Thomas Wagner Interlanguage Morphology Irregular Verbs in the Mental Lexicon of German-English Interlanguage Speakers Language in Performance LiP Interlanguage Morphology 42 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 (Hannover), Andrea Sand (Trier), Hans-Jörg Schmid (München), Josef Schmied (Chemnitz) and Edgar W. Schneider (Regensburg) Thomas Wagner Interlanguage Morphology Irregular Verbs in the Mental Lexicon of German-English Interlanguage Speakers Bibliografische Information der Deutschen Nationalbibliothek Die Deutsche Nationalbibliothek verzeichnet diese Publikation in der Deutschen Nationalbibliografie; detaillierte bibliografische Daten sind im Internet über <http: / / dnb.d-nb.de> abrufbar. Vorliegende Arbeit wurde als Dissertation an der Universität Siegen angenommen. © 2010 · 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 Printed in Germany ISSN 0939-9399 ISBN 978-3-8233-6547-1 Mazes intricate, Eccentric, intervolv'd, yet regular Then most, when most irregular they seem. John Milton (1608-1674): Paradise Lost, v. 622-4 To my parents 7 Table of contents Figures and tables 9 Acknowledgements 16 Preface 18 1 Introduction 19 2 Irregular verbs - a linguistic challenge 25 2.1 Introduction 25 2.2 Irregular verbs in English and German - a classic chestnut 25 2.3 One or two routes - irregular verbs and the dual mechanism 34 2.4 Good verbs, better verbs, best verbs - irregular verbs and prototype theory 36 2.5 Nodes, connections, and weights - irregular verbs and connectionism 41 2.6 Exemplars, nothing but exemplars - irregular verbs and analogical models 46 2.7 Rules, ever more general rules - irregular verbs and ‘islands of reliability’ 49 2.8 frinked or frank - irregular verbs and the role of semantics 53 2.9 sing sang sung - vowel change classes or universal apophony 56 2.10 Irregular verbs and L2 acquisition 59 2.11 Desiderata and hypotheses 63 3 Experiment 1 67 3.1 Experimental design 67 3.2 Results 74 3.2.1 Experiment 1_G_L1 (Native speakers of German) 74 3.2.2 Experiment 1_E_L1 (Native speakers of English) 82 3.2.3 Experiment 1_E_L2 (German-English interlanguage speakers) 89 3.3 Summary and discussion 97 8 4 Experiment 2 103 4.1 Experimental design 103 4.2 Results 105 4.2.1 Experiment 2_G_L1a (Native speakers of German) 106 4.2.2 Experiment 2_G_L1b (Native speakers of German) 112 4.2.3 Experiment 2_E_L1a (Native speakers of English) 118 4.2.4 Experiment 2_E_L1b (Native speakers of English) 123 4.2.5 Experiment 2_E_L2a (German-English interlanguage speakers) 127 4.2.6 Experiment 2_E_L2b (German-English interlanguage speakers) 132 4.3 Summary and discussion 136 5 Experiment 3 141 5.1 Experimental design 142 5.2 Results 145 5.2.1 Experiment 3_G_L1_oral (Native speakers of German) 145 5.2.2 Experiment 3_E_L1_oral (German-English interlanguage speakers) 150 5.2.3 Experiment 3_E_L2_written (German-English interlanguage speakers) 154 5.3 Summary and discussion 156 6 Discussion and conclusion 159 References 167 Appendix 180 9 Figures and Tables Figures 2 Irregular verbs - a linguistic challenge title of figure 1. The dual-route model for verb inflection 35 2. Architecture of a simple feed-forward (PDP) model representing the past tense formation of the verb string 42 3. Architecture of a PDP model as used by Joanisse and Seidenberg (1999) 45 4. Architecture of a memory-based learner software, illustrated with an example for English verb morphology 48 3 Experiment 1 title of figure 1. Mosaic plots for German verbs cross-classified by past tense formation and the three constituents onset, nucleus, and coda in 1_G_L1 75 2. Mosaic plots for German verbs cross-classified by vowel change and the three constituents onset, nucleus, and coda in 1_G_L1 79 3. Mosaic plots for English verbs cross-classified by past tense formation and the two constituents onset and coda in 1_E_L1 82 4. Mosaic plots English verbs cross-classified by vowel change and the two constituents onset and coda in 1_E_L1 85 5. Classification analysis showing the interaction of onset and coda with the types of vowel change in 1_E_L1 86 6. Mosaic plots for interlanguage verbs cross-classified by past tense formation and the two constituents onset and coda in 1_E_L2 90 7. Mosaic plots for interlanguage verbs cross-classified by vowel change and the two constituents onset and coda in 1_E_L2 92 8. Classification analysis showing the interaction of onset and coda with the types of vowel change in 1_E_L2 93 9. Mosaic plot for English verbs cross-classified by past tense formation and the two types of speaker (interlanguage speaker, L1 speaker) 94 10. Mosaic plot for English verbs cross-classified by past tense and type of speaker in 1_E_L2 96 10 4 Experiment 2 title of figure 11. Mosaic plots for German verbs cross-classified by past-participlesequence and the three constituents onset, nucleus, and coda in 2_G_L1a 107 12. Classification analysis showing the interaction of participle inflection with the resulting past-participle-sequence in 2_G_L1a 108 13. Mosaic plots for German verbs cross-classified by infinitiveparticiple-sequence and the three constituents onset, nucleus, and coda in 2_G_L1b 113 14. Classification analysis showing the interaction of all three constituents with the resulting infinitive-participle-sequence in 2_G_L1b 114 15. Classification analysis showing the interaction of participle inflection with the resulting overall infinitive-participle-sequence in 2_G_L1b 115 16. Mosaic plots for English verbs cross-classified by past-participlesequence and the two constituents onset and coda in 2_E_L1a 119 17. Classification analysis showing the interaction of past tense and participle responses with the resulting past-participle-sequence in 2_E_L1a 121 18. Mosaic plots for English verbs cross-classified by past-infinitivesequence and the two constituents onset and coda in 2_E_L1b 124 19. Classification analysis showing the interaction of past tense and infinitive responses with the resulting past-infinitive-sequence in 2_E_L1b 125 20. Mosaic plots for interlanguage verbs cross-classified by pastparticiple-sequence and the two constituents onset and coda in 2_E_L2a 128 21. Classification analysis showing the interaction of past tense and participle inflection with the resulting past-participle-sequence in 2_E_L1a 129 22. Mosaic plot for English verbs cross-classified by past-participlesequence and type of speaker 131 23. Mosaic plots for interlanguage verbs cross-classified by pastinfinitive-sequences and the two constituents onset and in 2_E_L2b 133 24. Classification analysis showing the interaction of past tense inflection with the resulting past-infinitive-sequence in 2_E_L2b 134 25. Mosaic plot for English verbs cross-classified by past-infinitivesequence and type of speaker 136 11 5 Experiment 3 title of figure 26. Mosaic plot for nonce-inflections cross-classified by type of inflection and contexts in which the nonce was presented in 3_G_L1_oral 146 27. Mosaic plot for nonce-inflections cross-classified by type of inflection and contexts in which the nonce was presented in 3_E_L2_oral 151 28. Mosaic plots for participants’ inflection cross-classified by all three responses in 3_E_L2_oral 154 29. Mosaic plot for participants’ inflection cross-classified by contexts in which the nonce was presented in 3_E_L2_written 155 30. Mosaic plot of participants’ inflection cross-classified by the type of experiment (oral and written) for German-English interlanguage speakers 158 Tables 2 Irregular verbs - a linguistic challenge title of table 1. Selection of ablaut patterns for past tense and participle in English and German 27 2. Verbs featuring stem vowel [ ] along with their past tense formations and examples 28 3. Categorisations of German irregular verbs 29 4. Main results of Bybee and Moder (1983, English L1) and Plag (2000, German-English interlanguage) 60 3 Experiment 1 title of table 1. Ranking of possible schemas based on the analyses of Köpcke 1998 70 2. Prototypicality of potential German nonce-verbs 71 3. Ranking of potential English nonce-verbs 72 4. Main results of Bybee and Moder 1993, Plag 2000, and the pilot study 2003 73 5. Frequencies for types of past tense formation with or without vowel change in 1_G_L1 74 6. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda as predictors and types of past tense as response variable in 1_G_L1 77 7. Frequencies for the types of vowel change in 1_G_L1 78 12 8. Ranking of test items attracting most of the vowel-changing responses in 1_G_L1 79 9. Frequencies for types of past tense formation with or without vowel change in 1_E_L1 82 10. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda as predictors and types of past tense as response variable in 1_E_L1 83 11. Frequencies for the types of vowel change in 1_E_L1 84 12. Ranking of test items attracting most of the vowel-changing responses in 1_E_L1 84 13. Parallels between the morphological make-up of those test items attracting most of the vowel-changing responses and the statistically significant interactions of the constituents as revealed by a classification analysis in 1_E_L1 87 14. Frequencies for types of past tense formation with or without vowel change in 1_E_L2 89 15. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda as predictors (pooled) and types of past tense as response variable in 1_E_L2 91 16. Frequencies for the types of vowel change in 1_E_L2 91 17. Frequencies for vowel-changing test items in 1_E_L2 92 18. Frequencies for the types of vowel change in 1_E_L1 and 1_E_L2 95 19. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset, nucleus, and coda as predictors and types of past tense formation as response variable in Plag 2000 99 4 Experiment 2 title of table 1. Frequencies for the types of past-participle-sequence with or without vowel change in 2_G_L1a 106 2. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda (pooled) as interacting predictors, nucleus as single main predictor, and types of past-participle-sequence as response variable in 2_G_L1a 107 3. Frequencies for all vowel change patterns in the past-participlesequences in 2_G_L1a 109 4. Highest frequencies for responses with existing, novel, and no vowel changes in 2_G_L1a 110 5. Frequencies for types of past tense and participle formation to a given infinitive nonce in 2_G_L1a 110 13 6. Frequencies for the types of infinitive-participle-sequence with or without vowel change in 2_G_L1b 112 7. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with interacting onset and (pooled) coda as predictors and types of infinitive-participlesequence as response variable in 2_G_L1b 113 8. Frequencies for types of vowel change in the overall infinitiveparticiple-sequences in 2_G_L1b 116 9. Highest frequencies for responses with existing, novel, and no vowel changes in 2_G_L1b 116 10. Contingency table of infinitive by participle formation in 2_G_L1b 117 11. Frequencies for the types of past-participle-sequence with or without vowel change in 2_E_L1a 118 12. Beta coefficients and significance values for the fixed effects and their interactions of a generalised linear mixed model via PQL with onset and coda (pooled) as interacting predictors and the types of past-participle-sequence as response variable in 2_E_L1a 119 13. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda (all variants) as predictors and the types of past-participle-sequence as response variable in 2_E_L1a 120 14. Frequencies for types of vowel change in the overall pastparticiple-sequences in 2_E_L1a 121 15. Frequencies for types of past tense and participle formation in 2_E_L1a 122 16. Frequencies for the types of past-infinitive-sequence with or without vowel change in 2_E_L1b 123 17. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda (pooled) as interacting predictors and the types of past-infinitivesequence as response variable in 2_E_L1b 124 18. Frequencies for types of overall past-infinitive-sequences in 2_E_L1b 126 19. Contingency table of past tense by infinitive formation in 2_E_L1b 126 20. Frequencies for the types of past-participle-sequence with or without vowel change in 2_E_L2a 128 21. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda as predictors and the types of past-participle-sequence as response variable in 2_E_L2a 129 22. Frequencies for types of vowel change in the overall pastparticiple-sequence in 2_E_L2a 130 14 23. Frequencies for types of past tense and participle formation in 2_E_L2a 130 24. Frequencies for the types of past-infinitive-sequence with or without vowel change in 2_E_L2b 132 25. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda as predictors and the types of past-infinitive-sequence as response variable in 2_E_L2b 133 26. Frequencies for types of vowel change in the overall pastinfinitive-sequence in 2_E_L2b 134 27. Frequencies for types of past tense and infinitive formation in 2_E_L2b 135 28. Percentage of overall vowel change, non-apophonic vowel change, and the statistically significant effects of onset, nucleus, and coda in all three experiments 138 5 Experiment 3 title of table 1. Frequencies for past tense inflections across all four contexts in 3_G_L1_oral 146 2. Frequencies for replies to question 1 in 3_G_L1_oral (what - if any - word did the highlighted word frinken remind you of? ) 147 3. Frequencies for replies to question 2 in contexts 1 and 2 in 3_G_L1_oral (Name a word that rhymes with frinken as quickly as you can.) 147 4. Frequencies for replies to question 3 in 3_G_L1_oral (Would you normally associate frinking with ‘vodka’ and ‘the consumption of liquids’? ) 148 5. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with context and the three follow-up questions as predictors and the inflection of the nonce-verb frinken as response variable in 3_G_L1_oral 149 6. Selected inconsistencies in the overall response patterns in 3_G_L1_oral 150 7. Frequencies for past tense inflections across all four contexts in 3_E_L2_oral 151 8. Frequencies for replies to question 1 in 3_E_L2_oral (what - if any - word did the highlighted word frink remind you of? ) 152 9. Frequencies for replies to question 2 in contexts 1 and 2 in 3_E_L2_oral (Name a word that rhymes with frink as quickly as you can.) 152 15 10. Frequencies for replies to question 3 in all four contexts in 3_E_L2_oral (Would you normally associate frinking with ‘vodka’ and ‘the consumption of liquids’? ) 152 11. Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with context and the three follow-up questions as predictors and the inflection of the nonce-verb frink as response variable in 3_E_L2_oral 153 12. Frequencies for past tense inflections across all four contexts in 3_E_L2_written 155 13. Comparison of the degree of context sensitivity across the present three experiments and Ramscar’s study (2002a) 157 16 Acknowledgements After the completion of this book the moment has come to look back and pay tribute to the many helping hands to which these pages owe their existence. While it has only one author, there are many people whose contributions proved invaluable in various areas. Finding participants for the numerous experiments was one of these. For help with the pilot study, I would like to say thank you to all the students from the seminar The Organisation of Language at the University of Siegen. When preparing the various experiments, I was happy to have professional assistance from Holger Schmitz with the CD-recordings. For a pair of spectacular voices on the CDs, I am indebted to Elizabeth Schlömer and Alexander Atkins. Finding English speaking participants for the experiments was made possible through the help of Matthias Fiedler, Monica Ward, Regina Standun, Thomas Müller, Tim Fernando, Billy Kelly, John McKenna, Joseph van Genabith and Angela Leahy in Dublin as well as Andrea Pfeil and Bernd Koch in Glasgow, Anke Börsel in Hull, Kristina Notthoff in Leicester, Stephanie Groenke in Edinburgh and Thomas Schmiedel in Oxford. For getting hold of the German participants I am truly grateful to Alexander Atkins, Holger Schmitz and Sabine Lappe in Siegen as well as Marcus Callies in Marburg. Big thanks go to all the students who eventually agreed to serve as guinea pigs in the experiments. They provided insightful, stimulating, and at times funny data. Throughout the process of data analysis, Thomas Mockenhaupt’s assistance when computerising the tons of questionnaires was greatly appreciated. And all the statistical evaluation would not have been possible without Harald Baayen’s invaluable advice. I am also grateful for a brief introduction to the statistical software R by Wieke Tabak in Nijmegen. There have also been numerous critical and - at times - challenging comments and suggestions throughout various stages of the book’s genesis, which helped me to maintain and sometimes regain focus and momentum. I am grateful for Richard Wiese’s and Susanne Bartke’s comments on the experimental designs, as well as Mareile Schramm’s patient feedback and relentless proofreading. As usual, all remaining errors are mine. On a more personal note, I also want to express deep gratitude to the non-linguistic support I have been lucky to receive over the past couple of years. A very special thank you goes to my wife, whose encouragement is beyond further elaborations. 17 Finally, it almost goes without saying that I am deeply grateful to my academic supervisor Prof. Dr. Ingo Plag, who has always accompanied my work with invaluable criticism and encouragement. He will be happy to know that his name was immortalised in psycholinguistic research thanks to some participants’ responses in Michael T. Ullman’s study (Ullman 2001: 717) and through some responses to German test item 10/ 34 in the present empirical data. Apparently, some people plag things. Linz, Austria, December 2009 Thomas Wagner 18 Preface At the turn of the millennium, Ingo Plag, the supervisor of this thesis, presented a paper on the topic of irregular past tense formation in German- English interlanguage. Tentative and brief as the discussion of the results was, the paper constituted pioneering work in an area that had been given rather little attention by linguists so far, namely the cognitive representation of irregular inflectional paradigms in interlanguages. Plag’s results were remarkable and thought-provoking. They did, however, remain inconclusive to a considerable extent, and he closed his presentation summarising that the present study seems to raise more questions than it answers, but I hope to have shown that these questions are sufficiently interesting to merit further investigation […]. (Plag 2000: 148) In 2004, I embarked on some of these further investigations, the results of which are presented in this book. No sooner, though, did I start familiarising myself with the intricacies of irregular verb morphology and its organisation in the mental lexicon in English and German, than it became evident that this research area had practically exploded into what is now known as the so-called ‘past tense debate’. For a great number of linguists, psychologists, biologists and cognitive scientists alike, regular and irregular verb morphology, mostly but not exclusively in English, had become the prime object of investigation. One consequence was a massive and quickly growing body of evidence and counterevidence for the theories put forward. The other was that my further investigations turned out to be a much bigger challenge than assumed at the outset. In retrospect, though, the work in this area of research provided a stimulating, enriching, and more often than not indeed challenging experience. Overall, I hope that despite the abundance of ever more complex and challenging theories and models, the present work will contribute to the debate and prove to be sufficiently interesting to merit even further investigations. 19 1 Introduction English irregular verb morphology constitutes a rather pertinacious problem both in the area of linguistics and language acquisition. In language acquisition, irregular verbs, the bane of every language student (Pinker 1999: xi), have long been dreaded by learners because of their seemingly erratic and thus error-prone patterning. Generations of students across the world have met irregular verbs as an alphabetically arranged succession of notoriously unpredictable exceptions, which demanded to be rote-learned in a monotonous and tedious fashion. In linguistics, English irregular verbs have long been assumed to be cognitively represented as a complex list of idiosyncratic, atomic items (Marcus 1995, cf. Bloomfield 1933, Chomsky 1968), rote-learned, and stored separately and individually in the mental lexicon in early childhood. Such a rather simple and static view of irregular verb morphology, however, not only shows a charming disregard for the internal morphological patterning within this word class, but also for a learner’s as well as speaker’s ability to generalise from existing patterns to novel, unknown forms. It has been repeatedly observed that adult speakers (Bybee and Moder 1983) as well as children (Bybee and Slobin 1982) overgeneralise irregular verbs. Such processes occur far too frequently to be simply classified as careless slips by children or linguistically ignorant speakers. What kind of mechanism, then, could be responsible for such productivity? Over the last two decades, the internal organisation of irregular verb morphology has been addressed in a great number of publications (for overviews see Pinker 1991, Tomasello 1995, Pinker 1999, Westermann 2000, Ullman 2001a). Numerous theories and models have been put forward, refined, and revised in rapid succession. In fact, research concerned with irregular verb morphology has assumed such a prominent place in cognitive linguistics, that it is now known as the ‘past tense debate’. And yet, up to the present, it appears as though a number of fundamental problems still need further investigation. The current approaches in this research area can be roughly broken down into three camps. The first is the ‘all-rules-camp’. Within this framework English irregular verb morphology is perceived as strictly rule-like in nature. In the classical version of such an approach, Chomsky and Halle (1968) proposed that all unpredictable and idiosyncratic information, such as irregular plurals or past tense formations, were stored as exceptions in the mental lexicon, and that past tenses were derived by a fixed set of input-output-oriented ablaut rules. In a more recent approach, Albright and Hayes (2003) claimed that both regular and irregular verbs could be ac- 20 counted for solely by a set of rule-like micro-generalisations. Behavioural English L1 data have been modelled successfully in such a fashion, outperforming a strictly analogy-based competitor model. However, since Albright and Hayes admit that these rule-like generalisations arise, as it were, out of the similarities and analogies in the verbal paradigm, the question remains what cognitive status these generalisations are supposed to have: do they represent directional and deterministic rules, and thus expand the area of the classical rule-models, or do they primarily express inputas well as output-oriented analogical schemas (cf. Bybee 1988, Marcus 1995), and therefore could be understood as part of a pattern associator. Also rule-like in nature, but proposed as a universal property of vowel change, is the so-called apophonic path (Ségéral and Scheer 1998). According to this path ablaut across all human languages is universal. The path specifies how one particular vowel quality could be derived from another one. There are, however, two problems with such a generalised theory of ablaut. First, the theoretical evidence appears to rest on a somewhat idiosyncratic analysis of irregular verbal paradigms, such as German or Hebrew. And second, this theory lacks empirical evidence form behavioural data. So far, only a couple of languages, with an emphasis on German, have been discussed theoretically in order to substantiate these claims. In sum, the above ‘all-rules-approach’ has got two disadvantages. First, in order to account for the seemingly unrelated variety of all vowel changes possible in English, it necessitates an enormous degree of abstractness, or underspecification, of the underlying representations, as well as a huge number of distinct rules. And second, a mechanism restricted to input-output operations is bound to leave all the attested cases of productivity starting out from derived forms unexplained. In contrast, the so-called ‘associationist’ approaches are built on the assumption that English irregular verb morphology rests on lexical connections and generalisations, governed by prototypicality, similarity and analogy (Bybee and Moder 1983, Köpcke 1998, Derwing and Skousen 1994, Chandler 1998, Eddington 2000a, Hickey and Martin 2001, Eddington 2004). Within this framework one can find a number of different, and partly conflicting, models. The so-called connectionist modellers, for instance, claim that irregular verb morphology was best modelled by what they termed ‘parallel distributed processing’ (PDP). Explicitly dispensing with any kind of deterministic rule, they model verb morphology solely based on parallelly activated information nodes, connections, and weights (Rumelhart and McClelland 1986, Plunkett and Marchman 1991, Hare and Elman 1995, Nakisa and Hahn 1996, Plunkett and Juola 1999, Joanisse and Seidenberg 1999, Westermann 2000). According to another associationist approach, irregular verb morphology was best modelled through a similarityand analogy-based mecha- 21 nism (Derwing and Skousen 1994, Chandler 1998, Eddington 2000a, Hickey and Martin 2001, Eddington 2004). Similar to connectionism, the cognitive reality of rules is denied in such models categorically, too. However, although associationist researchers have been modelling verb morphology in ever more sophisticated simulations for more than 20 years, they apparently still fail to cover important behavioural phenomena from language production and language learning. Therefore, the cognitive plausibility of their models has been repeatedly called into question (Pinker and Prince 1988, Prasada, Pinker and Snyder 1990, Prasada and Pinker 1993b, Pinker and Ullman 2002a, Albright and Hayes 2003). Such criticism points out that a fully-fledged cognitive representation of irregular verb morphology needed both a rule-like and an associationist mechanism. Thus, while the two previous approaches relied on one single-route-architecture (either rule-like or similarity-based), there also is a third camp, which promotes the notion of a dual route to irregular verb morphology. This approach, also known as the ‘dual mechanism’, claims that in irregular verb morphology across many languages regular verbs undergo symbolic rule-application (high-frequency regulars may actually be stored in the lexicon), while irregulars are stored as structural lexical entries and processed in memory by a pattern associator (Pinker 1991, Wunderlich and Fabri 1995, Clahsen 1997, Clahsen 1999). Thus, in a way this approach could be regarded as an attempt to reconcile the classical rule-model with the associationist claims. However, there are two shortcomings this approach has been criticised for. First, employing two distinct inflectional routes in one model seems uneconomic from a cognitive point of view. Both the ‘all-rules’ and the associationist researchers claim that cognitively plausible models should be organised as parsimoniously as possible and thus rely on one route only (Derwing and Skousen 1994, Albright and Hayes 2003). Second, the dual-route model remains somewhat vague as to the precise architecture of the two routes. To begin with, the symbolic rulecompartment has often been criticised for being something like a ‘black box’, since the model does not specify what cognitive mechanism in fact derives output forms from input forms. Another matter of debate is the exact nature of the pattern-associator. Reference has been made to prototype theory (Pinker 1991, 1997), family resemblance (Rosch 1973), schemas (Bybee and Moder 1983), and analogy (Skousen 1989, 1992, Chandler 1993, Eddington 2004), but how for instance prototype theory is supposed to govern English irregulars still remains unclear (cf. Bybee and Moder 1983). Only recently, a whole new dimension has been added to the so far exclusively form-based analysis of irregular verb morphology. After two decades, throughout which proponents both of singleand dual-route models had considered the organisational principles to be sensitive to form 22 only, now it was revealed that irregular verbs differ from regular ones not only in their morphophonological, but also in their semantic characteristics (Ramscar 2002a, 2002b, 2002c, Baayen and Moscoso del Prado 2005). Moreover, there is evidence that the similarity in meaning of a given verb to another verb can serve as a semantic attractor and bias a speaker’s inflectional behaviour (Ramscar 2002a, Ernestus and Baayen 2004). So far, however, this bias has been shown to work only once phonological similarity constraints are met (Huang and Pinker 2005), so that up to the present we still lack evidence for purely semantic attraction. Thus, in sum the precise role of semantics for English irregular verb inflection still needs further investigation. As can be seen, the literature on irregular verb morphology in the area of the L1 is replete with inconclusive and conflicting evidence. As Albright (2002: 684) put it: To date, no model - single or dual route - has been implemented that can adequately capture the full range of data. Thus, it hardly comes as a surprise that irregular verbs in the interlanguage have been given very little attention. With reference to the dual route model, there are a couple of interlanguage paper publications, such as for the American-Russian interlanguage (Chernigovskaya and Gor 2000), and the Korean-English interlanguage (Flege, Yeni-Komshian and Liu 1999, Birdsong and Flege 2001). As far as the German-English interlanguage is concerned, irregular verb morphology has been explored in one paper-publication by Plag (2000), in which he pointed out parallels between the organisation of irregular verbs in English L1 (cf. Bybee and Moder 1983, Pinker and Prince 1994) and the German-English interlanguage. These parallels, as Plag argues, could be taken as evidence either for transfer processes from a German L1 prototype (cf. Köpcke 1998), or, alternatively, as evidence for the universality of Ségéral and Scheer’s apophonic path (1998). The problem with the transfer hypothesis, however, is that a single prototype governing a particular vowel change is hard to distil from production data, both in English L1 and the interlanguage. For German L1, production data supporting a prototypical organisation are missing altogether. In sum, Plag concluded that German-English interlanguage irregular verb morphology was still a largely unexplored area (Plag 2000: 134). The exploration of this area is what this book is concerned with. By means of fresh production data, including the interlanguage speakers’ L1, their interlanguage, and the target language, Plag’s conclusions were revisited and expanded. In three production experiments three hypotheses about the nature of irregular verb morphology in German-English interlanguage were tested. They are given in (1). 23 (1) (a) Hypothesis 1 German-English interlanguage irregular verb morphology is best modelled according to prototype theory. The morphophonological properties of L2 prototypes emerge through transfer from the L1 and serve as morphophonological schemas in the interlanguage. (b) Hypothesis 2 German-English interlanguage irregular verb morphology is best modelled as rule-like input-output processes that follow the apophonic path. (c) Hypothesis 3 German-English interlanguage irregular verb morphology is codetermined by a verb’s semantic similarity to existing verbs. These similarities work as attractors independent of phonological similarities to existing verbs. In order to test hypothesis 1, a number of elicitation tasks investigated whether morphophonologically specified prototypes govern past tense formation in German L1, English L1, and the interlanguage, and whether similarities between German L1 and the interlanguage suggest transfer processes. In these experiments, participants had to inflect given nonceverbs, whose morphophonological make-up varied systematically with respect to potential prototypes. Fresh data from English L1 were obtained in order to compare the situation in the interlanguage with the target language, because, as Eubank pointed out, a comprehensive investigation of transfer phenomena necessitates a closer look at both the NL grammar and the advanced state of the interlanguage (cf. Eubank 1996: 79). Second, in order to test hypothesis 2, six elicitation tasks explored in how far vowel changes in past tense and participle inflection reflect the predictions made by the apophonic path. In these experiments, some participants were given nonce infinitives to inflect both for past tense and participle, while others were given participles in order to inflect them in a backward fashion for past tense and infinitive. Again, this was done with respect to German L1, English L1, and the interlanguage. And finally, in order to test hypothesis 3, three elicitation tasks investigated whether varying semantic environments of a given nonce-verb could influence its inflection for past tense. These experiments were carried out for German L1 and the interlanguage only. The remainder of this book is structured as follows: Chapter 2 provides the theoretical background necessary for the interpretation of the experimental data. Chapters 3, 4, and 5 each deal with one of the three elicitation experiments designed to test the hypotheses. They both present the experimental design and report the results. Chapter 6 summarises and discusses all results, closing with conclusions and perspectives for further 24 research. A selection of English questionnaires and all the test items can be found in the appendix. 25 2 Irregular verbs - a linguistic challenge 2.1 Introduction This chapter provides an in-depth discussion of the theoretical background necessary for the discussion of the empirical data. First, an outline of some basic claims about irregular verb morphology that can be found in the pertinent literature will be given. A discussion of traditional as well as more recent morphological and psycholinguistic accounts will follow, including a brief section on the ‘past tense debate’. Then, the role that irregular verbs have been playing in L1 and L2 acquisition research will be discussed. Particular attention will be given to German-English interlanguage. Finally, the three hypotheses that guided the empirical work will be developed and discussed. 2.2 Irregular verbs in English and German - a classic chestnut Modern English and German both have between 180 and 200 so-called irregular verbs. The counts vary because some of these verbs are archaic and hardly used any more (cf. Clahsen 1999, Clahsen et al. 2001 for German, and Marslen-Wilson and Tyler 1998, Bybee and Slobin 1982: 265, and Pinker 1991 for English). The notion of ‘irregularity’ here means that the verbs’ inflectional behaviour for past tense and participle formation deviates from the regular verbs. This deviation is commonly held to be unpredictable to some extent, and it cannot be accounted for by input-outputoriented rules 1 . Irregular verbs have been under the linguistic microscope for many decades, but up to the present day, we have neither a commonly accepted morphological analysis nor an accurate psycholinguistic account of this paradigm. Standard references, such as the Handbook of Morphology (Spencer and Zwicky 2003) offer a rather limited exploration of this phenomenon. There are a number of possible reasons for this. The first reason relates to the verbs’ diachronic development. For one thing, irregular verbs in German and English are quite extraordinary lin- 1 Irregularity in English and German verb morphology comes in a variety of shapes and degrees, and for German, subcategories such as strong verbs, mixed verbs, and suppletives are well established. In this and the remaining chapters, however, any non-regular pattern in English and German will be subsumed under the heading ‘irregular verb’ 26 guistic fossils. According to Venneman (2000), a great number of these verbs show an incoherent etymology, lacking historically related counterparts in other Indo-European languages. Instead, as Vennemann argues, these verbs apparently developed throughout their diachronic development through language contact with a Hamito-Semitic superstratum. And ever since, they have been a vestigial and indeed endangered morphological class. Not only have they constantly been losing members to the realm of regular verb inflection up to the present, but also they have generally failed to attract new members. A small number of exceptions from this general trend can be found, though, like the class of verbs featuring vowel change [ ]-[ ]-[ ], which received 12 new members since Old English (cf. Jespersen 1942). From Old English, only about half of the irregular verbs survived in Modern English, but despite their shrinking type frequency (in Modern English only around 5%, cf. Marcus et al. 1995, Marslen-Wilson and Tyler 1998), these irregular verbs are surprisingly strong on the token front, being among the most frequently used verbs of the language (Bybee and Slobin 1982: 265). Amongst the 30 most frequent English verb types, for instance, we only find eight regulars. Moreover, a lot of irregular verbs are verbs of every day usage, denoting central human activities, such as be, have, do, say, make, go, take, come, see, and get (cf. Henzen 1965, Köpcke 1998: 48), and they feature prominently in motherese and caretaker-speech, too. For an overview of the diachronicity of English irregular verbs, see Juilland and Macris (1973), and Pyles and Algeo (1992). For German, the situation is similar. While Middle High German had about 400 irregulars, the number decreased to around 180-200 nowadays, yielding a type-frequency of roughly 4% of all verbs (cf. Köpcke 1998: 45, Bittner 1996: 111). However, as in English, these verbs show a remarkable token frequency, which roughly resembles that of the regulars (cf. Clahsen et al. 1997, Marcus et al. 1995). Out of approximately 4,000 German verbs, for instance, more than 50% of the 80 most frequent types are irregulars. And amongst the 1000 most frequent German word forms, there are 80 verbs, 33 of which are irregular (cf. Augst 1975, Ruoff 1981, Weyerts and Clahsen 1994). In fact, one of the most irregular verbs both in English and German, the suppletive verb to be (‘sein’), often displays the highest token frequency in text corpora 2 . For a comprehensive overview of the diachronic development of German irregular verbs, see Bittner (1985), and Bittner (1996: 114-174). The second reason for the absence of a commonly accepted model of irregular verb morphology relates to the verbs’ synchronic characteristics. 2 The remarkable discrepancy between the irregular verbs’ type and token frequency has been repeatedly regarded as the prime factor for their survival (cf. Augst 1975, Marcus et al. 1995). 27 Irregular verbs form an unusual inflectional paradigm. They are full verbs in both languages, but their past tense and participle inflection deviate from the default suffixation pattern, displaying various forms of vowelchanges instead. These are called ablaut, gradation (Quirk et al. 1988, Wiese 1996, Wiese 2008) or apophony (Ségéral and Scheer 1998) 3 . Ablaut, one of the classic chestnuts of morphological analysis (Anderson 1988: 157), refers in general to systematic stem-vowel alternations, marking morphological categories (cf. Paul 1917: 189, e.g. - , - , - ), and is a characteristic of some Indo-European languages, such as English, German, or Dutch, as well as many Afroasiatic languages, such as Classical Arabic (cf. Kuryłowicz 1956). Despite its frequent occurrence in typologically different languages, however, there is little consensus as to its morphological status. Some researchers claim that ablaut, both in present-day English and German, is entirely irregular and thus unpredictable. As a consequence, all the verb forms undergoing ablaut are therefore thought of as fully lexicalised (cf. Wiese 1996: 113 for German). Others, however, have questioned this alleged lack of predictability (Wiese 2008 for German, and Bybee 2001 for English). So how can this classic chestnut of morphology be accounted for in English and German? Consider table 1. It presents examples of ablaut patterns (past tense and participle) in irregular verbs from both languages. Table 1: Selection of ablaut patterns for past tense and participle in English and German patterns vowel changes infinitive past tense participle English a-b-a [ ] - [ ] - [ ] come came come a-b-b [ ] - [ ] - [ ] meet met met a-b-b + -en [ ] - [ ] - [ ] speak spoke spoken a-b-c [ ] - [ ] - [ ] sing sang sung no-change [ ] - [ ] - [ ] cut cut cut no-change + -en [ ] - [ ] - [ ] beat beat beaten dental suffix [ ] - [ ] - [ ] burn burnt burnt suppletion [ ] - [ ] - [ ] go went gone German a-b-a [ ] - [ ] - [ ] fallen fiel gefallen a-b-b [ ] - [ ] - [ ] bleiben blieb geblieben a-b-c [ ] - [ ] - [ ] singen sang gesungen mixed verbs [ ] - [ ] - [ ] brennen brannte gebrannt suppletion [ ] - [ ] - [ ] sein war gewesen As we can see in the above table, there is a great variety of patterns. From a diachronic point of view, the seemingly unstructured patterns constitute a residue of older, phonologically conditioned processes. Synchronically, 3 These vowel changes can take suffixation (-t in English, -te in German). 28 they appear to be an unpredictable and almost arbitrary assembly of patterns. Some verbs show three different vowel qualities (e.g. sing, singen ‘sing’), some two (e.g. meet, bleiben ‘stay’), while a considerably large group of English verbs does not change the vowel at all (e.g. cut-cut-cut). Consider table 2 now. There, verbs featuring stem vowel [ ] are listed along with their past tense formations and examples (data taken from Bloch 1947). Table 2: Verbs featuring stem vowel [ ] along with their past tense formations and examples stem vowel past tense vowel examples (a) [ ] [ ] bide-bode, drive-drove, ride-rode, rise-rose, shrive-shrove, smite-smote, stride-strode, strivestrove, thrive-throve, write-wrote dive-dove, shine-shone (some dialects) (b) [ ] [ ] bite-bit, hide-hid, light-lit, slide-slid (c) [ ] [ ] strike-struck (d) [ ] [ ] bind-bound, find-found, grind-ground, windwound (e) [ ] [ ] buy-bought, fight-fought (f) [ ] [ ] fly-flew As we can see in table 2, the categorisation of ablaut patterns seems to depend on the morphology of the past tense verbs. Consider pattern (2d), for instance. There, irregular verbs containing the rhyme [ ] always show past tense forms with the rhyme [ ]. If a new verb like gind were to enter the English language, it would be a good candidate for past tense formations along the vowel change [ ]-[ ]. Such generalisations of existing ablaut patterns to new verbs can actually be observed both with adult speakers (cf. Bybee and Slobin 1982, Prasada and Pinker 1993b) and with children (Marcus 1995, 1999, Xu and Pinker 1995). Although the observable irregularities of past tense formation are clearly patterned, albeit with restricted generalisation properties, they cannot be captured by deterministic and generative rules that merely translate a given input form into a new output form. The phonological properties of irregular verbs are neither necessary nor sufficient to predict the pattern to which a verb will conform. As Marcus (1995: 193) put it for English, irregular verbs are somewhere in between an idiosyncratic list and a product of freely applying rules. As far as German is concerned, Wurzel came to the same conclusion, claiming that it was simply impossible to derive a rule for German past tense formation from a given base form (cf. Wurzel 1970). Paradigms of such a nature are often called hybrid or semi-regular. A comprehensive theory of these exceptional and yet patterned paradigms faces two challenges. First, looking at the stem vowel qualities, a theory 29 would need to explain the relationship between all vowel changing forms involved (infinitive, past tense, and participle). Second, it would need to categorise the various vowel change patterns and explain whether and how they form natural classes and generalise to new forms. The second problem becomes immediately apparent, once we have a closer look at traditional approaches to the classification of ablaut series (cf. Jespersen 1942, Halle 1953, Henzen 1965, Hoard and Sloat 1973, Augst 1975, Fabricius-Hansen 1977, Hempe 1988, Quirk et al. 1988, Bloomer 1994, Bittner 1996, Wiese 1996, Köpcke 1998, Wiese 2008). While, for instance, one of the earliest comprehensive accounts of English irregular verb morphology reported 11 classes, along with dozens of subgroups (Jespersen 1942), only a couple of years later Bloch (1947) suggested 20 conjugation types. In comparison, Hoard and Sloat (1973) argued for the existence of seven classes, along with four ablaut rules. This classification re-emerged in an analysis by Quirk et al. (1988), who, however, argued for 42 subclasses. Accounts on German irregular verbs show an even greater variety of categorisations, as table 3 illustrates. Table 3: Categorisations of German irregular verbs source number of categories number of rules or subclasses Halle 1953 4 13 rules, 44/ 46 subclasses Koenraads 1953 29 none Ulvestad 1956 48 none Hook 1968 8 18 rules, 16 subclasses Wurzel 1970 3/ 15 19 rules, 26 subclasses Augst 1975 52 30 subclasses Kern and Zutt 1977 4 41 subclasses Fabricius-Hansen 1977 10 49 subclasses Wiese 1996 20 none Bittner 1996 8 31 subclasses Eisenberg et al. 1998 39 none Ségéral and Scheer 1998 7 39 subclasses Attempts to describe the morphological process of ablaut are no less controversial than the definition of verb classes. This will be demonstrated in the following, using the English irregular verb sing as an example. Traditional accounts of irregular verb morphology, such as the ‘itemand-arrangement’, or ‘item-and-process’ framework (cf. Hocket 1954) were categorical, concatenative, and rule-based in nature, reflecting an all-or-none 30 approach (cf. Bolinger 1961, for a critical review see Augst 1975, Köpcke 1998, and Wiese 2006). These approaches, however, proved inefficient when it came to deriving an ablauted vowel by rule application from an unsegmentable form such as sing. Its past tense sang, which differs from its present tense stem only with respect to vocalism, simply defies concatenative operations. Thus, such a form would have to be considered as derived from an underlying form, perhaps sing, and the morphological process of ablaut would have to replace the vowel [ ] with, or change it into, a replacive morph [ ]. In the related ‘item-and-arrangement’ paradigm, sang would be analysed as a so-called discontinuous allomorph of sing [ _ _ ], containing an infixed vocalic allomorph [ ] of the past tense morpheme, which usually surfaces as -(e)d and thus would appear as a zero-allomorph for tense marking in this example. However, [ ] seems to be quite an unsuitable candidate for allomorphic past tense marking, since there are other strong verbs for which [ ] does not code past but present tense (cf. hang). The same would apply to a German verb form, such as past tense band (‘bound’, ‘tied’), which simply does not lend itself to morphological segmentation, and hence does not feature a concatenative tense marker. A similar problem would occur with Chomsky and Halle’s famous ‘SPE-model’ (Chomsky and Halle 1968). This model assumes that only unpredictable and idiosyncratic information is stored in the lexicon. As they put it, the lexicon is home only to lexical items with their inherent properties, in particular, those phonological properties that are not determined by general rule (Chomsky and Halle 1968: 44). Ablaut in irregular verbs, however, is based on a set of phonological ablaut rules, in which both the uninflected stem and the inflected irregular are derived from the same underlying representation. Thus, the past tense formation of a verb such as sing would develop in the following steps. First, its lexical representation is retrieved from the lexicon. There, it was coded for its syntactic category, other nonderivable information, and marked for undergoing a specific form of vowel change [ v (marked for ablaut)]. Then it is transferred to the morphology unit, where usually stems and affixes are concatenated. For sing [ vpast + marked for ablaut], however, the marker indicates that the past morpheme for sing is not the default but an ablaut rule. And finally, as an underlying form, it is transferred to the phonology unit with all its great many ablaut rules. There, [ ] is rewritten as [ ], and the derived surface structure [ vpast ] emerges. A widely observed problem with this approach is the enormous degree of abstractness of the underlying representations. As Pinker once put it, they can be wildly different (Pinker 1999: 97). The problems that all three above approaches have with ablaut are obvious. First, their segmental and concatenative credo fails to account for forms that are morphologically complex in a non-concatenative way. Sec- 31 ond, they disregard obvious paradigmatic relationships between verb pairs such as sing-sang, ring-rang, and swim-swam. It seems as though these models lack a way of representing paradigms altogether, since each word form needs to be derived independently from a base form. However, as Wiese 2008 pointed out, an analysis of apophonic operations must explain the relations between the vowels involved. Third, the above approaches do not allow for a gradual transition between what was traditionally known as regular and irregular (cf. Bittner 1985, Köpcke 1998: 46). And finally, they lack the psycholinguistic underpinning derived from behavioural data (cf. Becker 1990). Alternative approaches, such as the ‘item-and-paradigm’ framework (cf. Hocket 1954, Bierwisch 1967, Lieb 1976, 1992) have been trying to overcome the above problems. First, they regard ablaut as surface morphology that does not countenance underlying forms (and hence no rules that could turn underlying forms into surface forms). Second, these approaches show sensitivity to paradigmatic relationships and are not limited to concatenative operations. Here, the form sang would be analysed as a portmanteau morph (sing + morpheme {past}) and at the same time as part of the paradigm sing-sang-sung. In the ‘word-and-paradigm’, or ‘stem-and-paradigm’ framework (Matthews 1991, Anderson 1992, Lieb 1992, Aronoff 1994, Blevins 2003, Stump 2005), morphemes are dispensed with altogether, and full stems (or words) are viewed as the basic units in the lexicon. Here, the paradigm’s organisation evolves out of exemplar-driven similarities, and morphological structures and classes are graded instead of being strictly categorical (cf. Hay and Baayen 2005). In what is known as ‘natural morphology’ (Dressler 1977, 1985a, 1985b, 1986), an approach that explicitly aspires to psychological reality for its constructs (Dressler 1986: 519), morphemes are fully transparent. This means that every morpheme has one form and one meaning, and every meaning corresponds to exactly one form (cf. Halle’s analysis, in which strong verbs have more than one form of the stem (1953: 46)). No abstract underlying forms or zero affixations exist. Consequently, such an approach would render allomorphy, being already lexically represented, superfluous. The past tense of sing would be represented in a so-called lexical entry, as given in (1). (1) [ ] In (1), all possible morphological alternations (sing-sang-sung) are stated directly in the lexicon, and the relation between them is expressed in socalled via-rules (cf. Vennemann 1972). These via-rules, visualised by the big square brackets, are supposed to represent morphophonological relations 32 and lexical correspondences of limited generalisability. They can be thought of as related sets of word forms rather than productive rules in their own right. As a consequence, it is the very word forms themselves, and not rules that capture similarities within the paradigm. These via-rules, however, fail to provide an explicit and psycholinguistically plausible account of how the three allomorph-like variations in (1) are actually connected. The possibly most explicit psycholinguistic account of irregular verbs comes from ‘minimalist morphology’ (Wunderlich 1996, Wunderlich and Fabri 1995). In this model, paradigmatic relationships play a central role, acknowledging both the word and the paradigm as an explicit part of the grammar (cf. Corbett and Fraser 1993, Wunderlich 1995: 94, Stump 2005, chapters 2 and 6). Hence, the status of one single verb form is primarily based on its role in predicting other forms within the paradigm. As far as inflection is concerned, minimalist morphology is a dual-route theory. While dispensing with abstract morphemes and zero affixes, it proposes structured lexical entries for irregular verbs and rule-like affixation for regulars as two qualitatively distinct mechanisms for past tense inflection. This corresponds with Wiese’s analysis for German (2000: 209), who pointed out that ablaut did not belong to the realm of phonological rules but had better dealt with in terms of entry-specific representations. In minimalist morphology, the inflectional properties of a verb would be encoded in a so-called ‘structured lexical entry’ (Wunderlich 1996: 95), a notion which is not entirely new and has had its antecedent in Chomsky’s notion of lexical redundancy rules (cf. Jackendoff 1975). Those structured lexical entries are organised as a so-called inheritance tree, which is an upside-down hierarchical tree structure that arranges alternating morphophonological forms of the verb as a paradigm. The verb sing, along with its inflections for past tense and participle, would have the following entry (2). (2) [ ]v [ … … ] past tense [ … … ] participle In each node in this tree, a phonological string, such as [ ], is paired with a morphological feature value (‘V’, ‘past tense’, ‘participle’). The nodes for past tense and participle are so-called subnodes, and they inherit all the information from their mother node ([ ] v ), except for those features that a daughter node replaces. The slots for inherited information are marked in (2) by the three dots […]. The subnode for past tense ([… …] past tense ), for instance, would inherit both the onset and coda structure as well as the 33 morphological feature (V) from the mother node [ ] v , and only the nucleus would be replaced. Subregularities outside the lexical entry as well as lexical relations among irregular past tense forms of different verbs are captured through so-called ‘templates’, in which various stem segments are associated. Thus, a subnode like ([… …] past tense ) would not only be part of the entry for sing but also for other phonologically similar verbs (ring, drink, etc.). For the German participles, Wunderlich and Fabri (1995) proposed two distinct mechanisms, a rule-affixation process for regular participle formation with -t (machen - ge-mach-t, ‘to make’), and structured lexical entries for irregular participles, as illustrated for the German verb werfen (‘to throw’) in (3). (3) [ ] v ‘to throw’ [ … ... ] 2./ 3.prs/ pres [ ... … ] pret [ … ... ] part [ ... ] imp [ ... ... ] subj Inheritance of information in this model works in a twofold fashion. One way of accounting for alternations is the simple addition of information to the dominating node, as illustrated by the [ ]-participle suffix. The other way involves the substitution of information, as is the case for all the vowel changes. To sum up so far, it seems obvious that traditional rule-based approaches have failed to produce convincing accounts of irregular verb morphology. Either they provide only sparse explanations for the paradigmatic relation of ablauting verbs, or they suggest rather cumbersome models as far as the creation of past tense forms by rules are concerned. Consider, for instance, the abstractness of zero-affixation, or the monstrous rule apparatus suggested by Chomsky and Halle (1968). In this context, it seems noteworthy that some theoretical analyses of irregular verb morphology make surprisingly little effort to substantiate their models with behavioural data. Bittner, for instance, even went so far as to deny the psycholinguistic validity of his own analysis of German irregular verbs (Bittner 1996: 43). In conclusion, the most promising account of irregular verb morphology seems to come from minimalist morphology. Both the categorical distinction between lexically based and rule-based inflection, a design feature of this theory, and the emphasis on paradigmatic relations reflect current 34 trends in psycholinguistic research (cf. Becker 1990, Baayen et al. 1997, Krott et al. 2001, Stump 2005). The structured lexical entries of minimalist morphology seem to provide the most adequate theoretical backdrop for German and English irregular verb morphology. However, the exact nature of the entries’ structure is not sufficiently specified, and it remains to be seen whether and how production data from German and English L1 as well as the German-English interlanguage can be accommodated by this model. 