of the verb prefix pattern more likely. The high number of particles in German (see Chapter 5) might also have contributed to the German advantage. The separability of particles from the verb base might have invited children to look for separate meanings of prefix/ particle and verb base <?page no="182"?> 169 (Behrens 1998). The high frequency with which linguistic information is preposed to verb bases might thus have shaped German-speaking children’s attentional patterns. German-speaking children might have developed a habitual and possibly anticipatory tendency to attend to the beginnings of verbs. Since the reduplication was also verb-initial, they might have noticed the reduplication faster and, due to the high number of particles, they might additionally have been more willing to attribute a separate meaning to it. It seems likely that the experience with our native languages thus shapes our attention in some ways and that it might facilitate the learning of some constructions while making others more difficult to learn. Whether, in the case of novel prefix and novel reduplication learning, the verb level is the accurate level of description remains open. It is possible that children’s attentional pattern to the beginning of verbs is in fact a more general pattern to the beginning of lexemes of all word classes, which in turn also goes back to the frequencies with which morphemes are preposed to any base (verbal or otherwise). It is, however, known that children of the age groups under investigation are able to distinguish, for instance, nouns and verbs (Lieven et al. 2003) and prefer specific knowledge over more general knowledge (Lakoff 1987: 147 about Wilensky’s Law; Wilensky 1983: 25, 145). For the time being, the attentional pattern is tentatively assumed to refer to verbs, but further studies might reveal its more general nature. Should the attentional explanation receive further support from future novel construction learning research, the role of attentional patterns must be integrated in the proposed steps of constructional learning (cf. 3.4.3). It has been suggested that children store several exemplars of an emergent construction, subsequently compare them in a structural alignment and mapping process (analogy), abstract a schema and then generalize it to new cases. Over time and with increasing experience in the native language, this process would become affected by attentional patterns that are formed in accordance with the frequencies of recurring linguistic constructions and patterns. These attentional preferences would then affect all stages of the proposed learning process, including early ones, e.g., storage of a form is facilitated if it reflects a familiar information structure. The final tentative hypothesis relating to pattern frequency explored potential differences between learning of the novel prefix and the novel reduplication constructions. Seeing that the prefix construction was based on a pattern that existed in both German and English, whereas the reduplication was not (i.e., underlying pattern frequency was higher for the prefix than the reduplication), the former was expected to be easier to learn than the latter. Analyses revealed significant effects of pattern in all tasks and for both languages. The levels of significant frequency effects shifted so <?page no="183"?> 170 that more tokens seemed to be necessary in reduplication learning and the age at which children significantly generalized the novel construction was higher for reduplication learning (see t-tests). Moreover, fewer children remembered all previously experienced verbs in the novel construction in reduplication construction learning and more children remembered none compared to children learning the novel prefix construction. The finding that learning the novel prefix construction was easier than learning the novel reduplication construction provides strong support for hypothesis III. It further supports the relevance of the pattern level to learning. There is, however, an additional or possibly alternative explanation for the pattern effect. Apart from their occurrence or non-occurrence in German and English, there was a further difference between prefix and reduplication constructions. The prefix remained stable in all realizations, whereas the reduplication changed with respect to each verb, i.e., the novel prefix construction was partially-filled, while the reduplication construction was entirely abstract. Learning the novel reduplication construction therefore required children to focus exclusively on the relational similarity between the reduplicated syllable and the verb base ([ve][verb]), while children learning the novel prefix construction could potentially have been guided by the concrete phonological similarity of taor va- (object similarity) into noticing the similarity of the relation between prefix and verb base (ta[verb]/ va[verb]). The same number of tokens might have resulted in stronger memory representations due to the reoccurrence of the stable prefix, which might additionally have made this construction more transparent, thus further facilitating learning (as shown for verb learning by Narasimhan and Gullberg 2011). Based on the present data, any conclusions must therefore remain preliminary and support for Hypothesis III is not unlimited. III Learning a novel construction exemplifying a familiar pattern is easier than learning a novel construction exemplifying an entirely unfamiliar pattern (i.e., learning the novel prefix construction is easier than learning the novel reduplication construction). Since pattern frequency might have been supported by object similarity or either of the two (pattern frequency or object similarity) might be solely responsible for the effect, additional research is necessary. Future studies can contribute to disentangling the two factors, for instance, by teaching the same constructions to children whose language has no prefixes but uses reduplication or to children whose language has neither of the two processes or both of them. Sign languages frequently employ reduplication, but they rarely use prefixes (Emmorey 2000: 320). While this is an interesting field for further investigations, the added difference in modality used to produce language (hands/ body versus speech organs) might make a direct comparison to the present study difficult. Many pidgin and creole <?page no="184"?> 171 languages also make use of reduplication (e.g., Nigerian Pidgin English, Jamaican Creole English, Seychelles Creole French, Berbice Dutch Creole; Rubino 2005: 23-24), whereas prefixation much more uncommon. However, the fact that at least speakers of pidgins are usually not monolingual constitutes a potential confound in comparison to the present study. Sorani Kurdish and Tajik are examples of languages that use reduplication (Rubino 2005: 24). A close assessment of whether reduplication is used in the area of derivational verb morphology and a similar exploration of prefixation in these languages would be necessary in order to determine whether they can serve as potential candidates for testing novel verb prefix and reduplication learning. Such a test would further need to be preceded by an assessment of the underlying pattern frequencies. Under these conditions, comparisons to the present study might be possible and shed light on the issue of potential effects of object similarity in the learning of novel prefix and novel reduplication constructions. 6.4.4 Token frequency effects at the constructional level Previous research has suggested that input token frequency strengthens the representation and leads to the entrenchment of types in children’s minds. Support for this idea comes from studies showing that children acquire frequent constructions speedier and earlier and make fewer errors with them an (Ambridge et al. 2008; Hart 1991; Gathercole and Hoff 2009; Huttenlocher et al. 1991; Huttenlocher et al. 2002; Maratsos 2000; Marchman 1997; Maslen et al. 2004; Matthews et al. 2005, 2007; Moerk 1978; Naigles and Hoff-Ginsberg 1998; Rice et al. 1994; Rowland and Pine 2000; Rowland et al. 2003; Theakston 2004; Theakston et al. 2003; Theakston et al. 2004). Input frequency effects have been revealed at numerous levels. One study also explored them with respect to a novel construction. Children’s degree of learning a novel word order was revealed to increase with input token frequency (Wonnacott et al. 2012). For novel morphology learning and in particular for a construction based on an entirely unfamiliar pattern, frequency effects had not been previously explored, but frequency was also expected to affect learning in terms of correct responses. For German-speaking children, there were effects of token frequency in all tasks, for English-speaking children there were frequency effects in the act-out task for both novel constructions and in the production task for prefix construction learning. Learning was thus indeed revealed to increase concomitantly with token frequency in the majority of analyses, providing considerable support for hypotheses I d and II d. Token counts further revealed that high-frequency verbs were the first ones to be stored in the novel construction, before there was evidence of storage of less frequent ones. <?page no="185"?> 172 I/ II d Performance increases concomitantly with token frequency. The finding that frequency effects were less pronounced for Englishspeaking children might have several reasons. In some cases, trends towards an effect of token frequency were suggested by the bar charts but not borne out statistically. Moreover, the age range (and with it sample size) was smaller than in German, which might have hidden frequency effects, in particular because overall performance was lower than in German, which means that frequency effects might have been more apparent in slightly older English-speaking children. Furthermore, there were more children who did not learn the novel construction in English than in German and whose performance was thus presumably unaffected by frequency. As a consequence, they might have lowered the frequency effect that may indeed have been present for English-speaking children who did learn the novel construction. This assumption is further in keeping with the finding that there were fewer frequency effects for English-speaking children in reduplication learning, which was more difficult, than in prefix learning. This higher degree of difficulty was also reflected in the frequency effects of German-speaking children in the production task. For the easier prefix construction, differences in token frequencies between previously heard verbs of different frequency levels had little impact on children’s performance. For the more difficult reduplication construction, these differences in token frequencies of familiar items were indeed reflected in children’s performance. Difficulty of the construction thus presumably affected the number of tokens necessary for correct performance. Higher input token frequencies are thought to increase the representational strength and entrenchment of the respective types in children’s minds, which are in turn held responsible for children’s increasing performance. Strong memory representation of a number of types of an emergent construction is expected to be necessary for the formation of analogical comparisons that yield a constructional schema and allow the formation of the constructional category. In this manner high input token frequency is thought to contribute to constructional learning in terms of the steps proposed in 3.4.3. 6.4.5 Type frequency effects at the constructional level The only other study that had explored the role of type frequency in novel construction learning is a study by Wonnacott and colleagues (2012). They varied the number of types children were exposed to. Exposure to a single type was insufficient for generalizations in 5-year-olds, whereas four different types of the novel construction in the input did result in generalizations. What remained unknown, however, is how well children remem- <?page no="186"?> 173 bered how many of these four types. Therefore, rather than using only a particular, low number of types in the input, the present study assessed how many previously experienced types children had retained in memory by testing their ability to act-out, comprehend and produce them before they showed first generalizations of the novel construction. Results revealed that children had not necessarily stored two types of the novel construction before generalizing. There were, however, relatively few cases where children had not. In the act-out task, 9 children generalized once (out of four possible times), when having stored one or none of the previously heard examples. It has to be borne in mind that children received corrective feedback in this task. It is thus possible that they memorized items after failing them because the experimenter then acted them out. This might have allowed children to generalize the novel construction once, even though it seemed that they had not memorized any types or only a single one. In the forced-choice task, 14 children generalized based on none or one stored example. These cases might go back to the task format, where correct pointing was not impossible without accurate constructional knowledge. In the production task, there were 2 children who generalized once based on a single example. Here, as well as in the other tasks, it cannot be excluded that children had in fact stored more examples but were unable to show this knowledge in their own uses. Despite these potential explanations, Hypothesis IV was not supported by the data. IV At least two types (i.e., different verbs in the novel construction) are stored before the novel construction is generalized to new cases. Nonetheless, it was revealed that the number of generalizations children formed increased from one stored type onwards, i.e., the more types children remembered the more they generalized, in all three tasks. The present study is the first to show this relationship between the number of stored exemplars and the number of generalizations for novel constructions, in this case novel morphological constructions. A high number of stored types is thought to benefit the representational strength and entrenchment of the respective constructional schema, which in turn presumably promotes generalizations (productivity) of the construction. 6.4.6 Conclusion The present study has brought forward evidence for language learning from the input and explored the role of input frequencies at pattern and constructional levels in this process. Children were able to learn two novel constructions from the realm of morphology to different degrees depending on their age, the task and input frequencies, supporting the idea that constructions of different levels of abstractness can be learned from the input. The examination of two languages lends support to the idea that the <?page no="187"?> 174 same, possibly general, cognitive processes are at work regardless of the language explored. Moreover, the various effects of input frequency and the progression with age provide insights about the learning process itself. A novel construction that is based on a familiar pattern is learned with greater ease than one using no familiar source, presumably because the underlying pattern frequency differs. The frequency of a pattern that is indeed present in a language further affects the learning of a novel construction that exemplifies it (facilitation with higher frequency), but may additionally cause the development of more general attentional patterns in native speakers. The input token frequency of types of a novel construction affects children’s memory of these forms, with higher frequencies causing earlier learning success. The storage of types is a first step towards later generalizations. Generalizations are more likely if children have stored a high number of types of a novel construction, presumably because the constructional schema is represented more strongly and the categorization of new members is thus facilitated. <?page no="188"?> 175 7 Experiment 2: The role of the shape of the input distribution in the learning of a novel partially-filled construction The present study explored the role that the relation between type and token frequencies in the input plays in language learning. More specifically, it examined how the distribution of tokens per verb type in the input affects novel construction learning. First, a brief summary of the theoretical background is given. The experimental materials and procedure are then outlined. Subsequently, the results are reported and discussed. 7.1 Background The role of type frequency is important for schema abstraction and generalization. In order to form a productive slot in a string, children require a certain degree of variation in the emergent slot position (Bybee 1995; Bybee 2010: 95-96). Previous research suggests that a single type in the input is likely insufficient for children to generalize abstract constructions (Wonnacott et al. 2012, cf. Chapter 6), presumably because it does not allow the necessary analogical comparison process. The higher the number of different types in a slot position (when a certain degree of openness is given) the more prone a construction is to generalizations to new cases, i.e., the higher its productivity (Bybee 2010: 95-96; including overgeneralizations, Guillaume 1927/ 1973; Marchman 1997; corpus study Chapter 5). Input token frequency also affects this process. A certain degree of representational strength is thought to be necessary for a type to participate in the comparisons leading to schema formation (slot formation) and subsequent generalizations. Conversely, extremely high token frequency is presumed to lead to the highly independent storage of a type, which in turn hinders the type’s participation in comparison and schema formation (Bybee 1995). Thus, certain input frequencies and the resulting degrees of representational strength seem to be more facilitative for learning abstractions and for making generalizations than others. These considerations suggest that type and token frequency are not entirely unrelated. This circumstance is accommodated in the type-token ratio, which describes the number of tokens per type in the input. Depending on whether type-token ratios are similar for different types in the input or not, the resulting input distribution becomes balanced, skewed or else. In the case of a partially-filled construction, balanced input has been proposed to be particularly conducive to slot formation. An example of a bal- <?page no="189"?> 176 anced input distribution is a situation where the exposure to 120 tokens of the construction Throw the [X] is equally divided between the three types Throw the bottle, Throw the ball and Throw the teddy so that each one is heard 40 times. In a skewed input distribution of the same construction, Throw the bottle would, for instance, be heard 118 times and Throw the ball and Throw the teddy once each (Matthews and Bannard 2010). Since predictability for the slot position is lower for the balanced input, slot formation is more likely. In the case of skewed input the filler of the potential slot position is highly predictable because it is nearly always filled with the same one, which makes actual slot formation unlikely. This relation between predictability and slot formation was examined by Matthews and Bannard (2010) in a word-string repetition task. They showed that 2and 3-year-old children were more likely to have formed a slot at the end of a four-word sequence if the last word was difficult to predict based on the input. The ease, with which children were able to fill the fourth position with a new filler, was taken as evidence for the abstraction of the slot. The proposed relation between slot formation and variability awaits further experimental support and has not yet been explored in a learning task with a partially-filled construction. What has been explored, however, is the effect of the input distribution on abstract construction learning (cf. 4.4.1). Casenhiser and Goldberg (2005) compared skewed and balanced input distributions in novel word order learning. Based on the finding that naturalistic input is skewed towards one verb for several constructions, e.g., go occurs in 39% of all [Subject] [Verb] [Location] constructions (Goldberg et al.’s 2004 reanalysis 1 of Bates et al. 1988), Casenhiser and Goldberg expected skewed input to be beneficial in novel abstract construction learning. The construction they used took the form [Noun Phrase theme ] [Noun Phrase location ] [Verb Phrase novel ] and the meaning ‘appearance’. Children’s exposure was either skewed toward one verb type that occurred disproportionately frequently (8-2-2-2-2) or it was more balanced (4-4-4-2-2). The overall number of types and tokens was the same in both cases. The comprehension of generalizations was assessed in a forcedchoice task. While both groups learned the novel construction in comparison to a control group, children in the skewed input condition performed significantly better than children in the balanced condition. Goldberg and colleagues (2004) showed similar results for adults, but Goldberg, Casenhiser and White (2007) found the advantage of skewed input over different control groups only when all tokens of the skewed verb occurred at the beginning of the training but not when they were interspersed with tokens of verb types with lower frequencies. Childers and Tomasello (2001) found 1 A problem with their analysis is that data from 15 mothers was collapsed, which might have led to an overestimation of the skewedness of the distribution towards frequent verbs. <?page no="190"?> 177 an advantage of skewed input for the learning of a familiar abstract construction in 2½-year-olds. They trained children on the transitive construction. In one condition, training sentences took pronouns and full noun phrases as subjects and objects; in the other condition, only full noun phrases were used. Children in the former condition, where input was skewed towards recurrent pronouns, profited more from the training. The presumed reason why skewed input is more conducive to learning abstract constructions than balanced input is that the highly frequent lexeme serves as an anchor for the entire construction. Such an anchor is particularly useful in this situation because it indicates which structures can be successfully aligned and mapped in the analogical comparison, when all the other lexical material varies in each instantiation. Given these contradictory findings, it is possible that the ideal input distribution is dependent on the level of constructional abstractness. In partially-filled constructions, where part of the linguistic material remains stable all the time, balanced input allows the comparison of types and subsequent slot formation earlier because variability in the relevant position becomes apparent sooner. In abstract constructions, on the other hand, all linguistic material varies per definition in each instantiation. Comparisons are thus more difficult because they must rely exclusively on relational similarities. If, however, one dominant item re-occurs highly frequently in the same position, it provides additional supportive object similarity, makes relational similarities more apparent and thus invites the relevant comparisons. Based on these assumptions, the present study investigated the effect of the input distribution on learning a partially-filled novel construction. To this end, children were taught a novel prefix construction. The construction was partially-filled, because the prefix remained stable in all instantiations whereas the verb position constituted the slot ([prefix][verb]: va[verb]). The overall type and token frequencies in the input were the same in all conditions, but the distribution of tokens per type varied. Three levels of input ‘skewedness’ were tested: balanced, skewed and semi-skewed. In the balanced input condition, children heard all verb types roughly equally frequently in the novel construction. In the skewed condition, they heard one verb type considerably more often than all others. The semi-skewed condition lay in-between; children heard two types much more frequently than the others. The semi-skewed condition was included for exploration and because of the continuous nature of the input frequency distribution. If the shape of the distribution affects learning of such a partially-filled construction, as might be predicted from Bybee (1995) and Matthews and Bannard (2010), children in the balanced distribution group should do best in a comprehension task that assesses their ability to generalize the novel construction to new cases. Performance of children in the skewed input <?page no="191"?> 178 group should be the lowest (Hypothesis I a/ b). For the semi-skewed condition, it was assumed that performance would either reflect the degree of skewing and lie between the balanced and skewed conditions or that the two highly frequent types would allow the comparison sooner than the single highly frequent type in the skewed condition, resulting in a difference from the skewed, but not necessarily from the balanced condition. Children in all three groups were expected to show some understanding of the novel construction compared to a control group that did not receive any training. I Input distribution a Children are able to learn the novel construction from the input. b Balanced input is more conducive to learning a partially-filled construction than skewed input. 7.2 Method 7.2.1 Participants One hundred and seventeen monolingual English-speaking children (63 girls and 54 boys) between 5; 0 and 7; 0 (M = 6; 2) were recruited from primary schools in Greater Manchester, Cheshire and Derbyshire. Children received stickers as a small reward after testing; schools were given book vouchers for participation. 7.2.2 Materials and procedure Children were tested individually in a quiet room in their school. Each testing session lasted 15 to 20 minutes. Children were trained and tested on a novel construction that was used with different familiar verbs. In a between-subjects design, participants were randomly assigned to one of four input frequency conditions: balanced, semi-skewed, skewed and control. The construction children learnt was a novel verb prefix that took the form va[verb] and carried the meaning ‘pretend to perform an action’, e.g., vadrink meant ‘pretend to drink’. The same novel construction was also used in the study reported in Chapter 6. The form vawas selected because it is not a morpheme in English and does not mean anything on its own. The meaning ‘pretence’ was used because children aged 5 to 7 could be expected to be familiar with it (Fein 1981; Harris and Kavanaugh 1993; Leslie 1987; Rakoczy, Tomasello and Striano 2004) and because it is not expressed by a prefix in English. Instead, the pretend to [verb] construction usually serves to convey this meaning. The experiment consisted of two parts: a training phase and a forced-choice comprehension task. <?page no="192"?> 179 Verbs Eighteen verbs were used in the experiment (Table 13). They were selected on the basis of the following criteria: It was possible to act out distinct real and pretend versions of each verb. (Note that for verbs like sleep or think it would be impossible to discern a difference between real and pretend actions.) All verbs were transitive (expressing accomplishments or involving affected patients). Verbs were one syllable long in the infinitive and in the third person singular present tense (except for four verbs ending in sibilants, where the third person ending added an extra syllable). Verbs that started with a vowel were excluded a priori, since the novel forms of such verbs would have been unusual for English phonotactics and might thus have been more difficult to learn, e.g., eat would have become vaeat. According to the literature, all verbs were familiar to 5to 6-year-olds (Dale and Fenson 1996; Masterson and Druks 1998; Masterson et al. 2008; Stadthagen-Gonzalez and Davis 2006; Székely et al., 2004). 