2.3 One or two routes - irregular verbs and the dual mechanism The categorical distinction between lexically based and rule-based inflection, as proposed in minimalist morphology, constitutes perhaps the most heated debate in contemporary cognitive science (Pinker 1999: xi) and has been widely discussed both in linguistic and psycholinguistic circles. The central issue of this debate is the question as to whether verb morphology relies on one or two cognitive mechanisms (routes). Over the last two decades, two major theoretical camps, the so-called ‘single-route’, and the ‘dual-route’ approach have emerged. The single-route camp breaks down into the ‘all-rules’ and the ‘allassociations’ approaches. While the ‘all-rules’ approach (Chomsky and Halle 1968, Albright 2002, Albright and Hayes 2003) claims that vowel changes, even idiosyncratic ones, should be conceived of as rule-based operations, deriving an output form from an input form, the so-called associationist camp considers irregular verb morphology as exclusively based on similarities and analogies. In the latter approach, both lexicon and grammar (irregulars and regulars) are captured by the same associative mechanism 4 . For a more detailed description of such models, see section 2.5. The dual-route approach (also known as dual-mechanism, entries-plusrules-model, words-and-rules-theory, hybrid model, cf. Clahsen 1997, 1999, Pinker and Ullman 2002b,) claims that verb morphology contains the two core characteristics of human cognition, namely regular, symbolic, as well as irregular, associative processing. Therefore, the mental lexicon and the mental grammar are thought to be distinct subsystems in the human brain, both essential for verb morphology. 4 There is an enormous, albeit controversial range of evidence in favour of this approach (Rumelhart and McClelland 1986, Stemberger and MacWhinney 1986, Plunkett and Marchman 1991, Seidenberg 1992, Derwing and Skousen 1994, Bybee 1995, Hare et al. 1995, Nakisa and Hahn 1996, Sereno and Jongman 1997, Chandler 1998, Nakisa 1996, Plunkett and Juola 1999, Joanisse and Seidenberg 1999, Goebel and Indefrey 2000, Westermann et al. 1999, Hickey and Martin 2001, Eddington 2004). 35 While regular, predictable, and productive processes are thought to be computed by deterministic rules, irregular morphology involves formspecific, structured representations and instance-based processing. Figure 1 illustrates the architecture of a dual-route model for verb morphology (cf. Pinker and Ullman 2002b: 457). verb inflection the lexicon the grammar structured lexical entries and a pattern associator symbol manipulation and rule application irregular forms are stored in structured lexical entries generalisations are possible by pattern associator the relationship between items is governed by prototype theory rule-manipulating device (part of the ‘LAD’) rules are of a symbolic nature input forms are processed as symbols output forms are generated in real time when needed holdv heldv past V V walk suffix -edpast Figure 1: The dual-route model for verb inflection In this model, regular forms are generated, whenever needed, by a mechanism of generally applicable symbolic rules, while irregular forms are generated and stored by an associative network outside the grammar. Hence, regular past tenses have no real representation in the lexicon, as they are produced online in the grammar. In contrast, irregular verbs contain lexical representations for sound and syntactic category. An impressive bulk of evidence in favour of this approach has been amassed over the last two decades from a variety of sources. It comprises a number of typologically different languages, amongst which are English, French, German, Hausa, Hebrew, Icelandic, Italian, Japanese, Norwegian, Russian, and Spanish. And thanks to modern non-invasive brain imaging technologies such as PET, fMRI, magnetoencephalography and eventrelated potentials (ERPs), a lot of real-time behavioural data from various languages on generalisation properties have been obtained (cf. Clashen and Rothweiler 1993, Prasada and Pinker 1993, Marcus, et al. 1995, Jaeger et al. 1996, Penke et al. 1997, Ullman et al. 1997, Weyerts et al. 1997, Gross et al. 36 1998, Münte et al. 1999, Lavric et al. 2001, Tyler et al. 2002, Beretta et al. 2003, Bartke et al. 2005, Lück 2006). However, not all the clinical evidence supports the dual-route model. A recent fMRI study (Desai et al. 2006) of past tense inflection in Spanish, for instance, supports a single-route account. And various reports by the research team around Vargha-Khadem produced similar counterevidence, too (cf. Vargha-Khadem et al. 1995, 1998). For a recent overview, see Ingram and Chenery (2007). In sum, the past tense debate is far from settled, and although the investigation of past tense morphology has advanced to something like a test case for cognitive science (cf. Fodor and Pylyshyn 1988), the literature still is replete with proposals and counterproposals. Even after 20 years of research, we still seem to lack unequivocal evidence for the accuracy of one single model. 2.4 Good verbs, better verbs, best verbs - irregular verbs and prototype theory As pointed out in section 2.2, traditional accounts of irregular verb morphology were categorical and rule-based, trying to define clear-cut vowel change classes of verbs. However, reflecting a current tendency in linguistics, and confirming a long-harboured suspicion, this categorical approach is probably just a convenient illusion (cf. Givón 1986, Bybee 2001, Aarts 2004, Aarts et al. 2004, López Rúa 2005, Hay and Baayen 2005, Itkonen 2005, Fanselow et al. 2006, Wanner 2006). As Sapir once put it, all grammars leak (1921/ 1957: 38/ 39). Aarts (2004: 1) remarks in this context that [m]any schools of modern linguistics generally adopt a rigid approach to categorisation by not allowing degrees of form class membership, degrees of resemblance to a prototype or overlaps between categories. […] The alternative view, prevalent amongst descriptive and cognitive linguists, is to posit grammars which pervasively display categorial vagueness, more usually called gradience. Interestingly, the notion of ‘gradience’, also referred to as non-discreteness, fuzziness, flexibility, clustering, or continua, has been widely adopted only recently (cf. Givón 1986 for a historical account, for critical remarks s. Croft 2007). The first attempts in linguistics go back to Eleanor Rosch’s works from the 1970s (Rosch 1973), where she proposed a prototypical organisation of semantic concepts. While in classical approaches all members of a semantic category were held to be equal, and while one could determine category members as well as category boundaries, Rosch’s class membership in contrast was gradable, with better and worse members as well as fuzzy boundaries. Instead of necessary and sufficient membership conditions, category members now shared characteristics via so-called ‘family 37 resemblance’. According to Wittgenstein (1953), a family resemblance relationship consists of a set of items of the form AB, BC, CD, DE, where each item has at least one element in common with one or more other items, while no or only a few characteristics are common to all items 5 . The degree of prototypicality of a particular class member was expressed through its ‘cue validity’ (Rosch and Mervis 1975). Such a prototypical approach allows for a non-discrete continuum space within as well as between categories. (Givón 1986: 79) More than 20 years ago, Bybee and Slobin (1982) suggested such a prototypical organisation for English irregular verbs, too. In a series of production experiments, they found evidence that English L1 speakers generalised irregular verbs as predicted by prototype theory, and that class membership was defined through family resemblance (Bybee and Slobin 1982: 281). In their experiments, speakers would, based on similarities to existing patterns, inflect given nonce-verbs by mapping existing vowel change patterns systematically onto new verbs, thereby creating novel pairs such as stringstrang or clink-clank. Bybee and Slobin (1982: 285) concluded that English irregular verbs are rote-learned and stored in the lexicon, rather than generated by rule from base forms. But the evidence indicates that at least some of the rote-learned forms are organized into classes according to the phonological shape of the past-tense form. These classes are described by schemas. Bybee and Slobin’s schemas can be conceived of as phonological generalisations over a number of a category’s class members. A schema for irregular verbs ranges over one, two, or all syllable constituents (onset, nucleus, and coda), which can have different weights in terms of their importance as part of a schema. Schemas link a verb’s phonological shape with the morphological changes it undergoes within a paradigm, thus forming lexical associations. It can be assumed that these associations facilitate access and retrieval and constitute the basis for new analogical formations. Bybee and Slobin conceded, however, that such schemas would not necessarily apply to all irregular verbs. For the [ - ] and [ - ] vowel change pattern, Bybee and Slobin proposed the two schemas given in (4a), while Bybee (1988: 35, 1995: 431) refined the schema for [ - ] patterns as shown in (4b). (4) (a) vowel change [ ] - [ ] schema [ # _ _ ( )] vowel change [ ] - [ ] schema [ # _ _ / / / / ] (b) vowel change [ ] - [ ] schema [ C(C)(C) _ velar nasal ] Note that the above schemas define a derived form (nucleus [ ] and [ ]), and thus are outputor productoriented (cf. Stanners et al. 1979, who already found evidence that access to an irregular past form could serve as 5 Note that Rosch and Mervis point out that a family resemblance structure need not be the only source of prototypicality (1975: 599). 38 a prime for the base form). Schemas of such a nature can be constructed for a number of irregular verb classes. Take, for instance, the class of verbs that feature the changes given in (5) (data again taken from Bloch 1947). (5) (a) [ ] [ ] buy-bought, fight-fought (b) [ ] [ ] bring-brought (c) [ ] [ ] think-thought (d) [ ] [ ] seek-sought (e) [ ] [ ] beseech-besought, teach-taught (f) [ ] [ ] catch-caught As we can see, the past tense rhyme [ ] is in fact related to a bewildering variety of infinitive stem vowels. A derivational approach to these changes, therefore, would require a number of rules, in effect almost as many as verbs are involved in the pattern. It would be, however, much more appropriate to recognise that the properties of the past tense forms are not necessarily derivable from properties of the associated stem forms; they are cohesive only from an output-oriented point of view via family resemblance (cf. Wurzel 1984a: 68). In a follow-up study, Bybee and Moder (1983) explored potential schemas and their psycholinguistic validity for English L1 speakers in more detail. They, too, reported evidence for graded, output-oriented prototypical structures in irregular verb classes. In addition to that, they proposed prototypical input schemas for potential new category members, as given in (6). (6) input prototype for the [ ] - [ ] vowel change [sCC _ _ ] Note that the prototype in (6) is the abstract formalisation of an input form, or of a best new member of a class, respectively, while Bybee and Slobin’s output-oriented schemas (cf. (4)) are generalisations about a derived output form. Schemas primarily formalise those morphophonological properties that render a paradigm coherent from an output-oriented point of view. Note, too, that the three syllable constituents of the prototype in (6) display varying degrees of cue validity, with coda being the most powerful predictor for class membership, followed by onset and nucleus. In sum, Bybee and Moder conclude that classes of irregular verbs, which traditionally have been considered a more or less higgledy-piggledy conglomeration of unrelated items, are to be defined as natural classes. They conclude, too, that [s]peakers of natural language form categorisations of linguistic items in the same way they categorise natural and cultural objects. (1983: 267) Similar insights re-emerged in a statistical analysis of German irregular verb morphology by Köpcke (1998), who also argued for an outputoriented prototypical organisation (cf. Bittner 1996: 45). Like in Bybee and 39 Moder’s study (1983) on English, Köpcke identifies certain syllable segments and segment combinations as prototypical for membership in a particular German ablaut class. A selection of these prototypical syllable segments is given in (7). (7) prototypical stem vowels (a) [ , ] (b) [ , ] (c) [ ] prototypical schemas (d) [# _ _ / ] (e) [# _ _ / ] (f) [# _ _ / ] (g) [# _ _ ( )] Köpcke called the forms given in (7d-g) prototypical schemas. Despite the terminological similarities to Bybee and Moder, Köpcke’s schemas are in fact input prototypes. They specify a best member for a class, as exemplified for instance by the German verbs schreiben (‘to write’) or sinken (‘to sink’, see schemas 7d and 7g). Note that the forms in (7) exhibit varying degrees of prototypicality. A schema like (7g), for instance, would exhibit a high degree of prototypicality for a particular irregular verb class, since there are hardly any regular monosyllabic verbs bearing the same phonological make-up. In contrast, a schema such as (7e) exhibits only medium prototypicality, as there are about as many regulars as irregulars with this phonological form. And finally, note that, when comparing the ablauttriggering prototypes from English (6) and German L1 (7g), we can see that they largely overlap (English L1 [sCC _ _ ] vs. [# _ _ ( )] for German L1). Despite the valuable insights that prototype theory has offered for German and English irregular verbs, there are a number of problems with this approach. In fact, the whole notion of prototypicality is far from consensual (cf. Bolinger 1977, Geeraerts 1988, 1989: 587, Mangasser-Wahl 2000). The following quote by Geeraerts (1989: 608) illustrates vividly how contradictions arise between the various versions and interpretations of prototypicality. While the classical Roschian position is to characterize prototype theory as non- Arestotelian and Wittgensteinian, Givón (1986) has argued that prototype theory is non Wittgensteinian […], but whereas Givón also describes prototype theory as non-Platonic, Wierzbicka […] precisely presents an explicitly Platonic version of prototype theory. Prototypicality itself seems to be a fuzzy notion (cf. Geeraerts 1989). Thus, it does not come as a surprise that, while some cognitive researchers emphasise the importance of prototype theory (cf. Storms 2001: 377), others, such as Ashby and Waldron (1999), dismiss prototypes as adequate models for categorisation altogether. 40 A second problem concerns the theory’s protagonist, the prototype itself. As it turns out, it is of a somewhat elusive nature. According to an early definition, prototypes are all those category members to which subjects compare items (Rosch and Mervis 1975), and they are a function of the total cue validity of its elements. In this sense, a prototypical category member maximally overlaps with other members of that category and minimally overlaps with members of other categories. This is in line with a number of other definitions. Pinker, for instance, argued that the best member […] is called the prototype […] and it sums up the category in people’s minds (1999: 304), while Bybee and Aitchison equate the prototype with best exemplar (Bybee 2001: 27, Aitchison 1992: 72). However, another definition, given by Hampton (1998) with reference to Rosch (1975), states that the prototype is not simply the most typical category member (1998: 128) but an idealized abstract member of a class (1998: 138). Hence, the question arises as to whether a prototype needs to comprise all prototypical features or not. This is by no means a trivial question, as the production data by Bybee and Moder (1983: 261) as well as those by Plag (2000: 144) show. In both studies, the form that attracted most of the inflections of a particular vowel change did not contain the statistically most prototypical syllable constituents. Such a discrepancy clearly needs explanation. A third problem of prototype theory concerns the role and the place of schemas. Are they truly output oriented, and if so, how can it be that a schema resembles a prototype (Bybee 1988: 135)? Or are there two types of schemas, a source-oriented and a product-oriented schema? And what precisely is a schema’s function? Is it an essential organisational requirement, does it primarily facilitate faster access, or is it an emergent byproduct of massive storage (cf. Bybee and Slobin 1982: 286, Bybee 1991: 87, Bybee 1995). Fourth, the fuzziness of category boundaries is still a matter of debate (cf. Labov 1973, Geeraerts 1988, 1989, Schmid 1998). It is unclear whether the gradient nature of prototypical categories refers to its members, its class boundaries, or both. A semantic category, such as the famous bird, for example, clearly shows degrees of typicality within the class: some birds are ‘better’ birds than others are. However, while Bolinger (1992) claims that ‘birdiness’ was a prime example of a fuzzy category with gradient borders, class membership itself is in fact far from being fuzzy. On the contrary, being a bird appears to be a rather discrete affair; as Taylor (2003: 180) put it, although penguins might not be representative of the bird category, they are birds nonetheless, not birds to a certain degree. In light of these inconsistencies, the prototypical organisation of irregular verbs still seems to be open to debate. Although prototype theory offers an interesting approach, the role 41 both of prototypical input forms and output schemas needs further investigation. 2.5 Nodes, connections, and weights - irregular verbs and connectionism One of the most prominent single-route models, and perhaps the one most frequently pitted against the dual mechanism model, is known as ‘connectionism’. This term emerged in psychology and neurology (cf. Sokolik 1990: 687) and was probably coined by Thorndike (1898) in his study on incremental learning in animals. Connectionism is a cover term for many different models, ranging from implementations such as ‘parallel distributed processing’ (PDP, Rumelhart and McClelland 1986), ‘massive storage’ (Bybee 1985, 1988) to ‘constructivist neural networks’ (Westermann 2000). For an introduction to connectionism see Bechtel and Abrahamsen (1990), and for a more recent overview see Marcus (1998), and Christiansen and Chater (1999). The demarcation of types of models and the terminology in this area can be confusing. While Eddington, for instance, classifies connectionism as a kind of prototype approach (2004: 851), prototype models in turn are sometimes assigned to the family of similarity-based models. In connectionism, language behaviour is computationally modelled in associative networks. Such networks are sub-symbolic in nature, that is, they dispense with symbol-manipulating algorithms and rule-like processes. Rules are actually held to be nothing but the manifestation of patterns of distributed memory activation within the network. In one of the earliest publications of such simulations, Rumelhart and McClelland (1986: 217) suggested that the mechanisms that process language and make judgments of grammaticality are constructed in such a way that their performance is characterizable by rules, but that the rules themselves are not written in explicit form anywhere in the mechanism. Hence, while for rule-based models language is nothing but rules all the way down, in connectionism it is memory all the way up (cf. Pinker 1999: 97). Inspired by cerebral structures, connectionist models claim a high degree of psycholinguistic validity. In contrast to the dual-route approach, whose rule-compartment is still criticised for its lack of explicitness, connectionism has been striving for explicit and transparent models in every possible aspect. As far as verb morphology is concerned, connectionism denies any categorical distinction (and physiological dissociation) between regular and irregular inflection. Instead, both processes are subjected to the same 42 mechanism. This mechanism is of a rather simple composition. It builds and organises associations between the sounds of verb stems and the sounds of past tense forms, and it generalises these associations onto new patterns if they are similar enough in form. The first (successful) connectionist model of verb morphology was devised by Rumelhart and McClelland (1986). They used a feed-forward pattern associator, sometimes called perceptron, and simulated past tense inflection both of English regular and irregular verbs. Consider figure 2, which illustrates how Rumelhart and McClelland’s pattern associator would represent the past tense of the English irregular verb string. input layer output layer distributed representations of verb stems by wickelphones distributed representations of past tenses by wickelphones “teacher” units Unit 1a ] ] Unit 1b Unit 2a ] ] Unit 2b Unit 3a [ t r [ t r Unit 3b Unit 4a [ p l [ p l Unit 4b correct output [ t r ] Unit 5a [ s p [ s p Unit 5b Unit 6a [ s t [ s t Unit 6b Unit 7a l ] l ] Unit 7b Unit 8a [ t r [ t r Unit 8b Unit 9a ] ] Unit 9b Unit … … … Unit … plus additional units to mark the beginning and end of a verb plus additional units to mark the beginning and end of a verb Figure 2: Architecture of a simple feed-forward PDP model representing the past tense formation of the verb string As we can see, the above perceptron comprises densely interconnected input and output layers, teacher units and connections between each of the units. An input unit represents parts of a verb’s stem, and an output unit represents parts of a past tense form. The lines in figure 2 symbolise synapse-like connections, which send either excitatory or inhibitory signals 43 from one unit to another, thereby activating corresponding units. The cells framed in bold face represent infinitive units for string and past tense units for strung. Typically for such a network, units are connected in a massively parallel fashion. Each unit of a particular layer is normally arranged so that it has connections to every unit of prior and subsequent layers in the network. The connections in such a simple model are usually unidirectional (cf. ‘ ’ in figure 2). The input of a pattern associator consists of an abstract representation of the sounds of a verb stem. These sounds are represented as feature matrices. In Rumelhart and McClelland’s model, a feature matrix consisted of a three-part array of feature representations of phonemes (a trigram of socalled ‘wickelphones’) as well as the word boundary symbols ‘[‘ and ‘]’. An illustration of the wickelphones that are needed to form a wickelfeature matrix for the verb string is given in (8). (8) verb string [ ] where ‘[‘ and ‘]’ represent the beginning and end of a verb wickelphones 1 2 3 In figure 2, the combination of units 1a, 6a, and 8a would represent the verb string, while the units 2b, 3b, and 6b would represent its corresponding past tense. As we can see, units as such do not represent a verb. In fact, a verb as a whole, such as string, is not represented explicitly anywhere in the model. It comes into existence only as long as the wickelphones for its sounds are activated. This means that the model’s knowledge of irregular verbs is scattered across all possible combinations of input and output units and their activation strengths. How does such a model learn how to build the past tense of string and generalise this one past tense formation onto appropriate new pairs? First, note that, in the above model, verbs that are similar in sound share units (wickelphones). Whenever, for instance, the three units for string are activated, parts of the verbs stink, cling, and bring are co-activated, too. Whenever an input unit is turned on, it sends a signal down through the input layer, where the incoming signals are multiplied by the strength of each connection and fed to the output units. Whether an output unit is eventually activated by such an impulse depends on a pre-defined threshold level. Only if this level is reached, output units are activated. In the beginning of a model’s learning procedure, all the threshold levels have the value zero and thus would not be able to activate any other unit in the model. Then, in a learning phase, the model is taught the necessary connection weights (threshold levels) which are necessary to activate those connections that represent the correct stem-sound - past-tense-sound cluster. This is done by the so-called teacher units, which represent correct 44 forms of a past tense, like strung, or sang. When the model is trained, it compares the possible output it produced with the teacher units, then adjusts connections and their weights if necessary, and thus eventually produces the correct output for a given verb. In this way, all these models need practice time on a pre-defined learning set (a number of existing irregular verbs and past tenses, for instance), until all units and connections display the best possible values so they can handle as many forms as possible correctly. In sum, learning for connectionist nets means adjusting the strengths of connections and the threshold-values, usually in a direction that minimises the discrepancy between an actual output and a desired output. A network of such a kind is said to learn without rules (Sokolik 1990: 685). Note that parallel activations of connections automatically strengthen them. Thus, patterns that are activated frequently will have stronger connections than those with rare activation. In this sense, the connectionist network is sensitive to token frequency. Patterns such as drink-drank or shrink-shrank, for instance, share input units with ring-rang. Therefore, generalisations such as [ - ] will be learned more easily by the network than a rather unorthodox vowel change. As a consequence, the model will be able to overgeneralise an ablaut pattern such as [ - ] to phonetically similar new words. It could in fact produce nonce inflections such as skring-skrang, too. In the next phase the model is presented with unknown verb forms. In Rumelhart and McClelland’s model, these were 86 existing new verbs that had been omitted in the training phase. Interestingly, once the network had been trained on existing verbs, it achieved quite remarkable results when handling new input. Overall, more than three quarters of all the new verb inflections were correct. Moreover, the network also mirrored common overgeneralisation phenomena as observed in children, so that even from an acquisitional perspective, the network seemed to provide a psycholinguistically plausible model. Hence, Rumelhart and McClelland (1986: 267) claimed to have demonstrated that a reasonable account of the acquisition of past tense can be provided without recourse to the notion of a “rule” as anything more than a description of the language. We have shown that, for this case, there is no induction problem. The child need not figure out what the rules are, nor even that there are rules. Over the last two decades, this first model of past tense inflection has been criticised repeatedly (cf. Pinker and Prince 1988), and additional models have been suggested in order to address such criticism, thereby growing into more sophisticated networks than Rumelhart and McClelland’s original design. An example for a more recent PDP application is given below in figure 3. 45 cleanup units output units 100 hidden units cleanup units semantic units input units Figure 3: Architecture of a PDP model as used by Joanisse and Seidenberg (1999) While in early models verb forms were represented phonologically only, later models, such as the above one by Joanisse and Seidenberg (1999), incorporated additional information. First, as can be seen in figure 3, semantic units are incorporated in order to represent a verb’s meaning. Second, so-called hidden units are located between the input and the output layer. These are expected to extract and represent regularities within the input (cf. Broeder and Plunkett 1997: 429). Joanisse and Seidenberg (cf. 1999: 7593), however, admitted that the semantic units did not succeed in capturing similarities in meaning between verbs in their model and hence were considered unimportant for associationist past tense formations. To sum up, connectionist research has accumulated a substantial body of evidence supporting single-route architectures of English verb inflection. And for more than two decades, connectionist networks have been used to model empirical data across many areas of linguistics (cf. Aureliu 2001, Christiansen and Chater 2001: 82). There are still, however, problems with this approach. As Pinker and Prince (1988) pointed out in their comprehensive analysis of Rumelhart and McClelland’s model, some of the connectionist networks seem to have a directionality problem. While they manage to produce meaningful past tense forms and generalise to unknown verbs, they would not be able to recognise past tense forms and generalise backwards to verb stems, because in these models the direction of the process cannot be reversed easily. This, however, as Pinker and Prince argue, is an essential part of human language competence and thus should be represented in a simulation. Second, connectionist networks seem to offer a rather cumbersome representation of speakers’ linguistic tacit knowledge about allophonic variation. Instead of recognising the systematic nature of allophonic variation across morphological categories (pronunciation of the regular past tense suffix, for instance), a connectionist network would have to compute every single pronunciation detail separately and repeatedly. And finally, due to the networks’ sole focus on phonetic form, they have problems with homophones. Since connectionist networks remain indifferent towards gram- 46 matical information such as ‘regular’ versus ‘irregular’, they have had difficulties handling inflectional differences of homophone verbs such as ring / wring, or break / brake. Despite these problems, connectionist models have succeeded in modelling a variety of behavioural phenomena. And although Pinker once described most of the network simulations in the tradition of Rumelhart and McClelland as a disappointment (Pinker 1999: 130), connectionist researchers argue that they can provide a better and psycholinguistically more plausible account of verb morphology than the one provided by the dual mechanism. They even claim that connectionism could be able to replace the currently dominant symbolic models of language structure and language processing altogether in the future (cf. Seidenberg and MacDonald 1999, Christiansen and Chater 2001). 2.6 Exemplars, nothing but exemplars - irregular verbs and analogical models Analogical models, another cover term for a variety of associationist applications, are similar in nature to connectionist networks. Analogical models include, amongst others, ‘memory-based learners’ (cf. TiMBL, Daelemans et al. 1994a, 1999, Daelemans and van den Bosch 2005), ‘analogical modeling of language’ (AML, Skousen 1989), ‘analogical natural language processing’ (Jones 1996), the ‘generalised context model’ (GCM, cf. Nakisa et al. 2000), and the ‘decision tree model’ (Ling and Marinov 1993). They all are explicit models, with stored and related information on a sub-symbolic level. Thus, similar to connectionism, rules are considered a mere epiphenomenom of analogical processing in action. In fact, as Skousen (1989) put it, language rules owed their existence entirely to the reseacher’s inspection. Accordingly, all operations in analogical models are to be conceived of as solely based on the stored instances’ similarity to new input, without any abstraction involved. Note that this is decisively different from prototype models (cf. Posner 1986) . As Eddington (2004: 851) put it, exemplaror memory-based models are founded on the idea that no sort of rule or prototypical representation needs to be generalised from the data and stored as a unit or entity separate from the data. Instead, generalizations exist within the stored lexical items themselves. Accordingly, linguistic processing is a matter of lexical access, and analogy to existing patterns found among lexical items. One of the key concepts in Eddington’s description is ‘analogy’, by now a common notion in linguistics. In a general definition, analogy is described as a kind of similarity in which the same system of relations holds across different objects (Gentner 1998: 107). At the heart of such a system of relations lies ‘analogical mapping’, a notion introduced in the 19 th century by the linguist 47 Hermann Paul (1880), who already pointed out the importance of phonological similarities within a paradigm (cf. Köpcke 1998: 51). Such similarities could - by means of proportional equations - facilitate the mapping of individual forms, such as sing, onto other individual forms, such as sang. From there, more abstract mappings could be derived, such as [# _ _ ]-[# _ _ ]), finally generalising to new and unknown forms (spling-splang, cf. Ullman 2001: 719, Bybee and Slobin 1982: 279). The rationale behind an analogical model is the assertion that linguistic processing involves analogy to past linguistic experiences, as stored in the mental lexicon. By analogy, one familiar case provides a model for making inferences about unfamiliar cases, and thus predicts the behaviour of a novel case (cf. Gentner 1983). From a computational perspective, analogical approaches seem to be the exact opposite of rule-based approaches. While rule-based generative systems have always demanded sparse representations of lexical entries, they were more than generous when it came to the complexity of the algorithms and the handling of exceptions. Astronomical dimensions of rule-governed machineries, as employed by Chomsky and Halle (1968) for instance, were easily tolerated. Analogical models, in contrast, operate with huge databases but rather simple algorithms (cf. Aha et al. 1991). Analogical models have a number of advantages. First, language behaviour can be predicted solely on the basis of language input. Second, these models are simple and transparent in design and thus can be easily compared and tested. Third, most models are robust in the sense that they can make predictions even if the linguistic input is sparse, noisy, corrupt or not well-formed (cf. Jurafski 2002). Over the past two decades, a lot of evidence for the role of analogy in morphology has been gathered from a variety of sources. For an overview of morphological applications, see Skousen (1992, 1998), Chandler (1993, 1995), Daelemanns et al. (1994b), Wulf (1996), Nakisa et al. (2000), Rytting (2000) and Lappe and Plag (2003). Not only did these models show good performances in their morphological domains, but they have been tested on production data from the dual-route studies, too. Eddington (2000a) tested two analogical models on Prasada and Pinker’s (1993) data, and it turned out that both of them handled the data in a satisfactory fashion. Eddington concluded that the evidence adduced by Prasada and Pinker was consistent with single-route approaches. English verb morphology has been addressed so far in a number of analogical models (Derwing and Skousen 1994, Chandler 1998, Eddington 2000a, Hickey and Martin 2001, Eddington 2004). Eddington, for instance, used the memory-based learner software TiMBL (Daelemans et al. 1999), whose architecture is given in figure 4 below; it uses a fictitious example from English verb morphology. 48 instance based memory sing V s -ablaut learning spin V sC -ablaut component string V sCC -ablaut run V C -ablaut blink V CC regular nearest neighbours input: drink V CC -ablaut plink sing V s -ablaut string V sCC -ablaut plink V CC ? performance component output: assign [ ]-ablaut and thus create [ ] Figure 4: Architecture of a memory-based learner software, illustrated with an example for English verb morphology As we can see in figure 4, the software consists of two primary parts, a learning component and a performance component. The learning component (a so-called ‘lazy learner’), the basis for all computational processes in the performance component, is a database full of feature matrix representations of the linguistic input. As we can see in figure 4, information is stored in fully specified matrices, containing representations of onset, nucleus, and coda, as well as information on a target feature, which in the above case is a verb’s past tense inflection (type of ablaut vs. regular inflection). Note that the performance unit of such a model basically carries out a classification task, assigning past tense inflections to new input. First, a new item is presented to the model (input plink). An algorithm, containing a distance metric for pre-defined features, searches the database for similar verbs, which are those items that share most of the features with the new input. The algorithm then selects the best exemplars, the so-called ‘nearest neighbours’, and, via majority choice, suggests a specific morphological treatment for the new input. In our case, plink was compared to drink, sing and string, and eventually classified as a new member of the [ ]-ablaut vowel change class (plink-plank). Despite the valuable insights that analogical studies have provided, there are some problems with this type of modelling, too. First, analogical formations need to be constrained in some way to prevent nonsensical or phonotactically ungrammatical formations (cf. Jespersen 1942: 33, Becker 1990: 24). Second, analogical equations need to be complemented by some sort of constraint in order to predict properly, when and under which circumstances analogical formations can and should occur (cf. Itkonen 2005, Albright 2009). Consider, for instance, the verbs strike and sneak, which are the two most recent additions to the English [ ]-ablaut class of irregular verbs (strike-struck, sneak-snuck). How could a proportional equa- 49 tion, on the mere basis of similarity, produce a novel form when no phonological template exists? Prior to the arrival of sneak and strike, when a pair such as [# _ ] - [# _ ] was simply not part of the class, a schema [# _ ] could not possibly have extended to a new verb such as strike (cf. Bybee and Moder 1983: 255). It would thus seem as if the classical analogical equation is not sufficient to explain the dynamic growth of a paradigm. Another problem concerns the necessary number of exemplars or prototypes for analogical processes to be activated. How many of them does it take, for instance, to draw analogies and include a new form based on existing forms? In other words, is there something like a critical mass required for analogical reasoning to take place? This is indeed a vital question, since some researchers, like Albright and Hayes (2003), reject analogy as a basis for verb morphology. Instead, they suggest rule-like microgeneralisations (see the following section), which can operate, if need be, on one single infinitive-past-tense-pair. If such micro-generalisations turned out to be psycholinguistically valid, there would be no need for critical masses in analogical models. In other words, if Albright and Hayes’ argument were accurate, similarity-based reasoning would need to be able to work even on extremely small groups of exemplars. 2.7 Rules, ever more general rules - irregular verbs and ‘islands of reliability’ For most of the past two decades, the nature of past tense morphology has been debated between two theoretical camps. As discussed in section 2.3, one camp argued for single-route, the other for dual route verb morphology. And while the latter one included rule-based operations, single route models usually relied only on associations and analogy. Recently, though, single-route models have found new company. Albright and Hayes (2003: 119) propose a third approach, which uses multiple stochastic rules and no analogy. We propose a model that employs inductive learning to discover multiple rules, and assigns them confidence scores based on their performance in the lexicon. This model, which was tested against an analogical learner and fresh production data, challenges not only associationist approaches but also the rule-route of the dual mechanism. Albright and Hayes’ model incrementally derives morphological generalisations through iterative rule application, called ‘minimal generalisation’. Rules, in this context, express morphological changes in a specific phonological environment. A general rule for regular past tense formation in English could be formalised as given in (9), where t represents all allomorphic variants of the regular past tense morpheme. 50 (9) Ø t / __ # In a first step of the minimal generalisation process, the model learns wordspecific rules, starting from single word pairs. Then, these rules are combined in order to derive more general rules (cf. Albright and Hayes 2003: 124). This procedure is illustrated in (10). (10) (a) Ø d / [ __ ] [+past] word-specific rule for sign Ø d / [ __ ] [+past] word-specific rule consign + strident Ø d / [ X + continuant ] [+past] general rule - voice (b) Ø d / [ __ ] [+past] word-specific rule for vote Ø d / [ __ ] [+past] word-specific rule for need + coronal Ø d / [ X + anterior ] [+past] general rule - nasal - continuant In a second step, the accuracy of each rule is assessed by a confidence value, expressing how accurately the rules derive forms from a learning data set. The vowel change from cling to clung, for instance, could be formalised by the general rule given in (11). (11) [ ] [ ] / C(C)(C) __ [+past] The specified context in (11) covered nine irregular verbs in Albright and Hayes’ learning data set (scope) and produced six correct formations (hits). Its confidence value (hits/ scope) hence was .67. This measure of accuracy was supposed to help identify the most accurate rules. In order to test the model’s accuracy, Albright and Hayes compared its performance to production data. In a study loosely based on Prasada and Pinker’s design (1993) participants either produced past tense forms or provided acceptability ratings for given past tense forms. The results of the production experiments provided two main insights. First, participants volunteered a majority of regular responses and rated given regular forms as more acceptable than their irregular counterparts. This preference, however, differed significantly between the two experiments, which Albright and Hayes ascribe to the different nature of the tasks. Second, participants’ ratings of existing regular verbs showed gradience. This is remarkable, since, according to the dual-route approach, 51 only irregulars should be susceptible to such effects. Apparently, though, both regulars and irregulars are capable of forming subgeneralisations. The gradient behaviour shown in the production data was successfully replicated in the rule-model. Since the model allowed co-occurrence of highly general with less general rules, it was capable of accommodating more than one rule for the same structural changes, differing only in the rule’s degree of confidence. Consider again the general (default) rule for regular past tense formation as given in (9). Now compare this rule with the less general, more specific, and hence more ‘successful’ rule given in (12). (12) - sonorant Ø / [ X + continuant ] [+past] - voice The above rule states that all English verbs ending in a voiceless fricative are suffixed with [ ]. While the rule in (9) applies to all English regular verbs, the rule in (12) is restricted to around 350 verbs in Albright and Hayes’ learning set. Despite this constraint, however, rule (12) would receive a higher confidence value than rule (9), and thus it would remain in the model. A rule as shown in (12) formalises a particular context in which the rule itself can apply with a high degree of confidence. Albright and Hayes referred to such a context as an ‘island of reliability’, a context in which a morphological process is especially robust (Albright 2002: 685). Recall that these islands appear in regular verb inflection, too, rendering the area of the default gradient. Recall, though, that according to the dual route approach the default is supposed to be categorical, ranging over an entire syntactic category irrespective of the input’s form or frequency. Apart from these neighbourhood effects, Albright and Hayes’ model also describes the prototypicality of a verb class. The degree of proximity towards a prototype can be represented on a scale of increasing levels of similarity, ranging from factor 0 to 1, as illustrated in (13) for the cling-clung class in English. (13) .0 .054 .183 .466 .629 .792 .880 1.00 wush trup fub spun strung stuck clung prototype stung flung Although Albright and Hayes point out that this continuum was not to be interpreted as a proportional measure of similarity towards the prototype, it nevertheless seems to reflect a gradually increasing degree of family resemblance within the cling-clung class. Overall, Albright and Hayes’ rule-model handles their production data reasonably well. Both the speakers and the model seem to be sensitive to islands of reliability effects in the area of regulars and irregulars alike (cf. 52 Ernestus and Baayen 2003 for the same conclusion about Dutch verb morphology, and Albright 2002 for Italian). As a consequence, Albright and Hayes advocate a set of rules (specific rules and sub-rules, not one contextfree default rule) as the appropriate means of dealing with morphological generalisations in English verb morphology. As far as irregular verb morphology is concerned, however, the question is whether this model really adds a new perspective to the debate. Note, for instance, that islands of reliability bear a striking resemblance to the prototypes put forward by Bybee and Moder (1983) and Köpcke (1998). A synopsis is given in (14). (14) (a) Albright and Hayes’ rule for the [ ]-[ ] vowel change (English): [ ] [ ] / __ ] [+past] (b) Bybee and Moder’s prototype for the [ ]-[ ] vowel change (English): [sCC __ __ ] (c) Köpcke’s prototype for the [ ]-[ ] vowel change (German): [# __ __ ( )] A closer look suggests that these forms seem to be just different ways of spelling out the same thing, namely phonological environments in which a particular morphological process is very likely to apply. This raises the crucial question as to the qualitative difference between Albright and Hayes’ micro-level subregularities on the one hand and analogical sets on the other. Bybee (1988: 135) denounced any such difference, pointing out that some rules showed properties commonly associated with representations, while some representations resembled rules. She goes on claiming that widely applicable morphological schemas (what are usually called “rules”), such as the regular suffixation for English past tense, appear to be free of the lexicon in the sense that they apply readily to new forms. Later on, she remarked that the difference between phenomena handled by schemas and phenomena handled by rules is mainly a quantitative one (cf. Bybee 1991: 86). In other words, rules could also be interpreted as highly reinforced representational patterns or schemas (cf. Bybee 1988: 135). In a similar vein, Pinker discussed the ambiguity of symbolic rules and similarity-based pattern associations with respect to the design of Rumelhart and McClelland’s perceptron. Pinker (1999: 118) claimed that each of the connections were a probabilistic microrule that states something like, ‘If the stem contains a stop consonant followed by a high vowel, the past-tense form is likely to contain a nasal consonant at the end.’ With 460 input units connected to 460 output units, we have 460 x 460 = 211,600 microrules in all. 53 Whether or not morphological generalisations had better be represented as rules or exemplar-driven analogy is a crucial question, though. First, rules are supposed to be directional, deriving an output form from a given input form, while schemas are not. Thus, in Albright and Hayes’s model every single past tense form is supposed to be derived directionally. If such a model were an accurate description of real language competence, speakers should be able to generalise from an input form to an output form, but not vice versa. It remains to be seen, though, whether behavioural data show such directionality. A second decisive difference, as Bybee (1995) already pointed out, concerns the emergence of the similarity-based analogies. In similarity-based models (Bybee 1995: 428), actual usage in terms of both type and token frequency plays an important role in establishing and maintaining representations. In Albright and Hayes’ study, however, no reference is made to the role of language use. And third, rule models and analogical models differ in terms of computational load. While in a rule model, elaborate algorithms work with categorical input, and fully stored exemplars represent lexicalised exceptions only, an analogical model would be based on extensive databases and rather simple algorithms. As far as cognitive simplicity and transparency is concerned, some researchers argue that it would be more parsimonious to assume that a mental lexicon was governed entirely by analogy. 2.8 frinked or frank - irregular verbs and the role of semantics Although proponents of singleand dual-route inflection vehemently disagree about the cognitive nature of the mechanism involved, one issue has been consensual: associations between irregular verbs are regarded to be exclusively morphophonological (be they represented by rules or analogies). In other words, past tense formation was long held to be sensitive to morphological form only. This, however, has turned out to be an oversimplification. While researchers, both dual-route (Pinker and Ullman 2002a) and single-route (Joanisse and Seidenberg 1999: 759) have explicitly excluded the role of semantics over decades, now there is evidence that a verb’s meaning can indeed play a crucial role in verb inflection (Ramscar 2002a). Regular and irregular verbs, for instance, differ not only in their morphophonological but also in their semantic dimensions, as revealed by corpus-based analyses of English, German, and Dutch verb morphology (Baayen and Moscoso del Prado Martín 2005, Tabak et al. 2005). Based on these insights, Baayen and Moscoso del Prado Martín (2005: 668) defined the role of semantics as a 54 conspiracy of subtle probabilistic (graded) semantic distributional properties that lead to irregulars having somewhat different semantic properties compared to regulars. Dutch irregulars, for instance, turn out to have more densely populated semantic neighbourhoods than regulars; they show a higher degree of lexical relations and semantic similarity amongst each other than regular verbs. This, in turn, could prove to be a confound for the attested differences in inflectional behaviour. Second, a verb’s semantics seems to influence the way speakers inflect it (Ramscar 2002a, 2002b, 2002c). In a series of experiments, Ramscar showed that both regular and irregular inflection in English could be affected by verb meanings. In one of his experiments, participants were asked to inflect a given nonce-verb (frink or sprink) for past tense in one out of three semantically varying contexts. Interestingly, their inflectional behaviour seemed to depend on the kind of context in which the verb was embedded. Participants exposed to the nonce to frink in a context priming the meaning of drink showed a significant bias towards irregular past tense inflection (frank), whereas in a context associated with the real verb blink, they tended to produce regular inflections (frinked). This is illustrated in (15). (15) nonce frink context 1 context 2 story about ‘drinking’ story about ‘blinking’ bias frank bias frinked Apart from producing inflections, participants were also asked to indicate the contexts’ priming qualities. Three follow-up questions checked whether participants had actually used semantics as the main attractor, that is, if they had mapped the meaning of the prime onto the nonce. The overall results of this experiment suggested that semantic contexts succeed in priming regular and irregular inflections. Ramscar (2002a: 59) concluded that [i]f people encounter a novel verb form that is phonologically close to two different existing verbs but semantically close to only one of them, then they are likely to inflect the nonce-verb in the same way as the semantically similar verb. It is worth pointing out, though, that Ramscar’s results merely substantiate the hypothesis of local, item-based interactions of phonological and semantic attraction (frink-drink/ blink). So far, the individual role of a verb’s mean- 55 ing for inflection turns out to be a rather small one and it remains to be seen whether semantic similarity can serve as an attractor for phonologically dissimilar primes. Pinker and Ullman (2002a) as well as Huang and Pinker (2005) commented on the conflation of semantic and morphophonological attraction, claiming that Ramscar’s test design was fundamentally flawed. Due to the extreme similarity of nonce and prime, one being a near-doppelgänger of the other (Huang and Pinker 2005: 1006), semantic and phonological attraction might be confounded. As Pinker and Ullman (2002a: 473) put it, people were in effect given the question ’Does the experimenter want me to treat frink as a distorted version of drink, or of blink. In other words, participants’ reactions might have been based on purely lexical generalisations. This conclusion, however, is at odds with Ramscar’s significant biases towards prime consistent responses (2002a: 58). According to Pinker and Ullman’s reasoning, Ramscar’s participants should have shown a near-chance distribution of regular (frinked) and irregular (frank) responses, which, as Ramscar’s data suggest, was clearly not the case. In a production experiment, in which phonology and semantics were varied independently, Huang and Pinker (2005: 1004) explored the interaction between the two parameters in more detail. Based on the results they concluded that subjects’ tendency to accept an irregular past tense increased as similarities to known irregular verbs increased. Results from the semantic manipulation suggest that subjects’ tendency to accept an irregular past tense remained resistant to variation in semantic similarity unless the meaning of the novel verb highly resembled that of known irregular verbs […]. These findings demonstrate that experiments on the role of meaning for inflection need to vary phonological as well as semantic similarity in a systematic fashion. And this is what Huang and Pinker tried to do in their experiments. However, their variation of semantic contexts appears to be rather unsystematic. Obviously, levels of semantic similarity are much harder to calculate than phonological similarities, but Huang and Pinker’s study provides no clues as to how they devised these levels. In sum, the role of semantics in past tense inflection still seems to be uncharted area, and we have only scratched its surface (Baayen and Moscoso del Prado Martín 2005: 695). 56 2.9 sing sang sung - vowel change classes or universal apophony As pointed out before, theories of verb morphology, be it in terms of rule derivation, prototypical organisation, similarity or analogy, have had their difficulties describing English and German ablaut in a satisfactory fashion. In the following section, an alternative, universal approach to ablaut will be discussed. Ségéral and Scheer (1998) developed the notion of the so-called ‘apophonic path’ and claimed that ablaut (apophony) across all types of languages followed the principles spelled out by this path. Traditionally, apophony describes context-free vowel alternations that express a grammatical opposition and is usually being thought of as unpredictable and hence fully lexicalised (all vowel qualities involved). Nevertheless, Ségéral and Scheer claim that their apophonic path formalises a structured, predictable, and in fact universal morphological process 6 . What this path does is relating different verb forms within a paradigm. This means that one form is basic and serves as input for a morphological process. Accordingly, as Ségéral and Scheer (1998: 29) claim, one form of the verb (present tense), together with the universal apophonic mechanism, contains all the information necessary in order to predict the other verbal forms (preterite, past participle). This would mean that only one base form of an irregular verb needed to be fully represented in the mental lexicon, while all other ablauting forms could be derived in real time by apophonic rules. The apophonic path is given in (16). (16) Ø i a u u This path defines three different possible basic vowel qualities (i, a, u) and four apophonic relations along with their directionality, marked by ‘ ’. Skipping elements in the path is ruled out. Note that the path represents the vowels’ markedness, with Ø being the most unmarked element, and [u] the most marked one. It is, of course, immediately obvious that the path in (16) does not cover the entire range of German vowel change series at a surface level. However, Ségéral and Scheer claim that some of the German vowels involved in apophony are in fact complex, and thus need to be re-analysed on an infrasegmental level. At this level, vowels can consist of ‘parasitic’ and ‘apo- 6 So far, this theory has been examined with respect to Classical Arabic (Guerssel and Lowenstamm 1996), Berber (Bendjaballah and Haiden 2002, 2005), Akkadian (Ségéral 2000), German (Ségéral and Scheer 1998) and Spanish (Boyé 2000). 57 phonic’ elements (cf. Kaye et al. 1985), only one of which - namely the apophonic element - can enter the apophonic mechanism. The nucleus [ ] of the German verb bergen (‘recover’, ‘salvage’), for instance, would, on an infra-segmental level, consist of the parasitic element ‘a’, and the apophonic ‘i’. Speakers, then, knowing implicitly that the ‘i’ can never be ‘parasitic’ and thus must be apophonic in German, would automatically feed ‘i’ into the apophonic path, thus yielding the correct form barg for past tense. Based on this admittedly challenging re-analysis, the three vowels of the apophonic path are actually supposed to account for all of the test items used in the German experiments in the present dissertation (cf. Ségéral and Scheer 1998: 45). A summary of the vowel change patterns used is presented in (17). (17) (a) [ ] [ ] [ ] singen-sang-gesungen (b) [ ] [ ] [ ] rinnen-rann-geronnen (c) [ ] [ ] [ ] biegen-bog-gebogen (d) [ ] [ ] [ ] liegen-lag-gelegen The alleged universality of apophony is intriguing, especially since both English and German provide additional supportive evidence from children’s nursery rhymes, onomatopoeic expressions, and expletives. Amongst these, apophonic vowel changes such as [ - ] and [ - ] are remarkably frequent. The pattern [ - - ] is in fact the most frequent type of a three-vowel-ablaut pattern in German (cf. Marchand 1969: 430, Augst 1975), and this type of vowel change has already been described as a triad of optimally discernible vowels, or a vocalic triangle (cf. Martinet 1955: 263-64). Some examples for onomatopoeic expressions are shown in (18). (18) German examples English examples Bi-Ba-Butzemann Mischmasch bibble-babble kit-cat bim-bam-bum piff-paff-puff chit-chat mishmash hicke-hacke ri-ra-rutsch click-clack rickety-rackety kribbel-krabbel Singsang dingle-dangle riffraff Krimskrams Zick-Zack drizzle-drazzle zigzag If this apophonic path were to be psycholinguistically real, it would seriously challenge connectionist and analogical models. However, Ségéral and Scheer’s theory has its problems. First, as far as their analysis of German is concerned, the empirical basis is somewhat artificial. While they claim that their apophonic path is fully applicable from a synchronic point of view, their pool of analysed verbs lacks the class of mixed verbs, which should only be excluded on diachronic grounds (cf. Paul et al. 1989: 257). Instead, they include forms which are either archaic or whose irregularity, again from a strictly synchronic point of view, is questionable (dingen, 58 erkiesen, kreischen, schnauben). Overall, their reference list of irregular verbs contains 21 more or less debatable cases. In order to support a claim for universality, however, the selection of items on which the analysis is built should be maximally transparent and representative. Second, their analysis is built on some profound simplifications of the German vowel change system. Note that the apophonic path can account for the following seven two-stage and three-stage vowel change patterns (19). (19) Ø i i a a u u u Ø i a i a u a u u As it turns out, though, not all English and German irregular verbs follow one of the above changes. This is a serious caveat to the path’s universality. Ségéral and Scheer compensate for this by reducing the vowel system to what they claim are the phonologically relevant properties only (1998: 37). They dispense, for instance, with length and ATR contrasts, represent diphthongs as vowels followed by off-glides ([ ] is restated as [ ]), and they re-analyse a number of surface vowels as derived from underlying structures. The infinitive of the verb stoßen (‘to push’) [ ], for instance, is supposed to be derived from the underlying form [ ], the verb’s alleged synchronic structure (Ségéral and Scheer 1998: 54). Thus, they seem to treat the long [ ] as though it were a surface product of an underlying, diachronically defined dipththong. Leaving this kind of reasoning unexplained to a large extent, Ségéral and Scheer end up with a whole range of underlying and derived forms in order to force as many German vowel change series as possible into their apophonic mechanism. A third problematic issue is the mechanism’s surprising flexibility. Evidently, it allows partial apophony, so that in a three-stage vowel change (a-b-c), for instance, only ‘a’ and ‘b’ need to be related as predicted by the path. And fourth, only 59 percent of German strong verbs, by their own count, fully fit the apophonic model. The remaining ones are either merely halfor non-apophonic. In total, around 9% of all German irregular verbs (containing high frequent forms such as kommen (‘to come’) and rufen (‘to call’)) cannot be accommodated or predicted by the apophonic path. In order to substantiate Ségéral and Scheer’s claims, behavioural data need to be obtained from ablauting languages. A first step in this direction has been taken in experiment 2 (chapter 4), in which English and German past tense and past participle formations were tested in tandem. If a uni- 59 versal ablaut path were psycholinguistically real, we should see speakers of English and German use solely those changes spelled out in the apophonic path. And we should be able to observe directionality effects. 2.10 Irregular verbs and L2 acquisition As the previous sections have illustrated, verb morphology in L1 is an intricate and challenging affair. The same seems to be true for verb morphology in L2. Although morphemes and their acquisition by L2 learners have been debated widely over the last three decades, a range of conflicting evidence can be found in the literature. On the one hand, there is evidence, based mainly on connectionist studies, which supports a single-route approach to L2 verb morphology, such as provided by Blackwell and Broeder (1992), Sokolik and Smith (1992), Chandler (1994), Broeder and Plunkett (1997), Seidenberg and MacDonald (1999), Nakisa et al. (2000), Indefrey (2006), and Colombo et al. (2006). On the other hand, studies suggest that L2 verb morphology corresponds with predictions made by the dual-route approach, too (Beck 1997, Flege et al. 1999, Murphy 2000, Birdsong and Flege 2001, Gor and Chernigovskaya 2003, Hahne et al. 2006, Kırkıcı 2007). One dual-route account, as proposed by Flege et al. (1999) for English irregular L2 verb morphology, was based on dissociations between regulars and irregulars in terms of grammatical accuracy. In their experiments, L1-like accuracy of irregular verb inflection turned out to be a function of age of arrival (AoA) of interlanguage speakers in a foreign country. The higher the AoA, the lower the accuracy. Interestingly, such a correlation could not be established for regular verbs. In another study, Birdsong and Flege (2001) found significantly differing frequency effects for regular and irregular verbs in L2, thereby suggesting a dual-route approach, too. Unfortunately, however, none of the above results addressed the internal organisation of irregular L2 verb morphology. This was probably first done by Plag (2000). Transferring Bybee and Moder’s (1983) study into the area of the interlanguage, he observed interesting similarities as well as differences between how English L1 and German-English interlanguage speakers treated nonce-verbs. The main results of Plag’s as well as Bybee and Moder’s study are summarised in table 4. 60 Table 4: Main results of Bybee and Moder (1983, English L1) and Plag (2000, German-English interlanguage) English L1 speakers German-English interlanguage speakers frequency of ablaut 83.0% 30.4% most frequent ablaut two-stage [ ]-ablaut three-stage [ ]-ablaut importance of segments coda onset nucleus nucleus coda onset highest ablaut frequency [sCC _ _ ] [sC _ _ ] prototype [sCC _ _ ] [sCC _ _ / ] First, table 4 illustrates that the two groups of speakers produced different proportions of regular versus non-regular responses. This raises the question as to whether the willingness to ignore the alleged default mechanism is specific to the type of speaker. Second, both groups show different tendencies towards ablaut formation strategies. Why, for instance, is the threestage-ablaut via [ ] prototypical for the interlanguage speakers, whereas English L1 speakers seem to prefer the two-stage [ ]-ablaut. Third, the importance of the three syllable constituents onset, nucleus, and coda differs, too, which could be an indication for two groups of speakers perceiving the constituents’ salience differently. And finally, in both studies the statistically analysed prototype does not overlap completely with the form that attracted most of the ablaut formations. As far as this lack of overlap is concerned, Plag (2000: 146) already remarked that it would be possible to simply take the form with the highest proportion of ablaut responses as the prototype, which, however, has the disadvantage of making the stronger vowel effect of / / against / / in Bybee and Moder’s experiment, and the stronger onset effect of sCC as against sC in our experiment miraculous. For English L1 speakers, the prototype and the verb with most of the ablaut formations differ with regard to the nucleus, while for German-English interlanguage speakers onset and coda do not overlap. Such a lack of complete congruency violates the classical definition of a prototype, according to which a prototypical form should comprise the highest possible number of category specific attributes. There is, however, a possible explanation for this nucleus discrepancy in English L1. In Bybee and Moder’s experiment, English speakers systematically avoided no-change responses, which in turn reduced chances for input-vowels to be repeated as output-vowels. Thus, an input structure containing one of the two major ablaut outputvowels would automatically increase the frequency of the output-vowel that it did not contain; and nucleus [ ] happened to be one of the two major ablaut vowels. In other words, [ ]-input forms were bound to produce 61 [ ]-outputs (like in sang-sung) in a disproportional way. Such behaviour skews the results in favour of this particular ablaut. Plag already suggests a variety of explanations for the interlanguage ablaut strategies, referring to input frequency, perceptual salience, transfer, (cf. Odlin 2000, Goldschneider and DeKeyser 2001) and the universal path (Ségéral and Scheer 1998). If input frequency played a major role in L2 irregular verb acquisition, then we would have to assume that German learners of English, for some reason, are more exposed to English [ ]-[ ] than to [ ]-[ ] vowel change patterns (cf. Plag 2000: 146), since they were obviously more attracted to three-stage patterns. Such prevalence, however, would be hard to prove, both in the input interlanguage which speakers are exposed to and in learner corpora. Although frequency and perceptual salience, along with other determinants such as regularity, syntactic and semantic complexity (Larsen-Freeman 1976), invariance of form (Hatch 1983), L1 transfer (Odlin 2000), and individual differences (Andersen 1978) have repeatedly been claimed to influence morphological acquisition, we lack clear and applicable definitions of these concepts (cf. Larsen- Freeman 1976). What exactly, for example, renders a morpheme (or a process) salient relative to other morphemes (processes)? And how does a learner make use of a salient structure? In this respect, phonological and morphological similarity seem to play an important role (cf. Gass and Selinker 2001: 126-132, Lowie 1998), but very little has been said about how exactly salience emerges and works (cf. Ellis 2006). As Plag already pointed out, empirical evidence for frequency effects would necessitate a systematic investigation of an interlanguage speaker’s input. Obviously, this will be hard to measure. As Larsen-Freeman (2002: 275) pointed out recently, L2 research simply lacks neurologically plausible models and technologically convenient means of measuring and testing frequency effects in input. Still, as relevant and important as a frequency factor is, it requires greater definition and qualification. For instance, L2 learners are agents of their own learning process. They do not merely record frequency; they categorize frequently occurring patterns, abstract, and generalize from them. A second possible source for the similarities and differences between L1 and L2 speakers, as illustrated in table 4, could be cross-linguistic influences (Kellerman 1986), or transfer (cf. Odlin 2000). However, these two concepts are far from consensual, too. Gass (1983), for instance, defines language transfer as the superposition of L1 structures onto L2 structures. However, what would such a superimposed structure look like? Certainly, as far as irregular verbs are concerned, interlanguage speakers do not just copy from their L1 ablaut patterns and thus replace the target structures. So, what exactly does superimpose mean then, and how can one find out if 62 an interlanguage structure is in fact the result of superimposed transfer? Odlin (2000: 27), in contrast, claims that transfer was the influence resulting from similarities and differences between the target language and any other language that has been previously (and perhaps imperfectly) acquired. Again, it remains unclear how such an influence would work in the case of irregular verb morphology and how it is constrained. As it turns out, a plethora of constraining conditions, such as markedness, frequency, salience, transparency, concreteness, language distance - amongst others - can be found in the literature (cf. Shirai 1992, Odlin 2000). The notion of ‘markedness’, for instance, predicts that the less marked form, which is usually acquired first, is more likely to be transferred (for a counterexample see Bardovi-Harlig 1987). The so-called ‘markedness differential hypothesis’ (Eckmann 1977) posits that difficulties in the L2 are supposed to occur in those areas that are different from, and relatively more marked than the L1. There is also evidence that, from a learner’s perspective, infrequent items can be considered marked and are therefore less likely to undergo transfer (cf. Kellerman 1983). Since markedness, however, seems to be linked with frequency (Selinker 1983: 50), positive evidence could only be established once frequency analyses revealed that a statistically significant trend in the speaker’s native language appears […] and is then paralleled by a significant trend toward the same alternative in the speaker’s interlanguage behavior. As pointed out above, though, measuring frequencies in L2 seems hard to accomplish. Moreover, we do not really know what exactly makes a structure more marked than another, and under which circumstances marked and unmarked structures are transferred (cf. Shirai 1992: 103). Kellerman (1983: 117), addressing this issue, developed the concept of ‘transferability’, which defines the probability with which a structure will be transferred relative to other structures in the L1. According to this notion, an L1 item is less transferable if it is perceived by the learner as a marked form. Marked forms, in this contexts, are usually irregular, infrequent, and semantically or structurally opaque. This includes not only grammatically marked forms but also lexical items that are perceived by the learner as non-prototypical or infrequent (cf. Kellerman 1986). What does this mean for German-English interlanguage irregular verb morphology? Plag claimed that his data could be taken as evidence for transfer processes. Let us assume that, commensurate with the principle of interlingual transfer (Shirai 1992: 100), L1 and L2 are similar in structure, and that hence L1-L2 mapping is likely to occur. Let us also assume that German learners of English perceive certain vowel change patterns as frequent, structurally transparent and salient in their L1, and hence superimpose 63 them onto similar structures in their L2. What would be the result? For the English sing-sang-sung class, such a situation is illustrated in (20). (20) potential German L1 prototype (Köpcke 1998) # ( ) # ( ) # ( ) potential English L1 prototype (Bybee and Moder 1983) sCC sCC sCC potential interlanguage prototype? sCC ( ) sCC / ( ) sCC / ( ) If German interlanguage speakers superimposed their segments in such a fashion, [ ]-[ ]-[ ] or [ ]-[ ]-[ ] vowel changes should emerge in the interlanguage data. However, as Plag points out, his data did not provide a single case for [ ]-ablaut. This is not surprising, since such a vowel change does simply not exist in Modern English. Consequently, a simple mapping of cross-linguistic equivalents is very unlikely to occur with advanced interlanguage speakers. Instead, Plag (2000: 147) suggested that learners transfer the prototypical three-stage ablaut of German and realize it in their English interlanguage in a way compatible with the output of class 1 items [[ - ] vowel changes, my addition ] in English. If German L2 speakers’ ablaut were constrained in such a way, their vowel change generalisations should be limited to existing patterns only. If, however, they produced patterns alien to the target language, the above transfer scenario would need to be revisited. 2.11 Desiderata and hypotheses From all the problems and questions emerging in the discussion of the theoretical background, only a small selection can be addressed in the following empirical studies. First, the psycholinguistic reality of a prototypical organisation of German-English interlanguage irregular verb morphology will be investigated. We know that some classes of irregular verbs show prototypical characteristics in English L1 (Bybee and Slobin 1982, Bybee and Moder 1983, Prasada and Pinker 1993). We also know that prototype effects have emerged in L2 data with German-English interlanguage speakers (Plag 2000). And there is evidence that German irregular verbs show a prototypical structure, too (Köpcke 1998). However, we are faced with conflicting evidence as to the 64 morphophonological details of a potential prototype. Therefore, we cannot be sure whether such an organisation is in fact prototype based, or maybe associationist in some other, similarity-based and analogical fashion. As a result, the following hypothesis has been formulated in order to address this issue (21). (21) Hypothesis 1 German-English interlanguage irregular verb morphology is best modelled according to prototype theory. The morphophonological properties of L2 prototypes emerge through transfer from the L1 and serve as morphophonological schemas in the interlanguage. In order to test this hypothesis, three elicitation experiments investigated first the morphophonological organisation of irregular verbs in German L1, English L1, and the German-English interlanguage, and second whether similarities between German L1 and the interlanguage could be accounted for by transfer processes. In these experiments, participants had to inflect given nonce-verbs, whose morphophonological make-up varied systematically with respect to potential prototypes as suggested in the literature. Fresh data from English L1 were obtained in order to revisit the reported inconsistencies in Bybee and Moder’s data (1983) and assess, whether it is possible to distil a prototype from the English L1 data. If prototype theory turned out to govern interlanguage irregular verb morphology, production data should reveal the presence of a clearly discernible prototype at the heart of an ablaut class. Second, the psycholinguistic reality of the apophonic path was explored. Should such a universal path be universal, it would be traceable in English and German L1 as well as German-English interlanguage data. The following hypothesis was adopted. (22) Hypothesis 2 German-English interlanguage irregular verb morphology is best modelled as rule-like input-output processes that follow the apophonic path. According to this hypothesis, speakers’ vowel changes should be unidirectional, going only from a given input (infinitive, for instance) to a derived output (past tense, for instance). Given the path’s directionality, vowel change creations in a reverse direction should not occur. Moreover, all vowel changes that do occur should conform to the predictions made by the apophonic path. Should speakers, however, create ablaut in an arbitrary fashion, or should the German-English interlanguage speakers invent ablaut patterns alien both to their L1 and to the target language, then the psycholinguistic validity of such a path could be rejected. In order to test hypothesis 2, six elicitation experiments examined in how far vowel changes in past tense and participle inflection reflect the predictions made 65 by the apophonic path. In these experiments, participants were given both nonce infinitives to inflect for past tense and participle, and participles to inflect in a backward fashion for past tense and infinitive. Again, this was done with respect to German L1, English L1, and the interlanguage speakers. Third, the role of semantics for irregular verb inflection in the German- English interlanguage was explored. The pertinent hypothesis is presented in (23). (23) Hypothesis 3 German-English interlanguage irregular verb morphology is co-determined by a verb’s semantic similarity to existing verbs. These similarities work as attractors independent of phonological similarities to existing verbs. In order to test hypothesis 3, three elicitation experiments, based on Ramscar’s experiments (2002a), tested whether varying semantic environments of a given nonce-verb influenced its inflection for past tense. These experiments were carried out for German L1 and the interlanguage. Note that one crucial aspect of hypothesis 3 is the independence of phonological and semantic attraction. This means that participants in this experiment should prove to be significantly biased by semantic contexts with a phonologically dissimilar prime. Recall that Ramscar’s semantic effects depended on phonological similarity between nonce-verb and prime in the first place (frinkdrink/ blink, cf. section 2.8). In the present replication, however, semantic contexts priming phonologically dissimilar verbs were examined, too (frink-ride). If semantic similarity were capable of serving as an individual attractor, participants should feel biased to inflect frink for frank in a context priming the real verb ride. If phonologically dissimilar nonce-verbs, however, fail to trigger such a bias, then the role of semantics is either inseparably coupled with phonology, or it is in general much less effective than assumed for irregular verb inflection so far. 67 3 Experiment 1 This chapter provides a detailed account of experiment 1. First the experimental design will be described, followed by a discussion of the results. Recall the hypothesis that guided this experiment. It is restated for convenience in (24). (24) Hypothesis 1 German-English interlanguage irregular verb morphology is best modelled according to prototype theory. The morphophonological properties of L2 prototypes emerge through transfer from the L1 and serve as schemas in the interlanguage. Given the assumption that irregular verb morphology in the interlanguage can be regarded as the result of segment mappings from the L1 onto the L2, it is necessary to investigate irregular verb morphology in German native speakers, speakers of the target language English, as well as speakers of the interlanguage. Only then can you investigate possible mapping scenarios in detail. As a consequence, behavioural data were obtained from three groups of speakers, with experiment 1 being carried out three times, as illustrated in (25). (25) types of speaker experiment code German L1 speakers experiment 1_G_L1 English L1 speakers experiment 1_E_L1 German-English interlanguage speakers experiment 1_E_L2 In the experiment code (e.g. 1_G_L1) the first number denotes the experiment (number 1), and the following capital letter along with the specification L1/ L2 denotes the participants’ language. Thus, 1_E_L1 contains the results of English L1 speakers going through experiment 1, and 1_E_L2 contains the results produced by the interlanguage speakers. 3.1 Experimental design Experiment 1 was a paper-and-pencil elicitation task. Participants had to provide past tense forms to given nonce-verbs, and the nonce’s syllable constituents varied systematically in order to test their influence on the type of past tense formation. The nonces, along with 15 distractors, were distributed in an excerpt of Anthony Burgess’ novel A Clockwork Orange (Burgess 1972). This novel was particularly suited to providing a textual 68 basis, as its prevalent artificial slang helped establish a semantically diffuse environment. This was vital to the task, as the nonce-verbs and the fillers were supposed to be perceived by the participants as unknown but certainly potential English slang verbs. The main text was preceded by a brief introduction, selling the entire experiment either as a literary study or a professional translation project. L1 speakers were told that a new glossary for Burgess’ obscure slang was about to be developed, and that their intuition as native speakers was considered very helpful for this project. Interlanguage speakers were told that their intuition as advanced language learners was instrumental for a new translation. These measures were supposed to distract participants’ attention away from the linguistic details and make the text look as natural and genuine as possible. Following the introductory text, a short section with two items for practice and the text proper with its gaps were presented. A sample of an English questionnaire is given in (26). (26) […] Please listen to the instructions, read along, and fill in the gaps: So we (1 shig) ____ him lovely, grinning all over our litsos, but he still went on singing. Then we (2 strow ) ____ him so he laid down flat and heavy, and a bucketload of beer-vomit came whooshing out. […] Please listen to the tape, read along, and fill in the gaps: We (1 smink) ____ townwards, my brothers, but just outside, not far from what they called the Industrial Canal, we saw that the fuel needle (2 skig) ____, like our own needles had. The auto made noises and (3 wred) ____ kashl kashl kashl. Not to worry overmuch, though, because a rail station (4 skring) ____ blue just near. The question was whether to leave the auto to be sobiratted by the rozzes or, us feeling like in a hate and murder mood, to give it a fair tolchock into the starry waters for a nice heavy loud plesk before the death of the evening. This latter we (5 sprin) ____ […] As one can see, Burgess’ special slang makes ample use of unusual verb and noun forms. Often onomatopoeic, and sometimes derived from Slavic languages, words such as sobiratted, rozzes, tolchock, kashl, and plesk occur throughout the whole book and help render the text semantically diffuse. The corresponding sample from the German excerpt is given in (27). (27) Nach dem Kampf rannten wir zurück zum Auto, und ich ließ Georgie manövrieren. Wir fühlten uns alle ein malenky bisschen schlapp und mürbe, aber wir fuhren weiter, ohne wirklich darüber nachzudenken, in welche Richtung wir uns eigentlich bewegten 7 . Schließlich (1 knießen) ____ wir stadtwärts, meine Brüder, aber knapp außerhalb und nicht weit vom Industriekanal (2 schrimmen) ____ wir, dass die 7 In contrast to the English version, the German text begins with two gap-less sentences. This was supposed to set the past tense (Präteritum) scene for the intermediate learners of German. 69 Benzinanzeigenadel auf Null (3 luschen) ____, wie unsere eigenen Nadeln, und das Auto machte Geräusche und (4 spinken) ____ houi houi houi. Das war kein Grund zu übergroßer Sorge, denn nicht weit weg (5 grießen) ____ das blaue Licht eines Bahnhofs der Vorortbahn. The entire procedure was accompanied by a CD, on which all the instructions as well as the whole text were read out. In terms of experimental modality, stimuli were presented both orally and in writing, and then written responses were recorded. The reason why participants were exposed to written and aural stimuli was to make sure that the nonce-verbs triggered analogies to existing forms by their written form and through their intended sound sequence. Overall, the predominantly written test design, which has proven meaningful and robust in many studies already, was adopted and replicated without major changes in modality simply to be able to make meaningful comparisons to existing data. The German test items for experiment 1 were created as follows. In a first step, based on the lexical statistics provided by Köpcke (1998), rankings of German syllable constituents prototypical for ablaut were extracted. They are given in (28). (28) (a) ranking of prototypical German vowels and their frequencies of ablaut based on Köpcke 1998 [ , ] 8 53 % [ , ] 47 % [ ] 39 % (b) ranking of prototypical German vowel-consonant combinations and their frequencies of ablaut based on Köpcke 1998 [# _ 9 _ ( )] 92 % [# _ 10 _ / ] 70 % [# _ _ / ] 67 % [# _ _ ] 52 % [# _ _ / ! / / ] 47 % [# _ _ " ] 38 % [# _ _ # ] 38 % [# _ _ / ] 32 % [# _ _ / ] 22 % 8 Köpcke observed the following correlation: The feature frontness correlates with the degree of probability for ablaut. Front vowels are very frequent within ablaut formation, whereas back vowels are much less represented in this morphological process. 9 The category [ ] collapses the long and short variants, since they occur in complementary distribution. Strong verbs with [ ]-nuclei always feature nasals in the coda, whereas [ ] always takes fricatives. 10 The nucleus [ ] is followed by obstruents in 98% of all irregular verbs in Köpcke’s set (cf. Köpcke 1998: 55). The only irregular exceptions are: scheinen, gedeihen, leihen, and speien. 70 As we can see in (5a), vowels alone are rather ineffective predictors for a particular type of past tense inflection (around 50%). However, in combination with certain post-vocalic consonants, as listed in (5b), the schema [# _ _ ( )] emerges as prototypical for German strong verbs. In a second step, individual distance-values, expressing a segment’s degree of prototypicality, were allocated, so that possible onset-vowel-coda combinations could be ranked according to their respective probability of triggering ablaut in German. In table 1, possible combinations around the stem vowels [ ] and [ ] are ranked. Percentages given in italics were derived directly from Köpcke’s analyses; the remaining ones were calculated separately. Table 1: Ranking of possible schemas based on the analyses of Köpcke 1998 [# _ _# ] [# _ _# ] constituent % distance value constituent % distance value (a) (C)(C) 27 % 0 (b) (C) 36 % 11 0 C 36 % 12 1 C (incl. ) 33 % 1 s 14 % 2 CC 30 % 2 CC 9 % 3 others 3 others 4 (c) [ ( )] 92 % 0 (d) [ ], [ ] 70 % 0 [ ] 52 % 1 [ ], [ ] 67 % 1 [ " ] 38 % 2 [ ], [ ! ], [ ] 47 % 2 others 3 [ # ] 38 % 3 [ ], [ ] 32 % 4 [ ], [ ] 22 % 5 others 6 The sum of the onset, nucleus, and coda distance-values was supposed to express the degree of prototypicality of a given schema. Table 2 illustrates which kind of degrees of prototypicality German nonces such as schlinken, ninnen, and treichen 13 would yield. 11 This percentage is the result of the following calculation: there are around 33 strong verbs containing [ ] as their stem vowel, twelve of which have the structure [ C _ _ # ]. This translates into roughly 36 %, and it renders C the most common onset within this structure. 12 Here, the input structure CC is considered to be prototypical, since it is the most frequent specific structure, followed by single C, which includes all remaining consonants. 13 Note that the suffix -en was not coded for prototypicality, as it is a mandatory and ubiquitous part of a German verb’s infinitive form. 71 Table 2: Prototypicality of potential German nonce-verbs constituents C # distance-values 0 0 0 1 0 1 2 1 3 0 2 6 ranking of the nonce 1 schlinken verbs’ prototypicality 2 ninnen 3 treichen Following this calculation, a nonce word such as schlinken would be considered to be most prototypical, because the sum of its individual segments’ distance values is zero. Thus, its proximity to a possible prototype has the highest possible degree. The English test items for experiment 1 were created as follows. Recall that one aim of experiment 1 was to re-examine the alleged prototype effects, as reported in Bybee and Moder (1983) as well as Plag (2000), with regard to German-interlanguage speakers. Thus, in a first step all test items from Bybee and Moder and Plag were revisited and ranked according to their prototypicality. The target was to filter out those schemas that allowed best to control the single and combined effects of allegedly prototypical constituents. First, a factor for the importance of all three parts of a schema (onset, nucleus, and coda) for triggering ablaut was derived from Bybee and Moder (1983) and Plag (2000). Each constituent in a schema received a weight-factor documenting its importance, or weight, in triggering ablaut. An onset structure such as CC, for instance, would receive the factor ‘2’, since it is two times less likely to be involved in ablaut than the prototypical onset (sCC). This weight-factor was complemented by distance-values (ranging from 0 to 3, as given in the above table) assigned to the concrete phonemes that would fill the slots in a schema, such as an onset in [ ], for instance. Onset [ ] would receive the distance value ‘2’, since so far it has been evaluated as two grades less prototypical for triggering ablaut than the prototypical onset structure in the literature. The product of the distance-values and the weight-factor was supposed to represent a constituent’s overall proximity to the assumed prototype. An example of how such a procedure ranked potential test items is given in table 3. 72 Table 3: Ranking of potential English nonce-verbs constituents " distance-values 0 0 0 1 4 3 2 3 3 weight-factors x2 x3 x1 x2 x3 x1 x2 x3 x1 product of distances and weight factors 0 0 0 2 12 3 4 9 3 of all three constituents 0 17 16 ranking of the nonce 1 skring verbs’ prototypicality 2 flug 3 smeeb As we can see in the above table, a nonce word such as skring would be regarded more prototypical than flug. In a final step, nucleus was set constant for all English test items ([ ] stem vowel). Thus, those schemas around nucleus [ ] that allowed controlling the effects of onset structures sCC, sC, CC, and C as well as the velar codas [ ], [ ], and [ ] as best as possible were taken as templates for the final set of test items. A list of all German and English test items can be found in the appendix. The overall design of experiment 1 was piloted with 45 German learners of English at the University of Siegen. The purpose of this pilot study was manifold. To begin with, some common issues related to the repeated measure design (‘wug’-design, cf. Berko 1958) were revisited. This design, widely used in psycholinguistic research, has recently been criticised by a number of researchers (cf. Prasada and Pinker 1993, Ramscar 2002a: 51-54). The repeated exposure of speakers to test items, it is claimed, would cause orderand priming-effects, which could render the entire test design ineffective. However, testing generalisation properties of words either means treating existing word forms as if they were unknown, as implemented in network models and simulations, or eliciting reactions from speakers and learners through morphologically controlled artificial words. In the present experiments, potential order effects were taken care of by two measures. As far as the experimental design was concerned, every group of participants was split up into two subgroups, each working with test items displaying the reverse order of the other subgroup. For the analysis, then, the two subgroups were merged again. As far as the analysis is concerned, potential exposure effects were factored out by applying statistical models that can handle both fixed, predefined effects, and so-called random, ‘withinsubjects’ effects (cf. Pinheiro and Bates 2000, Bates 2005). This type of statistical modelling is a rather recently developed and computationally implemented statistical technique especially suitable for analysing repeated measure designs. 73 Apart from these technical issues, the pilot study was supposed to test comprehensibility of the tasks to interlanguage speakers, the order of the stimuli, the timing of the entire experiment 14 , and the suitability of the final set of test items 15 . Table 4 summarises the main results of the pilot study, compared to what Bybee and Moder (1983) as well as Plag (2000) had already discovered. Table 4: Main results of Bybee and Moder 1993, Plag 2000, and the pilot study 2003 Bybee and Moder 1983 English L1 Plag 2000 German-English interlanguage pilot study German- English interlanguage frequency regular past 17.0% 59.2% 57.4% frequency irregular past 83.0% 30.4% 40.1% most frequent ablaut [ ]-ablaut [ ]-ablaut [ ]-ablaut importance of segments coda onset nucleus nucleus coda onset nucleus coda onset input form with the highest ablaut frequency [sCC _ _ ] [sC _ _ ] [sCC _ _ ] suggested prototype [sCC _ _ ] [sCC _ _ ( )] [sCC _ _ ] Table 4 shows that both English L1 speakers and German-English interlanguage speakers do not always inflect novel and unknown verbs by the default, but apply ablaut to a considerable extent. Table 4 shows, too, that the two groups of German-English interlanguage speakers both follow the same ablaut strategy ([ ]-ablaut) and rank the importance of the syllable constituents identically. And the pilot study showed that the probability of participants forming ablaut past tenses depends on the proximity of a given nonce-verb to a prototypical ablaut schema. However, we can also see a lack of congruency between the statistically traceable prototype effects of certain constituent variants and the morphological make-up of those verbs that attracted most of the ablaut formations ([sCC _ _ ] against [sCC _ _ ] in the pilot study). As a consequence, in the final version of experiment 1 the possible effects of precisely those constituent variants that did not overlap were given special attention (sC, sCC, [ ], [ ], [ ], and nucleus [ ]). 14 Both native and interlanguage speakers were given exactly the same time frame. One could argue that interlanguage speakers would generally need more time, but, to my knowledge, there are no studies available on how to calculate additional timing for non-native participant groups in a controlled fashion. 15 A similar test run was conducted with a number of English native speakers in Dublin shortly afterwards, the main purpose of which was to test the comprehensibility of the English instructions. A separate pilot study for German was deemed unnecessary. 74 3.2 Results In this section the results of the three versions of experiment 1 are presented and discussed. First, all participants’ responses were computerised and coded according to their syllable constituents. While all responses to the English nonce-verbs were broken down into the 4 categories ‘past tense formation without vowel change’, ‘[ ]-ablaut’, ‘[ ]-ablaut’ and a residual category ‘others’, 6 categories were necessary for German, namely ‘past tense formation without vowel change’ 16 , ‘[ ]-ablaut’, ‘[ ]-ablaut’, ‘[ ]ablaut’, ‘mixed conjugation’ (ablaut in combination with suffixation), and again the residual category ‘others’. 3.2.1 Experiment 1_G_L1 (Native speakers of German) Experiment 1_G_L1 contains the results of 50 native speakers of German, all undergraduate students from the University of Siegen, who went through experiment 1. The first main result concerns the participants’ willingness to provide vowel-changing past tense formations. This is illustrated in table 5. Table 5: Frequencies for types of past tense formation with or without vowel change in 1_G_L1 types of past tense frequency percent without vowel change 919 65.6 % with vowel change 481 34.4 % 1400 100.0 % As we can see, 65.6% of all past tense formations followed the default inflection and did not show vowel change. At the same time, however, a substantial 34.4% of all responses featured vowel changes. Note however that, given the close proximity of all nonces to the German prototypical schemas, this proportion is rather moderate. Consider figure 1 now, in which three mosaic diagrams plot the two types of past tense formation against the distribution of onset, nucleus, and coda. 16 Despite the obvious differences between an English past tense and the German preterite (both morphological and semantic), English as well as German inflections were labelled ‘past tense’ for readability and for comparison’s sake. 75 onset by past tense onset past tense C CC sch schC schCC no vowel change vowel change nucleus by past tense nucleus past tense long i short i no vowel change vowel change coda by past tense coda past tense fricative k g m n ng ngk no vowel change vowel change 2= 11.9775 (Pearson), df = 4, p < .05 2 = .0318 (Yates’ correction), df = 1, p = .8584 2= 18.8067 (Pearson), df = 4, p < .001 Figure 1: Mosaic plots for German verbs cross-classified by past tense formation and the three constituents onset, nucleus, and coda in 1_G_L1 Mosaic plots such as the above ones can be interpreted as follows: each column represents the past tense frequency of an onset, nucleus, or coda variant. The two different shades of grey represent the distribution of responses with or without vowel change. As we can see, the proportion of responses with and without vowel change varies across some of the constituents (darker against the lighter shades of grey). Those horizontal asymmetries, that is, the different heights of the shaded areas relative to each other, constitute in fact the most important insights a mosaic plots has to offer, since they indicate potential effects of the constituent variants on the types of past tense formation. The longer a bar is relative to its neighbouring bars, the greater is its effect on the type of past tense formation. In figure 1 above, left-hand panel, the mosaic plot shows that all onset variants display a greater proportion of past tenses without vowel change than forms including vowel change (darker shaded bars representing past tense formation without vowel change are bigger in size, as already illustrated in table 5). More interesting, though, are the horizontal asymmetries. Consider the onset structures C and [ ] / <sch>, for instance, in the lefthand panel. There, we can see that the height of the bars representing the ‘no vowel change’ responses for these two constituent variants is greater than for C and CC. Based on these asymmetries, we can suspect partial effects of C and [ ] on regular past tense formation, and partial effects of C and CC on vowel-changing past tense formation. Now consider the middle panel for nucleus. As we can see, there are hardly any asymmetries across the distribution of the two types of past tense formation. Looking at the right-hand panel for coda, we can see - as with onset - pronounced asymmetries in the distribution across the types of past tense formation. So here too we can suspect an influence of the constituent in general. More particularly, [ " , ] appear to favour regular past tense forma- 76 tion, while engma favours vowel-changing past tense formations. Note here that both coda [ ] and [ ] show similar effects on vowel-changing past tense formations. Despite the visual insights such mosaic plots offer, they tell us little about how significant the observable effects of the constituents in fact are and whether they interact. These details were examined using linear regression and mixed effects models. Various models were fitted to each data set, with the fitting and reduction following the following procedure. First, a generalised linear regression model with all three constituents as predictors and the type of past tense as response variable was fitted to the German L1 data in order to examine potential significant partial effects of each individual factor. An analysis of deviance of such a model revealed significant partial effects for onset (p < .001) and for coda (p < .0005), but only an insignificant effect for nucleus (p = .705) 17 . The insignificance of nucleus was confirmed through a ‘drop-single-term-deletion-procedure’ (cf. Chambers 1992). Thus, at this point, we can conclude that both onset and coda are important predictors of the type of past tense formation (regular versus vowel change), while nucleus is no significant factor. Second, possible interactions of the constituents as well as potential effects caused by the repeated measure design of the experiments were explored using a so-called mixed effects models (cf. Pinheiro and Bates 2000). Such models contain both fixed and random effects, the latter of which control for specific factors that are supposed to cause random variation in a predictor’s coefficient. In our case, such a random variable is the withinsubject variation, caused by the repeated measure design of the experiments. The size of such a random effect is measured through its degree of variance, and its estimates are reported as the standard deviation of the residuals. Since the variance of the random effect ‘subject’ turned out to be substantial across all subject groups in all versions of experiment 1, fixed effects were always extracted using mixed effects models. The mixed effects model used in the present and the following experiments was a model with multivariate normal random effects via penalised quasi-likelihood estimation (henceforth PQL, cf. Breslow and Clayton 1993). All analyses included subject as random effect, and onset, nucleus, and coda as fixed effects. The mixed effects model in this and the following chapters was fitted to the data omitting further stepwise model reduction and simplification, since the mixed effects model were always preceded by data analyses 17 In addition to that, an analysis of variance of a simple logistic regression model also confirms the significance of onset ( 2 = 14.35 (Pearson), df = 4, p < .001) and coda ( 2 = 20.29 (Pearson), df = 4, p < .0001). The insignificance of nucleus in this model was established by applying a ‘fast-backward-variable-selection-procedure’ (cf. Lawless and Singhal 1978), which retained only onset and coda in the final model. 