2 Table 13. Verbs used in the training phase and the forced-choice task. A B C build ** draw ** cut ** fold * push * throw * catch glue drink drill peel knit mix roll wash tie sweep wipe Note: Verbs in bold print were used for the acting out at the beginning of the training phase. Verbs marked with ** were used as the most frequent ones in the skewed and semi-skewed conditions. Verbs marked with * were used as the second most frequent ones in the semi-skewed condition. Two strategies were used to make actions pretend: either no objects were used (e.g., the actor or actress pretended to drink with just their hand), or a relevant prop was swapped for a toy block or similar. The reason why all actions were not performed without any objects was that some would have been very difficult to interpret and that children might have taken the novel form to mean ‘mime’ rather than pretence. Verbs were split up into 3 groups of 6 verbs each (sets A, B, C in Table 13). One set of 6 verbs was used during training and the other two appeared in the forced-choice task, one with pretend items and one with real items. The assignment of the sets A, B 2 For mix and glue no records could be found in the literature, but informal nursery teacher interviews suggested that even younger children would already know these words. <?page no="193"?> 180 and C to training items, pretend test items, and real test items was counterbalanced between children. Training Six verbs were used in this phase. First, the child and experimenter acted out real and pretend versions of two verbs (bold in Table 13) and named them appropriately. The child was asked to perform the real action of a verb (e.g., How do you draw something? ). If necessary, the experimenter demonstrated the action. Next, the experimenter showed the child how to perform the corresponding pretend action and said, Now we want to vadraw. This is how you vadraw. Can you vadraw as well? Finally the experimenter asked the child to repeat the novel form. If the child did not produce the form, the experimenter repeated it, so the number of tokens would be the same for all children. The same procedure was followed for the second verb. Next, children were shown 24 film clips, in each of which an actor or an actress performed a real or a pretend action involving one of the six verbs. The clips were presented together with pre-recorded audio descriptions of the scenes. For real items these were sentences such as He builds something. He builds (one repetition of both sentences). For pretend items the verbs were used in the novel construction He vabuilds something. He vabuilds 3 (one repetition). In order to make the training more efficient, after watching each film children were asked to repeat the sentences (accuracy of repetition was not used as a measure for constructional learning). To make the repetition task more attractive, children were shown a photo of a man at the beginning of the training. They were told that this was Johnny who never listens. Hence, Johnny probably would not listen during the films, which is why the children were asked to tell Johnny what they had heard during each film once the clip was over and Johnny was listening again. To make this situation clearer, a small picture of Johnny with his ears covered was shown while the clips were playing, and the picture changed to one where Johnny was listening, once each clip had finished. Children were given a microphone to speak into so as to make the repetition task more interesting. Film clips always came in pairs, with the pretend action of a particular verb always following the corresponding real one. There were 12 such pairs. In the balanced condition, each pair was played twice over the course of the training (2-2-2-2-2-2). In the semi-skewed condition, one pair was played 5 times, one 3 times and all other pairs once each (5-3-1-1-1-1). In 3 Parallel to the study reported in Chapter 6, the simple present rather than the present progressive was used in the descriptions so as not to make the verbs in the novel construction unusually long. <?page no="194"?> 181 the skewed condition, one pair was played 7 times, all others once each (7- 1-1-1-1-1). The order of pairs was pseudo-randomized for each condition, such that no more than two pairs of the same verb occurred consecutively. Pairs with female and male actors alternated. The training phase was preceded by a warm-up, in which the experimenter demonstrated the repetition game to the child and then the child was encouraged to try it himself or herself. Training items were not used again during the experiment. Given that during each clip children heard the novel construction four times and that two of the verbs were additionally produced five times each by the experimenter (or four times by the experimenter and once by the child) in the act-out training, the frequency distribution of the novel construction was 13-13-8-8-8-8 in the balanced condition, 25-17-4-4-4-4 in the semi-skewed condition, and 33-9-4-4-4-4 in the skewed condition. Children in the control condition did not complete this phase of the experiment. Forced-choice comprehension task There were 12 test items. For each of them, children were presented with two films simultaneously. In one of the films the male or female actor performed the pretend action and in the other one the same person performed the corresponding real action. Children were asked to point to the film that matched the pre-recorded auditory description given together with the clips. For half of the items, it contained the novel construction, e.g., He vabuilds something. He vabuilds, and for the other half of the items it did not, e.g., He builds something. He builds. Twelve new verbs were used in this phase (6 for real items and 6 for pretend items), i.e., only generalization was assessed. Verb order and verb form (real or pretend) were pseudo-randomized, such that no more than two consecutive verbs took the same form. Two parallel versions were constructed by altering the verb forms (real pretend, pretend real). The target side was pseudo-randomized, such that the target never occurred on the same side more than twice in a row. As in the training phase, the male and female actors alternated. Two warm-up items served to familiarize children with the pointing procedure and to ensure their points were unambiguous. During this phase only normal, real actions were used. The same actions/ verbs were not used again during the training or the test phase. 7.3 Results Figure 27 shows the proportions of correct responses with respect to the condition (balanced/ semi/ skewed/ control) and construction, i.e., <?page no="195"?> 182 whether the sentence involved the novel pretend construction (pretend) or not (real). For real items requiring normal English children’s performance was close to ceiling for all the conditions. For items containing the novel construction, there was a clear trend of a decreasing performance with increasing ‘skewedness’ of the distribution with a further substantial drop in the control group (who had received no input with the novel construction before testing). For both real and pretend items, children in the control group exhibited very similar tendencies to point to the real actions (~90%; this behaviour was classified as incorrect for pretend items) rather than showing performance at chance level for pretend items. Figure 27. Proportions of correct responses as a function of condition and construction type. For reasons given in 6.3.1 a mixed effects logistic regression model (Baayen et al. 2008) was fitted to the data instead of analyzing the data as proportions of correct responses per child. Traditional t-tests were not used to further assess learning in this study, because children could be compared to the control group. Correct versus incorrect response was used as the dependent variable in the regression model; condition (balanced/ semi/ skewed/ control) and construction (real/ pretend) were included as fixed factors. Children and verbs served as random factors. The model was re-fitted twice with re-ordered levels of condition so as to test differences between all levels of the factor condition. Estimates of the coefficients for the original and the re-fitted models are given in Table 14. balanced semi skewed control Condition <?page no="196"?> 183 Table 14. Fixed effects in the Condition x Construction model. B SE z p Cd = balanced (vs control) 5.75 0.65 8.87 < .001 Cd = semi (vs control) 5.19 0.64 8.11 < .001 Cd = skewed (vs control) 3.41 0.61 5.64 < .001 Cd = semi (vs balanced) -0.56 0.63 -0.90 <.371 Cd = skewed (vs balanced) -2.34 0.60 -3.92 <.001 Cd = semi (vs skewed) -1.78 0.59 -3.02 <.003 Cx = real (vs pretence) 5.47 0.45 12.13 <.001 Cd = balanced (vs control) x Cx = real (vs pretence) -4.53 0.61 -7.41 <.001 Cd = semi (vs control) x Cx = real (vs pretence) -3.04 0.70 -4.37 <.001 Cd = skewed (vs control) x Cx = real (vs pretence) -1.02 0.71 -1.44 <.152 Cd = semi (vs balanced) x Cx = real (vs pretence) 1.50 0.68 2.20 <.029 Cd = skewed (vs balanced) x Cx = real (vs pretence) 3.51 0.70 5.03 <.001 Cd = semi (vs skewed) x Cx = real (vs pretence) 2.02 0.77 2.61 <.010 Note: Condition is abbreviated as Cd, construction is shortened to Cx. Reference levels are given in brackets. Table 15. Fixed effects in the Condition model (Cx = pretence). Table 15 gives the coefficients for the regression model, where only pretend items were taken into account. The results confirmed the intuitions based B SE z p Cd = balanced (vs control) 10.92 1.65 6.62 < .001 Cd = semi (vs control) 10.33 1.61 6.41 < .001 Cd = skewed (vs control) 7.03 1.50 4.68 < .001 Cd = semi (vs balanced) -0.59 1.53 -0.39 <.700 Cd = skewed (vs balanced) -3.89 1.42 -2.74 <.007 Cd = semi (vs skewed) -3.30 1.37 -2.40 <.017 <?page no="197"?> 184 on inspecting Figure 27. There were significant effects of condition and construction as well as a significant interaction between the two. Children in all experimental conditions were significantly above controls on pretend items. Furthermore, children in the semi-skewed and the balanced conditions were further significantly better than children in the skewed condition. On real items all children performed roughly equally well. The difference between real and pretend items was significant for all conditions. 7.4 Discussion The present study revealed that children between 5 and 7 were able to learn a completely novel morphological construction on the basis of very limited input. Learning was more successful when input was balanced in comparison to skewed input. Interestingly, semi-skewed input caused a similar advantage over skewed input. The first finding supports Hypothesis I a and is further in keeping with previous research on novel construction learning. It supports the previously reported findings on novel morphological construction learning (Chapter 6), where children from at least 4 to 6 years of age showed learning of a novel prefix or a novel reduplication. In addition, it extends earlier studies on novel argument structure construction learning, which revealed that 5to 7-year-old children were able to learn novel word orders from the input (Boyd and Goldberg 2011a; Casenhiser & Goldberg 2005; Wonnacott et al. 2012). I Input distribution a Children are able to learn the novel construction from the input. Performance on pretend items nevertheless did not reach the level of real items, suggesting that children’s knowledge of the novel construction was emergent. Children’s high competence on real items with regular English was expected (children knew this construction), but is important for one reason. This finding revealed children’s ability to discriminate between the two actions and the corresponding constructions. It is thus apparent that children’s good performance on pretend items was not caused by indiscriminate pointing to the novel perhaps more unusual action regardless of the input, but was indeed based on the construction in the auditory input. Children without training (control condition) did not point to the films with pretend actions at chance level when the novel construction was used in the input, but at a much lower rate (lower than 10% correct). Instead they showed a strong tendency to point to the known, real actions, even when they heard the novel construction. A reason might be that the linguistic experience they brought to the task affected their behaviour. They <?page no="198"?> 185 might have justified points to real actions with the fact that the novel construction also contained ordinary verbs (e.g., vabuild). The finding that control children performed at a much lower level than chance further implies that they did not learn the novel construction based on mutual exclusivity, i.e. by understanding the match between real action and normal verb and concluding that the unfamiliar (pretend) action was to be related to the unfamiliar language. It is thus also unlikely that children in the experimental conditions relied on mutual exclusivity when learning the novel construction. 4 The second and main finding of the present study concerned the shape of the input distribution. Children who had received balanced input were significantly better on pretend items than children in the skewed input condition. This finding supports Hypothesis I b and is further in line with Matthews and Bannard’s (2010) proposal that slot formation benefits the most from balanced input (i.e., high entropy = low predictability of slot fillers). It also supports the finding that young children were more likely to form a productive slot in a word string if the filler for the potential slot position was difficult to predict (Matthews and Bannard 2010). I Input distribution b Balanced input is more conducive to learning a partially-filled construction than skewed input. The present finding does, however, stand in opposition to previous results in novel argument structure construction learning. Casenhiser and Goldberg (2005) had shown that children benefitted more from skewed than from balanced input. Learning was facilitated if one verb re-occurred disproportionately frequently in the novel construction. I propose that the present findings can be reconciled with earlier ones. The difference between Casenhiser and Goldberg’s and the present study is the level of constructional abstractness. In the former, the argument structure construction was entirely abstract. This entails that the meaning of the constructions was attached to something variable, i.e., the word order, and that every position was filled with different lexemes in each realization. In the present study, on the other hand, the construction was partially-filled. The meaning was attached to the prefix, which remained stable in all realizations. The tasks children had to perform in order to learn the novel word order and the novel prefix constructions were consequently very different. As suggested in the background section, the learning process might be conceptualized in terms of analogical comparisons over different types leading to schema abstraction and generalization. In this sense, children 4 In the experiment reported in Chapter 6 children were also tested on an act-out and a production task. Performance in these tasks also contradicts the idea of learning by mutual exclusivity. <?page no="199"?> 186 learning the novel word order had to rely on relational similarity alone in order to form the comparisons because of the variation in each position. If one verb (or, in Childers and Tomasello’s training study, the pronouns in the subject or object position) re-occurred frequently because of input skewing, children were provided with additional object similarity, which presumably guided them towards aligning different types of the emergent construction accurately. This idea receives support from the literature on analogies, where relational mappings were revealed to be more easily formed if they were supported by additional object similarity (DeLoache 1990; Gentner and Markman 1997; Gentner and Medina 1998; Gentner and Toupin 1986; Holyoak and Thagard 1995: 83-84; Keane 1987). By providing this supportive object similarity, the skewed item might have served as an anchor for the entire abstract construction. In the novel prefix learning study, on the other hand, children were automatically presented with supportive object similarity in the input 5 , because the prefix remained stable in all realizations. Consequently, relational similarity between different types was always supported by the object similarity of the prefix itself, which potentially facilitated the analogical comparisons. Balanced input was then more conducive to slot formation, because it made the slot become apparent sooner than skewed input, where the same verb re-occurred frequently, thus making variability less obvious. It is possible that the fact that the prefix construction was embedded in a relatively stable [Subject] [Verb Phrase] ([Object]) construction most of the time (where the subject position was filled by pronouns and the object slot was either filled by something or absent) was additionally conducive to learning. Regardless of this aspect, the construction children learned was partially filled one way or another and, crucially, the meaning was ascribed to the va[verb] rather than to the word order. There are two aspects that might have further made the learning of novel argument structure constructions in the Goldberg studies particularly difficult (Casenhiser and Goldberg 2005; Goldberg et al. 2004; Goldberg et al. 2007). According to the usage-based model, children form abstract constructions from stored, familiar item-based ones, e.g., the abstract [Agent] [Verb Phrase] [Patient] construction might be formed on the basis of the item-based constructions [Kisser] kisses [Kissee], [Hitter] hits [Hitee], [Kicker] kicks [Kickee] and so forth (Tomasello 2003: 144). In each novel argument structure construction learning study, however, children were required to learn the abstract construction based on individual examples of 5 Supportive object similarity has also been revealed to be conducive to the learning of a novel abstract construction. Casenhiser and Goldberg (2005) showed that children’s learning of the novel abstract construction in the skewed condition was facilitated by recurrent material: When they added the suffix {-o} to all the novel verbs they used, the rate of correct responses increased significantly from 68% to 78%. <?page no="200"?> 187 each item-based construction. They thus had to take two steps at a time and form the abstract construction directly from the examples, without abstracting item-based constructions first. (Or, if they did abstract itembased constructions first, they had to do so before schematizing the abstract construction.) Learning was further complicated by the fact that novel verbs (i.e., invented verbs that children were not familiar with) were used. These two aspects made novel argument structure construction learning different from naturalistic learning of argument structure, where children can rely on familiar verbs and previously stored less abstract constructions, and therewith possibly particularly difficult. The final question concerns the role that semi-skewed input played in construction learning. For the partially-filled prefix construction, there was only a non-significant descriptive difference between the balanced and semi-skewed conditions, with children in the balanced condition performing slightly better on pretend items. Semi-skewed input was thus also very conducive to novel prefix construction learning. In comparison to the skewed condition, variability did become apparent sooner because two different verb types rather than a single one re-occurred frequently. According to the model of constructional learning, a single type is insufficient for learning, because it does not allow a comparison. This is precisely what Wonnacott and colleagues (2012) found for novel construction learning in 5-year-olds. In the present study more than one type was used in all conditions. It is, nevertheless, possible that extreme skewing towards a single type complicated the detection of the variability in the slot position and with it learning at this level. The re-occurrence of two frequent types in the semi-skewed condition might in contrast have provided a useful combination of type variability and entrenchment. A further reason for the advantage of both balanced and semi-skewed conditions might be that higher variability in training allowed earlier or a higher number of analogical comparisons, which potentially served to establish the new constructional category sooner or more successfully. Support for this idea comes from non-linguistic categorization. Mather and Plunkett (2011) revealed in a study with 10-month-olds that the order of presentation of training stimuli affected the accuracy of subsequent categorizations of new members, even if the stimuli were the same in different conditions. If dissimilar examples succeeded each other in training, the categorization of new items was more accurate than when similar examples followed each other during training. In the present study similarity between successive exemplars in training was highest for the skewed input condition because the highly-frequent type necessarily occurred twice in a <?page no="201"?> 188 row at two points in training. 6 This was not the case in the semi-skewed and the balanced condition. The higher dissimilarity of successive types of the novel construction in training might thus have facilitated the formation of comparisons as well as subsequent schema abstraction and constructional category formation for the balanced and semi-skewed conditions. This might in turn have made generalizations (i.e., the categorization of new members) more accurate in these two conditions. In conclusion, the present study provides further evidence that novel constructions can be learned. Moreover, it brings forward first experimental evidence that balanced as well as semi-skewed input are particularly facilitative of novel construction learning at the partially-filled level. Further research is needed to explore in more detail how different degrees of skewedness affect construction learning at different levels of abstractness. A follow-up study parallel to the present study explored these effects on abstract construction learning (Cordes, Krajewski and Lieven 2013; cf. 8.3.2). 6 Note that this is a circumstance caused by the skewing. It would not have been possible to avoid these temporally close repetitions due to the high number of tokens for the skewed type. <?page no="202"?> 189 8 Summary and general discussion The main aim of the present work was the exploration of the role of frequency on construction learning. It was hoped that the research in this book would contribute to the extension of existing knowledge about the comparison, abstraction and generalization processes in construction learning and how they are affected by frequency. The considerable lack of research in this area was pointed out in the introduction (Chapter 1) with reference to quotes from Abbot-Smith and Tomasello (2006) and Tomasello (2003: 125). In order to ameliorate this situation, naturalistic and novel construction learning in the area of derivational morphology was explored in two languages - German and English. Frequency effects were examined at the constructional level as well as the newly-introduced more abstract pattern level. This final chapter starts out with a summary of the main results of the work. In the following step, the findings are related to each other and to previous research. Their role for the understanding of the constructional learning process is discussed. Subsequently, three more general theoretical issues that surfaced in the course of the book receive extra attention. Before the final conclusion, two additional, disputable issues are addressed and future research directions are proposed. 8.1 Summary of the work The theory section was tripartite. Chapter 2 provided the background for choosing the usage-based approach to language learning as the underlying model in this book. The question of whether language learning from the input is possible without innate language-specific structures and constraints was examined by contrasting the usage-based view with the more traditional generative, innatist theory. The issue of learning from the input is important to the examination of frequency effects, because frequency is a characteristic of the input. Contrary to generative claims as to the poverty of the input, empirical research revealed the input to be exceptionally wellformed, essentially error-free (Newport, Gleitman and Gleitman 1977; Hornstein and Lightfoot 1981: 11; Sampson 2005: 43) and extensive in quantity (Cameron-Faulkner et al. 2003; Hart and Risley 1995: 132). The input thus constitutes a rich source for children’s highly-developed social intention-reading skills and their cognitive abilities to work on. Together these skills go a long way towards explaining how children build their constructional inventories based on the input and how they avoid potential <?page no="203"?> 190 over-general grammars. Following this reasoning, the position that all language knowledge, conceptualized as an inventory of constructions, can indeed be learnt from the input in absence of language-specific structures or constraints was adopted. In the subsequent chapter (Chapter 3) the cognitive processes were zoomed in on. They are relevant to the study of frequency, because input frequency is thought to work its effect via the cognitive processes involved in learning. Categorization and analogy formation have been suggested to play a main role in the learning process (Tomasello 2003: 4), which is why they were the focus of this chapter. They were first explored more generally in the linguistic and psychological literature and then re-defined for their application to language learning. It became clear that both processes involve a comparison. In language learning this comparison necessarily involves the alignment of the parallel relations of the compared structures. It was proposed that the comparison is based on two or more types of the emergent construction that have previously been stored in memory. The comparison was further assumed to yield a constructional schema that is more abstract than the individual types and represents their commonalities. Schema and construction types were thought to be stored mentally as a constructional category so that they can subsequently serve as standards of comparison when the construction is generalized to new cases in comprehension and production. In Chapter 4 frequency effects were integrated into this model. The storage of construction types was linked to input token frequency and input type-token ratio. Generally, higher numbers of tokens per type were expected to be conducive to storage because they increase the representational strength and entrenchment of the respective type (Bybee 2006, 2010: 19-20). Type variation (type frequency) was seen as a prerequisite for comparisons that allow schema and category formation. A high number of types was linked to the productivity of a construction, i.e., its generalizability to new cases (Bybee 1995, 2010: 95). It was further thought to be conducive to the representational strength of the schema of a construction. Chapter 4 further served to review previous studies on frequency in language learning. Considerable gaps were revealed in the area of derivational morphology learning. The number of types and tokens necessary and conducive to the learning of such constructions had remained mostly unexplored. Moreover, it became apparent that the effect of type-token ratios (i.e., of the shape of the input distribution) had only been examined experimentally for entirely abstract constructions, but not with respect to partially-filled ones. Since theoretical considerations gave rise to the assumption that the shape of the input distribution might differentially affect constructions at different levels of constructional abstractness, a partially-filled morphological construction once more emerged as an ideal testing ground for the ex- <?page no="204"?