77 through generalised linear regression and logistic regression models in order to evaluate the significance of each individual factor 18 . In the mixed effects model that was fitted to the German L1 data codas had to be recoded and collapsed due to the non-orthogonal design of the data. Thus, coda entered with the three levels ‘fricative’, ‘single consonant’, and ‘velar nasal plus optional stop’ only. Interestingly, this model did not yield any significant interactions. Only the interaction of [ ] and single consonant codas approached a significant level; its coefficient showed a positive slope and thus an effect towards avoiding past tenses with vowel change. Due to the lack of significant interactions, the final trimmed model examined possible main effects only. Those are reported in table 6. Table 6: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda as predictors and types of past tense as response variable in 1_G_L1 constituent -coef s-value df t-value p-value sign. intercept -1.378 0.296 1342 -4.647 < .0001 *** onsets CC 0.517 0.223 1342 2.317 < .05 * 0.436 0.251 1342 1.736 = .08 ---- C 0.857 0.195 1342 4.403 < .0001 *** CC 0.978 0.256 1342 3.834 < .0005 *** codas k, g -0.330 0.229 1342 -1.440 = .1500 ---- m, n -0.809 0.189 1342 -4.277 < .0001 *** 0.328 0.190 1342 1.732 = .0835 ---- k -0.131 0.201 1342 -0.650 = .5161 ---- * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant As we can see in table 6, the trimmed model produced coefficients for 9 variants of onset and coda. This includes the intercept, which represents the reference level (or baseline) for the calculation of the other variants’ coefficient values. For the onset -values, for instance, it represents the effect of onset structure C, against which all remaining onset -values were calculated. The same holds for coda [ ]. The -coefficient values in this model can be interpreted like this: negative values (like with [m,n]) are an indication of the constituent’s effect towards avoiding vowel change, while positive values illustrate a tendency towards vowel change (all onsets). As we can see, significant effects emerge for onsets CC, C, and CC as well as for coda [m,n]. Overall, the probabil- 18 There is also a technical reason preventing a rigorous stepwise model simplification with PQL algorithms. R 2 -values to document possible redundancies, for instance, or other measures to document effect sizes are simply not calculated by this algorithm. And an analysis of variance of an entire model, often applied to help eliminate insignificant factors stepwise, was - due to technical limits - not always possible. 78 ity for vowel-changing responses seems to increase significantly with the complexity of the onset structures. While the simple [ ] shows no effect at all, CC is just about significant, and C as well as CC turn out to be highly significant. In contrast, the only significant coda effect [m,n] comes with a negative slope and thus favours past tenses without vowel changes. In a next step, possible significant interactions between the constituents were examined through two classification models. One creates a so-called recursive partitioning and regression tree, as implemented in the statistical software package R ( R Development-Core-Team 2007), the other one uses a classification and regression tree (CART), as implemented in the software package SPSS AnswerTree. Both classification algorithms apply what is known as tree-based segmentation in order to determine a set of conditions that allow accurate predictions, or classifications, of the cases in a data set (cf. Breiman et al. 1984). For the German L1 data, the classification analyses, however, showed that that the models’ predictive power did not benefit from the inclusion of interactions. Note that such a total lack of significant interactions is somewhat surprising on an intuitive level, since one could expect that if a model allows formerly significant partial main effects to interact, the interacting factors should turn out to be as significant as when they entered the model as mere single fixed effects. This, however, cannot be found in the data, as the regression and classification models have demonstrated. In sum, there are single main effects of certain syllable constituents for German L1 speakers, but they do not readily interact and combine to morphological prototypes. This problem will be discussed in more detail in section 3.3. Let us have a closer look now at what has so far been reported as past tenses with vowel change. As we can see in table 7, German speakers produced 4 different ablaut patterns, along with the residual category ‘others’ (patterns which do not follow any existing vowel change). Table 7: Frequencies for the types of vowel change in 1_G_L1 types of vowel change example frequency percent [ ]-ablaut spinken spank 170 12.1 % [ ]-ablaut spinken sponk 139 9.9 % [ ]-ablaut spinken spunk 36 2.6 % mixed (ablaut + suffix) spinken spankte 25 1.8 % others spinken spenk 55 3.9 % The most frequent types of vowel change occur with [ ]and [ ]-ablaut. Note that unorthodox responses (row ‘others’ in the table) are quite rare (3.9%). This means that only a few participants, once they had opted for vowel changes, deviated from existing patterns and invented patterns alien 79 to German. The test items that attracted most of the vowel change responses are ranked in table 8. Table 8: Ranking of test items attracting most of the vowel-changing responses in 1_G_L1 test item [ ]-ablaut [ ]-ablaut [ ]-ablaut 1. spingen 25 1 2 2. strinken 13 0 2 3. stingen 12 3 1 4. sprinken 12 2 2 1. strießen 0 15 0 2. triechen 4 14 0 3. grießen 4 10 0 As we can see, test item spingen received by far most of the [ ]-ablaut formations. Most [ ]-ablauts occur with strießen, closely followed by triechen. Interestingly, these test items show a tendency for mutually exclusive ablaut formation; high values in one pattern usually tend to result in a low (or zero) amount of inflections in the other patterns. Based on table 8, two prototypical vowel-changing morphological schemas for [ ]and [ ]-ablaut could be sugsgested (29). The arrow ‘ ’ reads: “can be interpreted as prototypical for”. (29) (a) [ C _ _ ] spingen [ ]-ablaut (b) [ CC _ _ ] strießen [ ]-ablaut Consider figure 2 now. It contains three mosaic plots that illustrate how the three syllable constituents interact with the different types of past tense formation. onset by vowel change onset vowel change C CC sch schC schCC a mix o others regular u nucleus by vowel change nucleus vowel change long i short i a mix o others regular u coda by vowel change coda vowel change fricative k g m n ng ngk a mix o others regular u Figure 2: Mosaic plots for German verbs cross-classified by vowel change and the three constituents onset, nucleus, and coda in 1_G_L1 As we can see, there are asymmetries in the distribution of all three constituents across all occurring types of vowel change, and they are statisti- 80 cally highly significant (all 2 p-values < .0001). We can see, for instance, visible effects of nucleus [ ] on [ ]-ablaut and nucleus [ ] on [ ]-ablaut (middle panel). As far as coda is concerned (right-hand panel), we can see that codas containing the velar nasal show a strong tendency towards [ ]ablaut. Note here that the effects of both [ ] and [ ] are rather similar. In order to explore how the constituents influence the choice between all five different vowel changes, a partitioning and a classification and regression model were fitted to the German data. These analyses, though, show neither significant single effects nor significant interactions. They do, however, illustrate a rather mild tendency for verbs of the shape [ (C)(C) _ vowel _ ( )] towards [ ]-ablaut. At this point, let us summarise the most important results for experiment 1_G_L1 in (30). (30) (a) 30.4% of all responses show vowel changes (b) more than 95% of all vowel changes follow existing patterns (c) onset and coda are significant predictors for the types of past tense (regular - ablaut) (d) onset, nucleus, and coda are significant predictors for the types of vowel change (e) the presence of coda [ " , ] is significant for no-vowel-changing past tenses (f) the probability of vowel change increases with the complexity of onset (g) [ ]and [ ]-ablaut are the preferred vowel change patterns (h) schema [ C _ _ ] attracted most of the [ ]-ablaut formations (i) schema [ (C)(C) _ v _ ( )] shows a statistical effect towards [ ]-ablaut (j) schema [ CC _ _ ] attracted most of the [ ]-ablaut formations Overall, German native speakers showed similarity-based effects of irregular verb morphology similar to the ones found in previous studies with English native speakers. German native speakers applied two different types of vowel change frequently to the given test items, namely [ ]and [ ]-ablaut (cf. table 7), and they did so in analogy to the morphological make-up of existing verbs. These analogy-based attractions seem to rests primarily on onset and coda as far as the choice between regular and vowel-changing past tense formations is concerned. When choosing between the different vowel change patterns, all three constituents serve as analogical attractors. It is not easy, though, to distil one single prototypical schema for vowelchanging past tense formations from the German L1 data. Let us start with the role of onset. Here, the degree of complexity seems to be a good predictor for the choice between regular and vowel-changing past tense formations. The mixed effects model suggested that the higher the complexity, the more likely it is that such onsets produce vowel changes. In line with 81 these statistical effects, the nonce-verbs that attracted most of the vowelchanging responses featured onsets C and CC. So here the individual statistical main effects and the shape of the prototypically irregular nonceverbs overlap rather well. Nucleus generally seems to play a significant role only when it comes to predicting the type of vowel change. In other words, nuclei [ ] and [ ] attract different types of vowel change. Concerning coda, those test items attracting most of the vowel-changing responses (cf. table 8) and the statistical effects from the regression models (cf. table 6) are slightly at odds. First, the only significant partial effect for coda emerged in the mixed effects model with the variant [ " , ], showing a tendency towards regular past tense formation. Second, neither single main effects nor any interactions favouring a particular type of vowel change could be distilled from the data. In general, all single partial effects turned out to be independent and not co-determined by each other. The only exception is a rather mild tendency of the schema [ (C)(C) _ vowel _ ( )] towards [ ]-ablaut. In sum, like in Bybee and Moder’s (1983) as well as Plag’s (2000) study, the morphological make-up of those nonce-verbs attracting most of the vowel-changing responses is not mirrored in the statistical effects of the models. What does this mean for the notion of prototypical schemas underlying irregular verb morphology in German L1? At this point, we can only state that certain onset, nucleus, and coda structures serve as statistically traceable morphological attractors towards vowel change patterns. It is questionable, however, if these single effects combine to prototypes in the sense that one single form, containing the most prototypical elements, is at the heart of the morphological organisation of irregular German verbs. Finally, let us turn to the question of whether Köpcke’s statistical findings are confirmed by the German L1 behavioural data. First, we must conclude that Köpcke’s prototypical coda [ ( )] for [ ]-ablaut was only partly born out in experiment 1. Although partial effects of [ ], and to some degree [ ], clearly emerged in the mosaic diagrams in figure 2, neither of the two coda variants came out as a significant predictor in the classification model. In sum, the present evidence suggests that the organisation of irregular verb morphology in the mental lexicon of German L1 speakers is sensitive to morphophonological similarities to existing verbs (mainly resting on onset and coda), and that certain single constituent variants - but no variant combinations - serve as analogical attractors for vowel-changing past tense formations. Instead of a prototype-based model, a model based on simple similarity-based analogies, possibly formalised as schemas, seems to be appropriate for the data from German L1. 82 3.2.2 Experiment 1_E_L1 (Native speakers of English) This section summarises the results of 16 native speakers of English going through experiment 1_E_L1. Table 9 illustrates the participants’ willingness to provide vowel-changing past tense formations. Table 9: Frequencies for types of past tense formation with or without vowel change in 1_E_L1 types of past tense frequency percent without vowel change 154 38.5 % with vowel change 246 61.5 % 400 100.0 % As we can see, a remarkable 61.5% of all past tense formations show vowel change, 80% of which follow existing patterns. Consider figure 3 now. It plots the two types of past tense formation against the distribution of the segments onset and coda. Recall that nucleus was constant in the experiments with English nonce-verbs 19 . onset by past tense onset past tense CC sC sCC regular vowel change coda by past tense coda past tense f g K m n ng ngk t regular vowel change 2 = 1.8543 (Pearson), df = 2, p = .3957 2 = 28.3501 (Pearson), df = 5, p < .0001 Figure 3: Mosaic plots for English verbs cross-classified by past tense formation and the two constituents onset and coda in 1_E_L1 Figure 3, left panel, shows only small asymmetries in the distribution of the types of past tense formation across onset. However, there are pronounced, asymmetries in the distribution of the types of past tense formation across 19 Nucleus [ ], whose prototypical effects had been reported repeatedly in the pertinent literature, was set constant for all English test items both for L1 and L2 in order to control the effects of the preceding onset structures sCC, sC, CC, and C as well as the following velar codas [ ], [ ], and [ ] as best as possible. 83 coda . Coda [ ], for instance, clearly encourages past tense formations with vowel change. Codas [ , ] and the fricative, in contrast, shows a clear tendency towards past tenses without vowel change. Again, first a generalised linear regression model, trimmed with a dropsingle-term-deletion-procedure, with onset and coda as predictors and the type of past tense as response variable revealed significant effects both for onset (p < .05) and coda (p < .0001) 20 . And second, as with the German native speakers, the interaction of onset and coda as well as potential random effects were explored using the PQL model. And here, too, due to the non-orthogonal design of the data, both onset and coda had to be pooled, leaving binary variants for onset (C[C] vs. sCC) and for coda (velar vs. non-velar). And as with the German L1 data, the model revealed no significant interactions so that the final model was fitted for single main effects only 21 . They are reported in table 10. Table 10: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda as predictors and types of past tense as response variable in 1_E_L1 constituent -coef s-value df t-value p-value sign. intercept -2.513 0.836 327 -3.007 < .005 ** onsets sC 1.227 0.391 327 3.143 < .005 ** sCC 0.855 0.355 327 2.405 < .05 * codas k, g 0.511 0.783 327 0.652 = .5148 ---- m, n 2.213 0.771 327 2.870 < .005 ** 2.854 0.767 327 3.721 < .0005 *** k 2.379 0.749 327 3.175 < .005 ** t 2.216 0.813 327 2.726 < .01 ** * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant As we can see in table 10, the model found main effects both for onset and coda. In total, we obtain 8 coefficient values, 7 of which are significant. Again, negative -coefficients are an indication of the constituent’s effect towards avoiding vowel change, as we can observe for onset CC and coda [ ], mapped onto the intercept. In contrast, those coefficients with high positive values, like onset sC as well as codas [ " , ], [ (k)] and [t] indicate a pronounced tendency towards vowel change. 20 An analysis of variance of a similar simple logistic regression model confirmed these effects, too (onset: 2 = 7.99 (Pearson), df = 2, p < .05, coda: 2 = 31.38 (Pearson), df = 5, p < .0001). Both factors were retained in the final logistic model after a fast-backwardvariable-selection-procedure. 21 A general lack of significant interactions was confirmed in classification analyses, too. 84 In sum, the choice between past tense formation with or without vowel changes seems to depend on the presence both of onset and coda, with coda being the strongest predictor, and with onset sC as well as codas [ " , , , , ] having an effect on vowel change. Let us split up all past tense responses into their different types of vowel change patterns, as illustrated in table 11. Table 11: Frequencies for the types of vowel change in 1_E_L1 types of vowel change example frequency percent [ ]-ablaut skring skrang 132 33.0 % [ ]-ablaut skring skrung 64 16.0 % others 50 12.5 % As we can see in table 11, both forms of ablaut following existing patterns taken together amount to almost 50% of all past tenses with vowel change. Within vowel change, [ ]-ablaut is the predominant pattern. Interestingly, a substantial 12.5% of all responses (coded as ‘others’) deviate from existing patterns. Thus, in general, English L1 speakers are rather willing to provide past tense inflections with vowel change, but, at the same time, are also rather inventive when it comes to modelling patterns. Table 12 presents a ranking of those test items attracting most of the vowel-changing responses. Table 12: Ranking of test items attracting most of the vowel-changing responses in 1_E_L1 test item [ ]-ablaut [ ]-ablaut 1. spim 9 3 2. strink 9 2 3. sprit 9 1 4. spling 6 6 5. skring 6 3 6. sprink 6 2 As we can read from this table, there are three test items that attracted an equal amount of [ ]-ablaut responses (spim, strink, and sprit). Note that they vary considerably as regards their phonological make-up. We can also see that spling received most of the [ ]-ablaut inflections, but, at the same time, shows the same frequency for [ ]-ablaut. Thus, contrary to the German L1 speakers, it seems as if the nonces for English L1 speakers did not produce mutually exclusive proportions of ablaut patterns. The morphological schemas of the three nonces attracting most of the vowel changes are given in (31). 85 (31) (a) [sC _ _ " ] spim [ ]-ablaut (b) [sCC _ _ ] strink [ ]-ablaut (c) [sCC _ _ ] sprit [ ]-ablaut (d) [sCC _ _ ] spling/ skring [ ]-ablaut Consider figure 4 now. It contains two mosaic diagrams that illustrate how onset and coda interact with the types of vowel change, excluding all the cases with regular past tenses. onset by vowel change onset vowel change CC sC sCC a ae others coda by vowel change coda vowel change f g,k m, n ng ngk t a ae others 2 = 3.5188 (Pearson), df = 4, p = .475 2 = 17.4039 (Pearson), df = 10, p = .06589 Figure 4: Mosaic plots English verbs cross-classified by vowel change and the two constituents onset and coda in 1_E_L1 There are visible but insignificant asymmetries in the distribution of both constituents across the occurring vowel changes. But we can still see how coda [ ], for instance, affects [ ]-ablaut and seems to avoid no-vowelchange formations, while we can also see a clear effect of coda [ ] towards regular past tense formations. Note that, contrary to German L1 speakers, for English L1 speakers coda [ ] shows a much stronger effect on [ ]-ablaut than [ ] does, which, in turn, seems to favour [ ]-ablaut. As far as the choice between vowel change patterns is concerned, a classification analysis revealed the presence of significant predictor interactions, as illustrated in figure 5 now, which shows the resulting classification tree. 86 Category % n regular preterite 38,50 154 ae-ablaut 33,00 132 others 12,50 50 a-ablaut 16,00 64 Total (100,00) 400 Node 0 Category % n regular preterite 62,50 50 ae-ablaut 15,00 12 others 11,25 9 a-ablaut 11,25 9 Total (20,00) 80 Node 2 Category % n regular pre terit e 32,50 104 ae-ablaut 37,50 120 others 12,81 41 a-ablaut 17,19 55 Total (80,00) 320 Node 1 Category % n regular preterite 36,25 58 ae-ablaut 41,88 67 others 12,50 20 a-ablaut 9,38 15 Total (40,00) 160 Node 4 Category % n regular preterite 50,00 24 ae-ablaut 35,42 17 others 10,42 5 a-ablaut 4,17 2 Total (12,00) 48 Node 10 Category % n regular preterite 30,36 34 ae-ablaut 44,64 50 others 13,39 15 a-ablaut 11,61 13 Total (28,00) 112 Node 9 Category % n regular preterite 28,75 46 ae-ablaut 33,13 53 others 13,13 21 a-ablaut 25,00 40 Total (40,00) 160 Node 3 Category % n regular preterite 35,42 34 ae-ablaut 30,21 29 others 11,46 11 a-ablaut 22,92 22 Total (24,00) 96 Node 8 Category % n regular preterite 18,75 12 ae-ablaut 37,50 24 others 15,63 10 a-ablaut 28,13 18 Total (16,00) 64 Node 7 past tense coda Improvement =0,0231 (g,k); (f ) (nk); (ng); (m, n); (t) coda Improvement =0,0075 (nk); (t) onset Improvement =0,0045 CC sC; sCC (ng); (m, n) onset Improvement =0,0036 sCC; CC sC Figure 5: Classification analysis showing the interaction of onset and coda with the types of vowel change in 1_E_L1 The above tree can be interpreted as follows. The splits 22 of the upsidedown tree and their branches illustrate a decision-procedure in order to determine the choice between past tenses with or without vowel changes as clear-cut as possible. The leaf-nodes specify non-overlapping subsets of the entire data set. For all subsets (nodes), the classification algorithm inspects all predefined predictors (here onset and coda) and selects the one that is most useful for distinguishing between the subsets. Such a distinc- 22 The splits in figure 5 have been created through a so-called ‘information gain’ algorithm (cf. L. Breiman et al. 1984), which measures the level of increasing purity in a mother node and the corresponding gain in purity in a daughter node. 87 tion always makes the daughter nodes ‘purer’ than their mother node, in a sense that the distribution across the vowel changes in the daughter nodes is more pronounced than in a mother node. Potential overfitting of the data in such a model can be avoided by pre-defining sensible stopping rules and by applying what is known as ‘cost-complexity-pruning’, an algorithm that will stop the procedure from creating subsets without predictive value. Hence, all the subgroups (nodes) in the above tree illustrate statistically significant splits, and hence significant effects of the two constituents on the choice of vowel change. Let us have a closer look at the first split in the above tree. This split has been created according to the effect of coda. This means that coda is the most important predictor for the distinction between the types of vowel change. Further down in the tree, there are also splits based on the effects of onset (splits on the bottom level), which means that onset is a significant predictor too. Consider nodes 7, 9, and 10 now. Node 7 comprises a subset defined by the presence of onset sC and codas [ ] and [ " , ]. For this subset, [ ]-ablaut shows the highest ablaut in the entire tree. Since both coda [ ] and [ " , ] affect past tenses with vowel change [ ], we must assume that schemas either of the shape [sC _ _ ] or [sC _ _ " / ] are significant attractors for [ ]-ablaut. In node 9, which is defined by the presence of [sC(C) _ _ / ], the proportion of [ ]-ablaut increases to 44.6%. Since again both codas show an effect towards [ ]-ablaut, we can assume that nonce-verbs either with the shape [sC(C) _ _ ] or [sC(C) _ _ ] attract [ ]-ablaut significantly. In contrast, node 10 shows that for verbs in [CC _ _ / ], regular past tense formation holds the majority (50%). If we compare this classification analysis with the nonce-verbs attracting most of the vowel-changing responses, we can see interesting parallels (table 13). Table 13: Parallels between the morphological make-up of those test items attracting most of the vowel-changing responses and the statistically significant interactions of the constituents as revealed by a classification analysis in 1_E_L1 nonces attracting most ablaut formations interaction effects from a classification analysis [ ]-ablaut [sC _ _ " ] --------------------- [sCC _ _ ] [sC(C) _ _ ] [sCC _ _ ] [sC(C) _ _ ] [ ]-ablaut [sCC _ _ ] [sC _ _ ] --------------------- [sC _ _ " / ] As we can see, there is a great deal of overlap between the statistically traceable interactions and the behaviour of the ablaut-triggering test items. At this point, let us summarise all results for experiment 1_E_L1 in (32). 88 (32) (a) 61.5% of all responses show vowel changes (b) 12.5% of all responses show unorthodox vowel change patterns (c) around 80% of all vowel changes follow existing patterns (d) onset and coda are significant for the types of past tense (e) onset and coda are significant predictors for the types of vowel change (f) onset sC as well as codas [ " , , , , ] show partial effects on the types of past tense formation (g) onsets sC, sCC, and codas [ " , ], [ ], and [ ] are significant predictors for the types of vowel change (h) [ ]-ablaut is the preferred vowel change pattern (i) the nonce-verbs that attracted most of the [ ]-ablaut formations have the morphological schemas [sC _ _ " ], [sCC _ _ ] and [sCC _ _ ] What do these results tell us about possible prototypes for English L1 irregular verb morphology? One the one hand, as Bybee and Moder already suggested, native speakers of English are capable of inflecting nonces through vowel change in analogy to existing irregular verbs. This is borne out by the fact that the English L1 speakers in this experiment applied vowel change to the great majority of all responses, almost 80% of which followed existing patterns. On the other hand, however, there is a remarkable amount of unorthodox inventions, too. This means that a number of speakers, when trying to apply ablaut to the nonces, did not follow existing patterns 23 . Overall, the choice between past tenses with or without vowel changes depends on the presence of onset sC as well as all coda variants except for the fricative and the velar stop. However, as with German L1 speakers, no statistically traceable interactions emerged from the data for the distinction between past tense with and without vowel change. In contrast, the choice of ablaut seems to rest on certain interactions, as given in (9i) above. Note that the statistically significant interactions overlap to a remarkable degree with the test items that triggered most ablaut patterns. Therefore, contrary to German L1, we could propose three schemas with prototypical qualities for the type of English L1 vowel change. Schemas [sC(C) _ _ ] and [sC(C) _ _ ] are prototypical for [ ]-ablaut, while [sC(C) _ _ ] would be prototypical for [ ]-ablaut. In sum, though, like with the German L1 data, the evidence for strictly prototypical organisation of irregular verb morphology in English L1 is scarce. Again, a model based on simple similarity-based analogies, possibly formalised as schemas, seems to be more appropriate for the data from English L1. 23 In fact, except for one participant, all participants formed the odd unorthodox vowel change, ranging from merely 1 to a remarkable 11 instances per participant. 89 Let us finally compare the main findings from this experiment with the results reported by Bybee and Moder in (33). (33) English L1 (Bybee and Moder) English L1 Experiment 1_E_L1 (a) preferred ablaut [ ]-ablaut [ ]-ablaut (b) strongest effect coda coda (c) effect for [ ] [ ]-ablaut [ ]-ablaut (d) potential prototypes [sCC _ _ ] [sC(C) _ _ ( )] [sC(C) _ _ ] As we can see, both studies show interesting parallels. They both feature, for instance, coda as the most important predictor for the type of vowel change, with [ ] favouring ablaut. There are, however, interesting differences, too. The English speakers in the present experiment preferred [ ]ablaut, whereas Bybee and Moder’s L1 speakers produced a majority of [ ]ablaut. This however, might be another side-effect of Bybee and Moder’s variety of nuclei in the test items (cf. section 2.10), and therefore will not be commented on any more at this stage. 3.2.3 Experiment 1_E_L2 (German-English interlanguage speakers) In experiment 1_E_L2, 48 German learners of English from the Universities of Siegen and Marburg went through experiment 1. Consider table 14 first. It shows the frequencies of the types of past tense formation that the interlanguage speakers produced in this experiment. Table 14: Frequencies for types of past tense formation with or without vowel change in 1_E_L2 types of past tense frequency percent without vowel change 612 51 % with vowel change 588 49 % 1200 100 % As we can see, the proportion of past tenses with and without vowel change is almost even (51% compared to 49%). Among the vowel change responses, more than 90% follow existing patterns. Consider figure 6 now. It plots the two constituents onset and coda against the types of past tense formation. 90 onset by past tense onset past tense CC sC sCC regular vowel change coda by past tense coda past tense t f g,k m,n ng ngk regular vowel change 2 = .9388 (Pearson), df = 2, p = .6254 2 = 100.2827 (Pearson), df = 5, p < .0001 Figure 6: Mosaic plots for interlanguage verbs cross-classified by past tense formation and the two constituents onset and coda in 1_E_L2 Figure 6 shows clear horizontal asymmetries only for coda, right panel. For coda [ ], for instance, learners clearly favoured vowel changes over regular past tense formations, while with codas [ ] and [ , ] there is a pronounced tendency towards regular past tense. The effect of onset, in contrast, is remarkably constant across all three variants. However, both an analysis of deviance of a generalised linear regression model and an analysis of variance of a similar simple logistic regression model show significant partial effects both for onset ( 2 = 7.50 (Pearson), df = 2, p < .05) and coda ( 2 = 95.32 (Pearson), df = 5, p < .0005). Potential interactions of onset and coda (pooled into binary variants, double against triple consonant onsets, and codas with engma against all other codas) and random effects were again examined in a PQL model, which yielded one highly significant interaction and two single main effects. The significant interaction of onset and coda occurs with schema [sCC _ _ ] ( = 1.064, t(1096) = 3.510, p < .0005), with a positive -coefficient illustrating an effect for vowel-changing past tense formations. The remaining single main effects of the final model are reported in table 15. 91 Table 15: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda as predictors (pooled) and types of past tense formation as response variable in 1_E_L2 constituent -coef s-value df t-value p-value sign. intercept -0.740 0.201 1096 -3.686 < .0005 *** onsets sCC -0.120 0.198 1096 6.691 = .545 ---codas 1.481 0.221 1096 6.691 < .0001 *** k 0.402 0.204 1096 1.971 < .05 * * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant Since the onsets had to be collapsed in this final model in order to check for possible interactions, the difference between CC and sC, as shown in the mosaic plot in figure 6, could not be accounted for. We can see, however, how the two structures [ ] and [ k] differ in their effects from the other codas. They are both significant and show a positive coefficient, which means they favour past tenses with vowel change. Recall now the highly significant interaction that appeared with schema [sCC _ _ ] in the above mixed effects model, and compare this schema now with the single main effects of table 15. Interestingly, while sCC as a main single effect dislikes vowel change, it favours vowel change in the interaction of onset and coda ([sCC _ _ ]. Thus it appears as though single main effects and interaction effects do not entirely match. This quandary was corroborated through a classification analysis across all onset and coda variants, yielding not a single significant interaction. Hence, in conclusion, at this point the question as to if and how onset and coda interact as predictors for vowel changing past tenses has to be left unanswered. Overall, we can state that for German-English interlanguage speakers, codas with engma show a clear partial effect on vowel-changing past tense formations. In addition to that, the schema [sCC _ _ ] is prototypical for vowel change, but only in contrast with double-consonant onsets and codas without engma. Consider table 16 now. It shows the frequencies for the types of vowel change that the German-English interlanguage speakers produced. Table 16: Frequencies for the types of vowel change in 1_E_L2 types of vowel change frequency percent [ ]-ablaut 414 34.5 % [ ]-ablaut 123 10.3 % others 51 4.3 % 92 In table 16 we can see that [ ]-ablaut constitutes the predominant vowel change pattern with 34.5% for German-English interlanguage speakers. Note that, compared to the results for English native speakers, the category ‘others’ (non-existing vowel change patterns) is rather small, making up only 4.3% of all responses. Consider table 17 now, which ranks the frequencies of those test items that received most of the ablaut responses. Table 17: Frequencies for vowel-changing test items in 1_E_L2 test item [ ]ablaut [ ]-ablaut 1. tring 27 3 2. strink 26 8 3. skring 25 3 1. pling 19 14 As we can read from this table, four test items show a rather high frequency of [ ]-ablaut responses. We can also see that pling received most of the [ ]-ablaut inflections, but, at the same time, scores relatively high on [ ]-ablaut, too. The morphological schemas of those verbs attracting most of the vowel changes are given in (34). (34) (a) [CC _ _ ] tring [ ]-ablaut (b) [sCC _ _ ( )] strink / skring [ ]-ablaut (c) [sC _ _ " ] spim [ ]-ablaut (d) [CC _ _ ] pling [ ]-ablaut Interestingly, the schemas in (11a) and (11d), although being of identical shape, can be prototypical for two different types of vowel change. Consider figure 7 now, which plots the types of vowel change against onset and coda. onset by vow el c h a ng e on set vowel change C C sC sC C a ae others regular coda by v owel cha nge c o da vowel change f g ,k m ,n ng ng k t a ae others regular 2 = 5.1462 (Pearson), df = 6, p = .5252 2 = 131.0548 (Pearson), df = 15, p < .0001 Figure 7: Mosaic plots for interlanguage verbs cross-classified by vowel change and the two constituents onset and coda in 1_E_L2 93 As with English L1 speakers, effects of onset on any type of vowel change are rather small for the interlanguage speakers. In contrast, clear asymmetries can be seen for coda. Contrary to English L1, though, a classification analysis revealed a general lack of significant interactions. It did, however, highlight the single main effects of the codas [ ] and [ ], as illustrated in the following tree diagram below (figure 8). It shows the classification tree with onset and coda as potential predictors and the types of vowel change as the response variable. Ca te g or y % n ae a b la u t 34 ,5 0 4 1 4 re g u la r 51 ,0 0 6 1 2 aa b la u t 10 ,2 5 1 2 3 oth e r s 4 ,2 5 5 1 To ta l (1 0 0 ,0 0 ) 1 2 0 0 No d e 0 Cate g o ry % n a e -a b la u t 1 5 ,8 3 38 r e gu lar 7 5 ,0 0 18 0 a ab lau t 5 ,8 3 14 o th e rs 3 ,3 3 8 To ta l ( 20 ,0 0 ) 24 0 No d e 2 Ca te g o ry % n a ea b la u t 3 9 ,17 3 7 6 re g u la r 4 5 ,00 4 3 2 a -a b la u t 1 1 ,35 1 0 9 o th e r s 4 ,48 4 3 Tota l (8 0 ,00 ) 9 6 0 No de 1 Ca te g or y % n ae a b la u t 35 ,5 7 2 3 9 re g u la r 49 ,5 5 3 3 3 aa b la u t 9 ,3 8 6 3 oth e r s 5 ,5 1 3 7 To ta l (5 6 ,0 0 ) 6 7 2 No d e 4 Ca te g o r y % n a e -a b la u t 4 7 ,5 7 1 37 r e gu lar 3 4 ,3 8 9 9 a ab lau t 1 5 ,9 7 4 6 o th e rs 2 ,0 8 6 To ta l ( 2 4,0 0 ) 2 88 No d e 3 p a s t ten s e c o d a I m p r o v e m e n t =0 ,0 23 6 ( g ,k); ( f ) (n k); ( ng ) ; (m, n ) ; (t) c o d a I m p r o v e m e n t =0 ,0 0 72 (n k) ; (m, n ) ; (t) (n g ) Figure 8: Classification analysis showing the interaction of onset and coda with the types of vowel change in 1_E_L2 As we can see in the tree diagram, coda has been detected as the strongest predictor (first split). Consider Node 3 now, which is the subset featuring the highest amount of [ ]-ablaut (47.57%). The subgroup is defined by the presence of coda [ ], which could be taken as evidence for the schema [# _ _ ] to be prototypical for [ ]-ablaut in the German-English interlanguage. The risk estimate for this model, however, is 0.443333, which means that less than half of the data have been classified correctly. Thus, caution is in place when interpreting the magnitude of effects in this analysis. Before we move on to comparing the results of the English L1 speakers and the German-English interlanguage speakers, let us summarise all results for experiment 1_E_L2 in (35). 94 (35) (a) 49% of all responses show vowel changes (b) onset and coda are significant predictors for the types of past tense (c) codas [ ] and [ ] favour vowel change (d) schema [sCC _ _ ] is prototypical for the distinction between past tenses with and without vowel change if onset and coda are coded binary (e) [ ]-ablaut is the preferred vowel change pattern (f) only 4.3% of all responses show unorthodox vowel change patterns (g) coda is the only significant predictor for the types of vowel change (h) the nonce-verbs that attracted most of the [ ]-ablaut formations follow the schemas [sC _ _ " ], [sCC _ _ ( )] , and [CC _ _ ] What do the above results mean for the German-English interlanguage morphology? First of all, like English L1 speakers, the learners, too, are capable and willing to provide vowel change past tenses in analogy to existing irregular verbs. They do so in about 50% of their responses, and their vowel changes are remarkably faithful to existing patterns. Coda shows clear partial effects on the type of past tense as well as on the type of vowel change. In addition to that, [sCC _ _ ] emerged as interacting terms for vowel change in general. However, as with German L1 speakers, from the data of German-English interlanguage speakers no significant interactions, and hence no prototypes, can be distilled for the type of vowel change. Thus, as with the two previous experiments, simple similaritybased analogical formations seem to be most appropriate to account for the interlanguage data. Let us have a closer look at how the responses of the English L1 speakers differ from those produced by the learners. Consider figure 9 first, which plots the type of past tense formation (regular versus vowel change) against the type of speaker (interlanguage speaker and L1 speaker). type of speaker by past tense type of speaker past tense interlanguage speaker L1 speaker regular vowel change 2 = 25.2849 (Yates’ correction), df = 1, p < .005 Figure 9: Mosaic plot for English verbs cross-classified by past tense formation and the two types of speaker (interlanguage speaker, L1 speaker) 95 As we can see, there is a significant asymmetry between how L1 speakers and interlanguage speakers choose between regular and vowel-changing past tenses. Accordingly, an analysis of deviance of a generalised linear model, with onset, coda, and type of speaker as predictors shows that for the choice of past tense formation both the syllable constituents and the type of speaker have significant partial effects. This was confirmed in a logistic regression model, which, after a fastbackward-variable-selection-procedure, retained all three factors as significant in the final model (onset 2 = 13.40 (Pearson), df = 2, p < .005, coda 2 = 121.24 (Pearson), df = 5, p < .0005, type of speaker 2 = 28.27 (Pearson), df = 1, p < .0001). Interestingly, though, the significance of the type of speaker changes dramatically once random effects of the within-subject variance are allowed into the model. A mixed effects model via PQL with both constituents as well as type of speaker as fixed effects, and subject as random effects, first of all revealed that the estimate of the random effect is substantial (s of the residuals = 0.968). At the same time, the partial effect of the type of speaker is insignificant in this model (post-hoc ANOVA, F(1,62) = 3.149, p = .081); and so is its single main fixed effect ( = -0.561, t(62) = - 1.775, p = .081). However, once the mixed effects model allowed for interactions between coda and the type of speaker, a significant effect emerged for the interaction between type of speaker and coda (post-hoc ANOVA, F(10,1422) = 15.277, p < .0001). In sum, the asymmetries in figure 9 only arise given an interaction of speaker and constituent. Thus, only with regard to the L1 speakers’ treatment of coda do native speakers of English show a greater tendency towards vowel-changing tenses than German-English interlanguage speakers. Let us compare the distribution of vowel change patterns across the two groups of speakers now (table 18). Table 18: Frequencies for the types of vowel change in 1_E_L1 and 1_E_L2 English L1 speakers interlanguage speakers frequency percent frequency percent [ ]-ablaut 132 33.0 % 414 34.5 % [ ]-ablaut 64 16.0 % 123 10.3 % others 50 12.5 % 51 4.3 % In table 18, we can see similarities as well as differences between English L1 and interlanguage speakers. Both groups of speakers feature [ ]-ablaut as their most frequent vowel change pattern. The proportion between the two types of vowel change ([ ]and [ ]-ablaut) is different, though. And interlanguage speakers do not produce as many deviating patterns (‘oth- 96 ers’) as the L1 speakers. Consider figure 10, which plots the types of vowel change against the types of speaker. past tense by type of speaker type of speaker past tense interlanguage speaker L1 speaker a ablaut ae ablaut others regular 2 = 50.8221 (Pearson), df = 3, p < .0001 Figure 10: Mosaic plot for English verbs cross-classified by past tense and type of speaker in 1_E_L2 Figure 10 shows significant asymmetries across the two types of speakers, with a pronounced effect of L1 speakers towards [ ]-ablaut and ‘others’. A classification analysis, however, with onset, coda, and type of speaker as predictors and the type of vowel change as response variable, revealed that the type of speaker is not significant, neither as a single main effect nor in any interactions. This means that, in general, the choice between different vowel change patterns is not group specific. Thus, although coda seems to play a different role for the two groups of speakers when it comes to choosing between regular and vowel-changing past tense formations, it is not a significant factor for the choice between the vowel changes used. In sum, there is no significant difference between the ways in which native speakers and learners choose their preferred type of ablaut ([ ]-ablaut). There is, however, a difference as regards the kind of morphological attraction that triggers these [ ]-ablaut formations. While for the English native speakers, two statistically traceable prototypical schemas emerged ([sC(C) _ _ ] and [sC(C) _ _ ]), no significant combination of constituents could be found for the interlanguage speakers. 97 3.3 Summary and discussion The results of the above experiments provide numerous insights into the organisation of irregular verb morphology in the learners’ L2 mental lexicon. Some of the insights seem to corroborate existing findings, while others challenge them. As regards the learners’ L1 (German), the phonological make-up of a verb seems to serve as an analogical attractor towards irregular past tense formation. More than a third of all German L1 responses showed vowel change, with the [ ]-[ ]-ablaut being the salient pattern for German irregular past tense morphology. The analogical attraction towards this type of ablaut seems to rest on three parameters. The first one is the complexity of onsets. Complex onset structures featuring the extra-syllabic [ ] seem to be prototypical for vowelchanging past tense formations in general. Second, the length of the stem vowel is critical for the type of vowel change speakers choose. While [ ] favours ablaut in [ ], nucleus [ ] shows a bias towards [ ]-ablaut. Third, the presence of the velar nasal in codas, although not significant for the distinction between regular and vowel-changing past tense formations, seems to be a good predictor for the type of vowel-change. Here, codas [ ] and [ ] are prototypical for [ ]-ablaut. Based on the observed effects, the learners’ L1 grammar can be expected to be sensitive to the following similarity-based affects: if a verb’s onset is complex, and if the stem vowel is short and fronted, and if the coda contains a velar nasal, then the probability of such a verb being inflected by vowel-change is very high. In other words, the three constituent variants sCC, [ ], and [ ( )] are perceived as single prototypical attractors for [ ]ablaut. However, recall that in the statistical models no significant interaction of the predictors could be found. Thus, despite the prototypicality of the above individual variants, we cannot claim at this point that the one combined form [sCC _ _ ( )] would in fact serve as a prototype for [ ]ablaut. Irregular verb morphology in the learner’s target language (English L1), too, has turned out to be sensitive to the morphological make-up of a given verb. More than 60% of all English L1 responses showed vowel change, with [ ]-ablaut being the most salient pattern. The similarity-based effects that can be found, however, are slightly different compared to German L1. While it is not the most complex onset any more, but sC that is prototypical for vowel change in general, now vowel change arises with all nasal codas and the obstruent [ ]. Interestingly, codas in the voiceless fricative [ ] never produced any effects towards vowel change. This is not really surprising since all existing English verbs ending in the voiceless fricative are in fact regular (cf. Albright and Hayes 2003). 98 However, the choice of the most frequent ablaut pattern in [ ] in English L1 can clearly be attributed to the schemas [sC(C) _ _ ] and [sC(C) _ _ ], while ablaut in [ ] depends on the schema [sC(C) _ _ ]. The salience of onset structure sCC in English is probably due to its special phonotactic status 24 . In sum, it seems as if onsets with extra-syllabic [ ] and codas with either simple engma or a combination of engma and an obstruent are prototypical constituent variants for English L1 speakers. In addition to that, the choice between different types of ablaut depends on three schemas. These findings for English L1 are, except for the preferred vowel change pattern, mostly in line with Bybee and Moder’s (1983) findings. If we take the similarity-based attractors of the learners’ mother tongue and their target language together, the following situation arises (36). (36) prototypical onset prototypical nucleus prototypical coda prototypical schema German (learners’ L1) sCC [ ] [ ( )] [sC(C) _ _ ] English (learners’ target) sC, sCC [ ] [ ( )] [ ] [sC(C) _ _ ] [sC(C) _ _ ] The prototypical constituent variants in the English and German L1 show remarkable similarities. But what does this mean for possible similaritybased effects in the interlanguage? First of all, we can state that in the German-English interlanguage, too, morphology serves as an analogical attractor for irregular verb formation. Almost half of all German-English interlanguage responses showed vowel change, with [ ]-ablaut being the predominant pattern. Note here that when interlanguage speakers applied ablaut, they did so with hardly any unorthodox formations. This degree of faithfulness to existing patterns even outranks English L1 speakers. The morphological attraction in the interlanguage seems to rest on three parameters, too. They are summarised in (37), extending the illustration in (13) on to the interlanguage. (37) prototypical onset prototypical nucleus prototypical coda prototypical input schema German L1 sCC [ ] [ ( )] [sC(C) _ _ ] interlanguage sCC [ ] [ ( )] [# _ _ ] English L1 sC, sCC [ ] [ ( )] [ ] [sC(C) _ _ ] [sC(C) _ _ ] 24 Segment [ ] is more sonorous than [t] and [k], but it can occur before them in monosyllabic words such as sky and stripe, resulting in a double sonority peak. 99 Looking at the similarities in (14), it appears as though irregular verb inflection in the German-English interlanguage was governed by the same cognitive mechanism as in German and English L1. The prototypicality of certain syllable constituents ranges across all three groups of speakers, and both the speakers and the learners of English prefer the same type of ablaut ([ ]). When comparing these interlanguage data to the results reported by Plag (2000), we can find similarities as well as differences. First, looking at the preferred type of vowel change, in experiment 1 both native speakers and interlanguage speakers prefer [ ]-ablaut. This means that if speakers and learners are confronted with the same selection of test items the formerly reported discrepancy between preferred [ ]-ablaut (English L1) and [ ]-ablaut (English L2) is no longer observable. In order to closely examine more of the effects reported Plag’s data, they were re-analysed using a mixed effects model via PQL with all three constituents as fixed effects and the subjects as random effects. Contrary to Plag’s classification analysis, but similarly to the present interlanguage experiment, this regression model did not find any significant interactions. The fixed effects, though, show a good degree of overlap with the effects from experiment 1. The fixed effects from Plag’s data are summarised in table 19. Table 19: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset, nucleus, and coda as predictors and types of past tense formation as response variable in Plag 2000 constituent -coef s-value df t-value p-value sign. intercept -5.008 0.741 1321 -6.755 = .0000 *** onsets CC -0.235 0.193 1321 -1.222 = .2220 ---- sC 0.321 0.181 1321 1.774 = .0762 ---- sCC 0.310 0.179 1321 1.735 = .0829 ---nucleus 2.593 0.524 1321 4.952 < .0001 *** 2.671 0.526 1321 5.075 < .0001 *** codas , 1.584 0.540 1321 2.933 < .005 ** -0.308 0.742 1321 -0.415 = .6779 ---- " , 1.644 0.535 1321 3.071 < .005 ** 2.574 0.579 1321 4.445 < .0001 *** k 2.369 0.548 1321 4.320 < .0001 *** , -0.360 0.678 1321 -0.531 = .5953 ---- , 2.891 0.571 1321 5.066 < .0001 *** * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant First note that, contrary to nucleus and coda, onset does not produce any significant -coefficients in Plag’s data. Second, there are seven significant coefficients for nucleus and coda, with all positive -coefficients illustrating a tendency towards vowel-changing past tense formations. This is most 100 pronounced, and significant, for both nuclei ([ ] and [ ]), as well as for the codas with velar nasals and [ , ]. This is partly reminiscent of the interlanguage results reported in table 15, where onset played also a minor role when it comes to the distinction between regular and vowel-changing past tense formations, and where codas featuring the velar nasal showed significant fixed effects towards vowel change. Overall, all three groups of speakers in experiment 1 show a remarkable willingness to provide non-default inflections to given nonce-verbs. They all show a considerable rate of faithfulness to existing patterns, and in all three groups onset and coda are significant predictors of the type of past tense (with or without vowel change). Moreover, coda is the strongest, and sometimes the sole predictor for the type of vowel change across the three groups. This observation could be taken as evidence for some sort of crosslinguistic prototypicality for word-final structures in general, and for the importance of the velar nasal in codas in particular. The question now is how the similarity-based organisation in an interlanguage mental lexicon can be accounted for, and how it develops with regard to the influence of the native language of the learners and the target language grammar. Recall that Plag had proposed cross-linguistic influences as one possible mechanism. Such a mapping of the segments could be envisaged as follows: first, the learners are aware of the prototypicality of a three-stage ablaut pattern in German and English. They are also aware that in both German and English certain syllable segment variants serve as similaritybased attractors towards vowel-changing past tense formations. Thus, they could transfer, for instance, the prototypicality of a complex onset, a short front vowel and the velar nasal coda from the mother tongue onto their interlanguage grammar. When they then form ablaut, they transfer the prototypical three-stage [ - - ] pattern of their mother tongue onto their interlanguage and, as Plag already suggested, realise the interlanguage ablaut pattern so that it is compatible with a similar three-stage ablaut in English ([ - - ]). Such a strategy might be supported by the parallelism in writing between German and English (singen-sang-gesungen - sing-sangsung) 25 . However, although the similarity-based attractions are similar across the three groups of speakers, there are two problems with the current analysis. First, in all three groups the statistically traceable effects never overlap entirely with the morphological make-up of those forms triggering most of the vowel-changing responses. This problem has re-occurred now in the production data of five nonce-verb experiments (Bybee and Moder 1983, Plag 2000, present experiments) and thus seems to reflect an inherent problem when statistically tracing possible prototypes. A second and 25 For further elaborations on the role of transfer for German-English interlanguage irregular verb inflection see chapter 6 (discussion and conclusion). 101 probably related problem is the lack of the predictors’ interactions. Recall that single syllable constituent variants can show statistically traceable prototypicality effects, while combinations of such constituents do not. So why do effects that are observable in the ablaut-triggering test items do not emerge as interacting effects in the statistical models? This again raises the question of whether it really is one clear-cut prototype governing a certain type of irregular verb inflection, or whether it is a more loosely organised form of similarity-based morphological attraction. In the light of this situation, let us review the hypothesis that guided this experiment. Although the present data suggest that the irregular verb morphology in the interlanguage rests on similarity-based analogies transferred from the learner’s L1 and adapted to the target language, a clear prototype could not be established in any of the three experiments. Hence, hypothesis 1 and the notion of a prototype governing German-English interlanguage will be rejected at this point. 103 4 Experiment 2 This chapter provides a detailed account of experiment 2. First, the experimental design will be discussed, followed by the results. Recall the hypothesis that guided this experiment. It is restated for convenience in (38). (38) Hypothesis 2 German-English interlanguage irregular verb morphology is best modelled as rule-like input-output processes that follow the universal apophonic path. According to hypothesis 2, participants’ vowel change formations, when inflecting given nonce infinitives for past tense and participle, should be in accordance with the apophonic path. Recall that the path defines, amongst others, the two vowel changes i a and a u as rule-like input-output oriented derivations of one vowel quality out of another. Such derivations are assumed to be directional; they are not expected to work in the reverse order. In other words, speakers should prove to be able to complement a fragmentary apophonic path such as [i x x], for example, with ease, whereas the reverse direction such as in [x x u] should prevent speakers from systematic vowel change inflections. 4.1 Experimental design In order to examine the psycholinguistic reality and directionality of an apophonic path, experiment 2 was carried out in two directions, one being the reverse of the other. In all three groups of speakers (German L1, English L1, German-English interlanguage), some participants were prompted to provide inflections for a given infinitive, while others were presented participles (or past tenses, see below for an explanation) to inflect ‘backwards’ for past tense and infinitive. In the analysis of the participants’ behaviour, the consistency between the two vowel changes and their faithfulness to the apophonic path were of prime importance. Since there were three groups of speakers and two complementing directions of inflection across the groups of speakers, experiment 2 was carried out six times. An overview is given in (39). (39) Experiments on vowel change series starting out with the infinitive (a) 2_G_L1a German native speakers (b) 2_E_L1a English native speakers (c) 2_E_L2a German-English interlanguage speakers 104 Experiments on vowel change series in the reverse direction (d) 2_G_L1b German native speakers (e) 2_E_L1b English native speakers (f) 2_E_L2b German-English interlanguage speakers As in chapter 3, each type of experiment has its individual code, denoting the number of the experiment, the type of speaker (G_L1, E_L1, E_L2), and in addition to that the order of the inflection (a = forward, starting from the infinitive, and b = backward). All questionnaires consisted of widely employed sentence-completion frames (cf. Clahsen 1999) with nonce-verbs displaying a systematically varying proximity to possible prototypes (cf. Bybee and Moder 1983: 258, Prasada and Pinker 1993a: 24, Albright and Hayes 2003). The sentencecompletion frames were all preceded by an introductory text, which informed participants that they were taking part in a genuine linguistic experiment. The nonce-verbs were sold as an extra convenience in order to render the whole procedure more interesting. An audio CD provided a 10second-timing for two gaps, and each experiment took 15 minutes to administer. Stimuli were presented both orally and in writing, and written responses were recorded by the experimenter. A sample of English and German sentence completion frames can be found in (40). (40) (a) English forward Sam likes to smink. Yesterday he ____. Have you ever ____? backward Have you ever smunk? Yesterday I saw Sam; he ___ the whole evening. I think, Sam likes to___. (b) German forward Sabine hat immer Lust zu knießen. Gestern ____ sie den ganzen Abend. Hast du auch schon mal ____? backward Gestern habe ich Sabine gesehen; sie knoss den ganzen Abend. Ich glaube, Sabine hat immer Lust zu ____. Hast du schon jemals ____? As we can see in (3), both English and German sentence completion frames have an almost identical structure. However, while in English for the experiments in the reverse order participles (such as smunk) are given, German questionnaires featured past tense nonces (knoss, for instance). Thus, English questionnaires prompted participants to complement past tense and infinitive formations backwards, whereas in German infinitives and participles had to be filled in. The rationale behind this difference is the following: In German, the choice of possible past tense inflections in a reverse order from a given participle is rather restricted. A given [ ] as a participle stem vowel, for instance, would almost automatically elicit [ ] in the past tense, since [ - - ] is the by far predominant German pattern featuring participle vowel [ ]. 105 In English, in contrast, a given [ ] as a participle vowel leaves a considerable choice for possible past tense pairs. It can, for instance, be complemented backwards with [ ], as in cling-clung-clung, hang-hung-hung, cut-cutcut, or with [ ], in analogy to sing-sang-sung or run-ran-run. In order to test possible preferences of speakers, English questionnaires featured the participle, whereas for German past tenses were given. For experiment 2, nonce participles and nonce past tenses were created on the basis of the items in experiment 1. For the English nonce participles, test items from experiment 1 were transposed into participles containing stem vowel [ ]. A word such as smink, for instance, would become the participle smunk. 15 Fillers were interspersed as distractors. They followed ablaut patterns other than the ones under investigation, such as nillt (in analogy to spill - spilt) or haught (in analogy to teach - taught). As in experiment 1, there were two different orders of nonce-verb presentation, which were merged at the end of the administration of an experiment. For the German nonce past tenses three different stem vowels were used. Recall that the set of German infinitive nonces from experiment 1 contained the two stem vowels [ ], such as in knießen, and [ ], such as in stinnen. Reflecting the distribution and frequency of existing ablaut patterns with those two stem vowels in German, three past tense vowels were used. They are given in (41). (41) (a) [ ] [ ] [ ] (plog) 26 (b) [ ] [ ] [ ] (knoss) (c) [ ] [ ] [ ] (schmann) Test item friegen, for example, from experiment 1 was changed into frog in analogy to fliegen (‘to fly’) - flog (‘flew’). Note that all of the past tense vowels in (4) are accounted for by the apophonic path. As in the English questionnaires, 15 fillers were interspersed as distractors, this time following the regular German past tense, such as hittelte or stitzte, and, as in the English questionnaires, there were two different orders of nonce-verb presentation. 4.2 Results All participants’ responses were computerised and coded according to syllable constituent, type of first response, type of second response and the overall vowel change series that emerged from both responses. Sometimes, problems arose for the coding of vowel changes that did not match existing 26 Potential confusion about the pronunciation of a verb such as plog (long or short [ ]) was ruled out by having each item read out on the audio CD. 106 patterns. Consider results such as [ - - ] or [ - - ] for English. In the first example, the central [ ]-vowel could be interpreted as a valid participle ablaut for [ ]-infinitives (cf. sing-sung), but the [ ] in the past tense is not modelled on any existing pattern. In the second example, one could take the sequence as a jumbled sing-sang-sung pattern, or count the [ ]-past tense and the [ ]-participle as valid vowel changes from the infinitive’s perspective, while the overall series, though, would constitute a novel vowel change creation. Another frequent pattern was the intermingling of regular inflection and ablaut. Patterns such as [ reg. - ] for English or [ - reg.] for German were not rare, and again, although the [ ]-participle or the [ ]-past tense could be counted as instances for ablaut, the ablaut series as a whole did not resemble existing patterns. In the following analyses, once ablauts did not match any existing pattern (including mixed patterns), the overall sequence was coded as ‘novel vowel change’. 4.2.1 Experiment 2_G_L1a (Native speakers of German) Experiment 2_G_L1a contains the results of 32 native speakers of German, all undergraduate students from the Universities of Siegen and Gießen. They produced a past tense form (Sabine hat immer Lust zu knießen. Gestern ___ sie den ganzen Abend) and a participle (Hast du auch schon mal ___? ) to a given infinitive nonce. In the following, the emerging overall vowel change sequence, called ‘past-participle sequence’, was analysed first. Afterwards, the consistency between the two responses was looked at. The two types of emerging past-participle-sequence are given in table 1. Table 1: Frequencies for the types of past-participle-sequence with or without vowel change in 2_G_L1a types of past-participle-sequence frequency percent without vowel change 400 43.9 % with vowel change 512 56.1 % 912 100.0 % As we can see, past-participle-sequences with vowel change are in the absolute majority (56.1%). This means that German L1 speakers were indeed prepared to produce inflections in analogy to existing patterns apart from applying regular default inflection. Consider figure 11 now, where the two types of past-participle-sequence are plotted against onset, nucleus and coda 27 . 27 Here and in the following diagrams the string <ch> denote either [ # ] or [ $ ]. 107 onset by past-participle-sequence onset past-participle-sequence C CC sch schC schCC no vowel change vowel change nucleus by past-participle-sequence nucleus past-participle-sequence long i short i no vowel change vowel change coda by past-participle-sequence coda past-participle-sequence g m,n ng ngk s,ch no vowel change vowel change 2= 6.1008 (Pearson), df = 4, p = .1917 2= .7722 (Pearson), df = 1, p = .3795 2= 27.5639 (Pearson), df = 4, p < .0005 Figure 11: Mosaic plots for German verbs cross-classified by past-participle-sequence and the three constituents onset, nucleus, and coda in 2_G_L1a As we can see in figure 11, only coda shows significant asymmetries across the types of past-participle-sequence. Accordingly, both a trimmed generalised linear regression model and a simple logistic regression model revealed a significant partial effect for coda only. And after a PQL model with pooled codas did not find any significant interactions 28 , the final mixed model, now disregarding interacting terms, revealed both significant partial effects for coda (post-hoc ANOVA F(4,871) = 11.721, p < .0001) and fixed main effects for all three constituents. All coefficients taken from the final model are reported in the table below. Table 2: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset, nucleus, and coda as predictors and the types of past-participle-sequence as response variable in 2_G_L1a constituent -coef s-value df t-value p-value sign. intercept -0.449 0.424 871 -1.059 = .2900 ---nucleus -1.109 0.452 871 -2.452 < .05 * onsets CC 0.400 0.272 871 1.473 = .1410 ---- 0.272 0.304 871 0.894 = .3714 ---- C 0.295 0.241 871 1.223 = .2218 ---- CC 1.062 0.324 871 3.277 < .005 ** codas " , 0.711 0.534 871 1.332 = .1832 ---- 2.221 0.549 871 4.048 < .0001 *** k 1.727 0.545 871 3.171 < .005 ** , # 0.666 0.288 871 2.314 < .05 * * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant 28 As in experiment 1, interactions and possible random effects were analysed using the PQL model, and as in experiment 1, all models showed a substantial estimate of the random effects. 108 The coefficients in the above table read as follows. Negative values illustrate a tendency towards past-participle-sequences without vowel change. This tendency is significant for nucleus [ ]. In contrast, higher positive coefficient values indicate a constituent’s tendency towards vowel changing sequences, significant for onset CC and the two codas with a velar nasal. Note that only five of the 10 coefficients are significant. This means that the choice for vowel-changing past-participle-sequences depends on the presence of onset CC, nucleus [ ] and codas [ (k)] and [ , # ]. Note too that the more complex the onset, the bigger the effect size and the higher its degree of significance. A classification analysis explored how consistent the relationship between the first response (past tense) and the second response (participle) actually is. It turns out that there is a high degree of consistency between how participants handled the inflection for participle and the emerging overall past-participle sequence. Figure 12 shows the resulting tree diagram. Figure 12: Classification analysis showing the interaction of participle inflection with the resulting past-participle-sequence in 2_G_L1a The above tree can be read as follows. The split of the upside-down tree and the two branches illustrate a decision-procedure in order to determine the choice between past-participle sequences with or without vowel changes as clear-cut as possible. The leaf-nodes specify non-overlapping subsets of the entire data set. For both subsets (nodes), the classification algorithm inspected all predefined predictors (here the two responses) and selected the one that is most useful for distinguishing between the two classification of overall sequence by both responses | Participle = past tense vowel + t no vowel change 398/ 0 vowel change 1/ 513 109 subsets. It turns out that in this model only the inflection of the participle is a significant predictor for the overall past-participle sequence. The mother node is headed with a decision rule, stating that for all cases in which the participle is inflected using the past tense vowel plus suffix -t (regular participles), we get a the new distribution of overall vowel change and novowel-change responses as given in the left hand node (bottom left corner), namely 398 cases with no overall vowel change, and no case with vowel change. If the participle is not inflected regularly, the right hand-side branch illustrates the new subgroup with its distribution (1 to 513). In other words, participants handled the participle inflection in a very coherent fashion. This finding was complemented by fitting a mixed effects regression model to the data in order to predict the overall vowel change series. All three constituents as well as past tense formation entered the model as fixed single effects and subject as random effects. In this model, past tense formations without vowel change serve as an excellent predictor for the types of overall past-participle sequence ( = -13.677, t(867) = -10.297, p < .0001). Thus, in sum we can say that the there is a high degree of consistency between the two responses. It appears as if participants treated both inflections in tandem and thus produced either regular inflections or vowel change patterns in a systematic fashion. The question remains, however, how consistent the vowel change patterns themselves are. Let us have a closer look at the vowel change patterns that German L1 speakers produced. Table 3 shows their frequencies. Table 3: Frequencies for all vowel change patterns in the past-participle-sequences in 2_G_L1a examples for existing triplets frequency percent no vowel change 400 43.9 % pattern 2 gießen-goss-gegossen [ - - ] 162 17.8 % pattern 5 singen-sang-gesungen [ - - ] 122 13.4 % novel vowel change 77 8.4 % pattern 6 rinnen-rann-geronnen [ - - ] 59 6.5 % pattern mixed bringen-brachte-gebracht [ - - ] 53 5.8 % pattern 1 biegen-bog-gebogen [ - - ] 26 2.9 % pattern 4 schinden-schund-geschunden [ - - ] 10 1.1 % pattern 3 liegen-lag-gelegen [ - - ] 3 .3 % 912 100.0 % While 43.9% of all sequences are regular (no vowel change), interestingly, amongst all emerging vowel changes only 8.4% are inventions (‘novel vowel change’). This means that only a few L1 speakers invented ablaut patterns alien to German. The past-participle-sequence with the highest frequency is the two-stage pattern [ - - ] with 17.8%. The strongest three- 110 stage pattern occurs with the [ - - ] vowel change, which could be taken as evidence for the salience of the three-stage [ - - ] pattern whenever participants inflect from a given [ ] stem vowel. Note that all patterns that emerged in this experiment, except for [ - - ], are accounted for by the apophonic path (see Ségéral and Scheer 1998: 41, where they argue that [ - - ] was non-apophonic anyway). Thus, there is evidence that once German L1 speakers construct past-participle-sequences with vowel change patterns, those are generally in accordance with the apophonic path; in total only 9.5% are not. The given infinitives with the highest frequencies of vowel changing past-participle-sequences are ranked in table 4. Table 4: Highest frequencies for responses with existing, novel, and no vowel changes in 2_G_L1a existing vowel change novel vowel change no vowel change [ - - ] [ - - ] [ -x-x ] [ - - ] fießen 19 stingen 17 schrimmen 7 stinnen 20 strießen 18 sprinken 16 friegen 5 stiechen 19 triechen 18 strinken 15 strimmen 5 linnen 19 As we can see, infinitives of the form [# __ __ / # ] and [ C(C) __ __ ( )] triggered most of the patterns that followed existing past-participlesequences in German L1. This is mostly in line with the statistical effects from the above models. Recall that effects were detected for onset CC, nucleus [ ], and coda [ ], [ ], and [ , # ]. These effects are clearly mirrored in the most frequent infinitive nonce-verbs triggering ablaut pattern [ - - ]. Let us now have a closer look at how coherent the two responses are (table 5). Table 5: Frequencies for types of past tense and participle formation to a given infinitive nonce in 2_G_L1a frequency past tense frequency participle no vowel change 414 no vowel change 398 [ ] [ ] 260 [ ] [ ]/ [ ]/ [ ] 237 [ ] [ ] 214 [ ] [ ] 188 [ ] [ ] 19 [ ] [ ] 10 others 5 others 79 912 912 As far as past tense formations are concerned, [ ]-ablaut is the most frequent one (260), followed by past tenses in [ ] (214). This is reminiscent of the results in experiment 1, where [ ]-ablaut was the most frequent vowel 111 change response, too, albeit with a considerably lower relative frequency. The most interesting result in table 5, though, is that, overall, inflections for past tense correspond rather well with the inflections for participle. This impression is supported by a 2 -test ( 2 = 2245.834 (Pearson), df = 32, p < .0001). However, note too that all three vowel change patterns do lose a number of cases to the residual category ‘others’ (74 cases). And note that this sort of erosion also affected regular past tense formations, which, surprisingly, in 16 cases resulted in vowel-changing participles (414 regular past tense inflections, but only 398 regular participles). The least consistent vowel change pattern in this respect is the sequence [ ]-[ ]-[ ], which, as mentioned before, was regarded to be non-apophonic by Ségéral and Scheer (of 19 past tense items in [ ] only 10 feature an [ ]-participle, too). In contrast, the other two sequences are built rather consistently, and the inconsistencies do not arise due to ungrammatical mixtures of vowel change patterns but through the unorthodox, non-apophonic participles mainly. Overall, despite the above inconsistencies, vowel changes were created with a high degree of faithfulness to existing patterns and a high degree of consistency amongst each other. Let us summarise all results for experiment 2_G_L1a in (42). (42) (a) German L1 speakers produced past-participle-sequences with vowel change in 56.1% of all cases (b) amongst all vowel changes produced, only 8.4% are novel patterns (c) the distinction between past-participle-sequences with and without vowel change mainly depends on coda (d) fixed effects for vowel-changing past-participle-sequences could be found for onset CC and codas [ ], [ ], and [ , # ]. No interactions were found. The more complex onset is, the bigger is the effect size and the higher its degree of significance (e) the most frequent vowel change pattern is the two-stage pattern [ - - ] (f) there is a high degree of consistency between the first and the second responses as regards the distinction between regular and vowelchanging past-participle-sequences (g) there is a high degree of consistency between the first and the second responses as regards the different vowel change patterns Overall, the results of experiment 2_G_L1a seem to confirm the presence of an input-output oriented apophonic path governing German vowel change patterns, as proposed by Ségéral and Scheer. Native speakers of German neither mixed regular and vowel-changing patterns, nor did they combine different patterns when creating vowel-changes. Recall that a mere 8% of all cases deviated from existing patterns (s. ‘novel vowel change’ in table 5), and that the non-apophonic [ ]-[ ]-[ ]-pattern is almost non-existent. This indicates the psycholinguistic validity of an apophonic path in the German L1 speakers’ minds as well as the speakers’ ability to use such a path in order to model vowel changes starting out from a given infinitive. 112 4.2.2 Experiment 2_G_L1b (Native speakers of German) In experiment 2_G_L1_b, 34 native speakers of German, all undergraduate students from the Universities of Siegen and Gießen, were exposed to a given past tense and were prompted to provide infinitives and participles of their choice. Table 6 summarises the two types of infinitive-participlesequence that emerged in this experiment. Table 6: Frequencies for the types of infinitive-participle-sequence with or without vowel change in 2_G_L1b type of infinitive-participle-sequence frequency percent without vowel change 544 57.1 % with vowel change 408 42.9 % 952 100.0 % Contrary to experiment 2_G_L1a, the reverse version of the experiment features a minority of vowel changes in the overall sequence. While 65.1% of all responses in the previous experiment featured vowel change, here it is only 42.9%. This is quite a remarkable observation, since all given past tenses were overtly marked as irregulars (featuring no suffixation). Thus the question arises why so many participants simply repeated a given irregular past tense when trying to form a matching infinitive. Note that, unlike English, where congruity between past tense and infinitive vowels is quite common for a substantial number of the so-called ‘no-change verbs’ (cut-cut-cut, hit-hit-hit), such a pattern is rather rare in German 29 . Nevertheless, the majority of the German native speakers translated the past tense vowel directly into the infinitive, thereby simply repeating the given form of the nonce (e.g. stroch-stroch). This is a first indicator of a rather random and unpredictable inflection backward from a given past tense, and it could be taken as evidence that constructing vowel changes from a given output form seems to be considerably harder than from a given input form. Consider figure 13 now, which plots the types of infinitive-participlesequence against the variants of the syllable constituents. 29 The only common modern German ‘no-change-verbs’ are mahlen (‘to grind’), salzen (‘to salt’), and spalten (‘to split, to chop’). Notice, however, that these verbs are also mixed verbs and combine ablaut with regular affixing. Thus, a past tense form would, for instance, exhibit suffix -te. The test items in this experiment, though, were not suffixed. 113 onset by infinitive-participle-sequence onset sequence C CC sch schC schCC no vowel change vowel change nucleus by infinitive-participle-sequence nucleus sequence long o short a short o no vowel change vowel change coda by infinitive-participle-sequence coda sequence g m,n ng ngk s,ch no vowel change vowel change 2= 2.3535 ((Pearson), df = 4, p = 0.671 2= 0.239 (Pearson), df = 2, p = 0.8874 2= 3.7602 (Pearson), df = 4, p = 0.4394 Figure 13: Mosaic plots for German verbs cross-classified by infinitive-participlesequence and the three constituents onset, nucleus, and coda in 2_G_L1b Note that both types of infinitive-participle-sequence are distributed evenly across onset and nucleus, with only minor asymmetries visible for coda. Accordingly, trimmed generalised and logistic regression models identified only coda as a significant predictor. However, despite this impression, a mixed effects model with subject as random effects, and onset, nucleus, and a pooled coda (single consonant versus codas with engma) as fixed effects produced three significant interactions of onset and coda. They are summarised in table 7. Table 7: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with interacting onset and (pooled) coda as predictors and the types of infinitive-participle-sequence as response variable in 2_G_L1b -coef s-value df t-value p-value sign. intercept -0.663 0.699 905 -0.948 = .3434 ---interactions _ ( ) 1.389 0.697 905 1.993 < .05 * C _ ( ) 1.699 0.552 905 3.078 < .005 ** CC _ ( ) 2.064 0.630 905 3.279 < .005 ** * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant We can see that all interaction effects involve the codas with engma [ ( )]; and they are significant. The single-consonant-codas, mapped onto the intercept, are not. Moreover, we can see an effect for onset similar to the one in the previous experiment; the more complex onset, the greater its effect size and significance. All interaction coefficients are positive and thus, in this model, show an effect towards vowel-changing infinitiveparticiple series. These interactions also emerged in a classification analysis, with all three constituents as predictors (coda pooled as above, named 114 ‘Coda 2’) and the emerging overall infinitive-participle-sequence as response variable. Figure 14 shows the resulting decision tree. Figure 14: Classification analysis showing the interaction of all three constituents with the resulting infinitive-participle-sequence in 2_G_L1b From this tree we can read the following: The terminal node to the right (headed vowel change) is defined through the presence of onsets other than [ ] (<sch>), C, CC, or C, and codas other than single consonants (‘single C’). In other words, the relatively majority of vowel-changing overall series in this node (36 cases of vowel changes compared to 32 regulars) is due to the presence of onset CC and codas [ ( )], which is exactly the strongest interaction effect we find in the above mixed effects model (table 8). Thus, so far it seems as though verbs of the shape [ CC_ vowel _ ( )] are effective output-oriented schemas for vowel change. Another classification analysis explored how consistent the relationship between the first and the second reply was with respect to the difference between regular and vowel-changing infinitive-participle-sequences. The resulting tree is given in figure 15. classification of overall sequence by constituents | Onset=C,CC,sch Coda2=single C Onset=schC no vowel change 283/ 192 no vowel change 159/ 112 no vowel change 70/ 66 vowel change 32/ 36 115 Figure 15: Classification analysis showing the interaction of participle inflection with the resulting overall infinitive-participle-sequence in 2_G_L1b The above tree shows that all participle inflections that merely re-used the past tense vowel (with either suffix) resulted in 100% overall series with no vowel change (541 cases, bottom left terminal node). In contrast, almost all other participles (i.e. those featuring vowel change in their inflection) resulted in vowel-changing overall series (406 cases, with only 3 regulars). Thus, we can say that again there is a strong and predictable relationship between how participants handled the second response, the inflection for participle, and the emerging overall past-participle sequence. How consistent, though, are the vowel change patterns that emerged in this experiment? Let us have a closer look at the different patterns in the overall sequence. Table 8 shows their frequencies. classification of overall sequence by both responses | participle = past tense vowel + en, past tense vowel + t no vowel change 541/ 0 vowel change 3/ 406 116 Table 8: Frequencies for types of vowel change in the overall infinitive-participlesequences in 2_G_L1b examples for existing triplets frequency percent no vowel change 544 57.1 % novel vowel change 247 25.9 % pattern 5 singen-sang-gesungen [ - - ] 60 6.3 % pattern 2 gießen-goss-gegossen [ - - ] 42 4.4 % pattern 9 flechten-flocht-geflochten [ - - ] 22 2.3 % pattern mixed bringen-brachte-gebracht [ - - ] 16 1.7 % pattern 1 biegen-bog-gebogen [ - - ] 11 1.2 % pattern 6 rinnen-rann-geronnen [ - - ] 8 .8 % pattern 11 sterben-starb-gestorben [ - - ] 2 .2 % 952 100.0 % In contrast to the previous experiment, around a quarter of all emerging vowel changes do not conform to any existing patterns (25.9%). Novel vowel changes, in fact, form the most frequent pattern. Overall, less than 17% of all vowel changes follow existing patterns, the most frequent of which is [ - - ], attracting a mere 6.3%. Thus, there is evidence that once German L1 speakers construct participle-infinitive-sequences with vowel change, the majority of the patterns is not in accordance with the apophonic path. The given past tense nonce-verbs with the highest frequencies of vowel changing participle-infinitive-sequences are ranked in table 9. Table 9: Highest frequencies for responses with existing, novel, and no vowel changes in 2_G_L1b existing vowel change novel vowel change no vowel change [ - - ] [x- / -x] [ - - ] stoch 11 sprank 12 kang 23 gross 9 schrang 12 schmann 23 foch 12 lann 23 schlog 12 stross 11 Past tense nonces of the shape [# _ _ fricative] triggered most of the infinitive-participle-sequences with vowel change in accordance with existing patterns. Note that, although test item sprank produced mainly novel overall vowel change patterns, it was also the item that triggered most of the infinitives in accordance with pattern [ - ]. Let us now have a closer look at the coherence of the two inflections that the participants produced. Recall that they were first prompted to fill 117 in an infinitive and then the participle to a given past tense nonce. Contingency table 10 summarises all pairs of infinitive and participle formations. Table 10: Contingency table of infinitive by participle formation in 2_G_L1b infinitive formation patterns past vowel [ - ] [ - ] [ - ] [ - ] [ - ] others participle formation patterns mixed 2 21 9 3 35 others 13* 8* 15* 51* 15* 41* 64 207 [ - ] 1* 1 [ - ] 1* 1 [ - ] 8 2 10 [ - ] 1 1 [ - ] 59 1* 2* 62 [ - ] 4* 36 19 17* 76 [ - ] 11 3* 4* 18 past v. + -en 17 17 past v. + -t 523 1* 524 559 19 51 140 37 55 91 952 In the above table, all framed cells illustrate expected pairs of participle and infinitive inflection, that is, pairs that are modelled upon existing patterns. All cells marked with an asterisk reflect pairs that do not follow the apophonic path and thus are unpredictable observations. Empty cells illustrate the lack of examples for such pairs. Overall, a 2 -test suggests that infinitive and participle formation depend on each other ( 2 = 1879.864, df = 48, p < 0.001). As we can see, from 559 regularly formed infinitives (first column, bottom row), for instance, a total of 540 cases (523+17) finish as regularly formed and thus predictable participle formations (94%). This percentage can be taken as a sign for a high degree of coherence as far as regular responses are concerned. There are, however, some inconsistencies, too. The past-infinitive pattern [ - ], for instance, shows only 11 out of 19 predictable responses (58%), and from 140 [ - ] infinitives, only 88 finish in a predictable fashion (63%). It seems as though the given [ ]-past tense made some participants combine different ablaut strategies, so that in the end ungrammatical sequences evolved. We can also see in the above table that all infinitive formations show a tendency to pair up with participles in an unorthodox way (row ‘others’). This shows, too, that some participants had difficulties matching their first and their second vowel change. All important results of this experiment are summarised in (43). 118 (43) (a) German L1 speakers produced infinitive-participle-sequences without vowel change in more than 50% (b) amongst the vowel changes produced, around 60% are novel patterns (c) the most frequent vowel change occurs with a three-stage pattern [ - - ] (d) the distinction between infinitive-participle-sequences with and without vowel change significantly depends on the schema [ CC _ ( )] (e) the consistency between both responses is very high with respect to the distinction between regular and vowel-changing overall series (f) the consistency between both inflections is low amongst the vowel changes. To sum up, in contrast to the previous experiment, a high degree of consistency between both responses is observable mainly for regular inflections. There is evidence that the apophonic path works much less efficiently once it hast to be re-constructed in a non-input-output fashion. First, German L1 speakers reacted to given past tenses by avoiding vowel change responses to a great extent. Overtly marked irregular nonce-verbs were inflected for infinitive as if they were regular in almost 60% of all cases. Second, 25.9% of all overall vowel change series did not follow any existing pattern. And third, amongst those vowel changes that did follow existing patterns, the degree of consistency between infinitive and the participle formation is much lower than in the previous experiment. Evidently, a substantial amount of vowel-changing infinitives paired up unexpectedly with participles. This, too, is an indicator for the directionality of apophonic processes; they simply seem to work best if applied in the direction suggested by Ségéral and Scheer. 4.2.3 Experiment 2_E_L1a (Native speakers of English) Experiment 2_E_L1a contains the results of 140 native speakers of English, all undergraduate students from the Universities of Dublin and Maynooth (Ireland). These speakers produced two responses to each test item, namely a past tense form (Sam likes to smink. Yesterday he __.) as well as a participle (Have you ever __? ). First, the types of past-participle-sequence that emerged from the data were analysed. Afterwards, the consistency of the two responses was dealt with. Table 11 summarises the frequencies. Table 11: Frequencies for types of past-participle-sequence with or without vowel change in 2_E_L1a types of past-participle-sequence frequency percent without vowel change 690 19.7 % with vowel change 2810 80.3 % 3500 100.0 % 119 As we can see, there is an astounding proportion of past-participlesequences with vowel change (80.3%). Figure 16 illustrates the relationship between the types of past-participle-sequence and onset and coda (recall that nucleus was held constant for English). onset by past-participle-sequence onset past-participle-sequence CC sC sCC no vowel change vowel change coda by past-participle-sequence coda past-participle-sequence fricative g,k m,n ng ngk t no vowel change vowel change 2 = 23.5211 (Pearson), df = 2, p < .0001 2 = 86.8518 (Pearson), df = 5, p < .0001 Figure 16: Mosaic plots for English verbs cross-classified by past-participle-sequence and the two constituents onset and coda in 2_E_L1a In figure 16, asymmetries between the variants across the past-participlesequences are visible. They are more pronounced for coda, in particular for codas [ ], with a bias towards vowel-change, and a bias of the fricatives towards avoiding vowel change. Onset sCC (left panel), in contrast, shows a slightly stronger bias towards past-participle-sequences with vowel change than the other two onsets. A mixed effects model, with subject as random effects and binary onsets (with or without extra-syllabic [ ]) and codas (with or without engma) as interacting effects revealed one significant interaction and three significant main effects. These are summarised in the table below. Table 12: Beta coefficients and significance values for the fixed effects and their interactions of a generalised linear mixed model via PQL with onset and coda (pooled) as interacting predictors and the types of past-participle-sequence as response variable in 2_E_L1a constituent -coef s-value df t-value p-value sign. intercept 1.330 0.159 3357 8.375 < .0001 *** onset sCC 0.529 0.122 3357 4.327 < .0001 *** coda (k) 0.622 0.112 3357 5.574 < .0001 *** interaction sCC_ (k) 0.505 0.204 3357 2.478 < .05 * * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant 120 As we can see, onset sCC, when pitted against onsets without extra-syllabic [ ], shows a main effect towards vowel change, and it occurs in an interaction with an effect towards vowel change, too. The interaction in this model is actually nothing but the combination of the two main effects. Note here, though, that the interaction in the above table is just about on a significant level and did not emerge in a classification model. Therefore, all potential fixed effects, as visible in the above mosaic diagrams, were examined by a second mixed effects model. This time it was fitted to the data with both constituents as single fixed effects (with all variants) but without interactions, and with subject as random effect. This model confirmed the partial effects as shown in the mosaic plots (post-hoc ANOVA, onset: F(2,3353) = 15.318, p < .0001, coda: F(5,3353) = 30.644, p < .0001). All single main effects are summarised below. Table 13: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda (all variants) as predictors and the types of past-participle-sequence as response variable in 2_E_L1a constituent -coef s-value df t-value p-value sign. intercept -0.196 0.283 3353 -0.694 = 0.4879 ---onsets sC 0.201 0.145 3353 1.379 = 0.1680 ---- sCC 0.998 0.157 3353 6.348 < .0001 *** codas g,k 0.931 0.238 3353 3.914 < .0001 *** m,n 1.748 0.246 3353 7.101 < .0001 *** 2.438 0.241 3353 10.110 < .0001 *** 1.887 0.234 3353 8.065 < .0001 *** t 1.479 0.236 3353 6.253 < .0001 *** * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant As we can see, the model produced 7 coefficients, 6 of which are highly significant. For coda, all positive coefficient values indicate a bias towards vowel change. Apart from onset structure sC, the presence of all other constituent variants in the above table is significant for the choice between regular and vowel-changing past-participle-sequences. Thus, again it is the most complex onset that seems to be prototypical for vowel-changing verb inflection. A classification analysis revealed that the overall past-participlesequence can be predicted by the two inflections that the participants provided. In other words, there is a good degree of consistency between the first and the second response. The resulting tree is given in figure 17. 121 Category % n no vowel change 1 9,71 6 90 vow el change 8 0,29 2810 Total (100,00) 3500 Node 0 Category % n no vow el change 9 8,40 6 77 vow el change 1 ,60 1 1 Total (19,66) 688 N ode 2 Category % n no vow el change 8 ,33 1 vow el chang e 9 1,67 1 1 Total (0,34) 12 N ode 4 Category % n no vow el change 1 00,00 676 vow el change 0 ,00 0 Total (19,31) 676 N ode 3 Category % n no vow el chang e 0 ,46 1 3 vow el change 9 9,54 2799 Total (80,34) 2812 N ode 1 participle I mprovement =0,3030 past vo wel pasttense Improvement =0,0057 ae-ablaut; a-ablaut others; regular others; ae-ae; o: -o: ; i-i; a-a; a-ae; infinitive v owel Figure 17: Classification analysis showing the interaction of past tense and participle responses with the resulting past-participle-sequence in 2_E_L1a As we can see, participle formation is the strongest predictor for the distribution of the types of past-participle-sequences. In Node 1, we can see that those participles featuring vowel change almost always show vowel change in the overall sequence, too (99.54%). In contrast, if the participle just repeats the past tense vowel, that is, if it is regular, the proportion for non-vowel-change responses is at 98.4%. Note that past tense is a significant, albeit less strong predictor, too (second split). In Node 3 we see that those responses that show regular past tense and a past tense vowel in the participle quite logically result in overall responses without vowel change. In sum, the tree illustrates a very good degree of consistency between first and second response, with an amount of correctly predicted data of 99.4%. How consistent, though, are the vowel change patterns that emerged? Table 14 summarises their frequencies first. Table 14: Frequencies for types of vowel change in the overall past-participlesequences in 2_E_L1a examples for existing triplets frequency percent novel vowel change 1974 56.4 % no vowel change 690 19.7 % pattern 1 sing - sang - sung [ - - ] 393 11.2 % pattern 2 cling - clung - clung [ - - ] 225 6.4 % pattern 7 sit - sat - sat [ - - ] 180 5.1 % others 38 1.1 % 3500 100.0 % 122 As we can see, the great majority of all vowel changes produced does not reflect any existing patterns (‘novel vowel change’, 56.4%). Among the existing patterns, three-stage [ - - ] is clearly the predominant one (11.2%). Thus, there is evidence that if English L1 speakers construct vowelchanging past-participle-sequences from a given input form, the majority of the patterns is not in accordance with the apophonic path. Let us now have a closer look at the consistency of the two inflections that the English L1 speakers produced. Table 15 compares the frequencies for the two responses. Table 15: Frequencies for types of past tense and participle formation in 2_E_L1a frequency past tense frequency participle no vowel change 1283 no vowel change 838 [ ] [ ] 1083 [ ] [ ] / [ ] 577 [ ] [ ] 748 [ ] [ ] 214 others 386 others 1871 3500 3500 In the above table we can see that 1083 past tenses show [ ]-ablaut, most of which were triggered by nonce infinitive sprit. Note, too, that from these 1083 vowel-changing past tenses, only around 50% (577) finish the vowel change series modelled upon existing patterns. In a similar vein, past tenses starting with vowel change [ ]-[ ] end up with unorthodox pastparticiple pairs (‘others’) in more than 50% of the cases. These participles neither correspond to regular inflection nor to a recognisable vowel change pattern. Thus, like with German L1 speakers, faithfulness to apophonic vowel changes suffers greatly in the second response. On the other hand, a number of speakers first opted for regular inflection for past tense (1283 cases of regular past tenses) but then apparently altered their strategy midway and produced vowel changes in the participle (only 838 cases of regular participles). Overall, this results in a rather peculiar situation for the English L1 speakers. Despite a generally good degree of consistency as far as the distinction between regular and vowel changing overall series is concerned (cf. figure 7), the vowel change patterns themselves show remarkable inconsistencies. Although the English L1 speakers produced an astounding majority of vowel changing patterns (80.3%), amongst these more than 50% are actually inventions, mostly due to unorthodox participle formations. Compare this situation with the results of the English L1 speakers in experiment 1. There, too, the proportion of vowel change was remarkably high (61.5%), but, at the same time, the rate for unorthodox creations was relatively high, too (12.5%). Let us summarise all results for this experiment in (44). 123 (44) (a) English L1 speakers produced infinitive-participle-sequences with vowel change in more than 80% of all cases (b) amongst the vowel changes produced, around 56% are novel patterns (c) the most frequent vowel change occurs with a three-stage pattern [ - - ] (d) the distinction between infinitive-participle-sequences with and without vowel change depends on the presence of the schema [sCC_ (k)] (e) the consistency between both responses is very high with respect to the distinction between regular and vowel-changing overall series (f) the consistency between both inflections is very low amongst the vowel changes To sum up, based on the abundance of ungrammatical vowel changes and the low degree of consistency amongst the vowel change patterns, the English L1 data do not seem to confirm the hypothesis of an input-outputoriented apophonic path governing irregular verb morphology. 