> 191 pected effects. The literature review further revealed that frequency effects had predominantly been explored at the level of the particular construction in question (e.g., Hart 1991; Maratsos 2000; Marchman 1997; Maslen et al. 2004; Moerk 1978; Naigles und Hoff-Ginsberg 1998; Theakston et al. 2004). However, some studies suggest that related constructions also affect constructional learning (e.g., Abbot-Smith and Behrens 2006; Dressler 1997; Huttenlocher et al. 2002; Singley and Anderson 1989: 3). For this reason, it was assumed that frequencies of constructions that are related to the construction that is being learned at a more abstract constructional level (i.e., the pattern level) might also affect learning. In the experimental section frequency effects that were explored in one corpus and two experimental studies were reported. The corpus study (Chapter 5) investigated frequency at the newly-introduced pattern level. Child (2; 0-4; 11) and caregiver speech were analyzed in two dense childspeech corpora, a German and an English one. The derivational verb prefix pattern was examined in terms of verb prefix construction types, prefix verb types and prefix verb tokens. The availability of the pattern in the input in terms of all three counts was significantly higher for Germanspeaking Leo than for English-speaking Thomas. Leo also produced significantly more construction types, prefix verb types and prefix verb tokens than Thomas. Higher numbers of construction types, prefix verb types and prefix verb tokens in the input were related to higher output frequencies in all three areas. The frequency of the derivational verb prefix pattern thus differed between the two languages and higher input pattern frequencies were reflected in higher frequencies in children’s uses (in support of research questions I a and b; research questions are repeated at the end of this summary for reference). The latter finding suggests that frequencies at the abstract pattern level, which represents an abstraction over several similar, partially-abstract constructions, are relevant to language learning. The former finding provided necessary background information for the formulation of predictions in Experiment 1. Interestingly, input and output frequencies in the corpora were not only related at the pattern level but also at the constructional level. Construction types with higher type/ token frequency in the input had higher type/ token frequency in the output. Additionally, the analysis of creative uses of both children revealed that the number of creative generalizations of prefix constructions was much higher for Leo than for Thomas, who used only one form at best. Leo overgeneralized three to five (some cases were not unambiguous overgeneralizations) of the seven prefix constructions with the highest type frequencies in his input and output. Experiment 1 (Chapter 6) explored different frequency effects at the pattern level and at the constructional level. In a training study Germanspeaking and English-speaking children between 3 and 8 were exposed to <?page no="205"?> 192 one of two novel constructions from the area of derivational morphology. The frequency of the underlying pattern was shown to affect novel construction learning. German-speaking children were more successful than their English-speaking peers at learning a novel verb prefix construction, which was based on the verb prefix pattern that had been shown to be more frequent in German than in English (in support of research question II a ii). A novel construction based on an unfamiliar pattern, a verb-initial reduplication, was more difficult to learn for children of both languages. Interestingly, however, the advantage for German-speaking children reoccurred for this second novel construction, despite the fact that the pattern did not occur in German (or English; contrary to research question II b ii). For both novel constructions, children’s performance tended to increase with the token frequency of the respective types (in support of research question III) and children generalized more rampantly the more types of the novel construction they had stored. The storage of two types seemed to be conducive to but not mandatory for the formation of generalizations (contrary to research question IV). All effects were based on the finding that German and English children were able learn either of the two completely novel constructions from the input. Their ability to do so increased with age (in support of research questions II a i and II b i). Reliable learning in terms of generalizations to new cases was evident from 4 to 6 years onwards, depending on the construction and the language. At even younger ages children were able to use previously experienced items in the novel construction, which suggests that memory of familiar types precedes the generalization to new types. In Experiment 2 (Chapter 7) the effects of the shape of the input distribution on construction learning were examined. To this end, Englishspeaking children between 5; 0 and 7; 0 were trained on a novel partiallyfilled construction. The token frequency per type was manipulated in the input distribution, such that input during training was either balanced (similar number of tokens for each type), skewed (one type occurred disproportionately frequently) or semi-skewed (two types occurred more frequently than all other types). Compared to a control group the novel construction was learned successfully in all three conditions. Crucially, balanced input was more facilitative than skewed input (in support of research question V). It was further revealed that semi-skewed input was also more conducive to the learning of the novel partially-filled construction than skewed input. <?page no="206"?> 193 Research questions: I Pattern frequency in naturalistic language learning a Are input and output frequencies of the derivational verb prefix pattern related? b Are there differences in the frequency of the derivational verb prefix pattern between German and English in the language to and of children? II Novel construction learning and pattern frequency a Pattern present in native language; frequency varies between languages i Learnability: Can German-speaking and English-speaking children learn a novel construction based on a pattern that is used in their languages and does the success of learning increase with age? ii Frequency: Is learning a novel construction easier for children in whose native language the underlying pattern is more frequent? b Pattern absent from native language i Learnability: Can German-speaking and English-speaking children learn a novel construction based on a pattern that is absent from their languages and does the success of learning increase with age? ii Frequency: Is learning a novel construction equally difficult for all children given that the underlying pattern is absent from both languages? III Token frequency in novel prefix and novel reduplication learning Does token frequency increase novel construction learning, i.e., correct responses? IV Type frequency in novel prefix and novel reduplication learning Do children use at least two previously experienced types of the novel construction correctly before forming correct generalizations? V Type-token ratio in partially-filled construction learning Is balanced input more conducive to learning a partially-filled construction than skewed input? 8.2 Frequency effects in morphological learning In this section the results of the different studies are related to each other and to previous research. After briefly discussing children’s ability to learn from the input, frequency effects are debated - first at the level of individual constructions, then at the more general pattern level. <?page no="207"?> 194 8.2.1 Construction learning based on the input Two experiments revealed German-speaking and English-speaking children’s ability to learn two novel morphological constructions from the input. This finding extends previous knowledge in several ways. Previous research had shown that children from 5 years on are able to learn novel word order constructions with the form [Noun Phrase 1] [Noun Phrase 2] [Verb Phrase novel ] and the meaning of ‘appearance’ or ‘approach’ (Boyd and Goldberg 2011a; Casenhiser and Goldberg 2005; Wonnacott et al. 2012). The present studies revealed novel construction learning for a novel prefix construction and a novel reduplication construction. Children were thus able to learn both a partially-filled and an entirely abstract morphological construction from their input. They were successful at generalizing these morphological constructions from as young as 4 years onwards, extending earlier studies that had only shown learning from 5 years on. Even children who were younger than 4 showed some memory of previously encountered types of the novel prefix construction. Children’s performance increased with age, and an additional study of adults found that learning continues to increase for participants older than 8 years of age, since adults’ performance was at ceiling for all items, regardless of whether they were new or familiar (cf. Appendix II.4). The age effect in Experiment 1 corroborates findings by Boyd and Goldberg (2011a), who had shown it for children between 5 and 7 in novel word order learning. In contrast to two of the three previous studies, learning was assessed by a range of three tasks 1 (instead of a single forced-choice task) in Experiment 1, which allowed the reliable evaluation of children’s comprehension and production abilities after very brief exposure. It further rendered the hypothesis that children learned the novel construction by mutual exclusivity in the forced-choice task incorrect. The ability to learn a novel construction was also revealed not to be limited to a certain language. Previous studies had exclusively examined English; Experiment 2 showed novel construction learning in native speakers of German as well. The novel reduplication construction differed from all novel constructions that had previously been used in one additional and important way. It was more genuinely novel. Goldberg’s studies had made use of novel word orders. While the combinations of form and meaning that were used were indeed novel, the fact that word order carries meaning in English was familiar to the English-speaking participants. The same is true for the novel prefix construction used in Experiments 1 and 2. Children did not know the particular newly-invented combination of form and meaning, but they knew that other prefix constructions carry meaning in their language. This 1 In Wonnacott et al. (2012) children were also tested on three tasks. In a production task the experimenter supported children by providing the respective verb. <?page no="208"?> 195 is to say that they were familiar with the underlying pattern in all these cases. This was not true for the novel reduplication construction. Since systematic reduplication is absent from German and English, the formmeaning pairing was more genuinely novel in this case. Regardless of this fact, children were able to learn this novel construction. This finding is the first to show children’s ability to learn a more genuinely novel construction. All these findings support the usage-based assumption that it is possible to learn constructions of different degrees of abstractness from the input. Novel reduplication learning is particularly interesting because the construction and the underlying pattern were completely absent from children’s native language. Goldberg’s studies additionally underpin that such learning does not involve innate structures or constraints 2 because they made use of forms that violate a proposed universal linking rule. The second simple noun phrase in Goldberg’s novel constructions expresses a location, which is in violation of the linking rule that a locative argument is to be expressed by an oblique complement (Goldberg 2006: 83; Naigles, Gleitman and Gleitman 1993; Pinker 1989: 75-82). Children’s ability to learn the constructions in spite of this violation suggests that the ability of mapping form and meaning does not require Universal Grammar (Goldberg 2006: 83). 8.2.2 Effects of token frequency, type frequency and type-token ratio at the constructional level Token frequency In Experiment 1 the number of tokens for each type was varied experimentally so as to influence the representational strength of types in children’s minds. Performance in terms of act-out, comprehension and production revealed effects of this manipulation in most cases: The higher the input token frequency the better children’s performance. A closer look at performance at different ages further revealed that young children’s early successful performance tended to be on types with high token frequencies. Moreover, a more detailed inspection of the verbs children remembered best based on the total number of verbs they memorized in the novel construction showed that high-frequency types had the tendency to be remembered best. 2 Generative theories also entail that children learn the idiosyncrasies of their language from the input. However, this position differs from the usage-based standpoint in the assumption that the highly-abstract rule system that generativists think is responsible for the formation of sentences cannot be learned from the input alone, which makes innate language-specific structures and constraints necessary for syntax learning in generative view. <?page no="209"?> 196 There was one result in the corpus study that further supports the effect of input token frequency on learning. For the German corpus, input and output token frequencies correlated for individual constructions, that is, constructions with a higher number of tokens in the input also had higher token frequencies in the output. 3 In this case, it was not the number tokens per type but the overall number of tokens of a construction that were related in input and output. These findings in novel and naturalistic construction learning extend and corroborate previous research on the effects of input token frequency. Earlier studies had shown effects on the speed of construction learning at levels other than that of derivational morphology (Ambridge et al. 2008; Hart 1991; Matthews et al. 2005; Matthews et al. 2007; Theakston 2004; Theakston et al. 2004) and on the number of errors in inflectional morphology learning, i.e., higher input token frequency was related to fewer errors (Marchman 1997; Maratsos 2000; Maslen et al. 2004). One novel construction learning study had additionally revealed that token frequency facilitated children’s comprehension of an abstract word order construction (Wonnacott et al. 2012). The relation between input and output frequencies had been shown primarily for the distributions of lexemes, particle verbs and word classes (Behrens 2003, 2006; Huttenlocher et al. 1991). In previous research as well as in the present studies, the effects of token frequency on performance were thought to come about through intermediary representations. In line with previous assumptions and the proposed steps of constructional learning, input token frequency is thought to strengthen the representation and entrenchment of the respective types in speakers’ minds, i.e., in their constructicon, 4 which in turn facilitates their future use and thus brings about the reported effects. Type frequency In the corpus study, the German-speaking child formed creative overgeneralizations of three to five of the seven most frequent constructions in terms of input and output type frequency. Moreover, input and output type frequencies were revealed to be correlated at the constructional level, i.e., constructions with higher type frequency in the input were also higher in type frequency in the output. In Experiment 1 generalizations of the novel construction to new cases were more difficult than uses of previously experienced types and usually 3 Descriptively, the same tendency was present for the English corpus. Correlations were, however, not calculated in this case, because the number of different constructions was too low. 4 As previously described, constructionist approaches assume language knowledge to be represented in constructional inventories. These are sometimes referred to as the constructicon in analogy to the lexicon (Fillmore, Lee-Goldman and Rhodes 2012). <?page no="210"?> 197 followed the storage and usage of the first familiar types. The number of generalizations children formed increased with the number of familiar types they used accurately and there were only few cases where children generalized before showing evidence of comprehending or producing any previously experienced exemplar correctly. These findings corroborate the relationship between type frequency and generalization and further suggest that input and output type frequencies are related. Previous research on overgeneralizations in inflectional morphology had shown that children tend to overgeneralize constructions or schemas with high type frequency (Bybee 1995; Bybee 2010: 95-96; Guillaume 1927/ 1973). For derivational morphology on nouns, studies had also revealed an influence of productivity on innovative uses (Clark and Cohen 1984; Clark and Hecht 1982). The present corpus study was the first to demonstrate the relation between input type frequency and creative uses in the same child for several derivational verb prefix constructions. The findings of Experiment 1 further extend previously reported type frequency effects on generalizations to novel morphological constructions, though here the number of stored types rather than input type frequency was determining. In all cases high type frequency, that is, a large constructional category with numerous exemplars promoted productivity. This effect might be related to the fact that high type frequency increased the representational strength and entrenchment of the constructional schema, perhaps because high type frequency leads to a high number of analogical comparisons in construction learning and in the categorization of new members. In line with the proposed steps of constructional learning, the combination of a strongly represented, deeply entrenched schema and a large constructional category might be particularly inviting for the formation of generalizations, i.e., the formation of additional analogical comparisons. Type-token ratio Experiment 2 revealed the beneficial effect of balanced and semi-skewed input distributions in contrast to skewed input on learning a novel partially-filled construction. This finding is the first to show effects of the input distribution for a novel partially-filled construction. The present result corroborates Matthews and Bannard’s proposal (2010) that balanced input or, in their words, high entropy, will be more facilitative of slot formation in a partially-filled construction than skewed input, because variability becomes apparent sooner. This apparent variability is particularly helpful because the lexically-filled portion of the construction already provides learners with an anchor and thus facilitates the alignment of types in the analogical comparison. In contrast, skewed input obscures variability of the slot position in this case and is, in fact, not necessary to guide children <?page no="211"?> 198 towards the accurate alignments because of the stable material in one or more other position(s). The most reliable stable position was the prefix. It is possible that the fact that it was embedded most of the time in a relatively stable [Subject] [Verb Phrase] ([Object]) construction, where the subject slot was filled by pronouns and the object position was either filled by something or absent, contributed to the relative stability of the construction. Regardless, the construction children learned was partially filled. More crucially, the meaning of the novel construction was ascribed to va[verb] rather than the word order. Additional variability in the subject and object positions or the use of the prefix construction in a number of different constructions might have made learning more difficult, but presumably so in all input conditions. The present finding is thus in support of the view that the ideal shape of the input distribution depends on the degree of abstractness of the construction in question. As had been previously shown, skewed input is conducive to novel abstract construction learning (Boyd and Goldberg 2011a; Casenhiser and Goldberg 2005). The presumed reason is that skewed input provides a recurring anchor in a situation where each position of the construction varies per definition in each instantiation and thus helps children align constructional types in the comparison. The increased obviousness of the variability in the emergent slot position might also be reason for the beneficial effect of semi-skewed input. Dissimilarity between successive types in training was higher for balanced and semi-skewed input than for skewed input, because in the latter case the same verb sometimes occurred twice in a row (making for highly similar successive types). Support for this view comes from a study of nonlinguistic categorization showing that infants were more successful at categorizing new members if training items were arranged such that consecutive exemplars were fairly dissimilar than when the same training items were arranged such that similar exemplars followed each other (Mather and Plunkett 2011). An alternative or additional reason why balanced and semi-skewed conditions were more conducive to learning the novel partially-filled construction than skewed input is that two or more types occurred fairly frequently. This fact might have facilitated the storage of at least two types, which was required for the analogical comparison that is necessary for constructional learning. Both potential explanations might have resulted in a higher number of comparisons or in earlier comparisons, which might in turn have facilitated or accelerated the abstraction of the variable slot. 8.2.3 Effects of type and token frequency at the pattern level It is crucial to keep in mind that frequencies at the pattern level are also type or token frequencies of constructions that are related to the one that is being learned or, more precisely, summed-up type or token frequencies of <?page no="212"?> 199 all the constructions that the pattern abstracts over. For the naturalistic prefix construction, pattern frequency was assessed using the sum over all real prefix construction types (e.g., un[base] verb , dis[base] verb , re[base] verb , over[base] verb ), the sum over all types instantiating each of these construction types (e.g., undo, untie, uncork, disagree, rearrange, rewrite) and the sum of tokens for each of these types in the examined corpora. The constructional type frequency informs about the range of variation of the pattern, the prefix verb type frequency gives information about the variability and diversification of the pattern at a lower level and the prefix verb token frequency reflects the pattern’s overall frequency of occurrence in a language. Input pattern frequency and output pattern frequency in naturalistic learning The corpus study revealed that input and output frequencies were related at the more general pattern level. Higher input pattern frequencies were reflected in higher output pattern frequencies in terms of summed-up counts. This finding extends previous research that had revealed the inputoutput relation for lexemes, particle verbs and word classes (Behrens 2003, 2006; Huttenlocher et al. 1991) 5 to derivational verb prefix constructions. Input pattern frequency and novel construction learning Experiment 1 revealed effects of the underlying pattern frequency on children’s learning of novel constructions. Specifically, learning a novel verb prefix construction was easier for German-speaking children than for their English-speaking peers. This was predicted based on results from the corpus study that had shown the frequency of the verb prefix pattern to be higher in the input and in the output in German than in English. Contrary to expectation, the advantage for German-speaking children was also present in the learning of the novel reduplication construction, which was not based on a pattern that was used in either German or English and thus had an underlying pattern frequency of zero in both languages. Finally, and in line with predictions, learning of the novel prefix construction was easier than learning of the novel reduplication construction, which potentially reflects the underlying pattern frequencies. One explanation for how the effect of underlying pattern frequency in novel prefix learning came about is that German children were faster to categorize the novel prefix construction as one example of the verb prefix 5 In Behrens’ studies on particle verbs and word classes, it was the input and output distributions that were compared. Huttenlocher and colleagues correlated the frequencies of pairs of words in the input and the output. In the present case, only raw frequencies were assessed. <?page no="213"?> 200 pattern. They were faster in recognizing it as ‘one of a kind’ because their pattern category was larger in terms of previously stored, familiar construction types, verb types and verb tokens. In particular the higher construction type frequencies were potentially beneficial to generalization (cf. 8.2.2; Bybee 1995; Bybee 2010: 95-96; Clark and Cohen 1984; Clark and Hecht 1982; Guillaume 1927/ 1973). This rather rich interpretation of children’s behaviour entails that they have schematized and stored the pattern level, which is evident in the categorization of the novel construction. There is, on the other hand, also a leaner interpretation of the same finding in terms of the attentional shaping. According to this view, pattern frequency affected children’s attention allocation. For German-speaking children, the higher frequency of the verb prefix pattern and perhaps that of the particle verb pattern as well contributed to this effect. For Englishspeaking children the same effect was much weaker because the frequency of the verb prefix pattern was considerably lower and the English particle pattern cannot have contributed to the effect because particles follow the verb in English. Due to the more frequent positioning of meaningful information at the beginning of verbs in German, it is possible that Germanspeaking children develop an anticipatory tendency to attend to this position, which facilitates the storage of linguistic information occurring in this position. Moreover, not only the frequency of the related particle verb pattern but also the separability potentially plays a facilitative role. The insights that particle and verb are separable and that the particle itself carries a clearly delimited meaning in German possibly invites children to separate prefix and base and look for a meaning of the prefix (Behrens 1998). As a consequence, German-speaking children noticed the novel prefix in the verb-initial position sooner and were more willing to attribute a separate meaning to it than their English-speaking peers. This course of action thus led to the German advantage without German-speaking children necessarily having formed and retained a schematization at the pattern level. Based on the present data, it is difficult to determine the exact cause of the pattern frequency effect, that is, which interpretation is to be favoured. Section 8.3.1 provides additional reflections on the schematizations humans make and store in the constructicon. The second explanation is also capable of accounting for the German advantage in learning the novel reduplication construction. It is possible that the tendency to attend to the beginnings of verbs paired with the willingness to ascribe a separate meaning to linguistic material in this position was responsible for the fact that German-speaking children were more successful than their English-speaking peers when learning the novel reduplication. This interpretation entails that the frequency of a related pattern affected the learning of the novel construction. This idea is not entirely unlikely because the verb prefix pattern and the reduplication construction <?page no="214"?