4.2.4 Experiment 2_E_L1b (Native speakers of English) Experiment 2_E_L1b is the reverse of experiment 2_E_L1a and contains the results of 26 native speakers of English, all undergraduate students from Dublin City University and Trinity College Dublin. They were exposed to participle nonces and provided a past tense form (Have you ever smunk? Yesterday I saw Sam; he ____ the whole evening.) and an infinitive (I think, Sam likes to____ .). Table 16 summarises the frequencies of overall responses with and without vowel change. Table 16: Frequencies for the types of past-infinitive-sequence with or without vowel change in 2_E_L1b type of past-infinitive-sequence frequency percent without vowel change 239 38.4 % with vowel change 384 61.6 % 623 100.0 % As we can see, vowel-changing responses are clearly in the absolute majority (61.1%), albeit considerably lower than in the previous experiment (80.3%). And as in experiment 2_G_L1b, the high amount of responses without vowel change (past tenses and infinitives that simply repeated the given participle vowel) is surprising, since the given nonces participles again showed clear and overtly marked signs of irregularity. Consider figure 18 now. 124 onset by past-infinitive-sequence onset past-infinitive-sequence CC sC sCC no vowel change vowel change coda by past-infinitive-sequence coda past-infinitive-sequence fricative k,g m,n ng ngk t no vowel change vowel change 2 = 2.3475 (Pearson), df = 2, p > .05 2 = 33.4625 (Pearson), df = 5, p < .0001 Figure 18: Mosaic plots for English verbs cross-classified by past-infinitive-sequence and the two constituents onset and coda in 2_E_L1b Again, the asymmetries for the onset variants are rather small. For coda, however, the different effects across the constituents are much clearer. Note again a clear effect of [ ] and [ ] towards vowel change, and a bias for the fricative and coda [t] towards avoiding vowel change. When onset was re-coded into binary ‘double’ and ‘triple’ consonant clusters, and coda into ‘velar nasals’ against all other variants, a mixed effects model, with subject as random effects and onset and coda as interacting fixed effects yielded one significant interaction. All coefficients are reported in table 17. Table 17: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda (pooled) as interacting predictors and the types of past-infinitive-sequence as response variable in 2_E_L1b -coef s-value df t-value p-value sign. intercept 0.315 0.434 594 0.726 = .4679 ---onset sCC -0.436 0.284 594 -1.537 = .1249 ---codas ( ) 1.002 0.272 594 3.680 < .0005 *** interaction sCC _ ( ) 0.926 0.402 594 2.302 < .05 * * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant The triple and thus most complex onset shows a main fixed effect towards overall past-infinitive-sequences without vowel change (negative coefficient), but does not come out as a significant predictor. Thus, the distinc- 125 tion between C(C) and sCC onset structures does not play an important role. In contrast, the main effect for coda [ ( )] shows a significant tendency towards vowel change (positive coefficient). And the interaction, being the sum of both main effects, illustrates a tendency towards vowel change. In order to explore the significance of all onset and coda structures, a second mixed effects model, now disregarding interactions, was fitted to the data. This model contained all variants of onset and coda as fixed effects and subject as random effects. First of all, a post-hoc ANOVA showed a significant partial effect of coda (F(5,590) = 10.418, p < .0001) and thus confirmed the impression of the mosaic diagrams above. Second, the model again yielded significant fixed effects for both engma codas ([ ]: = 2.160, t(590) = 3.970, p < .0001) and [ ]: = 1.604, t(590) = 3.061, p < .005). Thus, in sum we can state that the distinction between past-infinitive-sequences with or without vowel change depends on the presence of codas with a velar nasal, and, based on the significant interaction in table17, in particular on the schema [ CC _ _ ( )]. As far as the relationship between the first and the second response is concerned, a 2 -test suggested a strong degree of dependence, ( 2 = 503.7163 (Pearson), df = 15, p < .00001), and a classification analysis yielded both past tense and infinitive inflection as significant predictors for the resulting overall past-participle-sequence. The tree is given in figure 19. Category % n vo wel c hange 6 1,64 384 no v ow el change 3 8,36 239 Total (100,00) 623 Node 0 Category % n vowelchange 8,51 20 novowelchange 91,49 215 Total (37,72) 235 N ode 2 Category % n vow el change 0 ,00 0 no vow el change 1 00,00 212 Total (34,03) 212 Node 6 Category % n vow el change 8 6,96 2 0 no vow el change 1 3,04 3 Total (3,69) 23 Node 5 Category % n vow el change 9 3,81 364 no vow el change 6 ,19 2 4 Total (62,28) 388 Node 1 Category % n vow el change 7 1,08 59 no vow el change 2 8,92 2 4 Total (13,32) 83 Node 4 Category % n vow el change 1 00,00 305 no vow el change 0 ,00 0 Total (48,96) 305 Node 3 vo wel change pattern past tense f o rmatio n Improvement =0,3419 no vo wel change infinitive for mation Improvement=0,0504 participle v owel pattern 2; pattern 9; 14 patter n 1,12; others; pattern 2,8,9,10,11 infinitive for mation I mprovement =0,0175 14; participle vow el pattern 1,7; pattern 5; pattern 2; pattern 8,10; pattern 9 Figure 19: Classification analysis showing the interaction of past tense and infinitive responses with the resulting past-infinitive-sequence in 2_E_L1b 126 Here we can see that past tense formation, the first response, is the best predictor for the distinction between the two types of past-infinitivesequence (first split). In Node 2 we can see that if the past tense shows no vowel change, 91.49% of the overall past-infinitive-sequence do not show vowel change either. Node 3 and 6 illustrate that, quite logically, vowelchanging past tenses and infinitives produce - taken together - vowelchanging overall series in 100% of all cases. Overall, the classification analysis confirms a high degree of consistency between the two responses, with an amount of correctly classified data of around 95%. Let us have a closer look at the different vowel changing responses that English L1 speakers produced in a backward fashion (table 18). Table 18: Frequencies for types of overall past-infinitive-sequences in 2_E_L1b examples for existing triplets frequency percent no vowel change 239 38.4 % pattern 1 sing - sang - sung [ - - ] 154 24.7 % novel vowel change 132 21.2 % pattern 2 cling - clung - clung [ - - ] 70 11.2 % pattern 9 hang - hung - hung [ - - ] 25 4.0 % pattern 8 cut - cut - cut [ - - ] 3 .5 % 623 100.0 % Dealing with the reverse direction of inflections, English native speakers produced around 60% vowel changes in total, amongst which around 20% do not match existing patterns. We can also see that the most frequent grammatical vowel change pattern used is the three-stage [ - - ] (24.7%), even outnumbering the novel vowel changes. Note, however, that this time an existing three-stage pattern outranks the novel creations. Let us now have a closer look at the consistency of the two inflections that the participants produced. Table 19 summarises all pairs that occurred. Recall that participants first provided a past tense and then an infinitive. Table 19: Contingency table of past tense by infinitive formation in 2_E_L1b past tense patterns regular others [ - ] [ - ] infinitive patterns others 6* 12 2* 5* 25 [ - ] 3* 16* 13* 25 57 [ - ] 1* 5* 3 9 [ - ] 1* 20* 21 [ - ] 14* 16* 41* 70 141 [ - ] 5* 88 1* 94 regular 212 5* 37* 22* 276 235 56 206 126 623 127 First note that 235 observations show regular backwards past tense formations, while 276 cases feature no vowel change for the infinitive. Thus, a number of participants must have started out employing vowel change first and then switched to regular inflections along the way. Moreover, as we can see, regular past tenses tend to pair up with regular infinitives in around 90% of all cases (212 out of 235). And past tenses in [ ] result in corresponding grammatical infinitives in 98 out of 126 cases (around 78%). In contrast, from the 206 [ ]-past tenses only 88 produce a grammatical vowel change series (42.7%). All results for English L1 speakers dealing with participle nonces are summarised in (45). (45) (a) English L1 speakers produced past-infinitive-sequences with vowel change in almost 62% (b) amongst the vowel changes produced, around 20% are ungrammatical (c) the most frequent vowel change pattern is the three-stage pattern [ - - ] (d) the distinction between infinitive-participle-sequences with and without vowel change depends on codas [ ] and [ ] and the schema [ CC _ _ ( )] (e) the consistency between both inflections is high among regular overall series (f) the consistency between both inflections is very low amongst the vowel changes In sum, the present data point towards an irreversible input-outputorientation of the universal path. First, in 38.4% of the cases participants chose to repeat an overtly marked irregular participle and treat it as if it were regular. Second, among the vowel change patterns that English L1 speakers constructed backwards, more than 20% do not mirror existing patterns. And while the degree of consistency is good for regular pastinfinitive sequences, it is very low among the vowel changes. 4.2.5 Experiment 2_E_L2a (German-English interlanguage speakers) Experiment 2_E_L2a contains the results of 36 German-English interlanguage speakers, all undergraduate students from the Universities of Siegen and Marburg. Like the English L1 speakers, they provided two responses to each test item, a past tense form (Sam likes to smink. Yesterday he ____.) and a participle (Have you ever ____? ). First, the resulting past-participlesequence will be analysed. Afterwards, the consistency of both inflections will be dealt with in detail. Table 20 summarises the frequencies for the two types of past-participle-sequence. 128 Table 20: Frequencies for types of past-participle-sequence with or without vowel change in 2_E_L2a types of past-participle-sequence frequency percent without vowel change 297 33.0 % with vowel change 603 67.0 % 900 100.0 % As we can see, vowel-changing responses are clearly in the majority (67%). Consider figure 20 now. There, the two types of past-participle-sequence are plotted against onset and coda. onset by past-participle-sequence onset past-participle-sequence CC sC sCC no vowel change vowel change coda by past-participle-sequence coda past-participle-sequence fricative g,k m,n ng ngk t no vowel change vowel change 2 = 15.5963 (Pearson), df = 2, p < 0.001 2 = 89.2584 (Pearson), df = 5, p < 0.001 Figure 20: Mosaic plots for interlanguage verbs cross-classified by past-participle-sequence and the two constituents onset and coda in 2_E_L2a There are visible asymmetries both of onset and coda across the type of past-participle-sequence. In particular, codas [ ], [ ] and [ ] seem to affect past-participle-sequences with vowel change. As far as potential interactions of the constituents are concerned, a mixed effects model via PQL with subject as random effects and onset and coda (pooled into binary variants) as interacting effects revealed no significant interactions of onset and coda. Only the interaction [sCC _ _ ( )] approached significance ( = 0.0635, t(861) = 1.882, p = 0.0602). A final model, with all onset and coda variants as fixed non-interacting effects was significant both for onset (post-hoc ANOVA F(2,857) = 7.596, p < .0001) and coda (F(5,857) = 17.541, p < .0001). The following main single effects emerged (table 21). 129 Table 21: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda as predictors and the types of past-participle-sequence as response variable in 2_E_L2a constituent -coef s-value df t-value p-value sign. intercept -1.671 0.462 857 -3.613 < .0005 *** onsets sC 0.112 0.237 857 0.473 = .6364 ---- sCC 0.773 0.252 857 3.068 < .005 *** codas g, k 1.057 0.431 857 2.452 < .05 ** m, n 2.037 0.437 857 4.663 < .0001 *** 3.021 0.433 857 6.973 < .0001 *** 2.447 0.423 857 5.784 < .0001 *** t 2.210 0.431 857 5.133 < .0001 *** * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant The coefficients in the above table can be interpreted as follows: For onset, the negative value of [ ], mapped onto the intercept, indicates the variant’s bias towards overall sequences without vowel change. For coda, all high positive values (especially the velar nasal) illustrate a bias towards vowel change. Note that among the 8 coefficients, only the onset effect for sC is not significant. A classification analysis suggested a good degree of consistency between the two responses, with participle formation being a significant predictor for the type of past-participle-sequence. The tree is given in figure 21. Category % n vow el change 6 7,00 603 no v owel chang e 33,00 297 Total (100,00) 900 Node 0 Category % n vow el change 1 00,00 603 no v owel change 0 ,00 0 Total (67,00) 603 Node 2 Category % n vow el change 0 ,00 0 no v owel change 1 00,00 297 Total (33,00) 297 Node 1 vo wel change pattern participle Improvement =0,4422 pattern 1; pattern 2,8,9,10; pattern 3; pattern 7; others; pattern 6,13,14,15 past tense vo wel + ed Figure 21: Classification analysis showing the interaction of past tense and participle inflection with the resulting past-participle sequence in 2_E_L2a As we can see, the model produced two perfectly complementary splits according to participle formation. While Node 1 illustrates that from all the regular participles (past tense vowel plus -ed), 100% show no vowel change in the overall response, Node 2 shows the opposite behaviour. All partici- 130 ples showing vowel change patterns result in vowel changes in the overall sequence, too (100%, Node 2). The predictions of this model are very accurate, too, with a risk estimate of only 0.003. This means that more than 99% of the data were classified correctly. Thus, there is a very strong relationship between the way participants distinguished between past tenses and participles with or without vowel change. Consider table 22 now, which summarises the frequencies for all vowel changes as produced by German- English interlanguage speakers. Table 22: Frequencies for types of vowel change in the overall past-participlesequence in 2_E_L2a examples for existing triplets frequency percent novel vowel change 118 13.1 % no vowel change 297 33.0 % pattern 1 sing - sang - sung [ - - ] 376 41.8 % pattern 2 cling - clung - clung [ - - ] 24 2.7 % pattern 7 sit - sat - sat [ - - ] 62 6.9 % others 23 2.5 % 900 100.0 % Apparently, German-English interlanguage speakers show a clear preference for one particular vowel change pattern (41.8%), namely [ - - ]. At the same time, L2 speakers are much less prone to creating novel and ungrammatical patterns than English L1 speakers (13.1% inventions for L2 compared to 56.4% for L1). Let us now have a closer look at the consistency of the two inflections that the English L2 speakers produced. Recall that they were first prompted to fill in a past tense form and then the participle to a given infinitive nonce. Table 23 summarises the frequencies of the two responses. Table 23: Frequencies for types of past tense and participle formation in 2_E_L2a frequency past tense frequency participle no vowel change 300 no vowel change 297 [ ] [ ] 504 [ ] [ ] / [ ] 445 [ ] [ ] 29 [ ] [ ] 24 others 67 others 134 900 900 English L2 speakers produced an astounding majority of [ - ] past tenses (504 out of 900), mostly triggered by test item spling. In comparison, past tenses without vowel changes occur in only a third of all cases (300 out of 900). As far as participles are concerned, 445 out of 900 responses follow 131 the [ - ] pattern. Overall, table 24 illustrates that there are only minor inconsistencies between first and second response. As in all the previous experiments, some vowel-changing past tense inflections end up as ‘others’, but in general the degree of consistency is remarkably high for German-English interlanguage speakers. All results for the interlanguage speakers dealing with given infinitives are summarised in (46). (46) (a) German-English interlanguage speakers produced past-participlesequences with vowel change in 67% (b) amongst the vowel changes produced, around 13% are ungrammatical (c) the most frequent vowel change pattern is by far the three-stage pattern [ - - ] (d) the distinction between infinitive-participle sequences with and without vowel change depends on onset and coda, and for coda it depends on all its variants (e) the consistency between both inflections is high among regular overall series (f) the consistency between both inflections is high amongst the vowel changes In sum, the interlanguage data from experiment 2_E_L2a point towards the presence of an input-output-oriented apophonic path. First, the learners produced vowel change patterns in 67% of all cases. Second, among the vowel change patterns that the learners produced, only around 13% do not mirror existing patterns. And the degree of consistency is very high both among regular and among vowel-changing past-participle-sequences. Let us now have a closer look at the differences between how English L1 speakers and the interlanguage speakers produced overall past-participlesequences with and without vowel change. Consider figure 22. It plots the types of past-participle-sequence against the types of speaker. type of speaker by past-participle-sequence type of speaker past-participle-sequence L1 speaker L2 speaker no vowel change vowel change 2 = 73.3107 (Yates’ corrected), df = 1, p < .0001 Figure 22: Mosaic plot for English verbs cross-classified by past-participle-sequence and type of speaker 132 The horizontal asymmetry suggests that the type of past-participlesequence is not independent from the type of speaker. And a mixed effects model, with subjects as random effects and type of speaker as fixed effect confirmed a significant partial effect (post-hoc ANOVA F(1,4259) = 28.834, p < .0001). In a mixed effects model with subjects as random effects and both the syllable constituents and the type of speaker as fixed effects, a significant main effect emerged for the type of speaker (interlanguage speakers: = -0.578, df = 4252, t = -5.529, p < .0001). In addition to that, a significant interaction emerged for the interlanguage speakers in combination with coda [ ] ( = 0.992, df = 4245, t = 2.047, p < .05), which was confirmed by a classification analysis. Thus we can say that, in general, the way past-participle-sequences were created is group-specific, and that the way interlanguage speakers treated coda [ ] is significantly different from how English L1 speakers dealt with this coda variant. 4.2.6 Experiment 2_E_L2b (German-English interlanguage speakers) In this section we turn to the English L2 speakers that dealt with given nonce-participles featuring the stem vowel [ ]. This group had to provide inflections in a reverse direction, first providing past tenses (Have you ever smunk? Yesterday I saw Sam; he ______ the whole evening.) as well as an infinitive inflection (I think, Sam likes to _____). The group consisted of 56 German learners of English, all undergraduate students from the Universities of Siegen and Marburg. Table 24 summarises the frequencies for the two types of past-infinitive-sequence. Table 24: Frequencies for the types of past-infinitive-sequence with or without vowel change in 2_E_L2b types of past-infinitive-sequence frequency percent without vowel change 422 30.1 % with vowel change 978 69.9 % 1400 100.0 % Similar to experiment 2_E_L2a, we get more than twice as many pastinfinitive-sequences with vowel change than without vowel change (69.9% compared to 30.1%). This is in fact the first time that the majority of participants treat the given participles, again clearly marked as irregulars, as expected. Consider figure 23 now, where the two types of past-infinitivesequence are plotted against onset and coda. 133 onset by past-infinitive-sequence onset past-infinitive-sequence CC sC sCC no vowelchange vowel change coda by past-infinitive-sequence coda past-infinitive-sequence fricative g,k m,n ng ngk t no vowelchange vowel change 2= 1.3316 (Pearson), df = 2, p = 0.5139 2 = 25.9973 (Pearson), df = 5, p < .001 Figure 23: Mosaic plots for interlanguage verbs cross-classified by past-infinitive-sequences and the two constituents onset and in 2_E_L2b As we can see in figure 23, we find pronounced asymmetries only for coda variants. Interestingly, this time codas [ " , ] and [ ] show the strongest partial effects on past-infinitive-sequences with vowel change. A mixed effects model with onset and coda as interacting effects and subjects as random effects yielded no significant interactions. The final model with both constituents as single fixed effects was significant for coda only (posthoc ANOVA F(5,1337) = 14.239, p < .0001), and showed the following fixed effects (table 25). Table 25: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with onset and coda as predictors and the types of past-infinitive-sequence as response variable in 2_E_L2b constituent -coef s-value df t-value p-value sign. intercept -0.080 0.543 1337 -0.147 = .8832 ---codas g, k 1.490 0.402 1337 3.712 < .0001 *** m, n 2.785 0.401 1337 6.947 < .0001 *** 2.279 0.389 1337 5.865 < .0001 *** 2.031 0.380 1337 5.346 < .0001 *** t 2.760 0.428 1337 6.450 < .0001 *** * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant According to the mixed effects model, all coda structures, except for the fricative, influence the choice between responses with or without vowel changes significantly. A negative coefficient value illustrates an effect to- 134 wards avoiding vowel change, as can be seen for the fricative (mapped onto the intercept). The high positive coefficient values of [m,n], the velar nasals, and [t] show a pronounced bias towards vowel changes. A 2 -test ( 2 = 1282.830 (Pearson), df = 12, p < .00001) and a classification analysis suggested a strong degree of dependence between first and second response, with inflection for past tense (the first response) as a significant predictor for the resulting overall past-infinitive-sequence. The tree is given in figure 24. Category % n vow el change 6 9,86 9 78 no v owel change 3 0,14 4 22 Total (100,00) 1400 Node 0 Category % n vow el chang e 1 ,43 6 no vow el chang e 9 8,57 4 14 Total (30,00) 420 Node 2 Category % n vow el change 9 9,18 9 72 no vo wel change 0 ,82 8 Total (70,00) 980 Node 1 vo wel change pattern past tense Improvement =0,4014 no vo wel change pattern 1,12; pattern 2,8,9,10,11; others Figure 24: Classification analysis showing the interaction of past tense inflection with the resulting past-infinitive-sequence in 2_E_L2b In Node 1, we can see that from 980 responses showing vowel changing past tenses (patterns 1 to 12), almost all cases result in an overall vowel change pattern (99.18%). The opposite holds for all those responses not showing any vowel-changing past tense (Node 2). From the 420 past tense responses without vowel change, 98.57% have no overall vowel change either. The above tree suggests a very high degree of consistency between first and second response as far as the distinction between overall series with and without vowel changes is concerned. Table 26 summarises the frequencies for all vowel change patterns. Table 26: Frequencies of types of vowel change in the overall past-infinitive-sequence in 2_E_L2b examples for existing triplets frequency percent no vowel change 422 30.1 % pattern 1 sing - sang - sung [ - - ] 470 33.6 % novel vowel change 90 6.4 % pattern 2 cling - clung - clung [ - - ] 85 6.1 % pattern 9 hang - hung - hung [ - - ] 107 7.6 % pattern 8 cut - cut - cut [ - - ] 123 8.8 % pattern 10 sit - sat - sat [ - - ] 3 .2 % pattern 12 100 7.1 % 1400 100.0 % 135 In table 26 we can see that, similar to the L1 speakers, the interlanguage speakers favour the [ - - ] pattern (33.6%), this time outranking both regular and novel patterns. In contrast to English L1, though, interlanguage speakers prove to be less inventive when using vowel change (only 6.4% novel vowel changes). This means that interlanguage speakers mirrored existing patterns in the vowel changes they used with a high degree of faithfulness. Let us now have a closer look at the consistency of the two inflections that the participants produced (table 27). Table 27: Frequencies for types of past tense and infinitive formation in 2_E_L2b past tense frequency infinitive frequency no vowel change 420 no vowel change 610 [ ] [ ] 604 [ ] [ ] 557 [ ] [ ] 337 [ ] / [ ] / [ ] [ ] 161 others 39 others 72 1400 1400 The preferred vowel changing backformation for interlanguage speakers goes, similar to L1 speakers, from the given [ ]-participles to [ ]-past tenses, and from these [ ]-past tenses to [ ]-infinitives. The participle triggering most of these responses was trung. Compare the number of regular past tenses (420) and regular infinitives (610), and note that a number of participants must have started out with vowel change patterns in the past tense and then switched to regular infinitives along the way. However, amongst the vowel changes the degree of coherence is relatively good, but both patterns show a tendency to end up in regular infinitives (second response). There is not much mixing of patterns, though, and the inconsistencies arise mainly due to the switching from irregular past tenses to regular infinitives. Hence, all the resulting three-step vowel changes that actually occurred are in fact rather consistent and, above all, in line with the apophonic path. All results for the interlanguage speakers dealing with given participles are summarised in (47). (47) (a) German-English interlanguage speakers produced past-infinitivesequences with vowel change in almost 70% of all cases (b) amongst the vowel changes produced, only 6.4% are ungrammatical (c) the most frequent vowel change pattern is by far the three-stage pattern [ - - ] (d) the distinction between infinitive-participle-sequences with and without vowel change depends on all the coda variants except for the fricative (e) the consistency between both inflections is relatively high amongst the vowel changes 136 Let us now have a closer look at the differences between how English L1 speakers and interlanguage speakers produced overall past-infinitivesequences with and without vowel change. Consider figure 25. type of speaker by past-infinitive-sequence type of speaker past-infinitive-sequence L1 speaker L2 speaker no vowel change vowel change 2= 12.8714 (Yates’ corrected), df = 1, p < .0001 Figure 25: Mosaic plot for English verbs cross-classified by past-infinitive-sequence and type of speaker We can see only a small asymmetry between the distribution of L1 and interlanguage speakers across the types of past-infinitive-sequence. Accordingly, a mixed model with subjects as random effects and both constituents and the types of speaker as fixed effect showed no single fixed effect for the type of speaker. Only when interacting with coda, the interlanguage speakers are significant, in particular in combination with coda [ " , ] ( = 1.568, df = 1952, t = 2.047, p < .01) and coda [ ] ( = 2.085, df = 1952, t = 3.396, p < .001). Thus, we can state that the way past-infinitive sequences were created is in general not group-specific, only as far as the reaction to the two coda structures [ " , ] and [ ] is concerned. 4.3 Summary and discussion What do the above results tell us about the psycholinguistic validity of a universal apophonic path? First, the morphophonological properties that were responsible for vowel-changes in the six experiments will be discussed, followed by the discussion of whether the produced vowel-changes substantiate the notion of a universal apophonic path or not. When German L1 speakers were prompted to create vowel changes from a given infinitive, the great majority of all responses adhered to the apophonic path, as suggested by Ségéral and Scheer (1998), while only around 8% did not follow existing patterns. The test items that triggered most of the existing vowel changes had the shape [ C(C) _ _ ( )], which 137 closely matches the statistically significant prototype effects. As in experiment 1, the most complex onset as well as velar nasal codas seem to be prototypical for vowel changing overall sequences. The vowel changes that were applied in the two responses show a very high degree of consistency, with the non-apophonic [ - - ] pattern being the only exception. In contrast, the group of German speakers that had to derive infinitives from given past tenses showed unpredictable patterns in two respects: First, overtly marked irregular past tenses were inflected for infinitive using the same stem vowel. This is remarkable, since hardly any of the existing irregular German verbs show vowel congruity between infinitive and past tense. Second, the great majority of vowel changes that were created do not mirror existing patterns. The test items that triggered most of the backwards vowel changes had the shape [ C(C) _ ( )], which again mirrored the statistically significant prototype effects. Overall, the consistency among the vowel changes created backwards is rather low. In sum, for German speakers the apophonic path seems to be relevant only when vowel changes are applied in the conventional, input-output oriented fashion. This could be interpreted as evidence for the irreversible rule-like, or input-output oriented nature of the apophonic path in German. In comparison, the results for English L1 speakers are much less coherent. Participants provided a very high percentage of vowel-changing past tenses and participles to a given infinitive, which were mostly, as with German L1 speakers, triggered by the most complex onset as well as the velar nasal codas ([sCC _ ( )]). However, amongst those vowel changes more than 50% came out as novel creations. Hence, the degree of consistency between the two vowel change responses is rather low. It appears as though English L1 speakers did not adhere to an apophonic path in a conventional, input-output oriented fashion. When prompted to complement the apophonic path in a reverse direction, the willingness to provide vowel changes was still comparably high (62%, triggered by the schema [sCC _ _ ( )]), and this time only 20% of the vowel changes turned out to be ungrammatical. In fact, the most frequent vowel change series that occurred in the reverse experiment was the existing [ - - ] pattern. The consistency between all the vowel changes, though, is again rather low, mainly due to the fact that a number of participants first provided a regular past tense and then switched to a vowelchanging infinitive. Recall that the English L1 speakers reacted to an overtly marked irregular verb by simply repeating the stem vowel in 38.4% of all cases. In sum, the evidence coming from English native speakers is rather confusing, and in the light of the above inconsistencies, the notion of an apophonic path underlying English irregular verb morphology has to be seriously questioned. 138 As far as German-English interlanguage speakers are concerned, they are also rather willing to apply vowel change when asked to inflect a given infinitive (67%). For the interlanguage speakers, too, the most complex onset as well as velar nasal codas ([sCC _ _ ( )]) play a significant role in triggering overall vowel change series. Interestingly, only 13% of all the vowel changes produced are ungrammatical, and the degree of consistency between the two responses is rather high. When applying the apophonic path in a reverse order, the amount of vowel changes produced is similarly high (69.9%), and so is the degree of consistency between the two responses. Exceptions occur mainly due to switching from irregular past tenses to regular infinitives. Thus, for the interlanguage speakers, the apophonic path seems to guide both the conventional, rule-like application of vowel change and the reverse direction. When comparing the behaviour of the English L1 and the interlanguage speakers as regards the choice of vowel change against overall regular responses, we can summarise that the way both groups of speakers choose vowel changes is significantly different, especially with respect to coda [ ] in the given nonce-infinitives. As regards the willingness to provide vowel changes in a reverse order, both groups behave similarly overall, but with regard to codas [ " , ] and [ ] in the given participles interlanguage speakers differ significantly. Let us summarise the ablaut-triggering elements and in how far the different types of speakers adhered to the apophonic path across all six experiments (table 28). Table 28: Percentage of overall vowel change, non-apophonic vowel change, and the statistically significant effects of onset, nucleus, and coda in all three experiments experiment % of overall vowel change % of nonapophony statistically significant effects 1. 2_G_L1a 56.1 9.5 C(C) _ _ ( ) 2. 2_G_L1b 42.9 25.9 C(C) _ _ ( ) 3. 2_E_L1a 80.3 57.5 sCC _ _ 4. 2_E_L1b 61.6 21.2 sCC _ _ ( ) 5. 2_E_L2a 67.0 15.6 sCC _ _ ( ) 6. 2_E_L2b 69.9 6.4 CC _ _ There are three important insights illustrated in table 28. First, the proportion of overall vowel change seems to vary unsystematically across the experiments. What is more, even those experiments with the same directionality of inflection show a considerable amount of variation. Second, the amount of vowel changes running against the proposed apophonic path varies unsystematically, too. Recall that, given the inputoutput orientation of the apophonic path, the proportion of overall vowel 139 change should be lower in the reverse experiments. In fact, the English interlanguage data go against this expectation. At the same time, the proportion of non-apophonic vowel changes is expected to be higher in the reverse experiments. However, this is clearly not the case when we look at the English L1 and interlanguage data. Third, there is an astounding degree of morphological overlap between the nonces triggering vowel change across all six experiments. In sum, while the experiments seem to differ randomly with respect to the predictions made by the apophonic path, they show remarkable similarities with respect to the morphological properties of prototypical input and output schemas. This raises two important questions. First, if we regard the apophonic path as rule-like and thus irreversible, how did the participants manage to create all the non-random, predictable vowel changes from the given output forms? And second, do the inconsistencies among the vowel change patterns follow any systematic patterns and can they be accounted for? As regards the first question, the data clearly point towards the possibility of output-oriented schemas facilitating analogical inflections, as Bybee and Moder already suggested for English L1 speakers (Bybee and Moder 1983). And this seems to hold for German L1 speakers and German- English interlanguage speakers, too. As the data suggest, all three groups of speakers were able to generalise from a given output form to a possible input in analogy to existing verbs. All three groups of speakers showed some sort of tacit knowledge of morphophonological similarities within the paradigm of irregular verb formations, and they were able to extend this knowledge, via analogy, to new and unknown verbs in a bi-directional fashion. In order to answer the second question, let us review the inconsistencies between first and second response that emerged with all three groups of speakers. They mostly occur in the second response. In experiment 2_G_L1a, for instance, all vowel change patterns lose members in the second response (the participle); even regulars lose cases to unorthodox participles. In experiment 2_G_L1b, again it is the second response showing a disproportional amount of unorthodox creations. Thus, for German, the inconsistencies do not arise by randomly mixing vowel changes or by using vowels completely alien to the apophonic system, but by oscillating between regular and vowel-changing past tense formation strategies in the second response (the participle). One explanation for the unorthodox participles could be the participle’s morphological make-up. As Ségéral and Scheer (cf. 1998: 39-40) already pointed out, the relationship between the infinitive (or present tense) form and the past tense form is much closer than between a past tense and a participle. Present and past tense are much 140 more similar in shape, differing solely by stem vowel alternation, while the participle involves stem alteration and suffixation. In English L1 and L2 we can observe a similar kind of erosion from first to second response; all vowel change patterns as well as regular past tenses lose cases to unorthodox participles. While in English L1 this tendency is rather pronounced, in L2 it is much more moderate. Overall, when inflecting from a given input form, there seems to be a general tendency for the second response, although applying vowel change, to disregard the apophonic path. So the question arises why the faithfulness to existing patterns changes substantially from one response to the other. Moreover, and maybe even more importantly, do these unorthodox participles employ vowels that are completely alien to the ablaut system of the language, or do we find vowels that have simply been put into the wrong place by the participants? As it turns out, the percentage of vowels that is completely alien to the apophonic system is rather low across all groups of speakers, with the only exception of experiment 2_E_L1a. Here, 26.8% of the participles produced by the English native speakers to a given input form can neither be explained with reference to the given infinitive nor to the produced past tense. Thus, completely random vowel changes do occur in all groups of speakers, but they are generally rather rare in the data of experiment 2. Let us finally review the hypothesis that guided this experiment. It predicted that German-English interlanguage irregular verb morphology could be modelled as rule-like input-output processes that follow the universal apophonic path. In the light of the present evidence, this hypothesis cannot be upheld. In fact, the notion of an apophonic path seems neither necessary nor sufficient to explain the behavioural evidence in this experiment. Instead, simple analogical relationships by way of inputand outputoriented schemas seem to offer the most parsimonious model to fit the present data. 141 5 Experiment 3 This chapter provides a detailed account of experiment 3. First, the experimental design will be presented. Second, the results of three versions of this experiment, one with German L1 speakers and two with English L2 speakers will be discussed. Recall the hypothesis that provided the basis for this experiment. It is restated for convenience in (48). (48) Hypothesis 3 German-English interlanguage irregular verb morphology is co-determined by a verb’s semantic similarity to existing verbs. These similarities can work as attractors independent of phonological similarities to existing verbs. The hypothesis predicts that participants inflect nonce-verbs according to a given semantic context in which the nonce-verb is presented. Moreover, we would expect that the semantic context exerts semantic attraction (that is, primes an existing real verb along with the real verb’s past tense inflection) irrespective of the nonce-verb’s phonological similarity to the prime (real verb). If both phonological and semantic attraction were genuinely independent, then a given semantic context should be able to manipulate participants’ inflections of a nonce-verb even through a phonologically dissimilar prime. Recall that Ramscar’s (2002a) semantic effects always depended on phonological similarity (# _ ) between nonce-verb and prime, as can be formalised in so-called ‘four-part-analogies’, such as drink : drank : : frink : frank, blink : blinked : : frink : frinked). In the present experiment, however, the power of semantic contexts priming phonologically dissimilar irregular verbs was examined, too, like in ride : rode : : frink : frank. Note again that merely evoking a listed form through proportional analogy based on phonological similarity is ruled out in such a four-part-analogy. Instead, the analogal inference is expected to exceed mere phonological grounds and avail itself of the morphological property {past, irreg.}, as formalised in A : B {past irreg.} : : X : Y {past irreg.} . In experiment 3, German L1 and English L2 speakers were tested. There were, however, two versions of the English L2 experiments. The first one was a slightly modified replication of Ramscar’s oral interviews (2002a), in which 80 undergraduate students from the University of Siegen took part. Since only one data point was obtained per participant, the interviews resulted in a rather restricted pool of responses. Therefore, a second paperand-pencil experiment was conducted with 182 undergraduate students from the University of Siegen. 142 5.1 Experimental design There were basically two designs of experiment 3. The first one was a brief oral interview, in which participants were presented one out of four stories, featuring to frink for English and frinken for German as the test item for past tense inflection, along with three follow-up questions. The second design was a paper-and-pencil experiment in which the test item had to be inflected for past tense only. Both in the interview and in the written experiment, participants were given a story that primed an existing verb (either regular or irregular), and the participants’ task was to inflect the nonce-verb for past tense according to the semantic environment provided by the story. The oral interview was divided up into three phases. First, an experimenter informed the participants that they were about to take part in a linguistic experiment and that their intuition either as an L1 speakers or as an advanced learner was considered very helpful. Then, the participants were presented with a flashcard featuring a short text and one gap. The text prompted each participant to pluralise a given nonce-noun. This was supposed to introduce the basic design of the task and provide an item for practice. An example from the English and the German flash cards is given in (49). (49) (a) A single w u c t e r i u m can be very dangerous for humans. When they breed and multiply, a lot of _______________ can kill humans. (b) Ein kleines W u k t e r i u m kann für Menschen gefährlich sein. Wenn es sehr viele sind, können _______________ sogar tödlich sein. Participants had to provide an oral response, which was recorded in writing by the experimenter. Afterwards, participants were given one out of four different stories to read. Embedded in this story were the nonce-verb and a gap prompting the nonce’s inflection for past tense. After reading through the text on the flash card, each participant had to inflect the nonce verb for past tense once. Again, the experimenter recorded the oral response in writing. In a second step, each participant was asked three follow-up questions. An example for the English flashcards is given in (50). (50) Q1 What if any word did the highlighted word frink remind you of? Word: _____ / No word . Q2 Name a word that rhymes with frink as quickly as you can: _________ . Q3 Please assume that the use of frink in the little piece of text above is its correct usage. Would you normally associate frinking with ‘vodka’ and ‘the consumption of liquids’? Yes . No . 143 These questions were supposed to evaluate whether and how the participants were in fact primed by the contexts, and what existing verb the nonce had actually reminded them of. The interviews closed with these three questions. Recall that participants were given only one out of four different stories, which primed different real verbs through different semantic contexts. All four English and German stories (semantic contexts) are summarised in (51). The entire stories can be found in the appendix. (51) (a) story about drinking (English) and trinken (‘to drink’) (b) story about blinking (English) and hinken (‘to limp’) (c) story about riding (English) and reiten (‘to ride’) (d) story about meditating (English) and meditieren (‘to meditate’) As we can see, context (51a) is supposed to prime the irregular verb to drink / trinken, while context (51b) primes regular to blink / hinken. Note that these two contexts contain a prime that is phonologically similar to the nonceverb (frink-drink/ blink). Context (51c) was supposed to evoke analogies to the irregular verb to ride, and the final context suggests analogies to the regular verb to meditate. The latter two nonce-verbs obviously bear no phonological similarity to the contexts’ primes (frink-ride/ meditate). The expected priming scenarios for the English experiments are as follows (52). (52) nonce-verb to frink context 1 context 2 context 3 context 4 story about drinking story about blinking story about riding story about meditating four-part analogy drink : drank frink : frank four-part analogy blink : blinked frink : frinked four-part analogy ride : rode frink : frank four-part analogy meditate : meditated frink : frinked expected bias towards irregular expected bias towards regular expected bias towards irregular expected bias towards regular expected inflection based on semantic bias target target target target frank frinked frank frinked In the context about drinking, the test item frink would, if the hypothesis were accurate, have been identified as morphophonologically and semantically similar to drink and hence be inflected irregularly. Accordingly, the 144 target for this context would have been frank. Likewise, in a context featuring the prime blink (context 2), we would expect regular inflection due to the nonce’s semantic and phonological attraction to the regular prime blink (target frinked). Contexts 3 and 4 should result in different inflections. The ride-context should trigger irregular frank, whereas the meditate-context should result in regular frinked. Three of the four contexts used in the experiments (cf. 4a, b, d) are slightly altered versions of Ramscar’s contexts (Ramscar 2002: 56). Their priming quality had already been assessed by a so-called ‘latent semantic analysis’ (LSA). This is a procedure which calculates a contextual distribution for each word encountered in a corpus by counting the frequency with which it co-occurs with other words. [...] Thus LSA provided a simple, objective measure of whether these were contexts in which one given verb was more likely to occur than another. (54) The context priming the verb to ride (cf. 4c) was new, and its priming quality was tested orally with a number of native speakers, both German and English. Their responses confirmed that this new context was indeed associated with the prime (ride). As pointed out before, the rationale behind this additional context was Ramscar’s lack of one particular priming scenario, namely the phonologically dissimilar irregular prime (ride - frank / reiten - frank). Adding this context to the test design meant that all four possible combinations of morphophonological and semantic priming were exploited. This is outlined with regard to English in the following matrix (53). (53) expected analogical inflection irregular regular degree of morphophonological similarity similar context 1 drink -frank context 2 blink frinked dissimilar context 3 ridefrank context 4 meditate frinked Exploiting analogical inferences through phonologically dissimilar verbs has already been suggested by Huang and Pinker (2005), who had observed that, in order to test pure semantic analogies and their direct role in inflection, one would need to examine the effect of semantic similarity in cases where there is low and moderate phonological similarity between novel verbs and the existing verbs to which they are similar. By expanding comparisons to cases in which both phonological and semantic similarities are manipulated, one can see whether semantic similarity elicits a generalization gradient analogous to the generalization gradient already known to exist for phonological similarity. (1002) 145 The ride-context is supposed to serve exactly this purpose. It is, admittedly, a challenging hypothesis to expect pure semantic four-part-analogies in the complete absence of phonological similarities, but since Pinker and Huang have already opened this dimension on theoretical grounds, it seems only adequate to try and give this thesis some empirical underpinning. So what is the exact benefit of this new ride-context? Recall that we expect semantic attraction between frink and meditate to occur despite the clear lack of phonological similarity. Thus, according to the hypothesis, participants are expected to inflect the nonce regularly. Such a context sensitive response would, however, be identical to the non-context-sensitive default inflection (regular past tense), since both would result in frinked. Thus, it would be hard to differentiate between frinked as an instance of semantic analogy [frink meditate, meditate = regular, as a consequence inflect frink regularly] or the ordinary default [frink = unknown, thus by default inflect frink regularly]. In the ride-context, however, a context-sensitive response would be clearly discernable. As with meditate, the prime ride rules out phonology as an analogical attractor to frink. If despite this lack of morphophonological similarity, semantic attraction proved effective (frink ride, ride = irregular, as a consequence inflect frink irregularly), here a prime consistent response would clearly differ from the default, as it would yield the irregularly inflected target frank. The only possible explanation for the occurrence of the pair ride - frank would be genuine semantic priming through a phonologically dissimilar irregular verb. 5.2 Results First, the results of the two oral interviews (German L1 speakers, 3_G_L1_oral, and German-English interlanguage speakers, 3_E_L2_oral) are presented. Then follow the results of the paper-and-pencil experiment conducted with German-English interlanguage speakers (3_E_L2_written). 5.2.1 Experiment 3_G_L1_oral (Native speakers of German) Experiment 3_G_L1_oral contains the results of 90 native speakers of German, all undergraduate students at the University of Siegen. Consider figure 26 first, in which the inflections that the German participants produced are cross-classified by the four contexts in which the nonces were presented (excluding all unorthodox replies). 146 context by inflection context inflection hinken meditieren reiten trinken frank frinkte 2= 36.8889.8714 (Pearson), df = 3, p < .001 Figure 26: Mosaic plot for nonce-inflections cross-classified by type of inflection and contexts in which the nonce was presented in 3_G_L1_oral We can see in figure 26 that the two types of inflection (frinkte, frank) are not distributed evenly across the four contexts. Instead, there are distinct and significant asymmetries across all four semantic contexts, with context trinken (‘to drink’) showing a pronounced effect on irregular inflections (frank), which is in fact the expected direction of the effect. Note, though, that context reiten and trinken seem to differ significantly from each other. This difference would not be expected, since both contexts were supposed to trigger irregular inflection (frank). This was clearly not the case. Table 1 presents the absolute and relative frequencies of the inflections in the varying contexts. The percentages in the framed cells illustrate the expected, context sensitive inflection. Table 1: Frequencies for past tense inflections across all four contexts in 3_G_L1_oral frequency percent context 1 (target: frank) regular responses frinkte 2 6.7 % irregular responses frank 28 93.3 % context 2 (target: frinkte) regular responses frinkte 13 65.0 % irregular responses frank 7 35.0 % context 3 (target: frank) regular responses frinkte 13 65.0 % irregular responses frank 5 25.0 % context 4 (target: frinkte) regular responses frinkte 14 70.0 % irregular responses frank 4 20.0 % As we can see, contexts 1, 2 and 4 produced a majority of context sensitive responses. In context 1, for instance, participants were, as intended, clearly 147 biased towards irregular inflection (93.3%). For context 3, though, participants showed a deviating behaviour, producing a mere 25% of context sensitive responses. Note that in context 3 (reiten, ‘to ride’), we can see genuine semantic priming through a phonologically dissimilar irregular verb in only 25% of all cases, the lowest proportion of context sensitive responses across all four contexts. Consider table 2 now. It summarises the replies to question 1. Table 2: Frequencies for replies to question 1 in 3_G_L1_oral (What if any word did the highlighted word frinken remind you of? ) type of reply frequency percent reminded me of a rhyming word 43 47.8 % reminded me of the prime 40 44.4 % reminded me of no word at all 6 6.7 % reminded me of some word 1 1.1 % 90 100.0 % As we can see, only 44.4% of all participants stated that they were in fact reminded of - and thus primed for - the intended real verb (frink-trinken, for instance). This means that for almost half of all the responses, the assumed analogical attraction between nonce and prime failed to work. However, we can also see that 47.8% were at least reminded of a rhyming word. This means that, if we take together the percentages of those participants being reminded of either the prime or a rhyming word, in the great majority of all cases (more than 80%) phonological similarities seemed to govern the reply to question 1. Table 3 summarises the replies to question 2 for those contexts examining the role of phonological similarity. Since question 2 solely focussed on rhyming associations, context 3 and 4 (phonologically dissimilar primes) are naturally excluded from this summary. Table 3: Frequencies of replies to question 2 in contexts 1 and 2 in 3_G_L1_oral (Name a word that rhymes with frinken as quickly as you can.) type of response context 1 context 2 rhyme with prime 56.7 % 25.0 % rhyme with irregular word 36.7 % 65.0 % rhyme with regular word 6.7 % 10.0 % 100.0 % 100.0 % For context 1, the direct association of nonce and expected prime via rhyming worked in 56.7% of all cases, in which participants stated the prime as their first rhyming association (frink - trinken). For context 2, in contrast, only 25% of all responses show a direct association between nonce and prime (frink - hinken). Interestingly though, the majority of all the non- 148 prime rhymes in both contexts are irregular real verbs. Table 4 summarises the responses to question 3. Table 4: Frequencies for replies to question 3 in 3_G_L1_oral (Would you normally associate frinking with ‘vodka’ and ‘the consumption of liquids’? ) type of reply context 1 context 2 context 3 context 4 yes same meaning 43.3 % 10.0 % .0 % 5.0 % no different meaning 56.7 % 90.0 % 100.0 % 95.0 % 100.0 % 100.0 % 100.0 % 100.0 % What we can see is that contexts 2, 3, and 4 hardly ever, or not all, prompted participants to perceive the nonce-verb and the prime as semantically similar (no - different meaning in 90-100%). Only for context 1, 43.3% of the participants found such semantic similarity (yes - same meaning). This means that the great majority of the participants did not regard the nonce as a potential and suitable word in the context they were given. Overall, the replies to the follow-up questions challenge the assumed priming qualities of the four scenarios, even the one for context 1. The question remains whether, and how, phonological and also semantic similarity actually influenced participants’ past tense inflection in this experiment. In order to explore the influence of the four contexts on the inflectional behaviour and the follow-up questions in more detail, the data were analysed using simple logistic regression models and mixed effects regression models via PQL. A logistic regression revealed a general significance of the factor context ( 2 = 14.35 (Pearson), df = 3, p < .005, C = .9). The insignificance of the three follow-up questions was supported by a post-hoc fastbackward-variable-selection-procedure (cf. Lawless and Singhal 1978), which retained only the factor ‘context’ in the final logistic model. Note that the effect size of this model is considerable (C = .9), and that the amount of variance explained lies at 62%. With all non-significant factors eliminated from the final logistic model, the effect size decreased just slightly to C = .85. A post-hoc ANOVA of a mixed effects model via PQL, with the inflection of the nonce-verbs as the response variable and context and the three follow-up questions as fixed and subjects as random effects, confirmed the general significance of the factor context (post-hoc ANOVA F(3,74) = 5.065, p < .005) and produced the following main effects. 149 Table 5: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with context and the three follow-up questions as predictors and the inflection of the nonce-verb frinken as the response variable in 3_G_L1_oral factor levels -coef s-value df t-value p-value sign. intercept 15.755 4126.495 74 0.004 = .9970 ---contexts meditieren 1.812 1.121 74 1.615 = .1106 ---- reiten 1.387 0.944 74 1.470 = .1458 ---- trinken -5.690 1.700 74 -3.348 < .005 ** questions 1: none 0.078 4455.040 74 0.000 = 1.000 ---- 1: rhyme -16.267 4126.495 74 -0.003 = .9969 ---- 1: target -13.549 4126.495 74 -0.003 = .9974 ---- 2: rhyme reg. -0.999 1.855 74 -0.539 = .5917 ---- 2: rhyme tar. 0.294 1.095 74 0.268 = .7894 ---- 3: same 1.262 1.277 74 0.988 = .3263 ---- * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant As we can see, the significance of the factor context only holds for its variant trinken (‘drink’), with a pronounced negative coefficient, which in this case illustrates that the responses in the drink-context went into the expected direction (irregular inflection of the nonce). This is true even when all predictors were tested pair-wise. Thus, only the context featuring a phonologically similar irregular verb managed to influence the inflectional behaviour of the participants significantly and as expected. This, in turn, suggests that phonology and semantics do not seem to operate as individual attractors, but - if at all - in tandem. The question arises here, however, why we do not find a similar attraction with the phonologically similar prime hinken (‘to limp’). Recall that for context hinken, despite its phonological similarity to the nonce-verb, only 65% provided a context sensitive response. And in the above table of coefficients the effect for context hinken (mapped onto the intercept) is far from being significant. Note, too, that in the above table none of the follow-up questions seems to be a good predictor for the inflectional behaviour of the participants. Such a lack of significant dependencies between the follow-up questions and the provided inflection is a remarkable result. It either suggests that, methodologically, such follow-up questions fail to record the degree of phonological and semantic priming, or that the entire priming scenarios are much less predictable than assumed so far. Consider table 6 now, which gives a selection of examples for some of the inconsistencies between the inflection and the responses to the questions in the overall response patterns. 150 Table 6: Selected inconsistencies in the overall response patterns in 3_G_L1_oral partpicipant context prime target inflection reminded of rhyme (a) 1 trinken frank frank trinken stinken (b) 2 hinken frinkte frinkte hinken blinken (c) 3 reiten frank frank no word no word (d) 4 meditieren frinkte frinkte schlafen frinkte For participant (a), his inflection of the nonce (frank) matches with the response to question 1. The verb trinken (‘to drink’) was correctly associated. The response to question 2, though, which should have been trinken, too, and not stinken (‘to stink’), shows that the semantic attraction of the nonce frink did not go into the expected direction. Compare this pattern with participant (c), who inflected the nonce in a context sensitive fashion (frank, triggered by irregular reiten, ‘to ride’) but then stated being neither reminded of any other word (no word as response to question 1) nor being capable of providing a rhyme (question 2). These inconsistencies are another piece of evidence for a lack of dependency between the participants’ inflectional behaviour and their reaction to the follow-up question. Overall, experiment 3_G_L1_oral showed that when German L1 speakers were presented with nonce-verbs in varying semantic contexts, only one context influenced their inflectional behaviour significantly; this context primed an irregular phonologically similar verb (trinken). In comparison, the context priming a regular phonologically similar verb (hinken) managed to trigger context sensitive responses in only 65% of all cases and does not significantly influence the participants’ inflectional behaviour. Interestingly, the context priming an irregular phonologically dissimilar verb (reiten) failed entirely to serve as an analogical attractor, with merely 25% of context sensitive responses and an insignificant effect in the regression models. Moreover, the follow-up questions turned out to be inappropriate measures for the degree of priming quality. As far as the role of semantics is concerned, the data in the above experiment suggest that meaning-based analogies seem to work exclusively for phonologically similar irregular primes. 5.2.2 Experiment 3_E_ L2_oral (German-English interlanguage speakers) Experiment 3_E_L2_oral contains the results of 80 German-English interlanguage speakers, all undergraduate students from the University of Siegen. Figure 27 plots the nonce-verb inflection against the four contexts in which the nonces were presented (excluding all unorthodox replies). 151 context by inflection context inflection blink drink meditate ride frank frinked 2 = 3.5744 (Pearson), df = 3, p = .3112 Figure 27: Mosaic plot for nonce-inflections cross-classified by type of inflection and contexts in which the nonce was presented in 3_E_L2_oral We can see clear asymmetries for the distribution of the types of inflection across the four contexts, but according to a logistic regression model (posthoc ANOVA 2 = 3.37 (Pearson), df = 3, p = .3379) and a mixed effects model with the four contexts as fixed and subjects as random effects and the inflection of the nonce as response variable (F(3,60) = 1.053, p = 0.376), they are not significant. This is a first indication that for German-English interlanguage speakers the four contexts failed to serve as analogical attractors. Consider table 7 now, which shows the absolute and relative frequencies of the inflections by German-English interlanguage speakers across the four contexts. Again, the framed cells mark the expected, context sensitive inflections. Table 7: Frequencies for past tense inflections across all four contexts in 3_E_L2_oral frequency percent context 1 (target: frank) regular responses frinked 10 50.0 % irregular responses frank 8 40.0 % context 2 (target: frinked) regular responses frinked 7 35.0 % irregular responses frank 10 50.0 % context 3 (target: frank) regular responses frinked 8 40.0 % irregular responses frank 9 45.0 % context 4 (target: frinked) regular responses frinked 9 45.0 % irregular responses frank 3 40.0 % As we can see, the proportion of prime consistent responses is, while being distributed rather evenly, below chance in all four contexts. Context 1, for instance, elicited only 40% of prime consistent responses, and context 2 152 only 35%. Apparently, the phonological similarity between frink and drink / blink proofed to be almost entirely ineffective for German-English interlanguage speakers 30 . Table 8 summarises the replies to question 1 across all four contexts. Table 8: Frequencies for replies to question 1 in 3_E_L2_oral (What if any word did the highlighted word frink remind you of? ) type of reply frequency percent reminded me of no word at all 37 46.3 % reminded me of a rhyming word 32 40.0 % reminded me of the prime 8 10.0 % reminded me of some word 3 3.8 % 80 100.0 % The percentage of responses indicating correct priming is remarkably low (10%), which corresponds to the low proportion of context sensitive responses in table 7. Table 9 summarises the replies to question 2. Table 9: Frequencies for replies to question 2 in contexts 1 and 2 in 3_E_L2_oral (Name a word that rhymes with frink as quickly as you can.) type of reply context 1 context 2 rhyme with prime 55.0 % 5.0 % rhyme with irregular word 20.0 % 95.0 % rhyme with regular word 25.0 % .0 % 100.0 % 100.0 % In context 1, the majority of all responses show that the rhyme is actually the prime. For context 2, though, the great majority provided a non-prime irregular verb. Table 10 summarises the responses for question 3. Table 10: Frequencies for replies to question 3 in all four contexts in 3_E_L2_oral (Would you normally associate frinking with ‘vodka’ and ‘the consumption of liquids’? ) type of reply context 1 context 2 context 3 context 4 yes - same meaning 55.0 % 65.0 % .0 % .0 % no - different meaning 45.0 % 35.0 % 100.0 % 100.0 % 100.0 % 100.0 % 100.0 % 100.0 % 30 One might be tempted to infer that for interlanguage speakers the verb to blink might not be part of their active vocabulary, and thus they maybe did not know how to deal with it as a prime. However, even if interlanguage speakers had been unsure about the exact meaning of the verb to blink, they - as advanced learners - would certainly have treated this verb as a regular, since learners at this proficiency level can be assumed to be familiar with all common irregular verbs of English. 153 What we can see is that only in contexts 1 and 2, the majority of the participants seemed to perceive the nonce-verb and the prime (frinken-trinken) as semantically similar. On average, though, most of the participants did not regard the nonce as a potential and suitable word in the context they were presented. So, how do the three follow-up questions relate to the type of nonce-verb inflection in this experiment? Again, both a logistic and a mixed effects model were fitted to the data. The logistic model revealed a mild effect solely for the reply to question 1 ( 2 = 11.81 (Pearson), df = 2, p < .005), which, however, was not retained after a fast backward variable selection procedure, because it only approached significance. And although the predictive power of the above logistic model is fairly good (C = 0.873), the variance explained ranges around 50% only). A mixed effects model, however, with subjects as random effects, is also significant for question 1 and produced the following fixed effects (table 11). Table 11: Beta coefficients and significance values for the fixed effects of a generalised linear mixed model via PQL with context and the three follow-up questions as predictors and the inflection of the nonce-verb frink as the response variable in 3_E_L2_oral factor levels -coef s-value df t-value p-value sign. intercept 2.285 1.093 55 2.092 = .0411 ---contexts drink 0.365 1.369 55 0.267 = .7908 ---- meditate -0.778 1.269 55 -0.613 = .5423 ---- ride -1.598 1.150 55 -1.389 = .1703 ---questions 1: rhyme -2.742 0.861 55 -3.185 < .005 ** 1: target -4.580 1.559 55 -2.938 < .005 ** 2: rhyme reg. 17.465 2137.283 55 0.008 = .9935 ---- 2: rhyme tar. 1.027 1.519 55 0.676 = .5020 ---- 3: same -0.992 0.973 55 -1.019 = .3126 ---- * = p < .05, ** = p < .005, *** = p < .0005, ---- = not significant As we can see, only two answers to question 1 are significant predictors of the nonce-verb inflection. The fixed effects in the above table show that only for those participants that were actually reminded of - and hence primed for - the intended prime (1: target) or a rhyming verb (1: rhyme), the inflection was context sensitive and went into the expected direction (irregular). Interestingly, the effect of neither context is at a significant level. Even when tested pairwise, no effect emerged for any of contexts. Consider figure 28, which plots the inflection of the nonce-verb against the responses to the follow-up questions. 154 response 1 by inflection response 1 inflection none rhyme target frank frinked response 2 by inflection response 2 inflection irreg. reg. target frank frinked response 3 by inflection response 3 inflection different same frank frinked Figure 28: Mosaic plots for participants’ inflection cross-classified by all three responses in 3_E_L2_oral In figure 28 we can see large asymmetries in the distributions of the responses to all three follow-up questions. Post-hoc ANOVAs to three logistic regression models, however, show only for questions 1 ( 2 = 16.46 (Pearson), df = 3, p < .001) and 2 ( 2 = 19.25 (Pearson), df = 3, p < .001) there is a significant relationship between the questions and the inflection. Overall, it seems, as though for German-English interlanguage speakers the four semantic contexts did not significantly influence the way they handled the inflection of the nonce-verb. The majority of all past tense inflections was not context sensitive, and neither context is a significant predictor of the nonce-verb inflection. In sum, the above experiment provides evidence that neither phonological nor semantic similarities could serve as analogical attractors for interlanguage speakers. 5.2.3 Experiment 3_E_L2_written (German-English interlanguage speakers) Experiment 3_E_L2_written contains the results of 182 German-English interlanguage speakers, all undergraduate students from the University of Siegen. Figure 29 plots the two types of nonce-verb inflection against the four contexts in which the nonces were presented (excluding the responses coded as ‘others’). 155 context by inflection context inflection blink drink meditate ride frank frinked 2 = 8.3451 (Pearson), df = 3, p < .05 Figure 29: Mosaic plot for participants’ inflection cross-classified by contexts in which the nonce was presented in 3_E_L2_written We can see in figure 29 that for English L2 speakers going through the written experiment, the four contexts show significant asymmetries across all four contexts. Although contexts drink and ride do not show the expected majority of irregular and thus context-sensitive responses, they show an effect into the expected direction (i.e. towards irregular frank). Table 12 shows the frequencies of participants’ nonce-verb inflections across all four contexts (excluding unorthodox responses). Table 12: Frequencies for past tense inflections across all four contexts in 3_E_L2_written frequency percent context 1 (target: frank) regular responses frinked 36 48.6 % irregular responses frank 21 28.4 % context 2 (target: frinked) regular responses frinked 31 81.6 % irregular responses frank 5 13.20 % context 3 (target: frank) regular responses frinked 20 62.5 % irregular responses frank 10 31.3 % context 4 (target: frinked) regular responses frinked 29 76.3 % irregular responses frank 6 15.8 % As we can see in table 12, context 1, despite its bias towards prime consistent inflections, elicited only 28.4% of those, which is a surprisingly low proportion given that this is context with a phonologically similar prime. We can also see that context 2, in contrast, triggered more than 80% prime consistent responses. As with the German L1 speakers, the four different 156 contexts in the English L2 written experiment result in rather different degrees of context-sensitivity (28.4% - 81.6%). The effect of context on the inflection of the nonce verb was again tested using logistic and mixed effects regression models. The logistic regression model, with all four contexts as predictors and the inflection of the nonce verb as response variable did not find any partial effect for context in general, but it yielded a significant single main effect for context drink ( = - 1.286, z = -2.32, p < .05). Likewise, a similar mixed effects model did not reveal any significant partial effects for context in general (F(3,155) = 2.342, p = .0754). However, as in the German L1 experiment, fixed effects emerged for context variant drink ( = -1.286, t(155) = -2.288, p < .05). Like with German L1 speakers, the only context that seems to have a significant effect on the participants’ inflection is a context featuring a phonologically similar irregular verb. In sum, also in the written experiment semantics serves as a potential attractor only once nonce and prime are phonologically similar. 5.3 Summary and discussion The three above experiments provide two important insights. First, the data suggest that the semantic environment in which a nonce-verb is presented does only partially influence the way speakers inflect the nonceverb. Second, the analysis of the data suggests that there are fundamental problems with the experimental design. Let us first summarise the nature of the semantic influence on irregular verb morphology. As far as German L1 speakers are concerned, only the irregular and phonologically similar prime trinken served as an analogical attractor. This reveals two important aspects of the nature of morphophonological and semantic attraction. First, if semantics worked as an individual attractor, the context that was priming reiten should have attracted a significant majority of irregular nonce-verb inflections. However, this context remained insignificant for the nonce-verb inflection and produced only 25% of context sensitive responses. Second, if phonological similarity served as an individual attractor, the context priming hinken should, like trinken, significantly influence the participants’ nonce-verb inflection. However, as the regression model showed, this was clearly not the case. In contrast, morphology did serve as an analogical attractor, since participants clearly differentiated between regular and irregular primes, even if their phonology was almost identical. Overall, the different behaviour of verbs such as blink versus drink raises the question as to whether the data illustrate the speakers’ sensitivity mainly to morphology, phonology, or to both in tandem (cf. McClelland and Patterson 2002: 470). To sum up, what we 157 find with German L1 speakers appears to be an intricately intertwined dependency of phonology, morphology, and semantics. As far as the German-English interlanguage speakers are concerned, the semantic contexts in general failed to serve as analogical attractors. Only one mild single main effect could be observed for the context priming the irregular and phonologically similar prime drink. This would confirm the findings from the experiment with the German native speakers. A similar result has been reported by Huang Pinker (2005: 1006), who had found that people extend an irregular inflection to a word that sounds like and means the same as an existing irregular verb. However, we found that this extension was limited to cases where the new verb was a near-doppelganger of an existing one (i.e., being similar to it both in sound and in meaning), which leads people to treat the new verb as the existing one in disguise. Mere semantic similarity, unless it was both extreme in magnitude and accompanied by high phonological similarity, was not enough to evoke the stored irregular patterns. Thus, we could state that, in general, semantic similarity is much less capable of triggering analogical attraction than phonological similarity. In sum, hypothesis 3, claiming semantic similarity to operate independent of phonological similarity, has to be refuted. There is evidence that semantic attractions cannot be dissociated from morphological and phonological effects. Let us finally review the problems with the experimental design. The oral experiments revealed three shortcomings. First, the stories providing the semantic contexts triggered a substantial amount of unorthodox past tense inflections. One could conclude that, at least for some participants, the nature of the task remained obscure. This impression is confirmed by the variety of context-sensitive responses across the present three experiments as well as Ramscar’s study (2002a). Consider table 13. Table 13: Comparison of the degree of context sensitivity across the present three experiments and Ramscar’s study (2002a) German L1 interlanguage oral interlanguage written Ramscar 2002a context 1 93.0 % 40.0 % 28.4 % 77.5 % context 2 65.0 % 35.0 % 81.6 % 73.0 % context 3 25.0 % 45.0 % 31.3 % ---------context 4 70.0 % 45.0 % 76.3 % 70.0 % Bearing in mind that the participants in all four groups went through almost identical processes with exactly the same nonce-verb, the degree of variability between the context-sensitive responses is substantial, with Ramscar’s study producing an astounding degree of context-sensitivity. 158 The second shortcoming refers to the experiment’s three follow-up questions. Evidently, they provided in none of the three experiments any useful insights into the nature of potential priming processes in the speakers’ minds. As the regression analyses showed, there is little or no relationship between how the participants inflected the nonce-verb and their responses to the three questions. And there is also a certain lack of consistency as far as the response behaviour across the three questions is concerned (cf. table 6). And the thirdly, the experimental modality seems to influence the response behaviour of the participants. Consider figure 30 now, which plots the type of experiment against the inflection of the nonce-verb. type of experiment by inflection type of experiment inflection oral written frank frinked 2 = 7.2742 (Yates’ corrected), df = 1, p < .01 Figure 30: Mosaic plot of participants’ inflection cross-classified by the type of experiment (oral and written) for German-English interlanguage speakers The significant asymmetry in the above mosaic diagram illustrates a strong relationship between the type of experiment and the way participants inflected the nonce-verb. This impression was confirmed in a logistic regression model ( 2 = 7.16, df = 1, p < .005) and a mixed model with subjects as random effects. The mixed model showed a main effect for the written experiment towards more regular inflections ( = 0.897, t(218) = 2.807, p < .01). Thus, differences in the nonce-verb inflection could be (partly) accounted for by the difference in experimental design (oral and written). Given the above shortcomings, the validity of the entire experimental design has to be called into question, and further research will probably need to resort to different methods and experimental designs. 159 6 Discussion and conclusion The dissertation set out to investigate irregular verb morphology with regard to the German-English interlanguage. In particular, three experiments were supposed to test what role prototype theory, the apophonic path, and semantic attraction played in interlanguage morphology. In the following, the evidence, which turns out to be far from conclusive, will be discussed with respect to the three hypotheses; as a conclusion, some key characteristics of German-English interlanguage morphology will be suggested, with an outlook on how to model them in future research. Let us start with the results from experiment 1. The hypothesis stated that German-English interlanguage irregular verb morphology was best modelled according to prototype theory, and that the interlanguage prototype was the result of transfer from the learners’ L1 (German) onto the interlanguage. The quest for prototypes governing a particular vowel change, however, failed. In none of the experiments, and thus in none of the three languages involved, could a clear single prototype be distilled from the data. One problem was the lack of overlap between single main effects in the statistical models and the morphological make-up of the highfrequent ablaut-triggering nonce-verbs. The other one was, apart from some exceptions, a general lack of statistically significant interactions both in the regression and classification models. In other words, single main effects of constituents did not combine (or reinforce each other) once significant constituent variants co-occurred in one verb. This could be observed when re-analysing Plag’s (2000) data, too. When fitting a generalised mixed model allowing for random within-subject variation to his data, the interaction (prototype) he had found could not be reproduced. In sum, there simply was no most central single prototype for irregular verb inflection to be found anywhere in the data. Thus, there is no reason to assume that a prototype is actually transferred from the learners’ L1 to the interlanguage. As a conclusion, hypothesis 1 had to be rejected. There are, however, clear similarity-based effects, like such of codas in engma, or the complexity of onset structures. But the mechanism responsible for such effects appears to operate on analogies built along individual constituents and constituent combinations, and not on one single best member of a vowel change class. Recall hypothesis 2 now, which predicted the universal apophonic path to govern the vowel qualities of irregular verb morphology in all three groups of speakers. However, neither the path’s alleged universality nor its morphophonological predictions were borne out by the production data. 160 First, both speakers and learners partly modelled their ablaut sequences on vowel changes not accounted for by the apophonic path. This was particularly prevalent in the English L1 group, where more than 50% of all vowel changes triggered by given infinitives was ungrammatical. And second, both speakers and learners did prove to use grammatically correct ablaut even when inflecting from a given output form (past tense or participle). Such behaviour cannot be accounted for by the directionality of the apophonic path. In sum, the theoretical construct of such a path appears to be both insufficient and unnecessary for a concise description of the data. As a conclusion, hypothesis 2 had to be rejected, too. The results from experiment 3 are highly inconclusive. Recall the hypothesis for this experiment. It predicted the independence of semantic and phonological attraction in German-English interlanguage irregular verb morphology. Based on the data from experiment 3, this hypothesis, too, needed to be rejected. First, the independence of semantic and morphophonological effects could not be established. Apparently, the only statistically traceable semantic effect arose with the phonologically similar irregular prime with German L1 speakers. This runs against the hypothesis, since we would have expected semantic attraction both from the phonologically dissimilar primes and from the phonologically similar regular primes. This was clearly not the case. Second, there is a substantial problem with the experimental design. The follow-up questions failed to illustrate how genuine the semantic attraction actually was. For neither group of speakers that dealt with followup questions could a significant relationship between the response behaviour to the nonce-verb and the questions be established. Moreover, it turns out that the oral and written test designs significantly influenced the participants’ inflections. There are two possible explanations for these observations. First, the design both of the inflection tasks and the questions is simply not suited to monitoring the nature of semantic priming of the nonceverb. Or second, the arbitrariness of the response behaviour in the followup questions shows that the priming of the semantic environments in the four stories did not work out as expected. Either way, in the end the findings from Ramscar (2002a) could not be reproduced and compared in a meaningful way, and further experiments would probably need to resort to different methods and test designs. What, however, do the results from experiment 1 and 2 tell us? Although both hypotheses had to be rejected, the data provide a number of insights about the nature of German-English interlanguage irregular verb morphology. First, consider the evidence from experiment 1. It clearly points towards a similarity-based, albeit not prototypical organisation of the interlanguage 161 irregular verb paradigm. And it suggests that the mechanisms governing these paradigms in English and German L1 as well as in the interlanguage are highly similar. This confirms most of the evidence adduced by Bybee and Moder (1983) and Plag (2000). As far as the interlanguage speakers’ L1 (German) is concerned, the psycholinguistic reality of some of Köpcke’s (1998) lexical statistics could be confirmed, too. A complex onset, nucleus [ ] and codas with engma turned out to be similarity-based morphological attractors for irregular verb inflection in German. And the preferred vowel change with these constituents is the [ ]-[ ]-ablaut. For the English L1 speakers, onset and coda, too, are morphological attractors for vowel change. The complexity of onset and the role of codas with engma are not that clear, though. English L1 speakers, however, showed an interesting sensitivity to codas in [ ] and [ ]. The voiceless fricative, for instance, almost exclusively favoured regular past tense formation, which shows that these speakers are sensitive to the fact that in English all verbs ending in the voiceless fricative are in fact regular. And coda [ ] showed a significant bias towards vowel change, both in English L1 and in the re-analysis of Plag’s (2000) data. One reason for such behaviour might be the tendency of speakers across inflectional languages to avoid suffixing a past tense morpheme to a stem that already seems to contain this suffix (cf. Stemberger 1981, Bybee and Slobin 1982: 269). In other words, both English L1 and interlanguage speakers might have used so many vowel changes with nonces ending in [ ], because they perceived regular inflection for these nonces as doubling past tense morphs (two [ ]), and thus resorted to ablaut. It is worth pointing out that interlanguage speakers seem to adopt a habit which so far has occurred only with English L1 speakers. The German-English interlanguage speakers are mainly sensitive to coda structures, which generally appear to be the strongest predictor for the type of past tense inflection (cf. Bybee and Moder 1983), especially the ones with engma. Interestingly, the way interlanguage speakers treated coda was significantly different to the way L1 speakers did, but only as far as the distinction between regular and irregular inflection was concerned. This was not the case for differentiating between the types of vowel change both groups favour. Interestingly, the interlanguage speakers’ vowel changes are much more faithful to existing patterns than the ones produced by the English L1 speakers. Let us now turn to the evidence of experiment 2. The groups with given infinitive nonces always produced a majority of vowel change responses. This is first of all another piece of evidence for the participants’ willingness to provide vowel change responses based on the morphophonological make-up of the nonce-verb. However, the faithfulness to existing patterns 162 was a problem. While the German L1 speakers produced only around 8% of ungrammatical vowel changes, for the English L1 speakers it was more than 50%, mainly due to a low degree of consistency between the first and the second response. This clearly violates the predictions made by the apophonic path. The groups with a given output form (past tense or participle) provided conflicting evidence. From the apophonic path’s point of view, all reverse grammatical vowel changes used by the participants must be left unexplained. This is most pronounced for the interlanguage speakers, whose responses with vowel change amounted to 67%, with only 13% of ungrammatical forms. These data would suggest that the nonces given (past tense or participle) could in fact serve as output-oriented schemas, and that participants can inflect backwards, as it were, from a derived form to a base form. Such bi-directionality and output-orientation of vowel change inflections cannot be rule-based in nature, neither as suggested by Chomsky and Halle (1968), nor by Albright and Hayes (2003). However, from a schema point of view, all ungrammatical formations are problematic. And for the German L1 speakers, for instance, they were substantial. To begin with, there were only 43% vowel changes in this group, and amongst those, 60% did not follow the three vowel qualities predicted by the apophonic path. Moreover, all instances in which participants inflected an overtly marked irregular nonce participle (or nonce past tense) regularly cannot be accounted for by output-oriented schemas either 31 . Thus, the question remains in how far output-oriented schemas can account for the data in experiment 2 at all. First, note that the great majority of ungrammatical vowel changes did either arise due to conflating different vowel change strategies in both responses, or due to unorthodox second responses. Thus, vowels were not just used randomly. But how systematic were they? Consider first some of the unorthodox vowel changes as reported by Albright and Hayes (2003). In their data, they found changes such as from [ ] to [ ]. Obviously, [ ] is not a possible ablaut to a given [ ] in any English vowel change pattern, and thus not an instance of genuine apophony. And yet from a strictly output-oriented point of view this vowel change is not entirely arbitrary, either. One could argue that those speakers producing [ ] to a given [ ] know that a lot of verbs form their past tense by ablaut in [ ], like broke, spoke, told or sold. They might apply this knowledge about a possible output-oriented generalisation property to an unknown form with stem vowel [ ], merely disregarding the constraint that stem vowel [ ] occurs with the infinitive vowels [ ], [ ], [ ], and [ ] only, and never with [ ]. But in the end, they used a possible apophonic vowel to create a vowel change. 31 One possible explanation is a certain lack of attention or thoroughness on the participants’ side. 163 Now recall that the great majority of ungrammatical vowel changes in the reverse version of experiment 2 predominantly rested on forms that constitute an established part of the English stem-paradigm, too; the speakers, sometimes ignoring syntactic categories, just applied them in the wrong place. Thus, the question is in how far even the misplaced vowel changes follow some kind of pattern. Based on the evidence from experiment 2, it is possible that the assumed linear three-place apophonic mechanism, relating base forms, past tenses, and participles, might be ill-conceived. Recall that in three-place word-paradigm models, like minimalist morphology, an irregular English verb such as cling would come with a triple mental representation (cling for present tense, clung for past tense, and clung for participle). However, the vowel changes from the output-oriented inflections in experiment 2 seem to be at odds with such a representation. Instead, the data seem to suggest a two-place stem-and paradigm-model (cf. Lieb 1976, 1992 for English, and Wiese, in press, for German). Such a model would predict that speakers are aware of all possible vowel changes and all other morphophonological information of verb stems, but without syntactic categories pre-defined for a particular stem-form (cf. Lieb 1976: 21). For a stem-form such as clung, for instance, the paradigm leaves it open which syntactic category this form will adopt (cf. Halle 1953: 47). Thus, in a stem paradigm, only two forms of the verb to cling would be represented in the mental lexicon (cling and clung). This is illustrated and compared with the three-place apophonic path using the verb to sing in (54). (54) (a) three-place relation (b) two-place relation cf. Ségéral and Scheer 1998 cf. Wiese, in press cling - clung - clung cling - clung - clung i a u [ ] [ ] As can be seen in (2b), the two-place stem paradigm would spell out the two representations [ ] and [ ] only. And from a strictly output-oriented point of view, a given participle stem in [ ] could, based on all possible relations in English irregular verb morphology, easily trigger various input stem vowels, such as [ ], [ ], [ ], [ ], or [ ]. Based on the notion of a twoplace stem paradigm, the participants’ reaction to the given output form in experiment 2 could be visualised as follows ((55), English example). 164 (55) possible vowels given a two-place stem paradigm [ ], [ ], [ ], [ ], or [ ] Have you ever __ [ ] __ ? Yesterday I saw Sam; he [ _ ] the whole evening. I think, Sam likes to [ _ ]. possible vowels given a two-place stem paradigm [ ], [ ], [ ], [ ], or [ ] In sum, inputand output-oriented schemas, operating with nuclei organised in two-place stem-paradigms, could be appropriate models to account for the data from experiment 2. Let us conclude by bringing together the evidence from experiment 1 and 2 and distil some key characteristics for German-English irregular verb morphology. First, as shown before (cf. section 3.3), an interlanguage irregular verb paradigm is governed by essentially the same mechanism as in English L1 and German L1. Therefore, from an acquisitional point of view, the obvious parallels in inflectional behaviour are probably not so much based on transfer processes, but they are the result of a universal pattern associator, performing similarity-based generalisations across languages. Such a mechanism would not be specific only to natural languages, but share properties and characteristics with other types of human categorisation and reasoning (cf. Bybee 1995: 453). Second, the similarity-based generalisations of the pattern associator primarily rest on inputand output-oriented schemas (Bybee 1995). Such schemas, or lexical templates, are connections and generalisations across lexical representations in a speaker’s mental lexicon, with minimal or no abstractions, and they develop through language use and are sensitive to typeand token-frequency (cf. Bybee 1995, 2001). In language acquisition and use, they help establish analogical formations, that is generalisations from known forms to unknown forms, and, notably, from a target language onto an interlanguage. In contrast to prototypes, schemas can range over one, two, or all syllable constituents (onset, nucleus, and coda), and the constituents can have different weights in terms of their importance as part of a schema. These characteristics would be in line with predictions made by minimalist morphology (Wunderlich 1996, Wunderlich and Fabri 1995) and analogyand similarity-based lexical representation through schemas (cf. Bybee 1988, 1991, Prasada and Pinker 1993). Third, the schemas governing the interlanguage irregular verb paradigm would comprise onsets, nuclei, and codas. However, given the variability of vowel changes both in the present L1 and the interlanguage data, 165 and thus contrary to minimalist morphology, schemas should be associated in a two-place stem-paradigm, and the vowel quality of the lexical entry would not necessarily be pre-defined for categories such as ‘past tense’ or ‘participle’ (cf. Lieb 1992). For future research, either explicit similarity-based analogical models (cf. Skousen 1989, Eddington 2004) or connectionist networks 32 (Hickey and Martin 2001, Westermann 2000, Christiansen and Chater 2001) seem to be suitable to explore German-English interlanguage data in more detail. As shown in section 2.10, a whole range of language acquisition processes has already been modelled in such a fashion (cf. Blackwell and Broeder 1992, Sokolik and Smith 1992, Chandler 1994, Broeder and Plunkett 1997, Seidenberg and MacDonald 1999, Nakisa et al. 2000, Keulers and Daelemans 2007, and Keulers 2008). And recently, analogical models were also run with English nonceverbs (cf. Eddington 2004), exploring typeand token-frequency effects as well as different morphophonological representations of onset, nucleus, and coda. Such applications might be capable of modelling in more detail the subtle conspiracies of the English L1, German L1, and interlanguage grammar. In particular, they might be able to elucidate the nature of the connections between lexical representations and model the behavioural interlanguage data as reported in Plag (2000) and the present studies. A possible third approach might come with applications based on probabilities and statistical properties of language. They, too, have been tested in various domains of morphology (see, e.g. Redington and Chater 1997, Boersma and Hayes 2001, Albright 2002, Baayen 2003, Bod et al. 2003, Ernestus and Baayen 2003, Escudero and Boersma 2004, Yang 2004, Bod 2008), and they, too, might be able to model interlanguage characteristics with respect to the two L1 grammars involved. To conclude, the present dissertation has perhaps raised more questions than it has answered, and it was more occupied with discussing problems than with providing solutions. But, perhaps, this is not just a bad thing altogether. As Albert Einstein once remarked, the formulation of a problem is often more essential than its solution, which may be merely a matter of mathematical or experimental skill. 32 Some researchers, like Smolensky (1988, 1999) argue for an integration of connectionist and symbolic approaches, as already formalised in Optimality Theory (cf. Burzio 2002). See also Boersma and Levelt (2004), and Hancin-Bhatt (2008). 167 References Aarts, B. (2004): “Modelling linguistic gradience”. In: Studies in Language 28,1, 1-49. Aarts, B., D. Denison, E. Keizer, G. Popova (2004) (eds.): Fuzzy Grammar. A Reader. Oxford: Oxford University Press. Aha, D. W., D. Kibler, M. K. Albert (1991): “Instance-based learning algorithms”. In: Machine Learning 6, 37-66. Aitchison, J. (1992): “Good birds, better birds and amazing birds: The development of prototypes”. In: P. J. L. Arnaud and H. Bejoint (eds.): Vocabulary and Applied Linguistics. Basingstoke and London: Macmillan, 71-84. Albright, A. (2002): “Islands of reliability for regular morphology: Evidence from Italian”. In: Language 78, 684-709. Albright, A. (2009): “Modeling analogy as probabilistic grammar”. In: J. Blevins (ed): Analogy in Grammar: Form and Acquisition. Oxford: Oxford University Press. Albright, A., B. Hayes (2003): “Rules vs. analogy in English past tenses: A computational / experimental study”. In: Cognition 90, 2, 119-61. Andersen, R. W. (1978): “An implicational model for second language research”. In: Language Learning 28, 221-282. Anderson, S. R. (1988): “Morphological theory”. In: F. J. Newmeyer (ed.): Linguistics: The Cambridge Survey. Vol. I. Linguistic Theory: Foundations. Cambridge: Cambridge University Press, 146-191. Anderson, S. R. (1992): A-morphous Morphology. Cambridge: Cambridge University Press. Aronoff, M. (1994): Morphology by Itself. Cambridge, MA: MIT Press. Ashby, F., E. Waldron (1999): “On the nature of implicit categorization”. In: Psychonomic Bulletin and Review, 6, 3, 363-378. Augst, G. (1975): “Wie stark sind die starken Verben? Überlegungen zur Subklassifizierung der Nhd. Verben“. In: Untersuchungen zum Morpheminventar der deutschen Gegenwartssprache. Tübingen: Narr, 231-281. Aureliu, L. (2001): “Mapping dissociations in verb morphology“. In: Trends in Cognitive Science 5, 7, 301-308. Baayen, R. H. (2003): “Probabilistic approaches to morphology”. In R. Bod, J. Hay, S. Jannedy (eds.): Probabilistic Linguistics. Cambridge, MA: MIT Press, 229-288. Baayen, R. H., T. Dijkstra, R. Schreuder (1997): “Singulars and plurals in Dutch. Evidence for a dual parallel route model“. In: Journal of Memory and Language 37, 94-119. Baayen, R. H., F. Moscoso del Prado Martín (2005): “Semantic density and pasttense formation in three Germanic languages”. In: Language 81, 3, 666-698. Bardovi-Harlig, K. (1987): “Markedness and salience in second language acquisition“. In: Language Learning 37, 3, 385-407. Bartke, S., F. Rösler, J. Streb, R. Wiese (2005): “An ERP study of German ‘irregular’ morphology”. In: Journal of Neurolinguistics 18, 1, 29-55. Bates, D. M. (2005): “Fitting linear mixed models in R”. In: R News, 5, 27-30. Bechtel, W., A. Abrahamsen (1990): Connectionism and the Mind: An Introduction to Parallel Processing in Networks. Cambridge, MA: Blackwell. 168 Beck, M.-L. (1997): “Regular verbs, past tense and frequency: Tracking down a potential source of NS/ NNS competence differences”. In: Second Language Research 13, 2, 93-115. Becker, T. (1990): Analogie und Morphologische Theorie. München: Wilhelm Fink Verlag. Bendjaballah, S., M. Haiden (2002): “Meaningful vowels”, in: P. Bottari (ed.): Proceedings of the XXVIII Incontro di Grammatica Generativa. Lecce: Congedo Editore. Bendjaballah, S., M. Haiden (2005): “Die Grammatik der Präpositionen im Berber”. In: Folia Linguistica, 39, 3-4, 319-361. Beretta, A., C. Campbell, T. Carr, J. Huang, L. M. Schmitt, K. Christianson, Y. Cao (2003): “An ER-fMRI investigation of morphological inflection in German reveals that the brain makes a distinction between regular and irregular forms”. In: Brain and Language 85, 67-92. Berko (Gleason), J. (1958): “The child’s learning of English morphology“. In: Word 14, 150-77. Bialystok, E. , K. Hakuta (1994): In Other Words: The Science and Psychology of Second- Language Acquisition. New York: Basic Books. Bierwisch, M. (1967): “Syntactic features in morphology: general problems of socalled pronominal inflection in German”. In: To Honor Roman Jakobson. Vol. 1. The Hague/ Paris: Mouton, 239-270. Birdsong, D., J. E. Flege (2001): “Regular-irregular dissociations in L2 acquisition of English morphology”. In: BUCLD 25. Proceedings of the 25th Annual Boston University Conference on Language Development, Volume 1, 123-132. Bittner, A. (1985): “Wie schwach sind die starken Verben? Überlegungen und Vorschläge zu einer Klassifizierung der neuhochdeutschen Verben”. In: W. U. Wurzel (ed.): Studien zur Morphologie und Phonologie I. Studia Grammatica 21. Berlin: Akademie Verlag, 51-74. Bittner, A. (1996): Starke ‚schwache’ Verben, schwache ‚starke’ Verben. Deutsche Verbflexion und Natürlichkeit. Tübingen: Stauffenburg. Blackwell, A., P. Broeder (1992): “Interference and facilitation SLA: A connectionist perspective”. Paper presented at the seminar on Parallel Distributed Processing and Natural Language Processing, San Diego, UCSD. Blevins, J. P. (2003): “Stems and paradigms”. In: Language 79, 737-767. Bloch, B. (1947): “English verb inflection”. In: Language 23, 399-418. Bloomer, R. (1994): System-congruity and the Participles of Modern German and Modern English: A study in Natural Morphology. Hamburg: Buske Verlag Bloomfield, L. (1933): Language. Revised from the 1914 edition. New York: Holt. Bod R., J. Hay, S. Jannedy (eds.) (2003): Probabilistic Linguistics. Cambridge, MA: MIT Press. Bod, R. (2008): “From exemplar to grammar: Integrating analogy and probability in language learning”. Submitted ms (http: / / staff.science.uva.nl/ ~rens/ analogy.pdf). Boersma, P., B. Hayes (2001): “Empirical tests of the Gradual Learning Algorithm”. In: Linguistic Inquiry 32, 45-86. Boersma P., C. Levelt (2004): “Optimality Theory and phonological acquisition”. In: L. M. Santelmann, M. Verrips, F. Wijnen, C. Levelt (eds.): Annual Review of Language Acquisition. Volume 3 (2003). Amsterdam: Benjamins. Bolinger, D. L. (1961): Generality, Gradience and the All-or-none. The Hague: Mouton. Bolinger, D. L. (1977): Meaning and Form. London: Longman. 169 Bolinger, D. L. (1992): “About furniture and birds”. In: Cognitive linguistics 3, 1, 111- 117. Boyé, G. (2000). Problèmes de Morpho-Phonologie Verbale en Francais, en Espagnol et en Italien. Doctoral thesis. Université Paris 7. Breiman, L., J. H. Friedman, R. A. Olshen, C. J. Stone (1984): Classification and Regression Trees. Pacific Grove, CA: Wadsworth and Brooks/ Cole Advanced Books and Software. Breslow, N. E. , D. G. Clayton (1993): “Approximate inference in generalized linear mixed models”. In: Journal of American Statistical Association, 88, 9-25. Broeder, P. and K. Plunkett (1997): “Connectionism and second language acquisition”. In: N. C. Ellis (ed.): Implicit and Explicit Learning of Language. London et al.: Harcourt Brace Jovanovich, 421-454. Burgess, A. (1972): A Clockwork Orange. London: Penguin. Burzio, L. (2002): “Missing players: Phonology and the past-tense debate”. In: Lingua 112, 157-199. Bybee, J. (1985): Morphology. Amsterdam: John Benjamins. Bybee, J. (1988): “Morphology as lexical organization”. In: M. Hammond, M. Noonan (eds.): Theoretical Approaches to Morphology. San Diego: Academic Press, 119-141. Bybee, J. (1991): “Natural morphology: The organization of paradigms and language acquisition”. In: T. Hübner, C. A. Ferguson (eds.): Crosscurrents in Second Language Acquisition and Linguistic Theory. Amsterdam: Benjamins, 67-91. Bybee, J. (1995): “Regular morphology and the lexicon”. In: Language and Cognitive Processes 10, 425-455. Bybee, J. (2001): Phonology and Language Use. Cambridge Studies in Linguistics 94. Cambridge: Cambridge University Press. Bybee, J., D. I. Slobin (1982): “Rules and schemas in the development and use of the English past tense”. In: Language 58, 265-289. Bybee, J., C. L. Moder (1983): “Morphological classes as natural categories”. In: Language 59, 251-270. Chambers, J. M. (1992): “Linear models”. In: J. M. Chambers, T. J. Hastie (eds.): Statistical Models in S. Pacific Grove, CA: Wadsworth and Brooks/ Cole. Chandler, S. (1993): “Are rules and modules really necessary for explaining language? ” In: Journal of Psycholinguistic Research 22, 593-606. Chandler, S. (1994): “An exemplar-based approach to language acquisition”. Paper presented at the Workshop on Cognitive Models of Language Acquisition, University of Tilburg, The Netherlands, April 1994. Chandler, S. (1995). “Non-declarative linguistics: Some neuropsychological perspectives”. In: Rivista di Linguistica 7, 233-247. Chandler, S. (1998): “Instance-based reference for past-tense verb-forms: An experimental study”. Paper presented at the First International Conference on the Mental Lexicon, Edmonton, Alberta, Canada. Chernigovskaya T., K. Gor (2000): “The complexity of paradigm and input frequencies in native and second language verbal processing: evidence from Russian. In: Language and Language Behavior 3, 2, 20-37. Chomsky, N., M. Halle (1968): The Sound Pattern of English. New York: Harper and Row. Christiansen, M. H., N. Chater (1999): “Connectionist natural language processing: the state of the art“. In: Cognitive Science 23, 4, 417-37. 170 Christiansen, M. H., N. Chater (2001): “Connectionist psycholinguistics: Capturing the empirical data”. In: Trends in Cognitive Science 5, 2, 82-88. Clahsen, H. (1997): “The representation of participles in the German mental lexicon: Evidence for the dual-mechanism model”. In: G. Booij, J. van Marle (eds.): Yearbook of Morphology 1996, 73-95. Clahsen, H. (1999): “Lexical entries and rules of language: A multidisciplinary study of German inflection”. In: Behavioral and Brain Sciences 22, 991-1060. Clahsen, H., M. Rothweiler (1993): “Inflectional rules in children’s grammars: Evidence from the development of participles in German”. In: G. Booij, J. van Marle (eds.) Morphology Yearbook, 1992, 1-34. Clahsen, H., S. Eisenbeiß, I. Sonnenstuhl-Henning (1997): “Morphological structure and the processing of inflected words”. In: Theoretical Linguistics 23, 201-249. Clahsen, H., S. Eisenbeiss, M. Hadler, I. Sonnenstuhl (2001): “The mental representation of inflected words: An experimental study of adjectives and verbs in German”. In: Language 77, 510-543. Colombo, L., I. Stoianov, M. Pasini, M. Zorzi (2006): “The role of phonology in the inflection of Italian verbs: A connectionist investigation”. In: The Mental Lexicon 1, 1, 147-181. Corbett, G. G. , N. M. Fraser (1993): “Network Morphology: a DATR account of Russian nominal inflection. In: Journal of Linguistics 29, 113-42. Croft, W. (2007): “Beyond Aristotle and gradience: A reply to Aarts”. In: Studies in Language, 31, 2, 409-430. Daelemans, W., S. Gillis, and G. Durieux (1994a): “Skousen's analogical modeling algorithm: A comparison with lazy learning”. In: D. Jones (ed.): Proceedings of the International Conference on New Methods in Language Processing. Manchester: UM- IST. Daelemans, W., S. Gillis, G. Durieux (1994b): “The acquisition of stress: a dataoriented approach”. In: Computational Linguistics 20, 421-451. Daelemans, W., J. Zavrel, K. van der Sloot, A. van den Bosch (1999): TiMBL: Tilburg memory based learner, version 2.0, reference guide. Induction of Linguistic Knowledge Technical Report. Tilburg, Netherlands: ILK Research Group, Tilburg University. (http: / / ilk.kub.nl/ ~ilk/ papers/ ilk9901.ps.gz, Zugriff am 01.09.2008). Daelemans, W., A. van den Bosch (2005): Memory-Based Language Processing. Studies in Natural Language Processing. Cambridge: Cambridge University Press Derwing, B. L., R. Skousen (1989a): “Real-time morphology: Symbolic rules or analogical networks”. In: Proceedings of the Fifteenth Annual Meeting of the Berkeley Linguistics Society, Berkeley Linguistics Society, Berkeley, California, 48-62. Derwing, B. L., R. Skousen (1989b). “Morphology in the mental lexicon: A new look at analogy”. In: G. Booij, J.van Marle (eds.): Yearbook of Morphology 1989, 55-71. Derwing, B. L., R. Skousen (1994): “Productivity and the English past tense: Testing Skousen’s analogy model”. In: S. D. Lima, R. L. Corrigan, G. K. Iverson (eds.): The Reality of Linguistic Rules. Amsterdam: Benjamins, 193-218. Desai, R., L. L. Conant, E. Waldron, J. R. Binder (2006): “fMRI of past tense processing: The effects of phonological complexity and task difficulty”. In: Journal of Cognitive Neuroscience 18, 278-297. Dressler, W. U. (1977): Grundfragen der Morphonologie. Wien: Verlag der Österreichischen Akademie der Wissenschaften. Dressler, W. U. (1985a): Morphonology. Ann Arbor: Karoma. 