> 201 (as well as the particle verb pattern) do share commonalities at a more abstract level, i.e., [morpheme 1][morpheme 2] ‘morpheme 1 modifies morpheme 2’. If the attentional shaping account holds, it is not necessary (though possible) for speakers to actually form and retain this abstraction. But the commonalities at this more abstract level illustrate the basis for why attentional preferences based on the verb prefix pattern should affect the learning of the novel reduplication construction. Future research can help determine exactly the way in which pattern frequency affects constructional learning. In other areas there is some evidence that the native language shapes children’s expectations about the distribution of information in a structure. For example, children were shown to understand the structure of root compounds their language used. English-speaking children expected modifier-head ordering from 2 years onwards, whereas Hebrew-speaking children expected head-modifier ordering at the same age (Berman and Clark 1989; Clark, Gelman and Lane 1985). It seems possible that frequency plays a role in shaping such expectations. If future research reveals that attentional expectations about the distribution of information are shaped by pattern frequencies, these aspects need to be integrated into the proposed steps of constructional learning. Attentional patterns presumably develop over time with increasing amounts of input. Consequently, constructional learning probably becomes increasingly affected by such patterns. It is expected that all steps are facilitated if a construction matches the attentional expectations of information distribution. The finding that the novel prefix construction was easier to learn than the novel reduplication construction could provide strong support to the assumption that the pattern level was in fact abstracted. Following this argument, the greater ease with which the novel prefix construction was learned goes back to the fact that children categorized the novel prefix construction ta[verb] or va[verb] as one novel example of the familiar verb prefix pattern [derivational prefix][base] verb , whereas they had no equally abstract pattern level construction in order to categorize the novel reduplication. 6 6 The reduplication construction might be categorized in terms of the more abstract [morpheme 1] [morpheme 1] ‘morpheme 1 modifies morpheme 2’ construction proposed earlier, if this is a construction that people abstract and represent mentally. This abstraction is not entirely unlikely, since the proposed construction for instance applies to determinative compounds (modifier-head). Speakers seem to use this abstraction in that case, i.e., they develop expectations as to which part modifies which part based on their experience (e.g., picture book) and rely on it in the interpretation of invented, new compounds (Clark, Gelman and Lane 1985). Regardless of this fact, this construction is considerably more general and more abstract than the verb prefix pattern and thus presumably more difficult to form and to make use of in categorization. <?page no="215"?> 202 There is, however, an entirely different potential explanation for the advantage of prefix over reduplication learning. The novel prefix and the novel reduplication constructions differed not only with respect to the frequency of the underlying pattern, but also with regard to the level of constructional abstractness and with it in similarity. While the novel prefix was a partially-filled novel construction, the reduplication was abstract. Relational similarity between different exemplars of the novel constructions was given in both cases, e.g., vadrink, vabuild, vawash va[verb] and dridrink, buibuild, wawash [ve][verb]. But only novel prefix learning was additionally supported by object similarity caused by the stable prefix that re-occurred in each instantiation and might thus have facilitated the alignment and mapping of exemplars in comparison. Facilitation of learning through object similarity has also been reported in the psychological literature on analogies (DeLoache 1990; Gentner and Markman 1997; Gentner and Medina 1998; Gentner and Toupin 1986; Holyoak and Thagard 1995: 83-84; Keane 1987). It remains open whether pattern frequency and similarity conspired to bring about the advantage of prefix learning or whether one of the reasons was solely responsible. As suggested in the discussion in Chapter 6 (and 8.4.3), future research with particular languages will be useful to tease these explanations apart. 8.2.4 Summary The three studies revealed effects of frequencies at constructional and pattern levels. The effects of token frequency on learning were extended to novel derivational morphological constructions and assumed to come into effect via the strengthening of representations in the mind. Type frequency was shown to affect productivity and thus the tendency to form (over)generalizations of both naturalistic and novel constructions in derivational morphology. It might additionally be possible that construction type frequency benefitted generalizations of the more general pattern. A balanced input distribution was first revealed to be conducive to the learning of a partially-filled novel construction. Input token and type frequencies were shown to be related to output frequencies for derivational verb prefix constructions. Input and output frequencies were also related when summed up over all construction types, verb types or verb tokens. Pattern frequency further affected novel construction learning. Higher underlying pattern frequencies facilitated the learning of a novel construction that exemplified the pattern; a novel construction based on an unfamiliar pattern was more difficult to learn. These effects might go back to the fact that pattern frequency affects children’s ability to categorize novel constructions. An alternative explanation holds that pattern frequency shapes children’s attention to certain preferred distributions of information structure. Such attentional patterns might influence constructional learning, e.g., by <?page no="216"?> 203 facilitating storage of morphemes in the attentional focus. Constructions in line with the attentional patterns that are formed on the basis of patterns with high frequency would consequently be easier to learn. This idea is able to account for the advantage for German-speaking children in reduplication learning, despite the fact that neither German nor English makes use of a reduplication pattern. 8.3 Broader theoretical implications Three more general theoretical issues that have been touched upon are examined in the following. The difficulty to decide between the rich and the lean interpretation of pattern frequency effects can be traced back to the first more general issue. In the present case it concerns the question whether the pattern level was schematized and stored in children’s mind. More generally, it questions which abstractions humans make and store. This matter is important in any constructionist account of language and language learning and therefore discussed in section 8.3.1. The finding that balanced input was more conducive to learning a partially-filled novel construction, whereas previous research showed skewed input to facilitate learning of an abstract novel construction leads to a second more general issue. In both cases anchoring by a fixed or recurring element is thought to facilitate learning. The idea of supporting object similarity has also been suggested in the psychological literature on analogies (DeLoache 1990; Gentner and Markman 1997; Gentner and Medina 1998; Gentner and Toupin 1986; Holyoak and Thagard 1995: 83-84; Keane 1987). The proposal that the processes involved in non-linguistic and linguistic processing are the same and therefore affected by the same factors, e.g., supporting object similarity, is discussed in section 8.3.2 about the potential domain-generality of the processes in language learning. The third issue surfaced indirectly a few times in this book. It was noted at different points that frequency is not the only factor influencing constructional learning. Additional factors including recency, similarity or context were proposed. In section 8.3.3 the question of whether all these factors work their effects by influencing the same more general factor of salience is discussed. 8.3.1 Abstractions in language The proposition that humans must form abstractions in order to use language in new, creative ways is undisputed in the usage-based, constructionist approach to language and language learning. The issue of precisely which abstractions are formed and stored enduringly in the mind is much less clear. <?page no="217"?> 204 It seems that this question is rarely discussed explicitly. Many authors focus on describing the representation of language in networks and the interrelations between constructions of different degrees of abstractness (e.g., inheritance hierarchies; cf. Croft 2001: 14-28, 53-56; Croft and Cruse 2004: 257-265; Goldberg 1995: 4-7, 74-81, 2006: 5-9; Kay and Fillmore 1999), but only a few state explicitly which structures they expect to be present in speakers’ minds. Two opposing positions are held by Kay and Croft. Kay (1997: 129) points out that “[n]o claim is intended that the internal representation in the mind of each speaker contains every generalization inherent in the data”. In stark contrast, Croft (2001: 3) asserts that all structures that are described in his Radical Construction Grammar have psychological reality. In language learning, general claims as to the presence or absence of representations of certain abstractions in children’s minds are rare as well. Instead, it is attempted to produce evidence as to the psychological reality of particular abstractions. That the production of such evidence necessarily entails a certain understanding or definition of the abstractions that are assumed to be formed is often left unacknowledged. Nevertheless, in Tomasello’s reflections, one particular criterion for abstractions becomes evident, even though he does not mention it explicitly. It is children’s understanding or production of new examples of a construction (2003: 314- 320), i.e., their generalizations of a construction, that is considered to mirror the abstraction of a constructional schema. Tomasello (2003: 316) proposes that structural priming in children is capable of revealing emergent abstractions whose representations are still weak. Structural priming shows that people are sensitive to previously experienced utterances in their input because they are more likely to use the same structures in their own subsequent speech (Bock 1986; see Pickering and Ferreira 2008 for a review). Six-year-olds were revealed to be primed by previously heard syntax and consequently more likely to produce sentence structures they had just heard previously (Savage, Lieven, Theakston and Tomasello 2003). These findings were first extended to 4and 5-year-olds in a sentence-repetition and picture-description task (Huttenlocher, Vasilyeva and Shimpi 2004) and later even to 3-year-olds (Shimpi, Gámez, Huttenlocher and Vasilyeva 2007). The constructions 3year-old children could be primed on were however more limited, reflecting the lower number of abstractions they had formed, and priming was stronger for older children. Apart from priming, the preferential looking paradigm has been suggested to be capable of revealing emergent weakly represented abstractions (Tomasello 2003: 316-317). More strongly represented abstractions can be shown in more demanding tasks. In such tasks, children’s abilities to generalize are assessed in comprehension or production with respect to particular constructions. For instance, children’s ability <?page no="218"?> 205 to generalize a familiar word order or a novel word order, or their ability to correct a weird word order, is assessed and taken as evidence for the abstraction of the respective argument structure construction (e.g., Akhtar 1999; Brooks et al. 1999; Casenhiser and Goldberg 2005; Matthews et al. 2005, 2007). The position that generalizations entail the previous abstraction of the respective constructional schema was also followed in this work (cf. 3.4.2). It was assumed that a schema is stored together with its exemplars as a constructional category, following analogical comparisons and schema abstraction. If the commonalities between exemplars were not stored in the form of a schema but considered a fleeting phenomenon of on-the-spot comparisons of exemplars, the category would lose its coherence and could not be perceived as such any more - the categoriness of the category would disappear (Ross and Makin 1999: 215). Even if only exemplars were ever used as standards of comparisons in generalizations of a construction, the schema would still be retained (and possibly also adjusted) because it holds the category together. According to the proposed conception humans can be credited with the abstract representation of a construction as soon as they categorize a new utterance as ‘one (example) of’ a familiar constructional category. Generalizations of a construction can thus be taken as evidence for the previous abstraction and representational presence of the respective constructional schema. Precisely how abstract this schema is is still difficult to determine. Based on the principle that “[m]ore specific knowledge takes precedence over more general knowledge” (Lakoff 1987: 147 about Wilensky’s Law, Wilensky 1983: 25, 145) the minimal assumption entails the existence of the least abstract schema that represents the generalization made. This can be illustrated by an example from the corpus study: Leo’s overgeneralization befärben be+’colour’ (‘colour with’) was taken to imply that he must have abstracted the constructional schema be 1 [base] verb ‘apply something; equip/ cover something with something’ rather than the more abstract [derivational prefix][base] verb ‘encode a contrast in action; encode a contrast to the normal, expected action’ or the even more abstract [morpheme 1][morpheme 2] ‘morpheme 1 modifies morpheme 2’. Of course, the fact that a speaker generalizes a construction once does not entail that the respective abstraction is stored for all time. If constructional schemas are never reinforced by the input, they are likely to be lost, as are members of a constructional category that never experience reinforcement (Bybee 2010: 66-71). 8.3.2 The domain-generality of cognitive processes in language learning The usage-based model holds that the cognitive processes involved in language learning are not specialized in this linguistic task, but of more a gen- <?page no="219"?> 206 eral nature. If this is true, the processes should be affected by the same factors in language learning and other non-linguistic cognitive tasks. This issue is examined for analogies in the following. Analogies are based on relational similarity between two situations. Relational similarity allows the successful alignment and mapping of the compared situations. Early in development children tend to attend to object similarity more than to relational similarity, which can result in nonanalogical matches in cases where object and relational similarity contradict. Over time, children develop a preference for relational similarity (relational shift), presumably because it entails greater depth and systematicity of the mappings that are formed (e.g., Gentner 1988, 2003; Gentner and Rattermann 1991; Kotovsky and Gentner 1996; Markman and Gentner 1993; Rattermann and Gentner 1998). Even after the relational shift, analogy formation is facilitated by additional object similarity that supports relational mappings (DeLoache 1990; Gentner and Markman 1997; Gentner and Medina 1998; Gentner and Toupin 1986; Holyoak and Thagard 1995: 83-84; Keane 1987). In the present work analogies were proposed to be part of constructional learning. Children were thought to form analogies over types of an emergent construction by aligning and mapping these types based on relational similarity. Children’s early linguistic analogies are often supported by additional object similarity, which can be achieved linguistically by repetitions of the same material. Eventually, children are able to form analogical comparisons based on relational similarity alone. This progression is mirrored in children’s early formation of concrete, item-based constructions and their emerging ability to form increasingly abstract constructions (Lieven et al. 2003; Lieven at al. 2009; Tomasello 2003: 6-7). The support of relational similarity by additional object similarity also became evident in Experiment 2 and the related reflections. In partiallyfilled constructions, such as the one explored in this study, supportive object similarity is automatically given by the lexically-filled position or positions. Consequently additional skewing of the input is not necessary for the structural alignment of examples and in fact hinders the speedy recognition of the variable position. In abstract constructions, on the other hand, only relational similarity is given. In this case, input skewing provides additional object similarity and thus facilitates the relational alignment of examples in the analogical comparison (Casenhiser and Goldberg 2005; Childers and Tomasello 2001). This argument receives additional support from a follow-up study to Experiment 2 (Cordes et al. 2013). A novel abstract word order construction with the meaning ‘pretence’ was taught to children. Due to the more abstract character of the construction, i.e., the meaning was ascribed to the variable word order rather than the <?page no="220"?> 207 stable prefix, balanced input that provided less object similarity in the form of recurring lexical material was less conducive to learning. The finding that linguistic analogies are affected by the same factors as non-linguistic ones (e.g., supporting object similarity) provides strong support for the likeness of linguistic and non-linguistic analogies and thus for the domain-generality of (at least one of) the processes involved in language learning. Even if the suggestion that derivational morphology is learned from the input using a domain-general mechanism could theoretically be integrated in those generative accounts that confine derivational morphology and potential rules of word formation to the lexicon (cf. lexicalist position, e.g., Aronoff 1976: 1-3), the findings proposing the same process to be at work in the learning of abstract word order constructions cannot be part of such an account. Such abstract constructions belong to the realm of syntax and their acquisition is consequently thought to involve innate structures in the generative account because of their complexity. However, the result that the cognitive process involved in morphological construction learning, word order construction learning and non-linguistic learning is influenced in the same manner by (supporting) object and relational similarity provides strong support for the proposal that it is in fact the same domain-general process of analogy formation that is at work in all these cases. This line of arguments thus supports the usage-based proposal that all language structures can be learned from the input using domaingeneral cognitive processes like analogy formation, probably with the help of humans’ social-cultural abilities. 8.3.3 The role of salience It is not frequency alone that has been proposed to facilitate constructional learning. A number of additional factors have been brought forward as well. Object similarity facilitates analogical comparisons (DeLoache 1990; Gentner and Markman 1997; Gentner and Medina 1998; Gentner and Toupin 1986; Holyoak and Thagard 1995: 83-84; Keane 1987). Linguistic labels and relational language highlight commonalities between objects, invite children to look for such commonalities or preserve similar relations linguistically, thus benefitting categorization or analogy formation (Gentner 2003; Gentner and Loewenstein 2002; Gentner and Rattermann 1991; Loewenstein and Gentner 1998, 2005; Kemler Nelson 1995; Sloutsky and Fisher 2004; Waxmann 1999; Waxman and Booth 2003; Waxman and Markow 1995). The recency effect facilitates the recall of most recently encountered information (Ebbinghaus 1913: 90-113). Semantic transparency benefits the acquisition of meanings, whereas opacity makes it more difficult (Gathercole and Hoff 2009; Kline and Demuth 2010; Narasimhan and Gullberg 2011). And additional context factors such as intonation facilitate <?page no="221"?> 208 learning, because they make certain characteristics of the situation stand out (Peters 1983: 19, 26). In spite of many differences, the proposed factors reveal a commonality: They all highlight certain information or make particular aspects stand out and easy to notice. They make something particularly salient. In cognitive linguistics two types of salience are distinguished by Schmid (2007: 119- 120). So-called cognitive salience refers to the activation of concepts in the mind during processing. It is achieved either by conscious, selective attention or by spreading activation from related concepts (Deane 1992: 34-35). Ontological salience, on the other hand, is understood as the relatively stable property of entities in the world to attract attention, e.g., if there is a dog in a field, the dog is usually more salient than the field (Schmid 2007: 120; see also figure-ground distinction in cognitive linguistics following Rubin 1921: 3). A slightly different definition of so-called situational salience highlights the fact that what is salient is context-dependent, “Salience defines the degree of relative prominence of a unit of information, at a specific point in time, in comparison to other units of information” (Chiarcos, Claus and Grabsi 2011: 2). An example is a man in a suit who is pushing a toddler in a stroller. In a business lunch situation the child in the stroller is salient, in a nursery school picnic situation the man’s formal clothing is salient (Smith and Mackie 2000: 66-67). I understand salience as a situational phenomenon (as proposed by Chiarcos et al.) that is related to cognitive and ontological salience. Based on this understanding, the factors of similarity, linguistic labels, relational language, transparency and context make certain aspects of a situation salient, which in turn affect cognitive salience and with it processing. Recency presumably influences cognitive salience more immediately, i.e., through residual activation. The relation between frequency and salience is two-fold. Repetitions over a short interval of time increase situational salience and cognitive salience in the short term. Repetitions in general also affect cognitive salience in the long term because they increase representational strength and entrenchment, thus lowering the levels necessary for activation more enduringly. Human language learning is thus influenced by salience. 7 The ways of heightening the salience of a concept, a linguistic form, a relation or else are manifold. Frequency seems to be one way of inducing salience. The fact that so numerous and diverse factors can be captured by salience suggests that salience is an important intermediary factor, which affects human processing. 7 Additional support comes from the use of the term salience in first volume of Slobin’s Crosslinguistic study of language acquisition (1985). Thirteen authors describing the acquisition of ten different languages invoke that concept of salience more than 20 times in order to account for early or speedy acquisition of certain structures. <?page no="222"?> 209 8.4 Open issues and future directions Finally, two issues require additional attention. The first one deals with the categorization of the constructions used in the present work as derivational rather than inflectional constructions. The second one concerns several definitions and classifications made in the framework of the corpus analysis. After the discussion of these two issues, future directions for research are proposed. 8.4.1 Inflectional versus derivational morphological constructions In the corpus study, the frequencies of derivational verb prefix constructions in English and German were compared. In the experiments, a novel verb prefix construction and a novel verb-initial reduplication were used. These constructions were categorized as derivational. The present section serves to underpin this categorization and discusses the potential alternative, the classification as inflectional. Traditionally, derivational morphology deals with the formation of new lexemes from a free lexical and a bound lexical morpheme. Inflectional morphology, in contrast, concerns the formation of new word forms of a lexeme by adding a bound grammatical morpheme to a free lexical one. An obvious difficulty is how to determine the types of morphemes involved or whether the resulting form is a new lexeme or only a new word form. This situation is referred to by Anderson (1982) as circular and has prompted numerous lists of criteria to distinguish inflection and derivation, which are often thought of as points on a continuum rather than a dichotomy (Bybee 1985: 5; Haspelmath and Sims 2010: 89-90, 98-105). The following criteria are frequently proposed (Anderson 1982; Bloomfield 1933: 223; Bybee 1985: 5, 99; Haspelmath and Sims 2010: 90): A higher relevance to syntax is attributed to inflection than to derivation (Anderson 1982; Bybee 1995: 99; Haspelmath and Sims 2010: 90). The obligatoriness of expression is associated with inflection but not derivation (Haspelmath and Sims 2010: 90). The applicability of inflection is considered unlimited compared to possibly limited applicability of derivation (Bybee 1985: 99; Haspelmath and Sims 2010: 90). Inflection is thought to leave the concept of the base intact, whereas derivation might result in a new concept (Bybee 1985: 5; Haspelmath and Sims 2010: 90). Inflectional meanings tend to be more abstract and compositional in comparison to the more concrete and possibly non-compositional meanings of derivational morphemes (Bybee 1985: 99 on transparency; Haspelmath and Sims 2010: 90). Derivations are usually expressed closer to the base than inflections (Bloomfield 1933: 223; Haspelmath and Sims 2010: 90). There is more base allomorphy in derivation (Haspelmath and Sims 2010: 90). Word-class changes are more common in inflection than in derivation (Haspelmath and Sims 2010: 90). <?page no="223"?> 210 Applying the criteria to the novel prefix {va- } / {ta-} with its meaning of ‘pretence’ provides the following picture. The prefix’s relevance to syntax is low, its expression is not obligatory (at least it is not for the invented forms in German or English - it is theoretically conceivable that there is a language where pretence is obligatorily marked) and its applicability is not unlimited (e.g., pretend breathing is impossible). These three points strongly support the derivational character of the novel prefix. For reasons related to the generalizability of the prefix in the experiments, it was necessary that it was compositional. While the criteria suggest that inflectional morphemes are more likely to be compositional, they do not exclude the possibility that derivational morphemes are compositional. In fact, all derivational prefixes that were counted in the corpus study had compositional character, e.g., the prefixes had clearly delimited meanings. The compositionality of the novel prefix does consequently not contradict its classification as derivational. The degree of abstractness of the prefix meaning is difficult to determine. To be sure, the ‘pretence’ meaning fits in nicely with a group of meanings that are expressed by other derivational prefixes in English or German: ‘deviations from the norm/ expectation with respect to different dimensions’, e.g., excessive size, unintentional manner, extreme attitudes (cf. 6.2.2), because it describes a deviation from the norm in terms of manner. The derivational character of the novel prefix is thus underlined by its meaning. The criterion of base allomorphy was said to be more common in derivation, but since it is not a prerequisite, its absence does not preclude a derivational categorization. Moreover, it is neither present in other derivational prefixes that use verbal bases. The position of the novel prefix relative to inflectional prefixes cannot be determined because of the absence of any inflectional prefixes in German and English. There is, however, one form that might be of interest in this respect: the German past participle. German verbs with derivational prefixes form the past participle without the gepart of the circumfix (e.g., er hat die Schachtel bemalt/ *gebemalt - ‘he has painted (on) the box’). This formation also sounds more natural than the alternative with gefor the novel prefix, e.g., er hat tabaut/ ? getabaut - ‘he has pretended to build’, again supporting a derivational classification. The fact that there is no word class change for the novel prefix was also experimentally given. In fact, in the production task, a few children creatively added the novel prefix to nouns, which is possible for real derivational prefixes, e.g., unhook, 8 suggesting that children considered word class changes that are not unlikely for derivations possible for the novel prefix construction. To sum up, there are numerous criteria that are only in line with a derivational interpretation. Several more criteria favour a derivational classifi- 8 Interestingly, many of the bases for English un[base] verb were originally nouns that are also used as verbs (conversion), e.g., undress, unzip. <?page no="224"?> 211 cation, in particular when other derivational prefixes are used as points of comparison. There is not a single criterion that clearly indicates the inflectional character of the novel prefix or contradicts its proposed derivational status. It is therefore concluded that the derivational interpretation was in fact appropriate. The derivational classification of the prefix is generalized to the novel reduplication (which could in theory also be derivational or inflectional), since it carried the same meaning and occurred in the same position as the prefix. 8.4.2 Issues regarding the corpus analysis There are three important issues regarding the manual corpus counts in Chapter 5 that require additional attention. The first one concerns the exclusion of prefix verbs from the counts and the second one bears on the attribution of the prefix construction meaning. The third issue concerns the assignment of prefix verbs to prefix construction types. In all three cases a certain degree of subjectivity was unavoidable, which is why the issues are discussed in the following. The selection of the prefix constructions, which were included in the manual pattern frequency counts in the corpus study (Chapter 5), followed numerous criteria. Some of these criteria ensured that the definition of the prefix constructions participating in the corpus counts was shared by the newly invented prefix construction used in the experiments. This was necessary because predictions about the learning of the novel prefix construction were based on the corpus counts. The transparency of the prefix construction meaning and the relatively high degree of compositionality are two examples of the criteria used. They were necessary because the novel prefix construction had to be ‘generalizable’, which would have been impossible if the meaning had been opaque and non-compositional. Nevertheless, these two criteria are not generally required of prefix verbs. Their application thus led to the exclusion of a number of potential prefix verbs in German (and very few in English), many of which are classified as prefix verbs in grammar books or dictionaries, e.g., verstehen ‘understand’. They may consequently be part of the category of prefix verbs in people’s minds. 9 It is thus possible that these types and their frequencies also play a role in novel prefix construction learning. Due to their dissimilarity from the novel prefix construction they were nevertheless excluded from corpus counts as a precautionary measure. These exclusions were deemed acceptable because they were much more frequent for German (they were mostly cases of high-frequency prefix verbs that had lost their transparency; cf. Bybee 2010: 46 for the relation between high frequency and the loss of 9 Their non-compositionality and opacity might however make them more peripheral members of this category. <?page no="225"?> 212 transparency), whereas there were hardly any for English. Consequently, the exclusions only led to a more conservative test of the hypotheses. The second issue concerns the attribution of meaning to the prefix constructions included in the counts. As listed in Appendix I.3, Tables a and b, these meanings were based on Quirk and colleagues (1984) and Marchand (1960) in English and Kühnhold and Wellmann (1973) and Barz (2006) in German. In English, one prefix always corresponded to one meaning, so that there was one prefix construction per prefix. In German, about half the prefixes carried more than one meaning and thus yielded several prefix constructions (form-meaning pairs). The exact number of constructions formed was necessarily subjective. The literature was used as a guideline together with the principle that very fine-grained distinctions were avoided, since it was considered unlikely that children would make them. On the other hand, children are said to develop more abstract constructional knowledge based on more local earlier generalizations, so that early distinctions might in fact be more fine-grained than later ones. One case where it was not clear whether different constructions should be assumed was German ver-. To one verconstruction (ver 3 -) the meaning ‘action or event resulting in a change of state’ was attributed. Because of the considerable differences between the verbs within this general category, three sub-categories were formed: ver 3a - ‘causative, to make or become (more) [adjective]’, e.g., verkürzen ‘shorten’, ver 3b - ‘completion of an action; action or event resulting in different, potentially more obscure state’, e.g., verhüllen ‘disguise’, verzaubern ‘bewitch’ and ver 3c - ‘action or event describing the relation between at least two components’, e.g., verknoten ‘knot together’. Due to the subjectivity that was necessarily involved in the definitions of the individual prefix constructions, it was considered to count prefixes rather than prefix constructions. This procedure would doubtlessly have resulted in lower numbers for Leo and his caregiver as illustrated in Table 16. Table 16. Verb prefix constructions and verb prefixes selected manually. Speaker Total verb prefix constructions Total verb prefixes Thomas 2 2 Caregiver Thomas 3 3 Leo 22 9 Caregiver Leo 26 11 The reason why these more conservative and more objective counts were not preferred is that they did not correspond to the theoretical assumptions that were made in this work. The constructional pattern level [derivational <?page no="226"?> 213 prefix][base] verb was abstracted from several more concrete prefix construction types, i.e., form-meaning pairs with a prefix as the stable part and an open slot in the base position such as un[base] verb and re[base] verb . The pattern level was not based on an abstraction over several prefixes that were not associated with specific meanings. The reason was that the frequencies of similar prefix constructions were thought to affect the learning of a novel prefix construction rather than the frequencies of prefixes with one or more meanings. This is why the more subjective categorization of prefix verbs into prefix constructions was implemented. Crucially, only the number of prefix construction types was affected, the overall counts of prefix types and prefix tokens remained the same. Moreover, even if prefixes rather than prefix constructions had been counted, all analyses would have remained significant (caregivers’ speech: Thomas’ caregiver vs Leo’s caregiver 2 (1) = 4.571, p < .033; children’s speech: Thomas vs Leo 2 (1) = 4.455, p < .035). Finally, the assignment of prefix verbs to prefix constructions was necessarily subjective, since there was no source available where all verbs that appeared the German corpus were classified with respect to the categories used. Inevitably some choices were more unambiguous than others. This problematic situation was ameliorated by the fact that an unbiased judge validated the assignments and conferred with the author on controversial cases until a classification was agreed upon. The classification of prefix verbs to prefix construction types did not affect the overall counts at the pattern level (cf. Table 8 in 5.2.2). Different classifications of individual verbs would only have affected the counts of verb types and tokens per construction (Appendix I.4, Table f). It is, however, unlikely that they would have changed the correlation between input and output verb types or verb tokens noticeably. 8.4.3 Future research directions In the discussion of the results and their more general implications several aspects surfaced that call for further research. Potential future directions are proposed in the following. The shape of the input distribution in terms of tokens per type was examined in Experiment 2. The results provided first implications that the ideal input distribution varies for different levels of constructional abstractness. This assumption awaits additional empirical support from studies using novel constructions of different degrees of abstractness within the same paradigm and from studies of naturalistic language learning. These studies may also explore the role of lexically-stable material in different positions of a more general construction that a particular novel construction is integrated in, e.g., the variability in subject or object positions when the novel construction is verbal. Future research might further shed light <?page no="227"?> 214 on the role of semi-skewed input on construction learning. At the moment, it is not entirely resolved how the effect of semi-skewed input in partiallyfilled construction learning came about. In particular, the question why skewing towards two types was more beneficial than skewing towards a single type in this case calls for further investigation. So far, a maximum of three degrees of skewing has been tested. Assessing more fine-grained degrees of skewing can also provide new insights into the role of the input distribution in learning. Another interesting area for future research is the nature of the abstractions children and adults make. This issue is generally not very wellresearched and studies exploring abstractions at different levels are necessary to provide a fuller picture of the constructional networks in people’s minds. In the present work it could be shown that children formed abstractions for the novel constructions they learned. Future studies of abstractions at more general levels might address the question of whether pattern level abstractions are formed as well. One more concrete issue invites further investigation. It is the question of whether pattern frequency or similarity was responsible for the differences between novel prefix and novel reduplication learning. As suggested in 6.4.3 teaching novel prefix and novel reduplication constructions to children whose language has no prefixes but uses reduplication or to children whose language uses neither of the two processes or both of them would make it possible to disentangle these two factors. Of course, any such investigation into novel construction learning would have to be preceded by an assessment of underlying pattern frequencies in the respective language in order to allow the formation of predictions for novel construction learning and to identify appropriate forms for the novel constructions. Finally, the pattern level itself invites future research. The pattern level has been revealed to be a useful more abstract level for the description of commonalities of constructions that are by definition partially filled. For such constructions the pattern level constitutes the first entirely abstract level of description. Frequency effects at the pattern level show that the frequencies of closely-related constructions (i.e., constructions that share commonalities with the construction that is being learnt at the more abstract pattern level) are relevant to the learning of the construction in question. This finding extends previous research where frequencies of more loosely-related constructions were revealed to affect learning (e.g., Abbot- Smith and Behrens 2006; Dressler 1997; Huttenlocher et al. 2002). Future studies may explore effects of pattern level frequencies on the learning of inflectional morphological constructions, other partially-filled constructions and perhaps also abstract constructions. The determination of construction types is the most unambiguous for constructions with one empty slot and one stable position at the construction level. In this case, the pat- <?page no="228"?> 215 tern level is entirely abstract, but construction types can be unambiguously described at the constructional level (abstract pattern level: [derivational prefix][base] verb ; partially-filled construction level un[base] verb 10 ; unambiguous determination of construction types possible). As soon as there are several empty and fixed positions, adjustments might become necessary when construction types are being determined. This is particularly true for constructions that are already entirely abstract by definition and where the pattern level would summarize their commonalities at a more abstract level. Future studies would have to provide suggestions as to how construction types are determined in this case, so that frequency counts at the construction type level are possible. Counts of token frequency at the pattern level are unaffected by this issue and can easily be formed and assessed even for abstract constructions that are summarized at a more abstract pattern level. Pattern frequency effects are not unlikely for more abstract constructions as studies with even more general structures suggest (Barnes et al. 1983; Hoff-Ginsberg 1998; Huttenlocher et al. 2002) and they potentially provide new insights into the role of frequency influencing constructional learning processes. 8.5 Conclusion This book contributes to the mounting evidence of frequency effects in children’s language learning. New evidence was obtained by using a methodological approach that combined corpus and experimental methods in order to achieve a balance between validity and control. The studies examined children’s learning of naturalistic and novel constructions with respect to frequencies at the constructional and the newly-introduced pattern levels. Derivational morphological constructions were investigated in two languages. Findings revealed first evidence that novel construction learning is not limited to specific languages. Previous research on novel argument structure construction learning was extended to morphological constructions, more specifically to a partially-filled prefix construction and an abstract reduplication construction. Frequency at the constructional level and the more abstract pattern level were shown to influence both naturalistic and novel construction learning. More precisely, input token frequency affected construction learning presumably by strengthening the representation of types, input type frequency was connected to productivity, the shape of the ideal input distribution depended on the level of constructional abstractness, input and output frequencies were related at con- 10 As previously stated, for prefix constructions that carry more than one meaning, several constructions (form-meaning pairs) can be postulated (cf. Chapter 5 and Appendix I.3 Table b, e.g., German ver[base] verb constructions). <?page no="229"?> 216 structional and pattern levels in natural language, and the frequency of the underlying pattern influenced novel constructional learning. These findings extend previous insights into the abstraction and generalization processes in construction learning and how they are affected by input frequencies of various kinds. The results provide support to the usage-based view of language learning by underlining the important role of input and language experience and by corroborating the idea that domaingeneral processes, in particular analogy formation, play a vital role in construction learning. <?page no="230"?> 217 References Abbot-Smith, Kirsten and Heike Behrens. 2006. “How known constructions influence the acquisition of other constructions: The German passive and future constructions”. Cognitive Science 30. 995-1026. Abbot-Smith, Kirsten and Michael Tomasello. 2006. “Exemplar-learning and schematization in a usage-based account of syntactic acquisition”. The Linguistic Review 23. 275-290. Aitchison, Jean. 2003. Words in the mind. An introduction to the mental lexicon. Oxford: Blackwell Publishing. Akhtar, Nameera. 1999. “Acquiring basic word order: Evidence for data-driven learning of syntactic structure”. Journal of Child Language 26. 339-356. Akhtar, Nameera. 2004. “Nativist versus constructivist goals in studying child language”. Journal of Child Language 31. 459-462. Akhtar, Nameera and Michael Tomasello 1997. “Young children’s productivity with word order and verb morphology”. Developmental Psychology 33. 952-965. Aksu-Koç, Ayhan A. and Dan I. Slobin. 1985. “The acquisition of Turkish”. The crosslinguistic study of language acquisition - Volume 1. Ed. Dan I. Slobin. Hillsdale: Lawrence Erlbaum. 839-878. Alishahi, Afra and Suzanne Stevenson. 2005. “A probabilistic model of early argument structure acquisition”. Proceedings of the Cognitive Science Society. 97-102. Ambridge, Ben and Elena V. Lieven. 2011. Child language acquisition. Cambridge: Cambridge University Press. Ambridge, Ben, Julian M. Pine, Caroline F. Rowland and Chris R. Young. 2008. “The effect of verb semantic class and verb frequency (entrenchment) on children’s and adults’ graded judgements of argument-structure overgeneralization errors”. Cognition 106. 87-129. Ambridge, Ben and Caroline F. Rowland. 2009. “Predicting children’s errors with negative questions: Testing a schema-combination account”. Cognitive Linguistics 20. 225-266. Anderson, John R. and Gordon H. Bower. 1972. “Recognition and retrieval processes in free recall”. Psychological Review 79. 97-123. Anderson, Lloyd. 1982. “The ‘perfect’ as a universal and as a language particular category”. Tense-aspect: Between semantics and pragmatics. Ed. Paul Hopper. Amsterdam: John Benjamnis. 227-264. Arnon, Inbal. 2010. Starting Big—The role of multi-word phrases in language learning and use. Doctoral dissertation, Stanford University. <http: / / purl.stanford.edu/ zs652gt8462>. Accessed 13.12.2011. Arnon, Inbal and Neal Snider. 2010. “More than words: Frequency effects for multiword phrases”. Journal of Memory and Language 62. 67-82. Aronoff, Mark. 1976. Word formation in generative grammar. Cambridge: MIT Press. Attneave, Fred. 1953. “Psychological probability as a function of experienced frequency”. Journal of Experimental Psychology 46. 81-86. Baayen, Rolf H., Doug J. Davidson and Douglas M. Bates. 2008. “Mixed-effects modeling with crossed random effects for subjects and items: Special Issue: Emerging Data Analysis”. Journal of Memory and Language 59. 390-412. <?page no="231"?> 218 Barnes, Sally, Mary Gutfreund, David Satterly and Gordon Wells. 1983. “Characteristics of adult speech which predict children’s language development”. Journal of Child Language 10. 65-84. Barz, Irmhild. 2006. “Die Wortbildung”. Duden - Die Grammatik. Eds. Dudenredaktion. Mannheim: Dudenverlag. 641-772. Bates, Elizabeth, Inge Bretherton and Lynn Snyder. 1988. From first words to grammar: Individual differences and dissociable mechanisms. New York: Cambridge University Press. Behrens, Heike. 1998. “How difficult are complex verbs? Evidence from German, Dutch and English”. Linguistics 36. 679-712. Behrens, Heike. 2003. “Verbal prefixation in German child and adult language”. Acta Linguistica Hungarica 50. 41-59. Behrens, Heike. 2006. “The input-output relationship in first language acquisition”. Language and Cognitive Processes. 2-24. Behrens, Heike. 2009. “Usage-based and emergentist approaches to language acquisition”. Linguistics 47. 383-411. Berlin, Brent and Paul Kay. 1969. Basic color terms: Their universality and evolution. Berkeley: University of California Press. Berman, Ruth A. and Eve V. Clark. 1989. “Learning to use compounds from contrast: Data from Hebrew”. First Language 9. 247-270. Bhatia, Tej K. 2000. Punjabi - Descriptive Grammars. London: Routledge. Bloomfield, Leonard. 1933. Language. New York: Holt, Rinehart and Winston. Bock, Kathryn. 1986. “Syntactic persistence in language production”. Cognitive Psychology 18. 355-387. Bohr, Niels. 1913a. “On the constitution of atoms and molecules. Part I”. Philosophical Magazine 26. 1-25. Bohr, Niels. 1913b. “On the constitution of atoms and molecules. Part II. Systems containing only a single nucleus”. Philosophical Magazine 26. 476-502. Bohr, Niels. 1913c. “On the constitution of atoms and molecules. Part III. Systems containing several nuclei”. Philosophical Magazine 26. 857-875. Borges, Jorge-Luis. 1964. Labyrinths. London: New Directions Publishing Corporation. Bowerman, Melissa. 1978. “Systematizing semantic knowledge: changes over time in the child’s organization of meaning”. Child Development 49. 977-987. Boyd, Jeremy K. and Adele E. Goldberg. 2011a. “Young children fail to fully generalize a novel argument structure construction when exposed to the same input as older learners”. Journal of Child Language 22. 1-25. Boyd, Jeremy K. and Adele E. Goldberg. 2011b. “Learning what not to say: Categorization and statistical preemption in ‘a-adjective’ production”. Language 87. 55- 83. Braine, Martin. 1971. “On two types of models of the internalization of grammars”. The ontogenesis of grammar. Ed. Dan I. Slobin. New York: Academic Press. 153- 186. Braine, Martin and Patricia J. Brooks. 1995. “Verb-argument structure and the problem of avoiding an overgeneral grammar”. Beyond names for things: Young children’s acquisition of verbs. Eds. Michael Tomasello and William Merriman. Hillsdale, NJ: Erlbaum. 353-376. Brooks, Patricia J. and Michael Tomasello. 1999. “How children constrain their argument structure constructions”. Language 75. 720-738. <?page no="232"?> 219 Brooks, Patricia J., Michael Tomasello, Delly Dodson and Lawrence B. Lewis. 1999. “Young children’s overgeneralizations with fixed transitivity verbs”. Child Development 70. 1325-1337. Brown, Allen. 1991. “A review of the tip-of-the-tongue experience”. Psychological Bulletin 109. 204-223. Brown, Ann L. and Mary J. Kane. 1988. “Preschool children can learn to transfer: Learning to learn and learning from example”. Cognitive Psychology 20. 493-523. Brown, E. Keith. 2004. “Generative grammar”. The linguistics encyclopedia. Ed. Kirsten Malmkjaer. London: Routledge. 171-192. Brown, Roger. 1973. A first language: The early stages. Cambridge: Harvard University Press. Brown, Roger and Camile Hanlon. 1970. “Derivational complexity and order of acquisition in child speech”. Cognition and the development of language. Ed. John R. Hayes. New York: Wiley. 11-53. Brown, Roger and Eric H. Lenneberg. 1954. “A study in language and cognition”. The Journal of Abnormal and Social Psychology 49. 454-462. Buttelmann, David, Malinda Carpenter and Michael Tomasello. 2009. “Eighteenmonth-old infants show false belief understanding in an active helping paradigm”. Cognition 112. 337-342. Bybee, Joan L. 1985. Morphology: A study of the relation between meaning and form. Amsterdam: John Benjamins. Bybee, Joan L. 1995. “Regular morphology and the lexicon”. Language and Cognitive Processes 10. 425-455. Bybee, Joan L. 2006. “From usage to grammar: The mind’s response to repetition”. Language 82. 711-733. Bybee, Joan L. 2010. Language, usage and cognition. Cambridge: Cambridge University Press. Bybee, Joan L. and Carol L. Moder. 1983. “Morphological classes as natural categories”. Language 59. 251-270. Bybee, Joan L. and Joanne Scheibman. 1999. “The effect of usage on degrees of constituency: the reduction of don’t in English”. Linguistics 37. 575-596. Bybee, Joan L. and Dan I. Slobin. 1982. “Rules and schemas in the development and use of the English past tense”. Language 58. 265-289. Cameron-Faulkner, Thea, Elena V. Lieven and Michael Tomasello. 2003. “A construction based analysis of child directed speech”. Cognitive Science 27. 843-873. Casenhiser, Devin and Adele E. Goldberg. 2005. “Fast mapping between a phrasal form and meaning”. Developmental Science 8. 500-508. Chiarcos, Christian, Berry Claus and Michael Grabski. 2011. “Introduction: Salience in linguistics and beyond”. Salience. Multidisciplinary perspectives on its function in discourse. Eds. Christian Chiarcos, Berry Claus and Michael Grabski. Berlin: De Gruyter. 1-30. Childers, Jane B. and Michael Tomasello. 2001. “The role of pronouns in young children’s acquisition of the English transitive construction”. Developmental Psychology 37. 739-748. Chomksy, Noam. 1957. Syntactic structures. The Hague: Mouton de Gruyter. Chomsky, Noam. 1959. “A review of B. F. Skinner’s Verbal behavior”. Language 35. 26-58. Chomsky, Noam. 1965. Aspects of the theory of syntax. Cambridge: MIT Press. <?page no="233"?> 220 Chomsky, Noam. 1968/ 2006. Language and mind. New York: Cambridge University Press. Chomksy, Noam. 1970. “Remarks on nominalization”. Readings in English Transformational Grammar. Eds. Roderick A. Jacobs and Peter S. Rosenbaum. Waltham: Gin and Company. 