171 Dressler, W. U. (1985b): “On the predictiveness of Natural Morphology”. In. Journal of Linguistics 21, 321-337. Dressler, W. U. (1986): “Explanation in natural morphology, illustrated with comparative and agent-noun formation”. In: Linguistics 24, 519-548. Eckman, F. (1977): “Markedness and the Contrastive Analysis Hypothesis”. In: Language Learning 27, 315-330. Eddington, D. (2000a): “Analogy and the dual-route model of morphology”. In: Lingua 100, 281-298. Eddington, D. (2000b): “Spanish stress assignment within analogical modeling of language”. In: Language 76, 92-109. Eddington, D. (2002a): “Spanish diminutive formation without rules or constraints”. In: Linguistics, 40, 395-419. Eddington, D (2002b): “Spanish gender assignment in an analogical framework”. In: Journal of Quantitative Linguistics 9, 49-75. Eddington, D (2004): “Issues in modelling language processing analogically”. In: Lingua 114, 7, 849-871. Eisenberg, P., H. Gelhaus, H. Henne. H. Sitta. H. Wellmann eds. (1998): Duden. Grammatik der deutschen Gegenwartssprache. 6 th edition. Mannheim et al.: Dudenverlag. Ellis, N. C. (2006): “Selective attention and transfer phenomena in L2 acquisition: Contingency, cue competition, salience, interference, overshadowing, blocking, and perceptual learning”. In: Applied Linguistics 27, 2, 164-194. Ernestus, M., R. H. Baayen (2001): “Choosing between the Dutch past tense suffixes -te and -de”, in: T. van der Wouden, H. Broekhuis (eds.): Linguistics in the Netherlands 2001. Amsterdam: Benjamins, 77-88. Ernestus, M., R. H. Baayen (2003): “Predicting the unpredictable: Interpreting neutralized segments in Dutch”. In: Language 79, 1, 5-38. Ernestus, M., R. H. Baayen (2004): “Analogical effects in regular past tense production in Dutch”. In: Linguistics 42, 873-903. Escudero, P., P. Boersma (2004): “Bridging the gap between L2 speech perception research and phonological theory”. In: Studies in Second Language Acquisition, 2, 551-585. Eubank, L. (1996): “Negation in early German-English interlanguage: more valueless features in the L2 initial state“, in: L. Eubank, B. D. Schwatz (ed.): Second Language Research. Special Issue on The L2 Initial State, 12, 1, 73-106. Fabricius-Hansen, C. (1977): “Zur Klassifizierung der starken Verben im Neuhochdeutschen“. In: Deutsche Sprache, 5, 193-205. Fanselow, G., C. Fery, M. Schlesewsky, R.Vogel (eds.) (2006): Gradience in Grammar. Generative Perspectives. Oxford: Oxford University Press. Flege, J. E., G. H. Yeni-Komshian, S. Liu (1999): “Age constraints on secondlanguage acquisition”. In: Journal of Memory and Language, 41, 78-104. Fodor, J. A., Z. W. Pylyshyn (1988): Connectionism and cognitive architecture: A critical analysis. Cambridge, MA: MIT Press/ Bradford Books. Gass, S. M. (1983): “Language transfer and universal grammatical relations”, in S. M. Gass, L. Selinker (eds.): Language transfer in language learning. Rowley, MA: Newbury House, 69-82. Gass, S. M., L. Selinker (2001): Second language acquisition: An introductory course (2 nd edition). Hillsdale, NJ: Erlbaum. 172 Gasser, M. (1990): “Connectionism and universals of second language acquisition”. In: Studies in Second Language Acquisition, 12, 179-99. Geeraerts, D. (1988): “Where does prototypicality come from? “. In: B. Rudzka-Ostyn (ed.): Topics in cognitive linguistics. Amsterdam: Benjamins, 207-229. Geeraerts, D. (1989): “Prospects and problems of prototype theory“. In: Linguistics 27, 587-612. Gentner, D. (1983): “Structure-mapping: A theoretical framework for analogy”. In: Cognitive Science. 7, 2, 55-170. Gentner, D. (1998): “Analogy”, in: W. Bechtel, G. Graham (eds.): A companion to cognitive science. Oxford: Blackwell, 107-113. Givón, T. (1986): “Prototypes: Between Plato and Wittgenstein”, in C. Craig (ed.): Noun classes and categorization. Proceedings of a symposium on categorization and noun classification, Eugene, Oregon. Amsterdam and Philadelphia: Benjamins, 77- 102. Goebel, R., P. Indefrey (2000): “A recurrent network with short-term memory capacity learning the German s-plural”, in: P. Broeder, J. Murre (eds.): Models of Language Acquisition. Inductive and deductive approaches. Oxford: Oxford University Press, 177-200. Goldschneider, J. M., R. M. DeKeyser (2001) “Explaining the ‘natural order of L2 morpheme acquisition’ in English: A meta-analysis of multiple determinants”. In: Language Learning, 51, 1-50. Gor, K., T. Chernigovskaya (2003): “Mental lexicon structure in L1 and L2 acquisition: Russian evidence”. Papers from the East of Eden Conference, a joint meeting of the Finnish Cognitive Linguistics Association and the Slavic Cognitive Linguistics Association. In: Glossos, 4, 1-31. Gross, M., T. Say., M. Kleingers, T. F. Münte, H. Clahsen (1998): “Human brain potentials to violations in morphologically complex Italian words”. In: Neuroscience Letters, 241, 83-86. Guerssel, M., J. Lowenstamm (1996): “Ablaut in Classical Arabic measure I active verbal forms”. In: Lecarme, J., J. Lowenstamm, U. Shlonsky (eds): Studies in Afroasiatic Grammar. The Hague: Hag, 123-134. Hahne, A., J. L. Müller, H. Clahsen (2006): “Second language learners’ processing of inflected words: Behavioral and ERP evidence for storage and decomposition”. In: Journal of Cognitive Neuroscience, 18, 1, 121-134. Halle, M. (1953): “The German conjugation”. In: Word, 9, 45-53. Hampton, J. A. (1998): “Similarity-based categorization and fuzziness of natural categories“. In: Cognition 65, 137-65. Hancin-Bhatt, B. (2008): “Second language phonology in Optimality Theory”. In: J. G. Hansen Edwards, M. L. Zampini (eds.): Phonology and Second Language Acquisition. Studies in Bilingualism 36. Amsterdam: John Benjamins, 117-146. Hare, M., J. L. Elman (1995): “Learning and morphological change”. In: Cognition, 56, 61-98. Hare, M. L., J. L. Elman, K. G. Daugherty (1995): “Default generalization in connectionist networks”. In: Language and Cognitive Processes, 10, 6, 601-630. Hatch, E. M. (1983): Psycholinguistics: A Second Language Perspective. Rowley, MA: Newbury House. Hay, J., R. H. Baayen (2005): “Shifting paradigms: gradient structure in morphology”. In: Trends in Cognitive Sciences, 9,7, 342-348. 173 Hempe, U. (1988): Die starken Verben im Deutschen und Niederländischen. Diachrone Morphologie. Tübingen: Niemeyer. Henzen, W. (1965): Deutsche Wortbildung. Tübingen: Niemeyer. Hickey, R. J., R. G. Martin (2001): “An instance-based approach to pattern association learning with application to the English past tense domain”. In: Knowledge- Based Systems, 14, 131-136. Hoard, J., C. Sloat (1973): “English irregular verbs”. In: Language, 49, 107-120. Hockett, C. F. (1954): “Two models of grammatical description”. In: Word 10, 210- 233. Hook, D. (1968): “Die Klassifizierung der starken Verben in der deutschen Gegenwartssprache”. In: Deutsch als Fremdsprache 5, 219-221. Huang, Y. T., S. Pinker (2005): “Semantic, phonological, and lexical influences on regular and irregular inflection”. Paper presented at the Annual Meeting of the Cognitive Science Society, Stressa, Italy, July 21-23, 1001-1006. Indefrey, P. (2006): “It is time to work toward explicit processing models for native and second language speakers”. In: Applied Psycholinguistics, 27, 1, 66-69. Ingram, J., H. Chenery (2007): Neurolinguistics. An introduction to spoken language processing and its disorders. Cambridge: Cambridge University Press. Itkonen, E. (2005): Analogy as Structure and Process. Approaches in Linguistics, Cognitive Psychology and Philosophy of Science. Human Cognitive Processing, 14. Amsterdam, Philadelphia: John Benjamins. Jackendoff, R. S. (1975): “Morphological and semantic regularities in the lexicon”. In: Language, 51, 639-671. Jaeger, J. J., A. H. Lockwood, D. L Kemmerer, R. D. Vanvalin, B. W. Murphy, H. G. Khalak (1996): “Positron emission tomographic study of regular and irregular verb morphology in English”. In: Language, 72, 3, 451-497. Jespersen, O. (1942): A modern English grammar on historical principles. Part IV: Morphology. Copenhagen: Munksgaard. Joanisse, M. F., M. S. Seidenberg (1999): “Impairments in verb morphology after brain injury: A connectionist model”. In: Proceedings of the National Academy of Sciences of the United States of America, 96, 13, 7592-7597. Jones, D. (1996): Analogical Natural Language Processing. Studies in Computational Linguistics. London: University College of London Press. Juilland, A., J. Macris (1973): The English Verb System. The Hauge: Mouton de Gruyter. Jurafsky, D. (2002): “Probabilistic modeling in psycholinguistics: Linguistic comprehension and production”. In R. Bod, J. Hay, S. Jannedy (eds.): Probabilistic Linguistics. Cambridge, Mass.: MIT Press. Kaye, J. D., J. Lowenstamm, J.-R. Vergnaud (1985): “The internal structure of phonological representations: a theory of Charm and Government”. In: Phonology Yearbook 2, 305-328. Kellerman, E. (1983): “Now you see it, now you don’t”, in: S. Gass, L. Selinker (eds.): Language transfer in language learning. Rowley, MA: Newbury House, 112-134. Kellerman, E. (1986): “An eye for an eye: Crosslinguistic constraints on the development of the L2 lexicon”. In: M. Sharwood Smith, E. Kellerman (eds.): Crosslinguistic Influence in Second Language Acquisition. Oxford: Pergamon Press, 35-48. Kern, P. C., H. Zutt (1977): Geschichte des deutschen Flexionssystems. Germanistische Arbeiten 22. Tübingen: Niemeyer. 174 Keuleers, E., W. Daelemans (2007): “Memory-based learning models of inflectional morphology: A methodological case study”. In: Lingue e Linguaggio 6 (2), 151- 174. Keuleers, E. (2008): Memory-based learning of inflectional morphology. Dissertation at the University of Antwerp. Manuscript. Kırkıcı, Bilal (2007): “The mental processing of L2 English lexical compounds: A developmental Dual-Mechanism account”. In: L. Roberts, A. Gürel, S. Tatar, L. Marti (eds.): Eurosla-Yearbook, 7, 7-25. Koenraads, W. H. A. (1953): Studien über sprachökonomische Entwicklungen im Deutschen. Amsterdam: Meulenhoff. Köpcke, K.-M. (1998): “Prototypisch starke und schwache Verben in der deutschen Gegenwartssprache”. In: Germanistische Linguistik, 141-142, 45-60. Krott, A., R. H. Baayen, R. Schreuder (2001): “Analogy in morphology: modeling the choice of linking morphemes in Dutch”. In: Linguistics, 39, 1, 51-93. Kuryłowicz, J. (1956): L’apophonie en Indo-Européen. Wrocław. Labov, W. (1973): “The Boundaries of words and their meanings”. In: C.-J. N. Bailey, R. W. Shuy (eds.): New Ways of Analyzing Variation in English. Washington: Georgetown University Press, 340-373. Lappe, S., I. Plag (2003): “Rules vs. analogy. Modeling variation in word-final epenthesis in Sranan“. In: I. Plag ed.: Phonology and Morphology of Creole Languages. Tübingen: Niemeyer. Larsen-Freeman, D. E. (1976): “An explanation for the morpheme acquisition order of second language learners”. In: Language Learning, 26, 125-134. Larsen-Freeman, D. E. (2002): “Making sense of frequency”. In: Studies in Second Language Acquisition, 24, 275-285. Lavric, A., D. Pizzagalli, S. Forstmeier, G. Rippon (2001): “Mapping dissociations in verb morphology”. In: Trends in Cognitive Science, 5, 7, 301-308. Lawless, J. F., K. Singhal (1978): “Efficient screening of nonnormal regression models”. In: Biometrics, 34, 318-327. Lieb, H.-H.(1976): “Grammars as theories: The case for axiomatic grammar (Part II)”. In: Theoretical Linguistics 3, 1-98. Lieb, H.-H. (1992): “Paradigma und Klassifikation: Explikation des Paradigmenbegriffs”. In: Zeitschrift für Sprachwissenschaft, 11. 3-46. Ling, C. X., M. Marinov (1993): “Answering the connectionist challenge: A symbolic model of learning the past tense of English verbs”. In: Cognition, 49, 235-290. López Rúa, P. (2005): Birds, colours and prepositions. The theory of categorization and its applications in linguistics. Lincom Studies in Theoretical Linguistics 33. München: Lincom. Lowie, W. M. (1998): The Acquisition of Interlanguage Morphology. PhD Thesis University of Groningen (http: / / dissertations.ub.rug.nl/ faculties/ arts/ 1998/ w.m.lowie/ , Zugriff am 01.08.2008). Lück, M., A. Hahne, H. Clahsen (2006): “Brain potentials to morphologically complex words during listening”. In: Brain Research 1077, 144-152. Mangasser-Wahl, M. (ed.) (2000): Prototypentheorie in der Linguistik. Anwendungsbeispiele, Methodenreflexion, Perspektiven. Tübingen: Stauffenburg. Marchand, H. (1969): The Categories and Types of Present-Day English Word-formation. München: Beck. Marcus, G. F. (1995): “The acquisition of the English past tense in children and multilayered connectionist networks”. In: Cognition 56, 271-279. 175 Marcus, G. F. (1998): The Algebraic Mind: Reflections on Connectionism and Cognitive Science. Cambridge Mass: MIT Press. Marcus, G. F. (1999): “Do infants learn grammar with algebra or statistics? Response”. In: Science, 284 (5413), 436-437. Marcus, G. F., U. Brinkmann, H. Clahsen, R. Wiese, S. Pinker (1995): “German inflection: The exception that proves the rule”. In: Cognitive Psychology, 29, 3, 189-256. Marslen-Wilson, W., L. K. Tyler (1998): “Rules, representations, and the English past tense”. In: Trends in Cognitive Science 2, 11, 428-35. Martinet, A. (1955): Économie des changements phonétiques. Traité de phonologie diachronique. Bern: Francke. Matthews, P. H. (1991): Morphology. 2 nd ed. Cambridge: Cambridge University Press. Münte, T. F., T. Say, H. Clahsen, K. Schiltz, M. Kutas (1999): “Decomposition of morphologically complex words in English: evidence from event-related brain potentials”. In: Cognitive Brain Research 7, 241-253. Murphy, V. A. (2000): “Compounding and the representation of L2 inflectional morphology”. In: Language Learning 50, 1, 153-197. Nakisa, R. C., U. Hahn (1996): “Where defaults don’t help: the case of the German plural system”. Paper presented at the 18 th Annual Conference of the Cognitive Science Society, San Diego, CoRR cmp-lg/ 9605020. Nakisa, R. C., K. Plunkett, U. Hahn (2000): “Singleand dual-route models of inflectional morphology”, in: P. Broeder, J. Murre (eds.): Models of Language Acquisition. Oxford: Oxford University Press, 201-224 Odlin, T. (2000): Language Transfer: Cross-Linguistic Influence in Language Learning. 7 th edition. Cambridge: Cambridge University Press. Paul, H. (1880): Prinzipien der Sprachgeschichte. 9 th edition 1975. Tübingen: Niemeyer. Paul, H. (1917): Deutsche Grammatik. Band 2. Halle (Saale): Niemeyer. Paul, H., P. Wiehl, S. Grosse (1989): Mittelhochdeutsche Grammatik. Tübingen: Niemeyer. Penke, M., H. Weyerts, M. Gross, E. Zander, H. Clahsen, T. F. Münte (1997). “How the brain processes complex words: an ERP-study of German verb inflections”. In: Cognitive Brain Research 6, 37-52. Pinheiro, J. C., D. M. Bates (2000): Mixed-effects Models in S and S-PLUS. New York: Springer. Pinker, S. (1991): “Rules of language”. In: Science 253, 530-534. Pinker, S. (1997): “Words and rules in the human brain”. In: Nature 387, 547-548 Pinker, S. (1999): Words and rules. The ingredients of language. London: Phoenix. Pinker, S., A. Prince (1988): “On language and connectionism: Analysis of a parallel distributed processing model of language acquisition”. In: Cognition, 28, 73-193. Pinker, S., A. Prince (1994): “Regular and irregular morphology and the psychological status of rules and grammar”. In: S. D. Lima, R. L. Corrigan, G. K. Iverson (eds.): The Reality of Linguistic Rules. Philadelphia: Benjamins, 321-352. Pinker, S., M. T. Ullman (2002a): “Combination and structure, not gradedness, is the issue”. In: Trends in Cognitive Sciences 6, 472-474. Pinker, S., M. T. Ullman (2002b): “The past and future of the past tense”. In: Trends in Cognitive Science, 6, 456-463. Plag, I. (2000): “Irregular past tense formation in English interlanguage”. In: I. Plag, K. P. Schneider (eds.): Language Use, Language Acquisition, and Language History. (Mostly) Empirical Studies in Honour of Rüdiger Zimmermann. Trier: Wissenschafticher Verlag, 134-149. 176 Plunkett, K., V. Marchman (1991): “U-shaped learning and frequency effects in a multi-layered perception”. In: Cognition 38, 43-102. Plunkett, K., P. Juola (1999): “A connectionist model of English past tense and plural morphology”. In: Cognitive Science, 23, 4, 463-490. Posner, M. (1986): “Empirical studies of prototypes”. In: C. Craig (ed.): Noun classes and categorization. Proceedings of a symposium on categorization and noun classification, Eugene, Oregon. Amsterdam and Philadelphia: Benjamins, 53-61. Prasada, S., S. Pinker (1993): “Generalization of regular and irregular morphological patterns”. In: Language and Cognitive Processes 8, 1-56. Prasada, S., S. Pinker, W. Snyder (1990): “Some evidence that irregular forms are retrieved from memory but regular forms are rule generated”. Paper presented at the Psychonomic Society meeting. New Orleans, LA, 16-18 November. Pyles, T., J. Algeo (1992): The Origins and Development of the English Language. 4 th edition. Fort Worth et al.: Harcourt Brace. Quirk, R., S. Greenbaum, G. Leech, J. Svartvik (1988): A Comprehensive Grammar of the English Language. New York: Longman. Ramscar, M. J. A. (2002a): “The role of meaning in inflection: Why the past tense does not require a rule”. In: Cognitive Psychology, 45, 45-94. Ramscar, M. J. A. (2002b): “Semantic grounding in analogical processing: An environmental approach”. In: Cognitive Science, 27, 41-71. Ramscar, M. J. A. (2002c): “When the fly flied and when the fly flew: the effects of semantics on the comprehension of past tense inflections”. In: Proceedings of the 24 th Annual Conference of the Cognitive Science Society, 768-773. Redington, M., N. Chater (1997): “Probabilistic and distributional approaches to language acquisition”. In: Trends in Cognitive Sciences, 1, 273-281. Rosch, E. (1973): “Natural Categories“. In: Cognitive Psychology 7, 532-547. Rosch, E. (1975): “Cognitive representations of semantic categories”. In: Journal of Experimental Psychology: General 104, 192-232. Rosch, E., C. B. Mervis (1975): “Family resemblance: Studies in the internal structure of categories”. In: Cognitive Psychology 7, 573-605. Rumelhart, D. E., J. L. McClelland (1986): “On learning the past tense of English verbs”, in: J. L. Mc Clelland, D. E. Rumelhart, PDP Research Group (eds): Parallel Distributed Processing, Volume 2. Cambridge MA: The MIT Press, 216-271. Ruoff, A. (1981): Häufigkeitswörterbuch gesprochener Sprache. Tübingen: Niemeyer. Rytting, C. A. (2000): “An empirical test of analogical modelling: the / k/ ~ Ø alternation”. In: A. K. Melby, A. R. Lommel (eds.): Lacus Forum XVII: The Lexicon. Fullerton, CA: Linguistic Association of Canada and the United States, 73-84. Sapir, E. (1921/ 1957): Language: An Introduction to the Study of Speech. New York: Harcourt, Brace and Company / Harvest Books. Schmid, H. J. (1998): “Zum kognitiven Kern der Prototypentheorie”, in: F. Ungerer (ed.): Kognitive Lexikologie und Syntax. Rostock: Universitätsverlag, 9-28. Ségéral, P. (2000): “Théorie de l´apophonie et organisation des schèmes en sémitique”. In: J. Lecarme, J. Lowenstamm, U. Shlonsky (eds): Research in Afroasiatic Grammar. Amsterdam: Benjamins, 263-299. Ségéral, P., T. Scheer (1998): “A generalized theory of ablaut: The case of modern German strong verbs”. In: R. Fabri, A. Ortmann, T. Parodi (eds.): Models of Inflection. Tübingen: Niemeyer, 28-59. 177 Seidenberg, M. S. (1992): “Connectionism without tears”. In: S. Davis (ed.): Connectionism: Theory and practice. New York: Oxford University Press, 84-122. Seidenberg, M. S., M. C. MacDonald (1999): “A probabilistic constraints approach to language acquisition and processing”. In: Cognitive Science, 23, 569-588. Selinker, L. (1983): “Language transfer”. In: S. Gass, L. Selinker (eds.): Language Transfer in Language Learning. Rowley, MA: Newbury House, 33-68. Sereno, J. A., A. Jongman (1997): “Processing of English inflectional morphology”. In: Memory and Cognition 25, 425-37. Shirai, Y. (1992): “Conditions on transfer: a connectionist approach”. In: Issues in Applied Linguistics, 3, 91-120. Skousen, R. (1989). Analogical Modeling of Language. Dordrecht: Kluwer Academic. Skousen, R. (1992): Analogy and Structure. Dordrecht: Kluwer Academic. Skousen, R. (1998): “Natural statistics in language modelling”. In: Journal of Quantitative Linguistics 5, 246-255. Smolensky, P. (1988): “On the proper treatment of connectionism”. In: Behavioral and Brain Sciences, 11, 1-74. Smolensky, P. (1999): “Grammar-based connectionist approaches to language”. In: Cognitive Science, 23, 589-613. Sokolik, M. E. (1990): “Learning without rules: PDP and a resolution of the adult language learning paradox”. In: TESOL Quarterly, 24, 4, 685-696. Sokolik, M. E., M. E. Smith (1992): “Assignment of gender to French nouns in primary and secondary language: A connectionist model”. In: Second Language Research, 8, 39-58. Spencer, A., A. Zwicky (eds.) (2003): The Handbook of Morphology. Oxford: Blackwell. Stanners, R. F., J. J. Neiser, W. P. Hernon, R. Hall (1979): “Memory representation for morphologically related words“. In: Journal of Verbal Learning and Verbal Behaviour 18, 399-412. Stemberger, J. P. (1981): “Morphological haplology”. In: Language, 57, 791-817. Stemberger, J. P., B. MacWhinney (1986): “Frequency and the lexical storage of regularly inflected forms”. In: Memory and Cognition 14, 17-26. Storms, G., P. de Boeck, W. Ruts (2001): “Categorization of novel stimuli in wellknown natural concepts: A case study”. In Psychonomic Bulletin and Review 8, 2, 377-384. Stump, G. T. (2005): Inflectional Morphology. A Theory of Paradigm Structure. Cambridge Studies in Linguistics, 93. Cambridge: Cambridge University Press. Tabak, W., R. Schreuder, R. H. Baayen (2005): “Lexical statistics and lexical processing: semantic density, information complexity, sex, and irregularity in Dutch”. In: S. Kepser, M. Reis (eds.): Linguistic Evidence - Empirical, Theoretical, and Computational Perspectives. Berlin: Mouton de Gruyter. Tayler, L. K., P. de Mornay-Davies, R. Anokhina, C. Longworth, Billi Randall, W. D. Marslen-Wilson (2002): “Dissociations in processing past tense morphology: neuropathological and behavioural studies”. In: Journal of Cognitive Neuroscience, 14, 1, 79-94. Taylor, J. R. (2003): Linguistic Categorization. Prototypes in Linguistic Theory. 2 nd edition. Oxford: Oxford University Press. Thorndike, E. (1898): Animal intelligence. New York: Macmillan. Tomasello, M. (1995): “Language is not an instinct“. In: Cognitive Development 10, 131-156. 178 Ullman, M. T., S. Corkin, M. Coppola, G. Hickok, J. H. Growdon, W. J. Koroshetz, S. Pinker (1997): “A neural dissociation within language: Evidence that the mental dictionary is part of declarative memory, and that grammatical rules are processed by the procedural system”. In: Journal of Cognitive Neuroscience, 9, 2, 266- 276. Ullman, M. T. (2001): “A neurocognitive perspective on language: The declarative / procedural model”. In: Nature Reviews. Neuroscience, 2, 10, 717-726. Ulvestad, B. (1956): “The strong verb conjugation system in German”. In: Word, 12, 91-105. Vargha-Khadem, F., K. E. Watkins, C. J. Price, J. Ashburner, K. J. Alcock, A. Connelly, R. S. J. Frackowiak, K. J. Friston, M. E. Pembrey, M. Mishkin, D. G. Gadian, R. E. Passingham (1998): “Neural basis of an inherited speech and language disorder”. In: Proceedings of the National Academy of Sciences of the United States of America, 95, 21, 12695-12700. Vargha-Khadem, F., K. E. Watkins, K. Alcock, P. Fletcher, R. Passingham (1995): “Praxic and nonverbal cognitive deficits in a large family with a genetically transmitted speech and language disorder”. In: Proceedings of the National Academy of Sciences of the United States of America, 92, 3, 930-933. Vennemann, T. (1972): “Rule inversion”. In: Lingua 29, 3-4, 209-242. Vennemann, T. (2000): “Zur Entstehung des Germanischen“. In: Sprachwissenschaft, 25, 233-269. Wanner, D. (2006): The Power of Analogy. An Essay on Historical Linguistics. Trends in Linguistics. Studies and Monographs 170. Berlin: Mouton de Gruyter. Westermann, G. (2000): Constructivist Neural Network Models of Cognitive Development. PhD thesis, Division of Informatics, University of Edinburgh, Edinburgh, UK. Westermann, G., D. Willshaw, M. Penke (1999): “A constructivist neural network model of German verb inflection in agrammatic aphasia”. In: Proceedings of the 9th International Conference on Artificial Neural Networks, 916-921. Weyerts, H., M. Penke, U. Dohrn, H. Clahsen, T. F. Münte (1997): “Brain potentials indicate differences between regular and irregular German plurals“. In: Neuroreport 8, 957-962. Weyerts, H., H. Clahsen (1994): “Netzwerke und symbolische Regeln im Spracherwerb: Experimentelle Ergebnisse zur Entwicklung der Flexionsmorphologie”. In: Linguistische Berichte, 154, 430-460. Wiese, B. (2008): “Form and function of verbal ablaut in contemporary standard German”. In: Robin Sackmann (ed.): Explorations in Integrational Linguistics. Four Essays on German, French, and Guaraní. Current Issues in Linguistic Theory 285. Amsterdam, Philadelphia: Benjamins, 97-151. Wiese, R. (1996): “Phonological versus morphological rules: on German ablaut and umlaut“. In: Journal of Linguistics, 32, 113-135. Wiese, R. (2000): The Phonology of German. Oxford: Oxford University Press. Wittgenstein, L. (1953): Philosophical Investigations. Oxford: Basil Blackwell. Wulf, D. (1996): “An account of German plural formation using an analogical computer model”. In: A.-M. Mineur (ed.): Eighth European Summer School in Logic, Language and Information. Proceedings of the First Student Session. Wunderlich, D. (1996): “Minimalist morphology: The role of paradigms“. In: G. Booij, J. van Marle (eds.): Yearbook of Morphology 1995, 93-114. 179 Wunderlich, D., R. Fabri (1995): “Minimalist morphology: An approach to inflection“. In: Zeitschrift für Sprachwissenschaft, 14, 2, 236-294. Wurzel, W. U. (1970): Studien zur deutschen Lautstruktur. Studia Grammatica 8. Berlin: Akademie Verlag. Xu, F., S. Pinker (1992): “Weird past tense forms”. In: Journal of Child Language, 22, 531-556. Yang, C. (2004): “Universal Grammar, statistics or both? ” In: Trends in Cognitive Sciences 8, 10, 451-456. 180 Appendix 1 Questionnaires 1.1 English questionnaire for experiment 1 Dear Participants, A Clockwork Orange, this 1962 masterpiece by the English author Anthony Burgess, which was made into a film by Stanley Kubrick in 1971, largely consists of a special slang language, called nadsat. In order to make this slang more transparent to readers, some people have developed a glossary for nadsat words (http: / / wso.williams.edu/ ~mhacker/ clockglossary. html). A group of researchers now wants to enhance this glossary and publish a printed version, so they investigate the slang words in detail. This is where this study comes in. We have designed a little questionnaire to check, how native speakers of English read and understand these slang verbs just by their first appearance in the text. Take an example. In Part 1, Chapter 3, it reads: While I was doing this, a [...] little dwarf [...] ittied in, selling the morning´s gazettas, [...] To “itty” would be one of those verbs in question, and we are investigating how a reader might be dealing with words like this. The following text contains passages from A Clockwork Orange, in which the 15-year-old protagonist Alex - amongst other things - reports about events in the past. Hence, the verbs in these passages have to be in the past tense. In front of the gaps, we distributed infinitive forms of slang verbs from the online glossary, and you will be asked to fill in past tense forms accordingly. Here is an example. You will see a sentence like this: So we ( shig ) ________________ him lovely, grinning all over our litsos, but he still went on singing. [Part 1, Chapter 2] Then you just fill in a form that you deem appropriate in this context. Please try to fill in all the gaps. Follow your intuition and the “sound” of the words. This experiment is not about grammaticality at all, thus there are no wrong forms in this study! In order to give you an idea of the verbs’ sound, a tape will be playing simultaneously, and the text will be read out. There will be a little pause at each gap for you to fill in the past tense form. Thanks a lot for your kind participation Two items for practice: 181 We will start with two short items for practice. In the box further down the page you can see two short sentences. These will be read out entirely in a moment, and right after the verbs in brackets, there will be a little pause (approximately 10 seconds). In these pauses you can fill in your past tense forms. Please listen to the tape, read along, and fill in the gaps: So we (1 shig) ________________ him lovely, grinning all over our litsos, but he still went on singing. Then we (2 strow ) ________________ him so he laid down flat and heavy, and a bucketload of beer-vomit came whooshing out. Now you should have filled in the gaps. If you have any questions, please let the instructor know, and we will stop the tape. If there are no further questions, we will proceed to the text proper overleaf. In the text, we distributed 40 of these verbs. As with the two items for practice, you will hear the whole text being read out, and in the pauses you can fill in your past tense forms. Please listen to the tape, read along, and fill in the gaps: We (1 smink) ______ townwards, my brothers, but just outside, not far from what they called the Industrial Canal, we saw that the fuel needle (2 skig) ______, like our own needles had. The auto made noises and (3 wred) ______ kashl kashl kashl. Not to worry overmuch, though, because a rail station (4 skring) ______ blue just near. The question was whether to leave the auto to be sobiratted by the rozzes or, us feeling like in a hate and murder mood, to give it a fair tolchock into the starry waters for a nice heavy loud plesk before the death of the evening. This latter we (5 sprin) ______ so we (6 vareet) ______ out and, the brakes off, all four (7 smick) ______ it to the edge of the filthy water that was like treacle mixed with human hole products, then one good horrorshow tolchock and in she (8 skrink) ______. We (9 smot) ______ back for fear of the filth splashing on our platties, but splussshhhh and glolp she (10 tring) ______ down and lovely. “Farewell, old droog”, called Georgie, and Dim (11 shat) ______ with a clowny great guff - “Huh huh huh huh.” Then we (12 platch) ______ the station to ride the one stop to Center, as the middle of the town was called. We (13 skrim) ______ our fares and (14 sprit) ______ strangely and quiet on the platform, old Dim fillying with the slot machines, his carmans being full of small malenky coin, and ready if need be to distribute chocbars to the poor and starving, though there was none such about, and then the old espresso rapido came lumbering in and we (15 cupet) ______ the train looking to be near empty. To pass the three-minute ride we (16 strink) ______ the upholstery, doing some nice horrorshow tearing-out of the seats’ guts and old Dim chaining the okno till the glass (17 crast) ______ and (18 glink) ______ in the winter air. We got out at Center and (19 stin) ______ slow to the Korova Milckbar, all going yawwwww a malenky bit and exhibiting to moon and star and lamplight our back fillings, because we were still only growing malchicks 182 and had school in the daytime, and when we got into the Korova, we found it fuller than we’d left it earlier on. But the chelloveck, who (20 fash) ______ away yesterday, in the land, on white and synthemesc or whatever, was still on at it. Anyway, Dim (21 spling) ______ in next to this veck and, with his big clown’s yawp that (22 creetch) ______ his hanging grape, he (23 skrit) ______ this veck’s foot with his own large filthy sabog. But the veck, my brothers, heard nought, being now all above the body. The devotchkas among them (24 strig) ______ these very lively litsos and wide big rots, very red, showing a lot of teeth, and smecking away and not caring about the wicked world one whit. And then the disc on the stereo (25 sprink) ______ off and out (it was Johnny Zhivago, a Russky koshka, singing Only Every Other Day), and in the like interval, the short silence before the next one (26 lovet) ______ on, one of these devotchkas - very fair and with a big smiling red rot and in her late thirties I’d say - suddenly (27 viddy) ______ with a burst of singing, only a bar and a half and as though she was like giving an example of something they’d all been govoreeting about, and it was like for a moment, O my brothers, some great bird had flown into the milkbar, and I (28 pling) ______ all the little malenky hairs on my plot standing endwise and the shivers crawling up like slow malenky lizards and then down again. Anyway, I (29 spick) ______. But old Dim, as soon as he’d slooshied this dollop of song like a lomtick of redhot meat plonked on your plate, (30 frink) ______ off one of his vulgarities, which in this case was a lip-trump followed by a dog-howl followed by two fingers pronging twice at the air followed by a clowny guffaw. I (31 plosh) ______ myself all of a fever and like drowning in redhot blood, slooshying and viddying Dim’s vulgarity, and I said: “Bastard. Filthy drooling mannerless bastard.” Then I (32 spriff) ______ across Georgie, who was between me and horrible Dim, and (33 sming) ______ Dim skorry on the rot. Dim (34 spink) ______ very surprised, his rot open, wiping the krovvy off his gobber with his rook and in turn looking surprised at the red flowing krovvy and at me. “What did you do that for? ”, he said in his ignorant way. Not many people (35 cheest) ______ what I did, and those who (36 spim) ______ cared not. The stereo (37 chumble) ______ on again and was playing a very sick electronic guitar veshch. “For being a bastard with no manners and not the dook of an idea how to comport yourself publicwise, O my brother.” Dim (38 krink) ______ a hound-and-horny look of evil, saying: “I don’t like you should do what you done then. And I’m not your brother no more and wouldn’t want to be.” He’d taken a big snotty tashtook from his pocket and (39 brosay) ______ the red flow puzzled, keeping on looking at it frowning as if he thought that blood was for other vecks and not for him. But that devotchka (40 sking) ______ now with her droogs at the bar, her red rot working and her zoobies ashine, not having noticed Dim’s filthy vulgarity. 183 1.2 English questionnaire for experiment 2 Dear participant, in this little experiment we are trying to investigate how words are stored and organised in our so-called mental lexicon. In particular, we want to see how people treat verbs. As you know, verbs can have different forms, and this experiment is about past tense and infinitive forms of English verbs. In order to make this study a bit more interesting, we constructed completely new verbs, sometimes called nonce-verbs, and we would like you to form the above-mentioned two forms from these new words. Take an example: The word fung is given, and then we will ask you to provide the infinitive and the past tense form. We would do this in a frame or short scene like: Have you ever fung? Yesterday I saw Sam; he ______________ the whole evening. I think, Sam likes to______________ . Here, the first gap should be filled with a past tense form, and the second gap should contain the infinitive. There is no wrong answer to this, just follow your intuition and focus on the sound of the words. There will be a tape running simultaneously, and you can hear the new verbs being read out, so that you have an idea what they might sound like. There will be 40 of these new verbs, and the whole procedure will not last longer than 20 min. Thanks a lot for your kind participation Two items for practice We will start with two short items for practice. In the box further down the page you can see three short sentences. In a moment, the number (1) and the new participle in bold print will be read out, and right after the participle in bold print, there will be a little pause (approximately 10 seconds). In this pause you can fill in your past tense form and the infinitive. Please listen to the tape, read along, and fill in the gaps: (1) Have you ever plunk? Yesterday I saw Sam; he ______________ the whole evening. I think, Sam likes to______________ . Now you should have filled in the gaps. If you have any questions, please let the instructor know, and we will stop the tape. If there are no further questions, we will proceed to the 40 mini texts overleaf. As with the two items for practice, you will hear the number and the participle being read out, and in the pauses you can fill in your past tense form and the infinitive. Please listen to the tape, read along, and fill in the gaps: 184 1. Have you ever smunk? Yesterday I saw Sam; he _____ the whole evening. I think, Sam likes to_____. 2. Have you ever skug? Yesterday I saw Abraham; he _____ the whole evening. I think, Abraham likes to_____. 3. Have you ever nillt? Yesterday I saw Bernadette; she _____ the whole evening. I think, Bernadette likes to_____. 4. Have you ever skrung? Yesterday I saw Bertram; he _____ the whole evening. I think, Bertram likes to_____. 5. Have you ever sprun? Yesterday I saw Angela; she _____ the whole evening. I think, Angela likes to_____. 6. Have you ever keft? Yesterday I saw Chris; he _____ the whole evening. I think, Chris likes to_____. 7. Have you ever smuck? Yesterday I saw Bridget; she _____ the whole evening. I think, Bridget likes to_____. 8. Have you ever skrunk? Yesterday I saw Clarence; he _____ the whole evening. I think, Clarence likes to_____. 9. Have you ever skit? Yesterday I saw Clara; she _____ the whole evening. I think, Clara likes to_____. 10. Have you ever trung? Yesterday I saw David; he _____ the whole evening. I think, David likes to_____. 11. Have you ever pought? Yesterday I saw Eleanor; she _____ the whole evening. I think, Eleanor likes to_____. 12. Have you ever gaught? Yesterday I saw Francis; he _____ the whole evening. I think, Francis likes to_____. 13. Have you ever skrum? Yesterday I saw Frederick; he _____ the whole evening. I think, Frederick likes to_____. 14. Have you ever sprut? Yesterday I saw Flora; she _____ the whole evening. I think, Flora likes to_____. 15. Have you ever belt? Yesterday I saw Graham; he _____ the whole evening. I think, Graham likes to_____. 16. Have you ever strunk? Yesterday I saw Georgina; she _____ the whole evening. I think, Georgina likes to_____. 17. Have you ever stat? Yesterday I saw Hugh; he _____ the whole evening. I think, Hugh likes to_____. 18. Have you ever glunk? Yesterday I saw John; he _____ the whole evening. I think, John likes to_____. 19. Have you ever stun? Yesterday I saw Jennifer; she _____ the whole evening. I think, Jennifer likes to_____. 20. Have you ever quet? Yesterday I saw Joan; she _____ the whole evening. I think, Joan likes to_____. 21. Have you ever splung? Yesterday I saw Jack; he _____ the whole evening. I think, Jack likes to_____. 22. Have you ever stought? Yesterday I saw Karen; she _____ the whole evening. I think, Karen likes to_____. 23. Have you ever skrut? Yesterday I saw Claudia; she _____ the whole evening. I think, Claudia likes to_____. 24. Have you ever strug? Yesterday I saw Leonard; he _____ the whole evening. I think, Leonard likes to_____. 185 25. Have you ever sprunk? Yesterday I saw Louise; she _____ the whole evening. I think, Louise likes to_____. 26. Have you ever wuilt? Yesterday I saw Mike; he _____ the whole evening. I think, Mike likes to_____. 27. Have you ever haw? Yesterday I saw Mary; she _____ the whole evening. I think, Mary likes to_____. 28. Have you ever pling? Yesterday I saw Michael; he _____ the whole evening. I think, Michael likes to_____. 29. Have you ever spuck? Yesterday I saw Matthew; he _____ the whole evening. I think, Matthew likes to_____. 30. Have you ever frunk? Yesterday I saw Mildred; she _____ the whole evening. I think, Mildred likes to_____. 31. Have you ever nept? Yesterday I saw Sam; he _____ the whole evening. I think, Sam likes to_____. 32. Have you ever spruff? Yesterday I saw Norris; he _____ the whole evening. I think, Norris likes to_____. 33. Have you ever smung? Yesterday I saw Nick; he _____ the whole evening. I think, Nick likes to_____. 34. Have you ever spunk? Yesterday I saw Nora; she _____ the whole evening. I think, Nora likes to_____. 35. Have you ever pit? Yesterday I saw Nina; she _____ the whole evening. I think, Nina likes to_____. 36. Have you ever spum? Yesterday I saw George; he _____ the whole evening. I think, George likes to_____. 37. Have you ever trept? Yesterday I saw James; he _____ the whole evening. I think, James likes to_____. 38. Have you ever krunk? Yesterday I saw Barbara; she _____ the whole evening. I think, Barbara likes to_____. 39. Have you ever haught? Yesterday I saw Christina; she _____ the whole evening. I think, Christina likes to_____. 40. Have you ever skung? Yesterday I saw Olga; she _____ the whole evening. I think, Olga likes to_____. 1.3 English questionnaire for experiment 3 Dear participant, please read the following little piece of text, and fill in the blank with the form of the highlighted word that you think is appropriate to the context. It is important that the form you choose matches the context of the sentence. Concentrate on how the new form of the novel word “sounds” in the context, not on how it might be spelled. Here is an example: A single wucterium can be very dangerous for humans. When they breed and multiply, a lot of _______________ can kill humans. In this example, you could fill the gap with wucteria or wucteriums, or anything else, depending on what seems appropriate to you. At a traditional party at a Russian hospital, the ill patients all frink in the onset of good weather, consuming a lot of vodka. Last year, patient Ivan Borovich _______________ around 35 glasses of vodka. He is better now. 186 In a classical symptom of Howson’s syndrome, patients all frink in their right eye if they are left handed or left eye if right handed, their eyelids opening and closing rapidly and uncontrollably. Last year, Howson’s patient Ivan Borovich _______________ around 35 times per minute for two days. He is better now. In a new therapy at a Russian hospital, the patients all frink in pairs of two like cowboys on horses. Last year, patient Ivan Borovich _______________ around on a rather energetic horse on his own, when he fell off and hurt himself. He is better now. In an alternative therapy at a Russian hospital, the patients all frink in the afternoons, going into a trance like state that lowers the heartbeat to alleviate pain. Last year, emitting a steady, low humming sound, patient Ivan Borovich _______________ around 3 hours. He is better now. 2 Test items Table 1: English nonce infinitive test items and fillers for the pilot study. Nonce infinitives taken directly from Plag (2000) appear in italics 1. kring 34. sprat 2. spling 35. slan 3. skrink 36. smip 4. sprink 37. ping 5. strink 38. ving 6. skrim 39. shink 7. sprin 40. tink 8. sking 41. glick 9. sming 42. trig 10. smink 43. clid 11. spink 44. tim 12. strig 45. vin 13. skrit 46. tang 14. sprit 47. tran 15. spim 48. trib 16. strank 49. sig 17. stin 50. wick 18. splam 51. prike 19. spriv 52. swike 20. pling 53. sprike 21. tring 54. sid 22. glink 55. stap 23. krink 56. trad 24. skig 57. tam 25. smick 58. kib 26. spick 59. krag 27. stid 60. gack 187 28. krin 61. spreen 29. plim 62. tad 30. skeen 63. skreek 31. spang 64. spak 32. glank 65. steeg 33. scrag Table 2: English nonce infinitive test items and fillers for experiment 1 order 1 order 2 order 1 order 2 1. smink sking 21. spling spling 2. skig krink 22. creech creech 3. wred wred 23. skrit stin 4. skring spim 24. strig glink 5. sprin spink 25. sprink strink 6. vareet vareet 26. lovet lovet 7. smick sming 27. viddy viddy 8. skrink spriff 28. pling sprit 9. smot smot 29. spick skrim 10. tring frink 30. frink tring 11. shat shat 31. plosh plosh 12. platch platch 32. spriff skrink 13. skrim spick 33. sming smick 14. sprit pling 34. spink sprin 15. cupet cupet 35. cheest cheest 16. strink sprink 36. spim skring 17. crast crast 37. chumble chumble 18. glink strig 38. krink skig 19. stin skrit 39. brosay brosay 20. fash fash 40. sking smink Table 3: English nonce participle test items and fillers for experiment 2 order 1 order 2 order 1 order 2 1. smunk skung 21. splung splung 2. skug krunk 22. stought stought 3. nillt nillt 23. skrut skrut 4. skrung spum 24. strug strug 5. sprun spunk 25. sprunk sprunk 6. keft keft 26. wuilt wuilt 7. smuck smung 27. haw haw 8. skrunk spruff 28. plung plung 9. skit skit 29. spuck spuck 10. trung frunk 30. frunk frunk 11. pought pought 31. nept nept 12. gaught gaught 32. spruff spruff 13. skrum spuck 33. smung smung 14. sprut plung 34. spunk spunk 15. belt belt 35. pit pit 16. strunk sprunk 36. spum spum 17. stat stat 37. trept trept 188 18. glunk strug 38. krunk krunk 19. stun skrut 39. haught haught 20. quet quet 40. skung skung Table 4: English nonce infinitive test items and fillers for experiment 2 order 1 order 2 order 1 order 2 1. smink sking 21. spling spling 2. skig krink 22. steach steach 3. nill nill 23. skrit stin 4. skring spim 24. strig glink 5. sprin spink 25. sprink strink 6. keave keave 26. wuild wuild 7. smick sming 27. hee hee 8. skrink spriff 28. pling sprit 9. skit skit 29. spick skrim 10. tring frink 30. frink tring 11. peek peek 31. neep neep 12. geach geatch 32. spriff skrink 13. skrim spick 33. sming smick 14. sprit pling 34. spink sprin 15. beal beal 35. pit pit 16. strink sprink 36. spim skring 17. stit stit 37. treap treap 18. glink strig 38. krink skig 19. stin skrit 39. heach heach 20. queet queet 40. sking smink Table 5: German nonce infinitive test items and fillers for experiment 1 order 1 order 2 order 1 order 2 1. knießen frinken 21. schliegen sprinken 2. schrimmen schminnen 22. pitschen pitschen 3. nisseln nisseln 23. schingen fießen 4. spinken schnießen 24. fießen schingen 5. grießen linnen 25. lichsen lichsen 6. pieben pieben 26. liepen liepen 7. stinnen friegen 27. sprinken schliegen 8. biechen kingen 28. fiechen spingen 9. witten witten 29. stingen strimmen 10. strinken pliegen 30. triechen stiechen 11. stiemen stiemen 31. bristen bristen 12. stitzen stitzen 32. strießen schringen 13. schiechen schinnen 33. schinnen schiechen 14. schringen strießen 34. pliegen strinken 15. miegeln miegeln 35. fritteln fritteln 16. stiechen triechen 36. kingen biechen 17. hitteln hitteln 37. friegen stinnen 18. strimmen stingen 38. krietschen krietschen 19. spingen fiechen 39. linnen grießen 20. sicken sicken 40. schnießen spinken 189 41. fietzen fietzen 42. schminnen schrimmen 43. frinken knießen Table 6: German nonce past tense test items and fillers for experiment 2 order 1 order 2 order 1 order 2 1. knoss frank 21. schlog sprank 2. schramm schmann 22. pitschte pitschte 3. nisselte nisselte 23. schang foss 4. spank schnoss 24. foss schang 5. gross lann 25. lichste lichste 6. piebte piebte 26. liepte liepte 7. stann frog 27. sprank schlog 8. boch kang 28. foch spang 9. wittete wittete 29. stang stramm 10. strank plog 30. troch stoch 11. stiemte stiemte 31. bristete bristete 12. stitzte stitzte 32. stross schrang 13. schoch schann 33. schann schoch 14. schrang stross 34. plog strank 15. miegelte miegelte 35. frittelte frittelte 16. stoch troch 36. kang boch 17. hittelte hittelte 37. frog stann 18. stramm stang 38. krietschte krietschte 19. spang foch 39. lann gross 20. sickte sickte 40. schnoss spank 41. fietzte fietzte 42. schmann schramm 43. frank knoss Table 7: German nonce infinitive test items and fillers for experiment 2 order 1 order 2 order 1 order 2 1. knießen frinken 21. schliegen sprinken 2. schrimmen schminnen 22. pitschen pitschen 3. nisseln nisseln 23. schingen fießen 4. spinken schnießen 24. fießen schingen 5. grießen linnen 25. lichsen lichsen 6. pieben pieben 26. liepen liepen 7. stinnen friegen 27. sprinken schliegen 8. biechen kingen 28. fiechen spingen 9. witten witten 29. stingen strimmen 10. strinken pliegen 30. triechen stiechen 11. stiemen stiemen 31. bristen bristen 12. stitzen stitzen 32. strießen schringen 13. schiechen schinnen 33. schinnen schiechen 14. schringen strießen 34. pliegen strinken 15. miegeln miegeln 35. fritteln fritteln 16. stiechen triechen 36. kingen biechen 17. hitteln hitteln 37. friegen stinnen 190 18. strimmen stingen 38. krietschen krietschen 19. spingen fiechen 39. linnen grießen 20. sicken sicken 40. schnießen spinken 41. fietzen fietzen 42. schminnen schrimmen 43. frinken knießen Table 8: Nonce infinitive test items as used by Bybee and Moder (1983), Prasada and Pinker (1993), and Plag (2000), organised according to schmas schema Bybee & Moder artificial words Bybee & Moder real words Plag 2000 artificial words Prasada & Pinker 1993, artificial words (1) sCC_ _ ( ) 1. spling 2. skrink 3. sprink 4. strink 1. skrink 2. sprink 1. spling 2. skring 3. sprink (2) sCC_V_ ( ) 5. strank 6. scronk 3. strunk (3) sCC_ _ / 7. strick 8. strig 4. strick 5. strig (4) sCC_V_ / 9. sprayk 10. spreyk 11. sprock 12. sprook 13. streyk 14. strug 15. skreek 16. scrag streak stroke 6. struk (5) sCC_ _ " / 17. skrim 18. sprin 7. sprin 8. skrim (6) sCC_ _C 19. spriv 20. skrit 9. spriv 10. skrit 11. sprit (7) sCC_ V_ C 21. sprat 12. splud 13. splug 14. skrug 15. strul 16. spreen 17. sprun 18. splun 19. skrun 20. scrun 21. strum (8) sCC_V_ " / 22. splam 23. spreen (9) sC_ _ ( ) 24. sming 25. sking 26. smink 27. spink sting slink 22. sming 23. sking 24. spink 191 (10) sC_V_ ( ) 28. spang 25. spunk (11) sC_ _ / 29. smick 30. skig stick 26. smick 27. skig (12) sC_V_ / 31. steeg 32. skeyk 33. stayk 34. spug 35. spock 36. spak stack 28. steeg 29. spuk 30. skug (13) sC_ _ " / 37. spim 38. stin skim 31. spim 32. stin 33. spum (14) sC_V_ " / 39. slan 40. skeen 34. slun 35. stun 36. skun 37. skeen (15) sC_ _C(C) 41. stid 42. smip skid 38. smip 39. stid (16) sC_V_C(C) 43. stom 44. stap 45. speeb 46. skeyp 47. skeep 40. stug 41. stul 42. speeb 43. slub 4. smeeb 5. smeeg 6. smeej 7. smeelth 8. smeenth 9. smaib 10. smaig 11. smeerg 12. smairg 13. smairph (17) CC_ _ ( ) 48. pling 49. tring 50. krink 51. glink fling 44. pling 45. tring 46. krink 14. fring 15. ning 16. frink (18) CC_V_ 52. glank (19) CC_ _ / 53. trig 54. glick trick 47. glick 48. trig (20) CC_V_ / 55. krag 56. treek 57. playk 58. glok 59. trook 60. treyk 61. kreeg (21) CC_ _ " / 62. krin 63. plim 49. krin 50. plim (22) CC_V_ " / 64. greem 65. tran 51. greem 52. prun 53. grun 54. trun 55. flum 17. greem 18. pleem 19. treem 192 (23) CC_ _C(C) 66. trib 67. clid trip clip 56. trib 57. clid 20. trisp 21. blip 22. plip 23. glinth 24. frilg 25. krilg 26. glip 27. brip 28. brilth 29. plimph 30. trilb (24) CC_V _ C(C) 68. treep 69. trad 58. treep 59. brug 60. plut 61. flug 62. flus 31. cleed 32. preed 33. queed 34. cleef 35. preek 36. ploab 37. ploamph 38. flape 39. blafe 40. gloke 41. proke 42. slace 43. ploag 44. ploanth (25) C_ _ ( ) 70. shink 71. tink 72. ping 73. ving link 63. shink 64. ping 65. tink (26) C_V_ ( ) 74. tang 66. lunk 67. funk (27) C_ _k/ g 75. gick 76. sig dig 68. gick 69. sig (28) C_V_g/ k 77. pook 78. deyk 79. gack 80. gok 81. seeg 70. seeg (29) C_ _n/ m 82. tim 83. vin 71. tim 72. vin (30) C_V_n/ m 84. tam 85. veem 73. rum 74. veem (31) C_ _C(C) 86. sid 87. kib bid tip 75. sid 76. kib 45. nist (32) C_V_C 88. toop 89. bive 90. peet 91. seyb 92. dop 93. tad 77. lug 78. bun 79. lun 80. lup 81. puv 82. thul 83. peet 46. queef 47. gleef 48. keeb 49. meep 50. goav 51. joam 52. nace 193 (33) C(C)_V 53. cloe 54. froe 55. plare 56. quare 57. foa 58. voa 59. jare 60. grare