184-221. Chomsky, Noam. 1975. Reflections on language. New York: Pantheon. Chomksy, Noam. 1980a. Rules and representations. New York: Columbia University Press. Chomksy, Noam. 1980b. “On cognitive structures and their development: A reply to Piaget”. Language and learning: The debate between Jean Piaget and Noam Chomsky. Ed. Massimo Piattelli-Palmarini. Cambridge: Harvard University Press. 35-54. Chomsky, Noam, 1981. Lectures on government and binding: The Pisa lectures. The Hague: Mouton de Gruyter. Chomsky, Noam. 1988. Language and problems of knowledge: The Managua Lectures. Cambridge: MIT Press. Chomsky, Noam. 1995. The minimalist program. Cambridge: MIT Press. Chomsky, Noam. 2000. The architecture of language. Oxford: Oxford University Press. Clahsen, Harald. 1996. “Introduction”. Generative perspectives on language acquisition: Empirical findings, theoretical considerations, and cross-linguistic comparisons. Ed. Harald Clahsen. Amsterdam: Benjamins. xv-xxvii. Clark, Eve V. 1993. The lexicon in acquisition. Cambridge: Cambridge University Press. Clark, Eve V. 1996. “Early verbs, event-types, and inflections”. Children’s language. Eds. John Gilbert and Carolyn E. Johnson. Hillsdale: Erlbaum. 61-73. Clark, Eve V. 2003. First language acquisition. Cambridge: Cambridge University Press. Clark, Eve V., Kathie L. Carpenter and Werner Deutsch. 1995. “Reference states and reversals: undoing actions with verbs”. Journal of Child Language 22. 633-662. Clark, Eve V. and Herbert H. Clark. 1979. “When nouns surface as verbs”. Language 55. 767-811. Clark, Eve V. and Sophia R. Cohen. 1984. “Productivity and memory for newlyformed words”. Journal of Child Language 11. 611-625. Clark, Eve V., Susan A. Gelman and Nancy M. Lane. 1985. “Compound nouns and category structure in young children”. Child Development 56. 84-94. Clark, Eve V. and Barbara F. Hecht. 1982. “Learning to coin agent and instrument nouns”. Cognition 12. 1-24. Clausner, Timothy C. and William Croft. 1997. “Productivity and schematicity in metaphors”. Cognitive Science 21. 247-282. Cordes, Anne-Kristin, Grzegorz Krajewski, Elena V. Lieven. 2013. “Effects of the shape of the input distribution on novel construction learning”. Manuscript in preparation. Crain, Stephen. 1991. “Language acquisition in the absence of experience”. Behavioural and Brain Sciences 14. 597-650. Croft, William. 2001. Radical construction grammar. Oxford: Oxford University Press. Croft, William and Alan Cruse. 2004. Cognitive Linguistics. Cambridge: Cambridge University Press. Culvalay-Haak, Martine. 1959. The verb in literary and colloquial Arabic. Berlin: Mouton de Gruyter. <?page no="234"?> 221 and James Street. 2006. “Individual differences in language attainment: Comprehension of passive sentences by native and non-native English speakers”. Language Sciences 28. 604-615. Dale, Philip S. and Larry Fenson. 1996. “Lexical development norms for young children”. Behavior Research Methods Instruments & Computers 28. 125-127. Deane, Paul D. 1992. Grammar in mind and brain: Explorations in cognitive syntax. Berlin: Mouton de Gruyter. DeLoache, Judy S. 1990. “Knowing and remembering in young children”. Emory Symposia in Cognition 3. 94-126. Demetras, M. J., Kathryn Nolan Post and Catherine Snow. 1986. “Feedback to first language learners: The role of repetitions and clarification questions”. Journal of Child Language 13. 275-292. Demuth, Katherine. 1989. “Maturation, continuity and the acquisition of Sesotho passive”. Language 65. 56-80. De Villiers, Jill G. 1980. “The process of rule learning in child speech: A new look”. Children’s Language: Volume 2. Ed. Keith E. Nelson. New York: Gardner Press. 1- 44. De Villiers, Jill G. and Peter A. de Villiers. 1973. “A cross-sectional study of the acquisition of grammatical morphemes in child speech”. Journal of Psycholinguistic Research 2. 267-278. De Villiers, Jill G. and Peter A. de Villiers. 1985. “The acquisition of English”. The crosslinguistic study of language acquisition - Volume 1. Ed. Dan I. Slobin. Hillsdale: Lawrence Erlbaum. 27-139. Diessel, Holger. 2004. The acquisition of complex sentences. Cambridge: Cambridge University Press. Diessel, Holger. 2007. “Frequency effects in language acquisition, language use, and diachronic change”. New Ideas in Psychology 25. 108-127. Donalies. Elke. 2005. Die Wortbildung des Deutschen: Ein Überblick. Mannheim: Narr Francke Attempto Verlag. Dressler, Wolfgang U. 1997. “Universals, typology, and modularity in Natural Morphology”. Language history and linguistic modelling: A festschrift for Jacek Fisiak on his 60 th birthday - Volume II. Eds. Raymond Hickey and Stanislaw Puppel. 1399- 1422. Dressler, Wolfgang U. 2003. “Degrees of grammatical productivity in inflectional morphology”. Rivista di Linguistica 15. 31-62. Dressler, Wolfgang U. 2005. “Morphological typology and first language acquisition: Some mutual challenges”. Morphology and linguistic typology. On-line Proceedings of the Fourth Mediterranean Morphology Meeting (MMM4) Catania 21-23 September 2003. Eds. Geert Booij, Emiliano Guevara, Angla Ralli, Salvatore Sgroi and Sergio Scalise. University of Bologna. 7-20. <http: / / morbo.lingue.unobo.it/ mmm/ >. Accessed 26.10.2011. Ebbinghaus, Hermann. 1913. Memory: A contribution to experimental psychology. New York: Teachers College Press. Eichinger, Ludwig M. 1989. Raum und Valenz im Verbwortschatz des Deutschen: eine valenzgrammatische Studie. Tübingen: Niemeyer. Eisenbeiß, Sonja. 2009. “Generative approaches to language learning”. Linguistics 47. 273-310. Eisenberg, Peter. 2006. Grundriss der deutschen Grammatik Band 1: Das Wort. Stuttgart: Verlag J. B. Metzler. <?page no="235"?> 222 Ellis, Nick C. 2002. “Frequency effects in language processing - A review with implications for theories of implicit and explicit language acquisition”. Studies in Second Language Acquisition 24. 143-188. Elman, Jeffrey L., Elizabeth A. Bates, Mark H. Johnson, Annette Karmiloff-Smith, Domenico Parisi and Kim Plunkett. 1996. Rethinking innateness: A connectionist perspective on development. Cambridge: MIT Press. Emmorey, Karen. 2000. “The confluence of space and language in signed languages”. Linguistics of American Sign Language: An introduction. Eds. Clayton Valli and Ceil Lucas. Washington: Gallaudet University. 318-346. Engel, Ulrich. 2004. Deutsche Grammatik. München: Iudicium Verlag. Fehlisch, Ulrike. 1998. “Zur Einordnung denominaler ein-Verben im deutschen Verbsystem”. Semantische und konzeptuelle Aspekte der Partikelverbbildungen mit ein-. Ed. Susan Olsen. Tübingen: Stauffenburg Verlag. 149-247. Fein, Greta G. “Pretend play in childhood: An integrative review”. Child Development 52. 1095-1118. Fernald, Anne. 1985. “Four-month-old infants prefer to listen to motherese”. Infant Behavior and Development 8. 181-195. Fillmore, Charles J., Paul Kay and Mary C. O’Connor. 1988. “Regularity and idiomaticity in grammatical constructions: The case of let alone”. Language 64. 501- 538. Fillmore, Charles J., Russell R. Lee-Goldman and Russell Rhodes. 2012. “The framenet constructicon”. Sign-based construction grammar. Eds. Hans C. Boas and Ivan A. Sag. Manuscript. <http: / / www.icsi.berkeley.edu/ pubs/ ai/ framenet constructicon11.pdf>. Accessed 10.03.2012. Fischer, Kerstin and Anatol Stefanowitsch. 2008. “Konstruktionsgrammatik: Ein Überblick”. Konstruktionsgrammatik I: Von der Anwendung zu Theorie. Eds. Kerstin Fischer and Anatol Stefanowitsch. Tübingen: Stauffenberg Verlag Brigitte Narr GmbH. 3-17. Fleischer, Wolfgang and Irmhild Barz. 1995. Wortbildung der deutschen Gegenwartssprache. Tübingen: Niemeyer. Gathercole, Virginia C. 1986. “The acquisition of the present perfect: Explaining differences in the speech of Scottish and American children”. Journal of Child Language 13. 537-560. Gathercole, Virginia C. 2002a. “Command of the mass/ count distinction in bilingual and monolingual children”. Language and literacy in bilingual children. Eds D. Kimbrough Oller and Rebecca E. Eilers. Clevedon: Multilingual Matters. 175- 206. Gathercole, Virginia C. 2002b. “Grammatical gender in bilingual and monolingual children: A Spanish morphosyntactic distinction”. Language and literacy in bilingual children. Eds D. Kimbrough Oller and Rebecca E. Eilers. Clevedon: Multilingual Matters. 207-219. Gathercole, Virginia C. 2002c. “Monolingual and bilingual acquisition: Learning different treatments of that-trace phenomena in English and Spanish”. Language and literacy in bilingual children. Eds D. Kimbrough Oller and Rebecca E. Eilers. Clevedon: Multilingual Matters. 220-254. Gathercole, Virginia C. and Erika Hoff. 2009. “Input and the acquisition of language: Three questions”. Blackwell Handbook of Language Development. Eds. Erika Hoff and Marilyn Shatz. Malden: Blackwell Publishing Ltd. 107-127. <?page no="236"?> 223 Gick, Mary L. and Keith J. Holyoak. 1980. “Analogical Problem-Solving”. Cognitive Psychology 12. 306-355. Gentner, Dedre. 1988. “Metaphor as structure mapping: The relational shift”. Child Development 59. 47-59. Gentner, Dedre. 2003. “Why we’re so smart”. Language in mind - Advances in the study of language and thought. Eds. Dedre Gentner and Susan Goldin-Meadow. Cambridge: MIT Press. 195-236. Gentner, Dedre and Kenneth J. Kurtz. 2006. “Relations, objects, and the composition of analogies”. Cognitive Science 30. 609-642. Gentner, Dedre and Jeffrey Loewenstein. 2002. “Relational language and relational thought”. Language, Literacy, and Cognitive Development. Eds. Eric Amsel and James P. Byrnes. Mahwah: Erlbaum. 87-120. Gentner, Dedre and Arthur B. Markman. 1997. “Structure mapping in analogy and similarity”. American Psychologist 52. 45-56. Gentner, Dedre and José Medina. 1998. “Similarity and the development of rules”. Cognition 65. 263-297. Gentner, Dedre and Laura L. Namy. 1999. “Comparison in the development of categories”. Cognitive Development 14. 487-513. Gentner, Dedre and Mary J. Rattermann. 1991. “Language and the career of similarity”. Perspectives on language and thought: Interrelations in development. Eds. Susan A. Gelman and James P. Byrnes. Cambridge: Cambridge University Press. 225- 277. Gentner, Dedre and Cecile Toupin. 1986. “Systematicity and surface similarity in the development of analogy”. Cognitive Science 10. 277-300. Gergely, György, Harold Bekkering and Ildikó Király. 2002. “Rational imitation in preverbal infants”. Nature 415. 755. Gick, Mary L. and Keith J. Holyoak. 1980. “Analogical problem solving”. Cognitive Psychology 12. 306-355. Goldberg, Adele E. 1995. Constructions: a construction grammar approach to argument structure. Chicago: University of Chicago Press. Goldberg, Adele E. 2006. Constructions at Work. The Nature of Generalisation in Language. Oxford: Oxford University Press. Goldberg, Adele E. 2011. “Corpus evidence of the viability of statistical preemption”. Cognitive Linguistics 22. 131-154. Goldberg, Adele E., Devin M. Casenhiser and Nitya Sethuraman. 2004. “Learning argument structure generalizations”. Cognitive Linguistics 15. 289-316. Goldberg, Adele E., Devin M. Casenhiser and Tiffani R. White. 2007. “Constructions as categories of language”. New Ideas in Psychology 25. 70-86. Goldinger, Stephen D. 1998. “Echoes of echoes? An episodic theory of lexical access”. Psychological Review 105. 251-279. Goldschneider, Jennifer M. and Robert M. DeKeyser. 2001. “Explaining the ‘natural order of L2 morpheme acquisition’ in English: A meta-analysis of multiple determinants”. Language Learning 51. 1-50. Goldstone, Robert L., Douglas L. Medin and Dedre Gentner. 1991. “Relational similarity and the non-independence of features in similarity judgments”. Cognitive Psychology 23. 222-264. Goodman, Nelson. 1972. “Seven strictures on similarity”. Problems and projects. Ed. Nelson Goodman. New York: The Bobbs-Merrill Co. 437-446. <?page no="237"?> 224 Goswami, Usha. 1991. “Analogical reasoning: What develops? A review of research and theory”. Child Development 62. 1-22. Goswami, Usha and Ann L. Brown. 1989. “Melting chocolate and melting snowmen: Analogical reasoning and causal relations”. Cognition 25. 69-95. Greenbaum, Sidney and Randolph Quirk. 1990. A student’s grammar of the English language. Harlow: Longman. Griffiths, Patrick. 1986. “Early vocabulary”. Language acquisition. Eds. Paul Fletcher and Michael Garman. Cambridge: Cambridge University Press. 279-306. Guillaume, Paul. 1927/ 1973. “Development of formal elements in the child’s speech”. Reprinted in Studies in child language development. Eds. Charles Ferguson and Dan I. Slobin. New York: Holt, Rinehart and Winston. 240-251. Hampton, James. 1997. “Psychological representation of concepts”. Cognitive models of memory. Ed. Martin A. Conway. Hillsdale, NJ: Lawrence Erlbaum Associates. 81-107. Harris, Margaret, Martyn Barrett, David Jones and Susan Brookes. 1988. “Linguistic input and early word meaning”. Journal of Child Language 15. 77-94. Harris, Paul L. and Robert D. Kavanaugh. 1993. “Young children’s understanding of pretense”. Monographs of the Society for Research in Child Development 58. v-92. Hart, Betty. 1991. “Input frequency and children’s first words”. First Language 11. 289-300. Hart, Betty and Todd R. Risley. 1995. Meaningful differences in the everyday experience of young American children. Baltimore: Paul H. Brookes Publishing Co. Hasher, Lynn and Walter Chromiak. 1977. “The processing of frequency information: An automatic mechanism? ” Journal of Verbal Learning and Verbal Behavior 16. 173-184. Hasher, Lynn and Rose T. Zacks. 1984. “Automatic processing of fundamental information - The case of frequency of occurrence”. American Psychologist 39. 1372- 1388. Hasher, Lynn, Rose T. Zacks, Karen C. Rose and Henrianne Sanft. 1987. “Truly incidental encoding of frequency information”. American Journal of Psychology 100. 69-91. Haspelmath, Martin and Andrea D. Sims. 2010. Understanding morphology. London: Hodder Education. Hauser, Marc D., Noam Chomsky and William T. Fitch. 2002. “The faculty of language: What is it, who has it, and how did it evolve? ” Science 298. 1569-1579. Herrmann, Esther, Maria Victoria Hernández-Lloreda, Josep Call, Brian Hare and Michael Tomasello. 2010. “The structure of individual differences in the cognitive abilities of children and chimpanzees”. Psychological Science 21. 102-110. Hirsh-Pasek, Kathy, Rebecca Treiman and Maita Schneiderman. 1984. “Brown & Hanlon revisited: Mother’s sensitivity to ungrammatical forms”. Journal of Child Language 11. 81-89. Hoff-Ginsberg, Erika. 1998. “The relation of birth order and socioeconomic status to children’s language experience and language development”. Applied Psycholinguistics 19. 603-629. Holyoak, Keith J. 2005. “Analogy”. The Cambridge Handbook of Reasoning and Thinking. Eds. Keith J. Holyoak and Robert G. Morrison. Cambridge: Cambridge University Press. 117-142. <?page no="238"?> 225 Holyoak, Keith J. 2012. “Analogy and relational reasoning”. The Cambridge Handbook of Reasoning and Thinking. Eds. Keith J. Holyoak and Robert G. Morrison. Cambridge: Cambridge University Press. 234-259. Holyoak, Keith J. and Paul Thagard. 1995. Mental Leaps: Analogy in Creative Thought. Cambridge: MIT Press. Hornstein, Norbert and David Lightfoot. 1981. Explanation in linguistics: The logical problem of language acquisition. London: Longman. Humboldt, Friedrich W. Freiherr von. 1836. Über die Verschiedenheit des menschlichen Sprachbaues und ihren Einfluss auf die geistige Entwicklung des Menschengeschlechts. Berlin: F. Dümmler. Hupp, Juli M., Vladimir M. Sloutsky and Peter W. Culicover. 2009. “Evidence for a domain-general mechanism underlying the suffixation preference in language”. Language and Cognitive Processes 24. 876-909. Huttenlocher, Janellen, Wendy Haight, Anthony Bryk, Michael Seltzer and Thomas Lyons. 1991. “Early vocabulary growth: Relation to language input and gender”. Developmental Psychology 27. 236-248. Huttenlocher, Janellen, Marina Vasilyeva, Elina Cymerman and Susan Levine. 2002. “Language input and child syntax”. Cognitive Psychology 45. 337-374. Huttenlocher, Janellen, Marina Vasilyeva and Priya Shimpi. 2004. “Syntactic priming in young children”. Journal of Memory and Language 50. 182-195. Jaccard, James. 2001. Interaction effects in logistic regression. Papers Series on Quantitative Applications in the Social Sciences. Thousand Oaks, CA: Sage. Jackson, Howard and Peter Stockwell. 2011. An introduction to the nature and functions of language. London: Continuum Publishing Group. Jaeger, T. Florian. 2008. “Categorical data analysis: Away from ANOVAs (transformation or not) and towards logit mixed models”. Journal of Memory and Language 59. 434-446. Jusczyk, Peter W. and Robert N. Aslin. 1995 “Infants’ detection of the sound patterns of words in fluent speech”. Cognitive Psychology 29. 1-23. Jusczyk, Peter W., Elizabeth A. Johne and Angela L. Bauman. 1999. “Infants’ sensitivity to allophonic cues for word segmentation”. Perception and Psychophysics 61. 1465-1476. Kay, Paul. 1997. Words and the grammar of context. Stanford: CSLI. Kay, Paul and Charles J. Fillmore. 1999. “Grammatical constructions and linguistic generalizations: the What’s X doing Y? construction”. Language 75. 453-481. Kay, Paul and Chad K. McDaniel. 1978. “The linguistic significance of the meanings of basic color terms”. Language 54. 610-646. Keane, Mark T. 1987. “On retrieving analogues when solving problems”. Quarterly Journal of Experimental Psychology 39A. 29-41. Keil, Frank C. and Nancy Batterman. 1984. “A characteristic-to-defining shift in the development of word meaning”. Journal of Verbal Learning and Verbal Behavior 23. 221-236. Kemler Nelson, Deborah G. 1995. “Principle-based inferences in young children’s categorization: Revisiting the impact of function on the naming of artifacts”. Cognitive Development 10. 347-38. Kirjavainen, Minna, Anna L. Theakston and Elena V. Lieven. 2009. “Can input explain children’s me-for-I errors? ” Journal of Child Language 36. 1091-1114. Kline, Melissa and Katherine Demuth. 2010. “Factors facilitating implicit learning: The case of the Sesotho passive”. Language Acquisition 17. 220-234. <?page no="239"?> 226 Kotovsky, Laura and Dedre Gentner. 1996. “Comparison and categorization in the development of relational similarity”. Child Development 67. 2797-2822. Krajewski, Grzegorz. 2011. “Regression coefficients and experimental design”. Unpublished manuscript. Kruschke, John K. 2005. “Category learning”. The Handbook of Cognition. Eds. Koen Lamberts and Robert L. Goldstone. London: Sage. 183-201. Kühnhold, Ingeborg and Hans Wellmann. 1973. Deutsche Wortbildung Typen und Tendenzen in der Gegenwartssprache. 1. Hauptteil. Das Verb. Düsseldorf: Pädagogischer Verlag Schwann. Labov, William. 1973. “The boundaries of words and their meaning”. New ways of analyzing variation in English. Eds. Charles-James Bailey and Roger W. Shuy. Washington: Georgetown University Press. 340-373. Lakoff, George. 1987. Women, fire and dangerous things. Chicago: University of Chicago Press. Lamprecht, Adolf. 1986. Grammatik der englischen Sprache. Berlin: Volk und Wissen Verlag und Cornelsen Verlag. Langacker, Ronald W. 1987a. Foundations of Cognitive Grammar Vol. 1: Theoretical Prerequisites. Stanford: Stanford University Press. Langacker, Ronald W. 1987b. “Nouns and verbs”. Language 63. 53-94. Langacker, Ronald W. 2000a. Grammar and conceptualization. New York: Mouton de Gruyter. Langacker, Ronald W. 2000b. “A dynamic usage-based model”. Usage-Based Models of Language. Eds. Michael Barlow and Suzanne Kemmer. Stanford: SLI Publications. 1-63. Leiss, Elisabeth. 2009. “Konstruktionsgrammatik versus Universalgrammatik”. Wie wir sprechen und schreiben. Festschrift für Helmut Glück zum 60. Geburtstag. Eds. Wieland Eins and Friederike Schmöe. Wiesbaden: Harrassowitz. 17-28. Lenneberg, Eric H. 1967. Biological foundations of language. New York: Wiley. Leslie, Alan M. 1987. “Pretense and representation: The origins of ‘theory of mind’”. Psychological Review 94. 412-426. Li, Ping and Yasuhiro Shirai. 2000. The acquisition of lexical and grammatical aspect. Berlin: Mouton de Gruyter. Lieven, Elena V. 2008. “Learning the English auxiliary: A usage-based approach”. Corpora in Language Acquisition Research: Finding Structure in Data (Trends in Language Acquisition Research, Vol.6). Ed. Heike Behrens. Amsterdam: Benjamins. 60-98. Lieven, Elena V. 2010. “Input and first language acquisition: Evaluating the role of frequency”. Lingua 120. 2546-2556. Lieven, Elena V., Heike Behrens, Jenny Speares and Michael Tomasello. 2003. “Early syntactic creativity: A usage-based approach”. Journal of Child Language 30. 333- 370. Lieven, Elena V., Dorothé Salomo and Michael Tomasello. 2009. “Two-year-old children’s production of multiword utterances: A usage-based analysis”. Cognitive Linguistics 20. 481-508. Lindgren, Kaj B. Über den oberdeutschen Präteritumschwund. Helsinki: Annales Academiae Scientiarum Fennicae. Loewenstein, Jeffrey and Dedre Gentner. 1998. “Relational language facilitates analogy in children”. Proceedings of the Cognitive Science Society 20. 615-620. <?page no="240"?> 227 Loewenstein, Jeffrey and Dedre Gentner. 2005. “Relational language and the development of relational mapping”. Cognitive Psychology 50. 315-353. MacDougall, Robert. 1904. “Recognition and recall”. Journal of Philosophy, Psychology and Scientific Methods 1. 229-233. MacWhinney, Brian. 1995. The CHILDES project: Tools for analyzing talk. Hillsdale: Erlbaum. MacWhinney, Brian. 2005. “A tale of two paradigms”. Language acquisition across linguistic and cognitive systems. Eds. Michèle Kail and Maya Hickmann. Amsterdam: John Benjamins Publishing Company. 17-32. MacWhinney, Brian. 2012. “The logic of the Unified Model”. The Routledge Handbook of Second Language Acquisition. Eds. Susan M. Gass and Alison Mackey. New York: Routledge. 211-227. Malmkjaer, Kirsten. 2004. “Behaviourist linguistics”. The linguistics encyclopedia. Ed. Kirsten Malmkjaer. London: Routledge. 59-63. Mandler, Jean M. 2000. “Perceptual and conceptual processes in infancy”. Journal of Cognition and Development 1. 3-36. Manning, Christopher. 2007. “Generalized linear mixed models (illustrated with R on Bresnan et al.’s dative data)”. Unpublished manuscript. Maratsos, Michael. 1983. “Some current issues in the study of the acquisition of grammar”. Handbook of Child Psychology Vol. III: Cognitive development. Ed. Paul Mussen. New York: John Wiley. 707-786. Maratsos, Michael. 2000. “More overregularisations after all: New data and discussion on Marcus, Pinker, Ullman, Hollander, Rosen & Xu”. Journal of Child Language 27. 183-212. Marchand, Hans. 1960. The categories and types of present-day English word-formation. Wiesbaden: Harrassowitz. Marchman, Virginia A. 1997. “Children’s productivity in the English past tense: The role of frequency, phonology, and neighborhood structure”. Cognitive Science 21. 283-304. Marcus, Gary F. 1993. “Negative evidence in language acquisition”. Cognition 46. 53- 85. Marcus, Gary F., Steven Pinker, Michael Ullman, Michelle Hollander, T. John Rosen and Fei Xu. 1992. “Overregularization in language acquisition”. Monographs of the Society for Research in Child Development 57. iii-178. Markman, Arthur B. and Dedre Gentner. 1993. “Structural Alignment in Similarity Comparisons”. Cognitive Psychology 25. 431-467 Maslen, Robert J., Anna L. Theakston, Elena V. Lieven and Michael Tomasello. 2004. “A dense corpus study of past tense and plural overregularization in English”. Journal of Speech, Language, and Hearing Research 47. 1319-1333. Masterson, Jackie and Judit Druks. 1998. “Description of a set of 164 nouns and 102 verbs matched for printed word frequency, familiarity and age-of-acquisition”. Journal of Neurolinguistics 11. 331-354. Masterson, Jackie, Judit Druks and Donna Gallienne. 2008. “Object and action picture naming in threeand five-year-old children”. Journal of Child Language 35. 373-402. Mather, Emily and Kim Plunkett. 2011. “Same items, different order: Effects of temporal variability on infant categorization”. Cognition 119. 438-447. <?page no="241"?> 228 Matthews, Danielle, Elena V. Lieven, Anna L. Theakston and Michael Tomasello. 2005. “The role of frequency in the acquisition of English word order”. Cognitive Development 20. 121-136. Matthews, Danielle, Elena V. Lieven, Anna L. Theakston and Michael Tomasello. 2007. “French children’s use and correction of weird word orders: A constructivist account”. Journal of Child Development 34. 381-409. Matthews, Danielle and Colin Bannard. 2010. “Children’s production of unfamiliar word sequences is predicted by positional variability and latent classes in a large sample of child-directed speech”. Cognitive Science 34. 465-488. McClelland, James L., Daniel E. Rumelhart and the PDP Research Group. 1986. Parallel Distributed Processing: Explorations in the microstructure of cognition. Volume 2: Psychological and Biological Models. Cambridge: MIT Press. Medin, Douglas L. and Marguerite M. Schaffer. 1978. “Context theory of classification learning”. Psychological Review 85. 207-238. Meltzoff, Andrew N. 1988. “Infant imitation after a 1-week delay: Long-term memory for novel acts and multiple stimuli”. Developmental Science 24. 470-476. Moerk, Ernst L. 1978. “Relationships between parental input frequencies and children’s language acquisition: a reanalysis of Brown’s data”. Journal of Child Language 7. 105-118. Morris, William C., Garrison W. Cottrell and Jeff L. Elman. 2000. “A connectionist simulation of the empirical acquisition of grammatical relations”. Hybrid neural symbolic integration. Eds. Stefan Wermter and Ray Sun. Berlin: Springer. 175-193. Naigles, Letitia R., Henry Gleitman and Lila Gleitman. 1993. “Children acquire word meaning components from syntactic evidence”. Language and cognition: A developmental perspective. Ed. Esther Dromi. New Jersey: Ablex Publishing. 104- 140. Naigles, Letitia R. and Erika Hoff-Ginsberg. 1998. “Why are some verbs learned before other verbs? Effects of input frequency and structure on children’s early verb use”. Journal of Child Language 25. 95-120. Narasimhan, Bhuvana and Marianne Gullberg. 2011. “The role of input frequency and semantic transparency in the acquisition of verb meaning: Evidence from placement verbs in Tamil and Dutch”. Journal of Child Language 38. 504-532. Nelson, Katherine. 1974. “Concept, word, and sentence: Interrelations in acquisition and development”. Psychological Review 81. 267-285. Newport, Elissa, Henry Gleitman and Lila Gleiman. 1977. “Mother, I’d rather do it myself: Some effects and non-effects of maternal speech style”. Talking to children: Language input and acquisition. Eds. Catherine E. Snow and Charles A. Ferguson. Cambridge: MIT Press. 109-149. Ninio, Anat. 1999. “Pathbreaking verbs in syntactic development and the question of prototypical transitivity”. Journal of Child Language 26. 619-653. Nosofsky, Robert M. 1986. “Attention, similarity, and the identificationcategorization relationship”. Journal of Experimental Psychology: General 115. 39- 57. Ochs, Elinor and Bambi B. Schieffelin. 2009. “Language acquisition and socialization: Three developmental stories and their implications”. Linguistic anthropology: A reader. Ed. Alessandro Duranti. 296-328. Malden: Wiley-Blackwell. Olsen, Susan. 1997. “Über den lexikalischen Status englischer Partikelverben“. Lexikalische Kategorien und Merkmale. Eds. Elisabeth Löbel and Gisa Rauh. Tübingen: Niemeyer. 45-71. <?page no="242"?> 229 Onishi, Kristine H. and Renée Baillargeon. 2005. “Do 15-month-old infants understand false beliefs? ” Science 308. 255-258. Peters, Ann M. 1983. The units of language acquisition. Cambridge: Cambridge University Press. Piaget, Jean. 1935/ 1970. The origins of intelligence in children. New York: Viking Press. Piaget, Jean, Jacques Montangero and J.-Bernhard Billeter. 1977. “Les correlates”. L’abstraction reflechissante. Ed. Jean Piaget. Paris: Presses Universitaires de France. 115-129. Pickering, Martin J. and Victor S. Ferreira. 2008. “Structural priming: A critical review”. Psychological Bulletin 134. 427-459. Pinker, Steven. 1984. Language learnability and language development. Cambridge: Harvard University Press. Pinker, Steven. 1989. Learnability and Cognition - The acquisition of argument structure. Cambridge: MIT Press. Pinker, Steven. 1995. “Language acquisition”. An invitation to cognitive science: Language. Eds. Lila Gleitman and Mark Liberman. Cambridge: MIT Press. 135-182. Pinker, Steven. 1999. Words and Rules: The Ingredients of Language. New York: Basic Books. Pinker, Steven. 2004. “Clarifying the logical problem of language acquisition”. Journal of Child Language 31. 949-953. Pinker, Steven and Alan Prince. 1988. “On language and connectionism: Analysis of a parallel distributed processing model of language acquisition”. Cognition 28. 73-193. Pinker, Steven and Michael Ullman. 2002. “The past and future of the past tense”. Trends in Cognitive Science 6. 456-463. Pizzuto, Elena and Christina Caselli. 1992. “The acquisition of Italian morphology”. Journal of Child Language 19. 491-557. Posner, Michael I. and Charles R. Snyder. 1975. “Attention and cognitive control”. Psychology of learning and motivation: Vol. 3. Eds. Gordon Bower and Janet T. Spence. San Diego: Academic Press. 55-85. Prasada, Sandeep and Steven Pinker. 1993. “Generalizations of regular and irregular morphology”. Language and Cognitive Processes 8. 1-56. Premack, David and Guy Woodruff. 1978. “Does the chimpanzee have a theory of mind? ” The Behavioral and Brain Sciences 4. 515-526. Pullum, Geoffrey K and Barbara C. Scholz. 2002. “Empirical assessment of stimulus poverty arguments”. Linguistic Review 19. 9-50. Quine, William V. 1969. “Natural kinds”. Ontological relativity and other essays. Ed. William V. Quine. New York: Columbia University Press. Quirk, Randolph, Sidney Greenbaum, Geoffrey Leech and Jan Svartvik. 1984. A grammar of contemporary English. Harlow: Longman Group Limited. Raaijmakers, Jeroen G. 2003. “A further look at the ‘language-as-fixed-effect fallacy’”. Canadian Journal of Experimental Psychology 57. 141-151. Rakoczy, Hannes, Michael Tomasello and Tricia Striano. 2004. “Young children know that trying is not pretending: A test of the ‘behaving-as-if’ construal of children’s early concept of pretense”. Developmental Psychology 40. 388-399. Radford, Andrew. 1990. Syntactic theory and the acquisition of English syntax: The nature of early child grammars in English. Oxford: Blackwell. <?page no="243"?> 230 Rattermann, Mary J. and Dedre Gentner. 1998. “More evidence for a relational shift in the development of analogy: Children’s performance on a causal-mapping task”. Cognitive Development 13. 453-478. Reali, Florencia and Morten H. Christiansen. 2005. “Uncovering the richness of the stimulus: Structure dependence and indirect statistical evidence”. Cognitive Science 29. 1007-1028. Regier, Terry. 1996. The human semantic potential: Spatial language and constrained connectionism. Cambridge: MIT Press. Rice, Mabel L., Janna B. Oetting, Janet Marquis, John Bode and Soyeong Pae. 1994. “Frequency of input effects of word comprehension of children with specific language impairment”. Journal of Speech and Hearing Research 37. 106-122. Roeper, Thomas and Jill G. deVilliers. 1992. “Ordered decisions in the acquisition of wh-questions”. Theoretical Issues in language acquisition: Continuity and change in development. Eds. Jürgen Weissenborn, Helen Goodluck and Thomas Roeper. Hillsdale, NJ: Lawrence Erlbaum. 191-236. Roeper, Thomas and Jürgen Weissenborn. 1990. “How to make parameters work: Comments on Valian”. Language Processing and Language Acquisition. Eds. Lyn Frazier and Jill G. deVilliers. Dordrecht: Kluwer. 147-162. Rosch, Eleanor. 1973. “On the internal structure of perceptual and semantic categories”. Cognitive development and the acquisition of language. Ed. Timothy E. Moore. New York: Academic Press. 111-144. Rosch, Eleanor. 1975. “Cognitive representation of semantic categories”. Journal of Experimental Psychology: General 104. 192-233. Rosch, Eleanor and Carolyn B. Mervis. 1975. “Family resemblances: Studies in the internal structure of categories”. Cognitive Psychology 7. 573-605. Rosch, Eleanor, Carolyn B. Mervis, Wayne D. Gray, David M. Johnson and Penny Boyes-Braem. 1976. “Basic objects in natural categories”. Cognitive Psychology 8. 382-439. Ross, Brian H. and Valerie S. Makin. 1999. “Prototype versus exemplar models”. The nature of cognition. Ed. Robert J. Sternberg. Cambridge: MIT Press. 206-241. Ross, Brian H., Susan J. Perkins and Patricia L. Tenpenny. 1990. “Reminding-based category learning”. Cognitive Psychology 22. 460-492. Rowland, Caroline F. 2007. “Explaining errors in children’s questions”. Cognition 104. 106-134. Rowland, Caroline F. and Julian M. Pine. 2000. “Subject-auxiliary inversion errors and wh-question acquisition: What do children know? ” Journal of Child Language 27. 157-182. Rowland, Caroline F., Julian M. Pine, Elena V. Lieven and Anna L. Theakston. 2003. “Determinants of acquisition order in wh-questions: Re-evaluating the role of caregiver speech”. Journal of Child Language 30. 609-635. Rubin, Edgar. 1921. Visuell wahrgenommene Figuren: Studien in psychologischer Analyse. Kopenhagen: Gyldendalske. Rubino, Carl. R. Galvez. 2001. “Pangasinan”. Encyclopedia of the World’s Languages: Past and Present. Eds. Jane Garry and Carl R. Galvez Rubino. New York: H.W. Wilson Press. 539-542. Rubino, Carl R. Galvez. 2005. “Reduplication: Form, function and distribution”. Studies on reduplication. Ed. Bernhard Hurch. Berlin: Mouton de Gruyter. 11-30. <?page no="244"?> 231 Rubino, Carl R. Galvez. 2008. “Reduplication”. The world atlas of language structure. Eds. Martin Haspelmath, Matthew S. Dryer, David Gil and Bernard Comrie. Oxford: Oxford University Press. 114-117. Rubino, Rejane B. and Julian M. Pine. 1998. “Subject-verb agreement in Brazilian Portuguese: What low error rates hide”. Journal of Child Language 25. 35-59. Rumelhart, Daniel E. and James L. McClelland and the PDP Research Group. 1986. Parallel Distributed Processing: Explorations in the microstructure of cognition. Volume 1: Foundations. Cambridge: MIT Press. Saffran, Jenny R., Richard N. Aslin and Elissa L. Newport. 1996. “Statistical learning by 8-month-old infants”. Science 274: 1926-1928. Sampson, Geoffrey. 1989. “Language acquisition: Growth or learning”. Philosophical Papers 18. 203-240. Sampson, Geoffrey. 2005. The 'Language Instinct' debate. London: Continuum. Savage, Ceri, Elena V. Lieven, Anna L. Theakston and Michael Tomasello. 2003. “Testing the abstractness of children’s linguistic representation: Lexical and structural priming of syntactic constructions in young children”. Developmental Science 6. 557-567. Schieffelin, Bambi B. 1979. “Getting it together: An ethnographic approach to the study of the development of communicative competence”. Developmental Pragmatics. Eds. Elinor Ochs and Bambi B. Schieffelin. New York: Academic Press. 73-108. Schmid, Hans-Jörg. 2005. Englische Morphologie und Wortbildung. Eine Einführung. Berlin: Erich Schmidt Verlag. Schmid, Hans-Jörg. 2007. “Entrenchment, salience and basic levels”. The Oxford Handbook of Cognitive Linguistics. Eds. Dirk Geeraerts and Hubert Cuyckens. Oxford: Oxford University Press. 117-138. Schmid, Hans-Jörg. 2011. English morphology and word formation. An introduction. Berlin: Erich Schmidt Verlag. Scholze-Stubenrecht, Werner, Birgit Eickhoff and Dieter Mang. 2001. Duden - Das Fremdwörterbuch. Mannheim: Dudenverlag. Schwitalla, Johannes. 1997. Gesprochenes Deutsch: Eine Einführung. Berlin: Erich Schmidt Verlag. Shannon, Claude E. and Warren Weaver. 1949. The mathematical theory of communication. Urbana: University of Illinois Press. Shimpi, Priya, Perla B. Gámez, Janellen Huttenlocher and Marina Vasilyeva. 2007. “Syntactic priming in 3and 4-year-old children: Evidence for abstract representations of transitive and dative forms”. Developmental Psychology 43. 1334-1346. Shirai, Yasuhiro and Roger W. Andersen. 1995. “The acquisition of tense-aspect morphology: A prototype account”. Language 71. 743-762. \ ^ _ $``~ Untersuchungen zum trennbaren Verb im Deutschen I. Prag: Karolinum. Singley, Mark K. and John R. Anderson. 1989. The transfer of cognitive skill. Cambridge: Harvard University Press. Shepard, Roger N. 1962a. “The analysis of proximities: Multidimensional scaling with an unknown distance function. Part I”. Psychometrika 27. 125-140. Shepard, Roger N. 1962b. “The analysis of proximities: Multidimensional scaling with an unknown distance function. Part II”. Psychometrika 27. 219-246. Skinner, Burrhus F. 1957. Verbal behavior. Acton: Copley Publishing Group. <?page no="245"?> 232 Slobin, Dan I. 1973. “Cognitive prerequisites for the development of grammar”. Studies of Child Language Development. Eds. Charles A. Ferguson and Dan I. Slobin. New York: Holt, Rinehart, and Winston. 175-176, 179-183, 191-194, 208. Slobin, Dan I. 1985. The crosslinguistic study of language acquisition - Volume 1. Hillsdale: Lawrence Erlbaum. Slobin, Dan I. 1997. “The origins of grammaticizable notions: Beyond the individual mind”. The crosslinguistic study of language acquisition - Volume 5: Expanding the contexts. Ed. Dan I. Slobin. Mahwah: Lawrence Erlbaum Associates. 265-323. Sloutsky, Vladimir M. and Anna V. Fisher. 2004. “Induction and categorization in young children: A similarity-based model”. Journal of Experimental Psychology 133. 166-188. Smith, Eliot R. and Diane M. Mackie. 2000. Social Psychology. Philadelphia: Taylor and Francis. Smoczynska, Magdalena. 1985. “The acquisition of Polish”. The crosslinguistic study of language acquisition - Volume 1. Ed. Dan I. Slobin. Hillsdale: Lawrence Erlbaum. 595-686. Stadthagen-Gonzalez, Hans and Colin J. Davis. 2006. “The Bristol norms for age of acquisition, imageability, and familiarity”. Behavior Research Methods 38. 598-605. Street, James and Ewa $ attainment: How much do adult native speakers of English know about passives and quantifiers? ” Lingua 120. 2080-2094. Suppes, Patrick. 1974. “The semantics of children’s language”. American Psychologist 129. 103- 114. Surian, Luca, Stefania Caldi and Dan Sperber. 2007. “Attribution of beliefs by 13month-old infants”. Psychological Science 18. 580-586. Szekely, Anna, Thomas Jacobsen, Simona D'Amico, Antonella Devescovi, Elena Andonova, Daniel Herron, Ching C. Lu, Thomas Pechmann, Csaba Pleh, Nicole Wicha, Kara Federmeier, Irina Gerdjikova, Gabriel Gutierrez, Daisy Hung, Jeanne Hsu, Gowri Iyer, Kathryn Kohnert, Teodora Mehotcheva, Araceli Orozco-Figueroa, Angela Tzeng, Ovid Tzeng, Analía Arévalo, Andras Vargha, Andrew C. Butler, Robert Buffington and Elizabeth Bates. 2004. “A new on-line resource for psycholinguistic studies”. Journal of Memory and Language 51. 247- 250. Taylor, John R. 2009. Linguistic Categorization. Oxford: Oxford University Press. Tennie, Claudio, Josep Call and Michael Tomasello. 2006. “Push or pull: Imitation vs. emulation in great apes and human children”. Ethnology 112. 1159-1169. Theakston, Anna L. 2004. “The role of entrenchment in constraining children’s verb argument structure overgeneralisations: A grammatical judgment study”. Cognitive Development 19. 15-34. Theakston, Anna L., Elena V. Lieven and Michael Tomasello. 2003. “The role of the input in the acquisition of their person singular verbs in English”. Journal of Speech, Language, and Hearing Research 46. 863-877. Theakston, Anna L., Elena V. Lieven, Julian M. Pine and Caroline F. Rowland. 2004. “Semantic generality, input frequency and the acquisition of syntax”. Journal of Child Language 31. 61-99. Thorndike, Edward L. 1906. Principles of teaching. New York: Seiler. Tomasello, Michael. 1992. First verbs: A case study of early grammatical development. New York: Cambridge University Press. <?page no="246"?> 233 Tomasello, Michael. 1995. “Joint attention as social cognition”. Joint attention: Its origins and role in development. Eds. Chris Moore and Philip J. Dunham. Hillsdale: Erlbaum. 103-130. Tomasello, Michael. 1999. The cultural origins of human cognition. Cambridge: Harvard University Press. Tomasello, Michael. 2000a. “Do young children have adult syntactic competence? ” Cognition 74. 209-253. Tomasello, Michael. 2000b. “First steps toward a usage-based theory of language acquisition”. Cognitive Linguistics 11. 61-82. Tomasello, Michael 2003. Constructing a Language. A Usage-Based Theory of Language Acquisition. Boston: Harvard University Press. Tomasello, Michael. 2005. “Beyond formalities: The case of language acquisition”. The Linguistic Review 22. 183-197. Tomasello, Michael and Patricia J. Brooks. 1998. “Young children’s earliest transitive and intransitive constructions”. Cognitive Linguistics 9. 379-395. Traugott, Elizabeth Closs. 2008. “Grammaticalization, constructions and the incremental development of language: Suggestions from the development of degree modifiers in English”. Variation, selection, development—probing the evolutionary model of language change. Eds. Regine Eckardt, Gerhard Jäger and Tonjes Veenstra. Berlin/ New York: Mouton de Gruyter. 219-250. Tversky, Amos. 1977. “Features of similarity”. Psychological Review 84. 327-352. Tversky, Amos and Itamar Gati. 1978. “Studies of similarity”. Cognition and categorization. Eds. Eleanor Rosch and Barbara B. Lloyd. Hillsdale: Erlbaum. 79-98. Underwood, Benton J. 1971. “Recognition memory”. Essays in neobehaviorism: A memorial volume to Kenneth W. Spence. Eds. Howard H. Kendler and Janet T. Spence. East Norwalk: Appleton-Century-Crofts. 313-335. Ungerer, Friedrich and Hans-Jörg Schmid. 2006. Introduction to cognitive linguistics. London, New York: Pearson Education. Valian, Virginia. 1990. “Logical and psychological constraints on the acquisition of syntax”. Language processing and language acquisition. Eds. Lyn Frazier and Jill de Villiers. Dordrecht, Holland: Kluwer. 119-145. Valian, Virginia. 2009. “Innateness and learnability”. The Cambridge Handbook of Child Language. Ed. Edith L. Bavin. Cambridge: Cambridge University Press. 15- 34. Wang, Shih-P. 2005. “Corpus-based approaches and discourse analysis in relation to reduplication and repetition”. Journal of Pragmatics 37. 505-540. Waxman, Sandra R. 1999. “Specifying the scope of 13-month-olds’ expectations for novel words”. Cognition 70. B35-B50. Waxman, Sandra R. and Amy Booth. 2003. “The origins and evolution of links between word learning and conceptual organization: new evidence from 11month-olds”. Developmental Science 6. 128-135. Waxman, Sandra R. and Dana B. Markow. 1995. “Words as invitations to form categories: Evidence from 12to 13-month-old infants”. Cognitive Psychology 29. 257- 302. Weinrich, Harald. 1993 (2007). Textgrammatik der deutschen Sprache. Mannheim: Bibliographisches Institut. Weist, Richard M. 1983. “Prefix vs. suffix information processing in the comprehension of tense and aspect”. Journal of Child Language 10. 85-96. <?page no="247"?> 234 Weist, Richar M. and Emilia Konieczna. 1985. “Affix processing strategies and linguistic systems”. Journal of Child Language 12. 27-35. Wellman, Henry M., David Cross and Julanne Watson. 2001. “Meta-analysis of theory of mind development: The truth about false belief”. Child Development 72. 655-684. Wellman, Henry M. and David Liu. 2004. “Scaling of Theory-of-Mind Tasks”. Child Development 75. 523-541. Wellman, Henry M., Sarah Lopez-Duran, Jennifer LaBounty and Betsy Hamilton. 2008. “Infant attention to intentional action predicts preschool theory of mind”. Developmental Psychology 44. 618-623. Wellmann, Hans. 1998. “Die Wortbildung”. Duden - Grammatik der deutschen Gegenwartssprache. Eds. Dudenredaktion. Mannheim: Dudenverlag. 408-557. Whorf, Benjamin L. 1956. Language, thought, and reality. Boston: MIT Press. Wilensky, Robert. 1983. Planning and understanding: A computational approach to human reasoning. Reading: Addison-Wesley. Wimmer, Heinz and Josef Perner. 1983. “Beliefs about beliefs: Representation and constraining function of wrong beliefs in young children’s understanding of deception”. Cognition 13. 103-128. Wittgenstein, Ludwig. 1958. Philosophical Investigations. Oxford: Blackwell. Wonnacott, Elizabeth, Jeremy Boyd, Jennifer Thomson and Adele E. Goldberg. 2012. “Input effects on the acquisition of a novel phrasal construction in 5 year olds”. Journal of Memory and Language 66. 458-478. World Atlas of Language Structure. <http: / / wals.info/ feature>. Accessed 01.02. 2012. Wright, William. 1859. A grammar of the Arabic language. Cambridge: Cambridge University Press. <?page no="248"?> 235 APPENDIX I Chapter 5 I.1 Excerpt from Leo corpus (Leo 4; 5; 3) *MUT: weiss nicht, <macht die>[/ / ] macht den Mund so (ei)n+bisschen gekraeuselt . %mor: $V: S: PRES: 1s|wissen$ADV: neg|nicht$VAUX: 01: PRES: 3s| machen$DET: def: AKK: SG|den$N: 08: m: AKK: SG|Mund$ADV|s o$ADV: quan|ein+bisschen $V: 01: PART: 0|kraeuseln . *CHI: ich mal da was . %mor: $PRO: pers: NOM: SG|ich $V: 01: PRES: 1s|malen $ADV|da $PRO: indef: AKK: SG|etwas/ was . *CHI: ja, is(t) das schon bemalt, da ? %mor: $INTER|ja $VAUX: S: PRES: 3s|sein $PRO: typ: NOM: SG|das$ ADV: temp|schon $V: 01: PART: PF|be+malen $ADV|da ? Translation *MOT: don’t know, <she makes>[/ / ] makes her mouth somewhat a little puckered . %mor: $V: S: PRES: 1s|know $ADV: neg|not $VAUX: 01: PRES: 3s|make $DET: def: AKK: SG|the$N: 08: m: AKK: SG|mouth$ADV|somewh at$ADV: quan|a+little $V: 01: PART: 0|pucker . *CHI: I paint something there . %mor: $PRO: pers: NOM: SG|I $V: 01: PRES: 1s|paint $ADV|there$PRO: indef: AKK: SG|something . *CHI: well, is that already pained on, there ? %mor: $INTER|well $VAUX: S: PRES: 3s|be $PRO: typ: NOM: SG|that $ADV: temp|already $V: 01: PART: PF|painted+on $ADV|there ? I.2 Excerpt from Thomas corpus (Thomas 4; 3; 2) *CHI: we're going to take off in a minute . %xmor: pro|we~v: aux|be&PRES v|go-PROG inf|to v|take ptl|off prep|in det|a n|minute . *CHI: I'm [/ ] I'm just going to do a quick knot . %xmor: pro|I~v: aux|be&1S adv|just v|go-PROG inf|to v|do det|a adj|quick n|knot . <?page no="249"?> 236 *CHI: then I'm going to untie it again . %xmor: conj: coo|then pro|I~v: aux|be&1S v|go-PROG inf|to v|un#tie pro|it adv|again . *MOT: are you trying to do a knot with the hose pipe ? %xmor: v: aux|be&PRES pro|you v|try-PROG inf|to v|do det|a n|knot prep|with det|the n|hose n|pipe ? I.3 Lists of manually selected verb prefix constructions and prefix verbs in English and German Table a. Manually selected verb prefix constructions and prefix verbs for English corpus. Prefix Prefix verbs Prefix meaning overoverheat overload to perform an action to a high or excessive degree 1 rerearrange to perform an action again/ anew/ back 2 rebuild recycle redress rename reuse rewind ununcork to reverse an action; to remove by an action 3 undecorate undo undress unhook unload unlock unopen unpack unpeel unplug unroll unset unsettle unstick untie untwist unwrap unzip <?page no="250"?> 237 1 Quirk et al. (1984: 986) describe the meaning of overas ‘too much’. 2 Quirk et al. (1984: 990) state the semantics of reas ‘again, back’. Marchand (1960: 188) refers to this meaning as ‘anew, again, back’. 3 Quirk et al. (1984: 983) mention two meanings of un-: ‘to reverse an action’ and ‘to deprive of/ release from an action’. These meanings were combined here. Marchand (1960: 204) reports that the original meaning of unwas ‘opposite’ and specifies the present-day English meaning as ‘undo the result of the verbal action’ and ‘cause the object of the verb to be no longer -ed’. Table b. Manually selected verb prefix constructions and prefix verbs for German corpus. Prefix Prefix verbs Translation Prefix meaning bebedrucken beschmeißen bespannen besprühen bestreuen ‘print on’ ‘throw at’ ‘cover’ ‘spray (on)’ ‘sprinkle (on)’ be 1 -: apply something; equip/ cover something with something; attach something to something; can be driven by a purpose 1 befeuchten ‘moisten’ besaufen ‘get pissed’ beschmieren ‘smudge’ besteigen ‘climb’ beziehen ‘cover (with)’ bekleckern ‘spill on’ betanken ‘fill up (with fuel)’ beschreiben ‘write on’ bestreichen ‘spread’ bewölken ‘cloud over’ beladen ‘load with’ bemalen ‘paint on’ bekleben ‘stick on’ belegen ‘put on’ bedecken ‘cover with’ beschriften ‘label’ begucken ‘look at’ bewinken befärben beschneien bezweifeln bestaunen besprechen bearbeiten bewachen bedenken beenden ‘doubt’ ‘marvel at’ ‘discuss’ ‘work on’ ‘guard’ ‘think about’ ‘stop’ be 2 -: reorganization of roles in an event; change of valency; (e.g., previously intransitive verbs, reflexive verbs or verbs with dative object now involve accusative object or become reflexive; actions are <?page no="251"?> 238 beeilen bedienen bedrohen ‘hurry’ ‘serve’ ‘threaten’ directed towards accusative object (transitivization/ reflexivation)) 2 beantworten ‘answer’ besetzen ‘occupy’ belohnen ‘reward‘ bewundern ‘admire’ bezahlen ‘pay’ bedanken ‘thank’ begrüßen beschützen benutzen begraben ‘greet’ ‘protect ‘use’ ‘bury’ be 3 -: more or less substantial extension, narrowing or intensification of the verb stem’s meaning 3 befühlen ‘feel’ befürchten ‘fear’ besagen ‘mean’ dedegenerieren ‘degenerate’ negation of original action denunzieren ‘denounce’ durchdurchleuchten durchlöchern durchfeuchten ‘shine through’ ‘perforate’ ‘soak’ to perform an action completely, thoroughly or extensively; to see through an action; causative; incorporation 4 eis- 5 eisbaden eisschwimmen ‘swim in water accessed through ice’ to perform an action in water accessed through a broken layer of ice ententgleisen entfernen ‘derail’ ‘remove’ ent 1 -: to remove to miss; to rid; also unintentionally 6 entspannen ‘relax’ entgehen ‘miss’ entleeren ‘empty’ entladen ‘unload’ entstauben ‘dust’ entwerten ‘devaluate’ entstehen entspringen ‘originate’ ‘emanate’ ent 2 -: the beginning of events, inchoative/ ingressiva 7 entzünden ‘ignite’ ererwachen erleuchten erzittern ‘wake up’ ‘enlighten’ ‘tremble’ er 1 -: ingressiva/ inchoative; describe (sudden) beginning or single occurrence of an <?page no="252"?> 239 erschrecken ‘startle’ event 8 erraten ersticken erdrücken erhängen erschlagen ‘guess right’ ‘suffocate’ ‘crush to death’ ‘hang’ ‘slay’ er 2 -: egressiva/ resultative verbs; achievement of a result describe the completion of an event; often related to death and how it eventuates 9 erwürfeln ‘play dice and win’ erkennen ‘recognize’ erneuern erweitern ‘renew’ ‘extend’ er 3 -: causative verbs; to make or become (more) [adjective] 10 erleichtern ‘facilitate’ erhöhen ‘raise’ gegebrauchen gefallen ‘use’ ‘please’ positive form in contrast to negative form involving missgelingen ‘succeed’ gestalten ‘design’ überüberbacken übertünchen überziehen überfahren überholen übersetzen überfliegen ‘gratinate’ ‘whitewash’ ‘cover’ ‘run over’ ‘take over’ ‘translate’ ‘scan’ über 1 -: to perform an action involving a localization and potentially a movement, e.g., passing something, passing above or across something; being located on top of something 11 übermalen ‘paint over’ übertreiben überleben überfressen ‘exaggerate’ ‘survive’ ‘overeat’ über 2 -: to perform to a high or excessive degree; crossing the line 11 übertreffen ‘exceed’ überschneiden ‘overlap’ überweisen übernehmen ‘transfer’ ‘take over’ über 3 -: action or event involving a transfer übertragen ‘transfer’ überlassen ‘relinquish’ überbringen ‘convey’ überreichen ‘hand over’ übergeben ‘deliver’ übermitteln ‘deliver’ überprüfen überreden ‘control’ ‘coax’ über 4 -: to wield authority in performing an action überwachen ‘oversee’ überzeugen ‘convince’ übersehen ‘overlook’ über 5 -: to perform an action <?page no="253"?> 240 überhören ‘fail to hear’ unsuccessfully or without the anticipated result 12 übernachten überwintern überstehen ‘sleep over’ ‘hibernate’ ‘survive’ über 6 -: to perform an action or event that lasts over a certain amount of time 13 umumhüllen umkreisen umgeben umkleben ‘envelope’ ‘circle’ ‘surround’ ‘cover with’ to perform an action with a certain direction, an action around an object or space; incorporation 14 umkurven ‘round’ umschwirren ‘buzz around’ umsegeln ‘circumnavigate’ ververlaufen verfahren verrutschen verschlucken ‘get lost’ ‘get lost’ ‘get out of place’ ‘choke’ ver 1 -: to perform an action erroneously; unintended event; contrast; to act in the wrong way 15 verlegen ‘misplace’ verschreiben ‘misspell’ vertun ‘err’ verdrücken ‘press the wrong button’ verschlafen ‘oversleep’ verpennen ‘oversleep’ verschütten ‘spill’ verwackeln ‘blur’ verhunzen ‘ruin’ verwechseln ‘mix up’ verwechslern ‘mix up’ verpassen ‘miss’ verwackeln ‘blur’ vermackeln ‘ruin’ verschlängeln verschneiden verkrickelkrackeln verreisen verbringen verschicken ‘travel’ ‘spend (time)’ ‘send’ ver 2 -: to perform an action covering certain (longer) period of time 16 ver 3 -: to perform an action or event resulting in a change of state (superordinate category) verlängern verstärken vertrocknen ‘lengthen’ ‘reinforce’ ‘wither’ ver 3a -: causative; to make or become (more) [adjective] (subcategory) 17 <?page no="254"?> 241 verkleinern ‘shrink’ verheilen ‘heal’ verfeinern ‘refine’ verkürzen ‘shorten’ verschönern ‘beautify’ versteifen ‘stiffen’ verwelken ‘wither’ vergrößern ‘enhance’ verdicken ‘thicken’ verkräuseln ‘curl’ verdickern verschwierigen verbrauchen verwandeln verbrennen verdecken ‘use up’ ‘transform’ ‘burn’ ‘cover’ ver 3b -: completion of an action; action or event resulting in more obscure/ different state (subcategory) 18 verhungern ‘starve’ verbauen ‘use up building’ verkleiden ‘dress up’ verändern ‘change’ verschieben ‘defer’ verzieren ‘adorn’ verbiegen ‘bend’ versenken ‘sink’ verspeisen ‘eat up’ versickern ‘seep away’ verarbeiten ‘process’ verdrücken ‘scoff’ verdursten ‘die of thirst’ versinken ‘sink’ vernichten ‘destroy’ verhüllen ‘veil’ verkriechen ‘hole up’ verzaubern ‘bewitch’ verpesten ‘contaminate’ versteinern ‘petrify’ vergraben ‘bury’ verwunden ‘wound’ verkommen ‘decay’ verputzen ‘scoff’/ ‘plaster’ verqualmen ‘fill with smoke’ verschnupfen ‘get a runny nose’ vertarnen verbinden verteilen ‘connect’ ‘distribute’ ver 3c -: to perform an action or event describing the relation <?page no="255"?> 242 vermischen verreiben ‘mix’ ‘rub’ between at least two components (subcategory) 19 verschmieren ‘smear’ verkeilen ‘become wedged’ verknoten ‘knot together’ verstreuen ‘disperse’ vermehren ‘increase’ verdrehen ‘twist’ vereinen ‘unify’ verkleben ‘stick together’ vermengen ‘blend’ verhaken ‘get entangled’ verrühren ‘mix’ zerzerreißen zerschneiden zerbrechen zerbröseln zerteilen zertrümmern ‘tear apart’ ‘cut up’ ‘break’ ‘crumble’ ‘dissect’ ‘shatter’ zer 1 -: to perfom an action or event resulting in the subdivision into smaller parts; resultative; potentially intensifying; potentially damaging 20 zerlegen ‘dismantle’ zerkleinern ‘mince’ zerfließen ‘dissolve’ zerschnippeln ‘chop up’ zerpflücken ‘pluck to pieces’ zersägen ‘cut to pieces’ zerfetzen ‘shred’ zerstückeln ‘dismember’ durchzerschneiden zerstören zerdrücken zerknittern zerstechen ‘destroy’ ‘crush’ ‘crinkle’ ‘sting’ zer 2 -: to perform an intentionally or unintentionally damaging action 21 zerkochen ‘cook to rags’ zerspringen ‘burst’ zerstürzen Please note that Leo’s non-canonical forms are discussed in 5.2.2. They are not translated in this table. Please further note that several meanings are difficult to differentiate or might be subsumed under a more general one (e.g., zer 1 and zer 2 -). This issue is discussed in 8.4.2. Most counts in the corpus analysis were unaffected by the precise classification of verbs to prefix constructions and the exact number of postulated verb prefix constructions (i.e., meanings). 1 Kühnhold and Wellmann (1973: 146) describe this meaning as signalling ‘contact’ (be- 1 in their classification). <?page no="256"?> 243 2 According to Barz (2006: 696-697), the prefix beserves to change the semantic and syntactic valency of the predominantly verbal bases. This is also a feature of some beprefix verbs that are found in other categories. In the present classification their semantics were weighted higher than the valency changes in these cases, resulting in respective categorization in the semantically richer cateogory. 3 Kühnhold and Wellmann (1973: 146, be- 3 verbs) describe the meaning of this prefix as ‘intensifying’. 4 Kühnhold and Wellmann (1973: 147) distinguish a focus on ‘direction’ durch- 1 (durchlaufen ‘traverse’), ‘completeness of action’ durch- 2 (durchforschen ‘(re-) search through/ thoroughly’) or ‘end state’ durch- 4 (durchfeuchten ‘wet through/ thoroughly’). Since, in contrast to Kühnhold and Wellmann’ classification, particle verbs, which make up a large proportion of exemplars in each category, are not taken into account in the present classification, the categories are merged under the common characteristic of the ‘thorough, complete or extensive execution of an action’. Barz (2006: 702-703) calls the resulting change in semantic roles and syntactic functions ‘incorporation’. 5 Eisenberg (2006: 267) understands eisas a particle verb bordering on syntax, giving the example of eislaufen. Considering the corpus examples, eisseems a hybrid of prefix and particle: Like prefixes it is not separated in main clauses with finite verbs, e.g., Er eisbadet. ? Er badet eis. However, the past participle is formed with (-)ge-, which is common with particle verbs, e.g., Er hat eisgebadet. *Er hat eisbaden./ *Er hat eisbadet. Despite its borderline status, eisis listed as a verb prefix construction to ensure completeness. 6 Kühnhold and Wellmann (1973: 148) also use a category signalling ‘removal’ (ent- 1 ). However, they (1973: 148) categorize the verb entfernen ‘remove’ itself as a member of a different, third category focusing on the ‘end state’. In contrast, the present account includes this verb in the ‘removal’category. Nevertheless, there is a semantic component entailing that a certain ‘end state’ will ensue for many of the verbs in this category. Barz’ corresponding category is termed ‘opposing event, withdrawal or negation of an event’ (2006: 704) . 7 This category corresponds to Kühnhold and Wellmann’s category ent- 2 (1973: 148), which expresses ‘beginning’. Barz (2006: 704) refers to such examples as ‘ingressive’. 8 This category corresponds to Kühnhold and Wellmann’s category er- 3 (1973: 148), which expresses ‘beginning’. Barz (2006: 704) refers to such examples as ‘ingressive’. 9 Kühnhold and Wellmann (1973: 148-149) subdivide this category into: ‘successful completion’ (their er- 1 , e.g., erarbeiten ‘develop/ work out’) and ‘complete realization and completion’ (their er- 5 , e.g. erwürgen ‘strangle’). Since the semantic differences are relatively small and Barz (2006: 704) refers to all such examples as egressive, this more inclusive category was adopted here. 10 According to Kühnhold and Wellmann’s er- 2 (1973: 148-149), this category ‘focuses on the end state’. However, this final state is more specifically described by the adjective involved in the prefix verb formation, or as Barz (2006: 705) puts it, by ‘making something [adjective]’. This notion is extended in the present category to ‘become (more) [adjective]’. Incidentally, the examples given by Kühnhold and Wellman also fulfil this criterion (erblinden ‘go blind’, erfrischen ‘refresh/ freshen’). <?page no="257"?> 244 11 The two present categories über 1 and über 2 are subsumed under one heading (über- 1 ) in Kühnhold and Wellmann (1973: 150). Their category signals the ‘localization above something else or exceeding a limit or superiority’. In the present systematization the ‘localization/ direction’ component and the ‘limitexceeding’ component (or aspect that is ‘divergent from norm’, Barz 2006: 704) serve to define the two separate categories. 12 According to Barz (2006: 704) this prefix expresses an ‘opposing event, the withdrawal or negation of an event’. 13 This category corresponds to Kühnhold and Wellmann’s über- 2 (1973: 150), which expresses ‘action over a certain time span’. 14 This category corresponds to Kühnhold and Wellmann’s um- 1 (1973: 150), which expresses ‘action or position around something’. Barz (2006: 702-703) calls the resulting change in semantic roles and syntactic functions ‘incorporation’. 15 This category corresponds to Kühnhold and Wellmann’s class ver- 4 (1973: 152), which expresses ‘wrong execution of an action’. It further reflects Barz’s interpretation as a ‘wrong, unsuccessful event’ (2006: 704), and Fleischer and Barz’s definition as ‘devious or faulty execution of an action’ (1995: 326). 16 This category differs from Kühnhold and Wellmann’s (1973: 152) classification. They categorize verbs like verreisen ‘travel’ as verbs signalling ‘removal’. The present classification instead focuses on the ‘extended period of time covered’ by all the verbs. 17 The verbs in this category resemble those in Kühnhold and Wellmann’s class ver- 2 (1973: 152), which focuses on the ‘end state’ of an action. This definition is reflected in the present supercategory ‘change of state’, to which ver 3a belongs. Barz (2006: 705) distinguishes ‘making something [adjective]’ and ‘becoming [adjective]’. In the present classification, these two categories are subsumed under the meaning ‘become (more) [adjective]’. Incidentally, the examples given by Kühnhold and Wellmann’s for their ver- 4 category are also derived from adjectives (veralten ‘become old/ age’, verdicken ‘thicken’) and thus match my definition. The present category in many cases also comprises semantic components of Kühnhold and Wellmann’s class ver- 3 , which features the ‘realization or occurrence of a contact’ as in vergolden ‘make golden’. 18 The verbs in this category closely resemble those in Kühnhold and Wellmann’s ver- 1 class (1973: 151), which expresses the ‘complete execution of an action’. The same understanding is found in Fleischer and Barz (1995: 325), who also extend it to include ‘complete, purposeful depletion of material’ in their definition. Barz (2006: 704) refers to such examples as ‘egressive’. 19 Many members of this category include a semantic component ‘realization or occurrence of a contact’ of Kühnhold and Wellmann’s class ver- 3 (1973: 152), e.g., verreiben ‘spread’, vereinen ‘unite’. According to Fleischer and Barz (1995: 325) verbs of this class express a ‘connection’. Extending both these definitions, the meaning of the present category is ‘action or event describing the relation between two (or more) components’. 20 Fleischer and Barz (1995: 327) characterize the prefix meaning of the present category zer 1 as ‘divide or shred’. If the meaning component of ‘apart/ into pieces’ is already incorporated in the simplex verb semantics, this meaning is intensified. Barz (2006: 704) refers to zermore generally as expressing a ‘destructive event’. <?page no="258"?> 245 21 Fleischer and Barz (1995: 327) refer to the meaning of the present prefix zer 2 as ‘damaging’. Verbs belonging in this category do not carry the semantic component of ‘division into pieces’ in their simplex form. A ‘damaging or destructive change of state’ is expressed. Barz (2006: 704) refers to zermore generally as expressing a ‘destructive event’. I.4 Search results for manually selected verb prefix constructions in English and German Table c. Prefix constructions, prefix verb types and tokens used by Thomas and his caregiver. Prefixes Caregiver Thomas Thomas over- 4 overload 2 overheat re- 32 rewind 2 rewind 12 recycle 7 recycle 4 reuse 2 rebuild 2 redress 2 rename un- 30 undo 8 undo 16 undress 14 unwrap 8 unplug 6 unload 6 unroll 6 untwist 6 unzip 5 unpack 2 unpack 4 undecorate 4 unsettle 2 unhook 2 unhook 2 unlock 2 unlock 2 unset 2 unstick 2 untie 2 unopen Note: The form marked in blue print is a non-canonical use by Thomas. <?page no="259"?> 246 Table d. Prefix constructions, prefix verb types and tokens used by Leo and his caregiver. Caregiver Leo Leo bebe 1 - 25 beladen 2 beladen 14 bekleben 1 bekleben 11 beschriften 4 beschriften 10 bemalen 3 bemalen 6 beschreiben 5 bestreichen 5 bedecken 5 belegen 2 belegen 4 bewölken 1 bewölken 3 besprühen 3 beziehen (einbeziehen) 2 bekleckern 1 bekleckern 2 betanken 1 betanken 2 beschmeißen 2 begucken 1 besteigen 2 besteigen 1 beschmieren 1 beschmieren 1 bedrucken 1 bespannen 1 bestreuen 1 besaufen 1 befeuchten 2 beschneien befärben be 2 - 26 bezahlen 7 bezahlen 16 bedanken 14 beenden 4 beenden 10 beeilen 4 beeilen 9 besetzen 7 besetzen 9 bewundern 2 bewundern 8 bedenken 4 bearbeiten 4 besprechen 3 bezweifeln 2 beantworten 2 bedienen 2 bedienen 2 belohnen 1 bedrohen 2 bedrohen 1 bestaunen <?page no="260"?> 247 1 bewachen be 3 - 79 benutzen 63 benutzen 14 befürchten 1 befürchten 13 begrüßen 2 beschützen 2 beschützen 2 begraben 2 befühlen 1 besagen bewinken de- 2 denunzieren 1 degenerieren durch- 3 durchlöchern 1 durchfeuchten 1 durchleuchten eis- 2 eisschwimmen 2 eisschwimmen 2 eisbaden entent 1 - 40 entgleisen 15 entgleisen 17 entfernen 6 entfernen 6 entspannen 4 entgehen 3 entleeren 1 entladen 1 entstauben 1 entwerten ent 2 - 8 entstehen 4 entstehen 1 entspringen 1 entspringen 1 entzünden erer 1 - 20 erschrecken 15 erschrecken 2 erzittern 2 erwachen 1 erleuchten er 2 - 44 erkennen (wiedererkennen) 9 erkennen 8 erraten 2 erraten 3 erwürfeln 2 erschlagen <?page no="261"?> 248 1 erdrücken 1 erhängen 1 ersticken 1 ersticken er 3 - 13 erweitern 5 erhöhen 1 erleichtern 2 erneuern ge- 82 gefallen 8 gefallen 10 gelingen 4 gebrauchen 4 gebrauchen 3 gestalten überüber 1 - 9 überbacken 6 überziehen 11 überziehen 6 überfahren 2 überfahren 5 übermalen 8 übermalen 5 überholen 2 überholen 2 übersetzen 2 übersetzen 1 übertünchen 2 übertünchen 1 überfliegen über 2 - 9 überleben 8 übertreiben 2 überfressen 2 übertreffen 2 überschneiden über 3 - 10 überweisen 2 überweisen 7 übernehmen 5 übertragen 2 übertragen 2 übergeben 2 übermitteln 2 überlassen 1 überreichen 1 überbringen über 4 - 3 überreden 1 überwachen 1 überzeugen 1 überprüfen über 5 - 6 übersehen <?page no="262"?> 249 1 überhören über 6 - 2 überwintern 2 überstehen 1 übernachten 1 übernachten um- 2 umgeben 1 umkreisen 1 umkreisen 1 umkurven 1 umsegeln 2 umschwirren 2 umhüllen 1 umkleben verver 1 - 18 verlaufen 9 verlaufen 16 verfahren 6 verfahren 13 verpassen 6 verpassen 10 verlegen 8 verrutschen 14 verrutschen 6 verschlucken 1 verschlucken 3 vertun 3 verschreiben 2 verschlafen 2 verpennen 2 verpennen 2 verhunzen 2 verwechseln 2 verwechslern 1 verschütten 1 vermackeln 1 verwackeln 1 verschlängeln verschneiden verkrickelkrackeln ver 2 - 9 verreisen 5 verreisen 3 verbringen 2 verbringen 2 verschicken ver 3a - 17 verlängern 4 verlängern 4 verstärken 2 verstärken 3 vertrocknen 3 vertrocknen 3 verkleinern 2 verheilen 1 verheilen 2 verfeinern <?page no="263"?> 250 2 verkürzen 2 verschönern 2 versteifen 2 verwelken 1 vergrößern 1 vergrößern 1 verdicken 1 verdickern 1 verkräuseln 2 verschwierigen ver 3b - 28 verbrauchen 19 verbrauchen 15 verwandeln (umverwandeln, zurückverwandeln) 2 verwandeln 12 verbrennen 8 verbrennen 10 verdecken 5 verdecken 9 verhungern 4 verhungern 7 verbauen 3 verbauen 7 verkleiden 2 verkleiden 6 verändern 2 verändern 6 verschieben 5 verzieren 1 verzieren 5 verbiegen 4 versenken (reinversenken) 1 versenken 3 verspeisen 3 versickern 2 verarbeiten 2 verarbeiten 2 verdrücken 2 verdursten 2 versinken 2 vernichten 1 verhüllen 2 verhüllen 1 verkriechen 2 verkriechen 1 verzaubern 1 verpesten 1 versteinern 1 vergraben 1 verwunden 1 verkommen 1 verputzen 1 verqualmen 1 verschnupfen vertarnen ver 3c - 33 verbinden 10 verbinden 30 verteilen 8 verteilen 6 vermischen 2 vermischen 4 verreiben 1 verreiben <?page no="264"?> 251 3 verschmieren 1 verschmieren 2 verkeilen 2 verknoten 1 verknoten 2 verstreuen 2 vermehren 1 verdrehen 1 vereinen 1 verkleben 1 vermengen 1 verhaken 1 verrühren zerzer 1 - 13 zerreißen 11 zerreißen 9 zerschneiden 6 zerschneiden (durchzerschneiden) 3 zerbrechen 1 zerbrechen 3 zerbröseln 1 zerbröseln 3 zerteilen 2 zertrümmern 2 zertrümmern 2 zerlegen 1 zerlegen 2 zerkleinern 1 zerkleinern 2 zerfließen 1 zerschnippeln 2 zerschnippeln 1 zerpflücken 1 zerpflücken 1 zersägen 1 zersägen 1 zerfetzen zer 2 - 9 zerstören 11 zerstören 2 zerdrücken 2 zerdrücken 2 zerknittern 1 zerknittern 1 zerstechen 1 zerstechen 1 zerkochen 2 zerspringen zerstürzen Note: Forms marked in blue print are non-canonical uses by Leo. Forms marked in purple are also non-canonical forms Leo used, but they occurred in additional files that were not included in the counts. They are nevertheless included in the list (without their number of occurrence though), because their non-canonicality is expected to be relevant to the development of prefix constructions. The forms are discussed in 5.2.2. For English translations of verb meanings, please see Table b. <?page no="265"?> 252 Table e. Counts of verb types and tokens per prefix construction as produced by Thomas’ caregiver (CG) and Thomas (excluding innovative uses). CG prefix verb types Thomas prefix verb types CG prefix verb tokens Thomas prefix verb tokens un[base] verb 16 4 115 14 re[base] verb 6 2 54 9 over[base] verb 2 6 Construction types 3 2 Table f. Counts of verb types and tokens per prefix construction as produced by Leo’s caregiver (CG) and Leo (excluding innovative uses). CG prefix verb types Leo prefix verb types CG prefix verb tokens Leo prefix verb tokens ver 3b [base] verb 30 13 ver 3b [base] verb 141 53 be 1 [base] verb 20 13 be 3 [base] verb 113 66 be 2 [base] verb 16 7 be 2 [base] verb 112 28 ver 1 [base] verb 16 6 be 1 [base] verb 104 22 ver 3c [base] verb 13 8 ge[base] verb 99 12 zer 1 [base] verb 13 10 ver 1 [base] verb 90 38 ver 3a [base] verb 12 6 ver 3c [base] verb 88 25 ent 1 [base] verb 8 2 ent 1 [base] verb 73 21 be 3 [base] verb 7 3 er 2 [base] verb 60 12 er 2 [base] verb 7 3 zer 1 [base] verb 43 27 über 1 [base] verb 7 7 ver 3a [base] verb 41 12 über 3 [base] verb 6 4 über 1 [base] verb 34 28 über 2 [base] verb 5 0 über 3 [base] verb 28 6 zer 2 [base] verb 5 5 er 1 [base] verb 25 15 er 1 [base] verb 4 1 über 2 [base] verb 23 0 ge[base] verb 4 2 er 3 [base] verb 19 2 um[base] verb 4 4 zer 2 [base] verb 15 17 durch[base] verb 3 0 ver 2 [base] verb 14 7 ent 2 [base] verb 3 2 ent 2 [base] verb 10 5 <?page no="266"?> 253 er 3 [base] verb 3 1 über 5 [base] verb 7 0 über 4 [base] verb 3 1 durch[base] verb 5 0 über 6 [base] verb 3 1 über 4 [base] verb 5 1 ver 2 [base] verb 3 2 über 6 [base] verb 5 1 de[base] verb 2 0 um[base] verb 5 6 über 5 [base] verb 2 0 de[base] verb 3 0 eis[base] verb 1 2 eis[base] verb 2 4 Construction types 26 22 <?page no="267"?> 254 II Chapter 6 II.1 One-sample t-tests assessing novel construction learning Table g. Act-out task. T-test values by age group and frequency level (German, prefix). Item frequency level Age t df p level a (introduced in training film) 3 2.92 15 < .011* 4 12.64 15 < .001* 5 9.50 15 < .001* 6 17.75 15 < .001* 8 47.00 15 < .001* level b (new in act-out task) 3 1.00 15 < .334 4 2.83 15 < .013* 5 3.62 15 < .003* 6 9.13 15 < .001* 8 16.19 15 < .001* Table h. Act-out task. T-test values by age group and frequency level (English, prefix). Item frequency level Age t df p level a (introduced in training film) 4 4.50 15 < .001* 5 4.94 15 < .001* 6 9.34 15 < .001* level b (new in act-out task) 4 5.58 15 < .001* 5 3.29 15 < .006* 6 8.77 15 < .001* <?page no="268"?> 255 Table i. Forced-choice task. T-test values by age group and frequency level (German, prefix). Item frequency level Age df t 0.5 p 0.5 t 0.1 p 0.1 level a (introduced in 3 15 -0.82 < .431 1.77 < .100 training film) 4 15 3.91 < .002* 7.27 < .001* 5 15 15.00 < .001* 24.45 < .001* 6 15 15.00 < .001* 24.45 < .001* 8 children at ceiling level b (introduced in act- 3 15 -1.00 < .334 2.15 < .049* out task) 4 15 2.76 < .015* 5.67 < .001* 5 15 6.71 < .001* 11.99 < .001* 6 children at ceiling 8 children at ceiling level c (new in forced- 3 15 -2.41 < .030* 0.84 < .414 choice task) 4 15 0.00 = 1 3.23 < .006* 5 15 3.87 < .002* 8.45 < .001* 6 15 5.00 < .001* 9.72 < .001* 8 15 15.00 < .001* 24.45 < .001* Note: t 0.5 refers to t-values in the comparison against 0.5, p 0.5 refers to the corresponding p-values. t 0.1 refers to t-values in the comparison against 0.1, p 0.1 refers to the corresponding p-values. Table j. Forced-choice task. T-test values by age group and frequency level (English, prefix). Item frequency level Age df t 0.5 p 0.5 t 0.1 p 0.1 level a (introduced in 4 15 -0.82 < .424 1.77 < .098 training film) 5 15 1.57 < .139 4.03 < .002* 6 15 3.48 < .004* 6.76 < .001* level b (introduced in act- 4 15 -1.69 < .111 0.97 < .346 out task) 5 15 1.32 < .207 3.82 < .002* <?page no="269"?> 256 6 15 2.76 < .015* 5.67 < .001* level c (new in forced- 4 15 -2.41 < .030* 0.84 < .414 choice task) 5 15 0.29 < .774 3.06 < .008* 6 15 1.78 < .100 5.13 < .001* Note: t 0.5 refers to t-values in the comparison against 0.5, p 0.5 refers to the corresponding p-values. t 0.1 refers to t-values in the comparison against 0.1, p 0.1 refers to the corresponding p-values. Table k. Production task. T-test values by age group and frequency level (German, prefix). Item frequency level Age t df p level a (introduced in training film) 3 3.09 15 < .008* 4 4.47 15 < .001* 5 12.12 15 < .001* 6 31.00 15 < .001* 8 31.00 15 < 001* level b (introduced in act-out task) 3 2.24 15 < .041* 4 4.14 15 < .001* 5 10.50 15 < .001* 6 15.65 15 < .001* 8 15.00 15 < .001* level c (introduced in forced-choice task) 3 1.00 15 < .334 4 3.47 15 < .004* 5 7.65 15 < .001* 6 6.33 15 < .001* 8 12.12 15 < .001* level d (new in production task) 3 1.00 15 < .334 4 1.78 15 < .100 5 3.50 15 < .004* 6 4.33 15 < .001* 8 11.62 15 < .001* <?page no="270"?> 257 Table l. Production task. T-test values by age group and frequency level (English, prefix). Item frequency level Age t df p level a (introduced in training film) 4 1.86 15 < .083 5 3.00 15 < .009* 6 5.22 15 < .001* level b (introduced in act-out task) 4 2.45 15 < .028* 5 3.66 15 < .003* 6 5.55 15 < .001* level c (introduced in forced-choice task) 4 2.15 15 < .049* 5 2.91 15 < .011* 6 3.66 15 < .003* level d (new in production task) 4 1.46 15 < .164 5 2.08 15 < .056 6 3.15 15 < .007* Table m. Act-out task. T-test values by age group and frequency level (German, reduplication). Item frequency level Age t df p level a (introduced in training film) 3 2.42 15 < .029* 4 5.22 15 < .001* 5 6.45 15 < .001* 6 21.10 15 < .001* 8 children at ceiling level b (new in act-out task) 3 children at floor 4 2.18 15 < .046* 5 2.55 15 < .023* 6 11.86 15 < .001* 8 14.35 15 < .001* <?page no="271"?> 258 Table n. Act-out task. T-test values by age group and frequency level (English, reduplication). Item frequency level Age t df p level a (introduced in training film) 4 4.46 15 < .001* 5 5.37 15 < .001* 6 4.42 15 < .001* level b (new in act-out task) 4 2.18 15 < .046* 5 3.34 15 < .005* 6 3.16 15 < .007* Table o. Forced-choice task. T-test values by age group and frequency level (German, reduplication). Item frequency level Age df t 0.5 p 0.5 t 0.1 p 0.1 level a (introduced in train 3 15 -1.86 < .083 1.07 < .303 ing film) 4 15 1.07 < .300 3.61 < .003* 5 15 1.29 < .217 4.34 < .001* 6 children at ceiling 8 children at ceiling level b (introduced in act- 3 15 -2.74 < .016* 0.49 < .628 out task) 4 15 0.62 < .545 3.55 < .003* 5 15 0.52 < .610 2.99 < .010* 6 15 10.25 < .001* 17.16 < .001* 8 15 15.00 < .001* 24.45 < .001* level c (new in forced-choice 3 15 -4.04 < .001* -0.22 < .826 task) 4 15 -2.78 < .014* 0.97 < .347 5 15 -0.32 < .751 2.74 < .016* 6 15 2.78 < .014* 6.53 < .001* 8 15 15.00 < .001* 24.45 < .001* <?page no="272"?> 259 Note: t 0.5 refers to t-values in the comparison against 0.5, p 0.5 refers to the corresponding p-values. t 0.1 refers to t-values in the comparison against 0.1, p 0.1 refers to the corresponding p-values. Table p. Forced-choice task. T-test values by age group and frequency level (English, reduplication). Item frequency level Age df t 0.5 p 0.5 t 0.1 p 0.1 level a (introduced in train 4 15 -3.48 < .004* -0.19 < .850 ing film) 5 15 -1.17 < .262 1.59 < .133 6 15 0.52 < .610 2.99 < .010* level b (introduced in act- 4 15 -5.20 < .001* -1.11 < .287 out task) 5 15 -2.15 < .049* 0.75 < .465 6 15 0.00 = 1 2.44 < .028* level c (new in forced-choice 4 15 -4.57 < .001* -0.65 < .529 out task) 5 15 -1.96 < .069 0.69 < .504 6 15 0.00 = 1 2.64 < .019* Note: t 0.5 refers to t-values in the comparison against 0.5, p 0.5 refers to the corresponding p-values. t 0.1 refers to t-values in the comparison against 0.1, p 0.1 refers to the corresponding p-values. Table q. Production task. T-test values by age group and frequency level (German, reduplication). Item frequency level Age t df p level a (introduced in training film) 3 1.73 15 < .104 4 3.66 15 < .003* 5 5.51 15 < .001* 6 21.96 15 < .001* 8 31.00 15 < 001* level b (introduced in act-out task) 3 1.00 15 < .334 4 2.74 15 < .016* <?page no="273"?> 260 5 3.15 15 < .007* 6 14.10 15 < .001* 8 15.00 15 < .001* level c (introduced in forced-choice task) 3 1.00 15 < .334 4 2.24 15 < .041* 5 2.08 15 < .056 6 7.32 15 < .001* 8 11.21 15 < .001* level d (new in production task) 3 1.46 15 < .164 4 1.86 15 < .083 5 1.38 15 < .189 6 4.47 15 < .001* 8 9.14 15 < .001* Table r. Production task. T-test values by age group and frequency level (English, reduplication). Item frequency level Age t df p level a (introduced in training film) 4 1.38 15 < .189 5 2.08 15 < .056 6 3.42 15 < .004* level b (introduced in act-out task) 4 1.46 15 < .164 5 1.38 15 < .189 6 3.31 15 < .005* level c (introduced in forced-choice task) 4 1.00 15 < .334 5 1.46 15 < .164 6 3.31 15 < .005* level d (new in production task) 4 1.00 15 < .334 5 children at floor 6 2.78 15 < .014* <?page no="274"?> 261 II.2 Proportions of correct responses on real items by task, pattern and language Figure a. Forced-choice task (German, prefix, reduplication). Proportions of correct responses on real items by age group. Figure b. Forced-choice task (English, prefix, reduplication). Proportions of correct responses on real items by age group. 3 yrs 4 yrs 5 yrs 6 yrs 8 yrs 3 yrs 4 yrs 5 yrs 6 yrs 8 yrs Age 4 yrs 5 yrs 6 yrs 4 yrs 5 yrs 6 yrs Age <?page no="275"?> 262 Figure c. Production task (German, prefix, reduplication). Proportions of correct responses on real items by age group. Figure d. Production task (English, prefix, reduplication). Proportions of correct responses on real items by age group. 3 yrs 4 yrs 5 yrs 6 yrs 8 yrs 3 yrs 4 yrs 5 yrs 6 yrs 8 yrs Age 4 yrs 5 yrs 6 yrs 4 yrs 5 yrs 6 yrs Age <?page no="276"?> 263 II.3 Memorization and generalization of the novel construction by task and pattern Table s. Act-out, forced-choice and production tasks. Number of children who memorized and generalized 0 to x verbs in the novel prefix construction. Act-out task Number of memorized verbs Number of generalizations 0 1 2 3 4 0 18 4 0 0 0 1 8 2 0 0 0 2 12 9 0 2 1 3 8 7 7 3 0 4 3 7 23 11 3 Forced-choice comprehension task Number of memorized verbs Number of generalizations 0 1 2 0 17 4 0 1 6 0 0 2 2 5 1 3 0 7 5 4 4 29 48 Production task Number of memorized verbs Number of generalizations 0 1 2 0 38 0 0 1 8 0 0 2 6 3 0 3 7 0 1 4 8 3 0 5 4 7 3 6 6 16 18 Note: N German = 80; N English = 48. <?page no="277"?> 264 Table t. Act-out, forced-choice and production tasks. Number of children who memorized and generalized 0 to x verbs in the novel reduplication construction. Act-out task Number of memorized verbs Number of generalizations 0 1 2 3 4 0 33 1 0 0 0 1 17 2 0 0 0 2 8 2 1 0 0 3 6 10 7 0 0 4 6 7 15 8 5 Forced-choice comprehension task Number of memorized verbs Number of generalizations 0 1 2 0 29 8 0 1 15 1 1 2 4 0 2 3 3 10 3 4 3 15 34 Production task Number of memorized verbs Number of generalizations 0 1 2 0 59 0 0 1 6 2 0 2 8 0 0 3 6 1 0 4 5 4 0 5 5 6 4 6 1 8 13 Note: N German = 80; N English = 48. <?page no="278"?> 265 II.4 Novel prefix and reduplication learning in Germanspeaking adults German-speaking adults’ ability to learn the novel prefix and the novel reduplication construction were assessed in order to determine whether performance would continue to increase with age and finally reach ceiling. Small adjustments in the design were made. Thirty-six undergraduates (8 males, 28 females), aged between 19 and 34 (M=23), were recruited from Ludwig-Maximilians-Universität, Munich, Germany. Adults were tested in two groups. The novel constructions (prefix or reduplication) were randomly assigned the groups. None of the adults spoke a foreign language that uses reduplication. The training film was projected to a classroom wall using a beamer. The act-out task was performed by the experimenter for adult participants to watch. Clips for the forced-choice comprehension and production tasks were also projected to the wall. Participants responded in writing. Proportions of correct responses in the forced-choice task and the production task are given in Tables u and v. T-tests were not performed, since adults were at ceiling, which entails that there was hardly any variance at all. Neither were mixed-effects logistic regression models fitted to the forced-choice data or to the production data. Again the reason was that all adults performed at ceiling, regardless of their age (in years), the construction they learned (prefix/ reduplication), the frequency level and the items (real/ pretence). It was concluded that adults are able to learn either novel construction from the input after very limited exposure. There was no difference between the two constructions. It was thus shown that performance in the learning of both novel morphological constructions continues to increase with age. Table u. Proportions of correct responses and standard errors (SE) for Germanspeaking adults in forced-choice task. Pattern Prefix Reduplication Construction Construction Frequency level pretence real pretence real Level a 1 (0) 0.969 (0.022) 0.975 (0.020) 1 (0) Level b 1 (0) 1 (0) 0.925 (0.043) 0.95 (0.027) Level c 0.969 (0.022) 1 (0) 0.975 (0.020) 0.975 (0.020) <?page no="279"?> 266 Table v. Proportions of correct responses and standard errors (SE) for Germanspeaking adults in production task. Pattern Prefix Reduplication Construction Construction Frequency level pretence real pretence real Level a 1 (0) 0.969 (0.022) 1 (0) 1 (0) Level b 1 (0) 0.938 (0.030) 1 (0) 0.975 (0.020) Level c 0.969 (0.022) 0.906 (0.036) 0.975 (0.020) 0.950 (0.027) Level d 0.938 (0.030) 0.969 (0.022) 0.975 (0.020) 1 (0) <?page no="280"?> According to usage-based, constructionist accounts the linguistic input in general and input frequencies in particular play an important role in children’s language learning. English-speaking children have been shown to be able to learn entirely novel, invented word order constructions from their input. This book aims to extend this line of research to the area of morphology. Two experimental studies investigate German-speaking and English-speaking children’s ability to learn novel morphological constructions from the input. The effects of input frequencies on this learning process are examined in detail. A corpus study provides the morphological background data for the invented constructions and presents additional support for frequency effects from naturalistic language learning. By combining two empirical methods and by exploring morphological learning in two different languages this book provides new insights into the cognitive processes that are assumed to be involved in children’s language learning and reveals how these processes are affected by different kinds of input frequencies.
