The semantic transparency of English compound nouns Martin Schäfer Morphological Investigations 3 language science press Morphological Investigations Editors: Jim Blevins, Petar Milin, Michael Ramscar In this series: 1. Trips, Carola & Jaklin Kornfilt (eds.). Further investigations into the nature of phrasal compounding. 2. Baechler, Raffaela. Absolute Komplexität in der Nominalflexion. 3. Schäfer, Martin. The semantic transparency of English compound nouns. The semantic transparency of English compound nouns Martin Schäfer language science press Martin Schäfer. 2018. The semantic transparency of English compound nouns (Morphological Investigations 3). Berlin: Language Science Press. This book is the revised version of the author’s habilitation, Friedrich-Schiller-Universität Jena, 2017 This title can be downloaded at: http://langsci-press.org/catalog/book/153 © 2018, Martin Schäfer Published under the Creative Commons Attribution 4.0 Licence (CC BY 4.0): http://creativecommons.org/licenses/by/4.0/ ISBN: 978-3-96110-030-9 (Digital) 978-3-96110-031-6 (Hardcover) DOI:10.5281/zenodo.1134595 Source code available from www.github.com/langsci/153 Collaborative reading: paperhive.org/documents/remote?type=langsci&id=153 Cover and concept of design: Ulrike Harbort Typesetting: Martin Schäfer Proofreading: Plinio A. Barbosa, Jose Poblete Bravo, Merlijn Breunesse, Stefan Hartmann, Martin Hilpert, Gianina Iordachioaia, Timm Lichte, Ahmet Bilal Özdemir, Steve Pepper, Katja Politt, Valeria Quochi, Edalat Shekari, Andrew Spencer, Carola Trips, Jeroen van de Weijer, Amr Zawawy Fonts: Linux Libertine, Arimo, DejaVu Sans Mono Typesetting software: XƎL A TEX Language Science Press Unter den Linden 6 10099 Berlin, Germany langsci-press.org Storage and cataloguing done by FU Berlin Dedicated to the next generation (in order of appearance): Charlotte, Henriette, Anton, Magdalena, Moritz, Henrike, Emma, Lene, Mathilde, Marie, Simon, Anne, Theo, Ole, Jakob Contents Acknowledgments xi Abbreviations xiii 1 Introduction 1 1.1 A first notion of semantic transparency . . . . . . . . . . . . . . 1 1.2 Compounds and complex nominals . . . . . . . . . . . . . . . . 3 1.3 Aims and Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1.4 Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2 Semantic transparency in psycholinguistics 7 2.1 Structure and lexical access . . . . . . . . . . . . . . . . . . . . . 7 2.1.1 Morpheme-based models . . . . . . . . . . . . . . . . . . 8 2.1.2 Models without morphemes . . . . . . . . . . . . . . . . 16 2.1.3 Models of conceptual combination . . . . . . . . . . . . 24 2.1.4 Conclusion: the different models . . . . . . . . . . . . . 29 2.2 Measuring semantic transparency . . . . . . . . . . . . . . . . . 29 2.2.1 Establishing semantic transparency . . . . . . . . . . . . 30 2.2.2 Summary: measuring semantic transparency . . . . . . 35 2.3 Psycholinguistic studies . . . . . . . . . . . . . . . . . . . . . . . 37 2.3.1 Priming paradigms . . . . . . . . . . . . . . . . . . . . . 37 2.3.2 Eye movement studies . . . . . . . . . . . . . . . . . . . 43 2.3.3 Experiments targeting conceptual combination . . . . . 46 2.3.4 Overview: experimental traces of semantic transparency 52 2.3.5 Conclusion: experimental traces of semantic transparency 52 2.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 3 Related phenomena and notions 59 3.1 Semantic transparency reflected in other linguistic phenomena . 59 3.1.1 Semantic transparency and outbound anaphora . . . . . 59 3.1.2 Semantic transparency and compound stress . . . . . . 64 3.1.3 Conclusion: semantic transparency and other phenomena 65 Contents 3.2 Other measures and notions . . . . . . . . . . . . . . . . . . . . 66 3.2.1 Quantitative measures . . . . . . . . . . . . . . . . . . . 66 3.2.2 Semantic overlap . . . . . . . . . . . . . . . . . . . . . . 66 3.2.3 Compositionality and literality . . . . . . . . . . . . . . 66 3.2.4 Semantic transparency as one dimension of idiomaticity 71 3.2.5 Semantic transparency and productivity . . . . . . . . . 72 3.3 Transparency in other domains . . . . . . . . . . . . . . . . . . 73 3.3.1 Phonological transparency . . . . . . . . . . . . . . . . . 73 3.3.2 Orthographic transparency . . . . . . . . . . . . . . . . 74 3.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 4 The semantic analysis of compounds 77 4.1 Set-theoretic approaches . . . . . . . . . . . . . . . . . . . . . . 78 4.1.1 Intersective modification . . . . . . . . . . . . . . . . . . 78 4.1.2 Subsective modification . . . . . . . . . . . . . . . . . . 79 4.1.3 Non-subsective modification . . . . . . . . . . . . . . . . 81 4.1.4 Problems for a set-theoretic classification of adjectives . 82 4.2 Relation-based approaches: the semantics of compounds . . . . 86 4.3 Levi (1978) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 4.3.1 Levi’s complex nominals . . . . . . . . . . . . . . . . . . 87 4.3.2 Levi’s recoverably deletable predicates . . . . . . . . . . 89 4.3.3 Predicate nominalization . . . . . . . . . . . . . . . . . . 93 4.3.4 Evaluating Levi’s approach . . . . . . . . . . . . . . . . 97 4.3.5 Conclusion: the enduring appeal of Levi’s system . . . . 99 4.4 Fanselow (1981) . . . . . . . . . . . . . . . . . . . . . . . . . . . 100 4.4.1 Compounds involving relational nouns . . . . . . . . . . 100 4.4.2 Determinative compounds . . . . . . . . . . . . . . . . . 103 4.4.3 Evaluating Fanselow’s approach . . . . . . . . . . . . . . 116 4.5 Mixed approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 117 4.5.1 Pustejovsky (1995) . . . . . . . . . . . . . . . . . . . . . 117 4.5.2 Extending the analysis to compounds 1: Jackendoff (2010) 121 4.5.3 Extending the analysis to compounds 2: Asher (2011) . . 122 4.5.4 Approaches using underspecification . . . . . . . . . . . 124 4.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 5 Previous models 127 5.1 Distributional semantics and word space models . . . . . . . . . 127 5.1.1 The basics of distributional semantics: a toy example . . 128 5.1.2 Design decisions . . . . . . . . . . . . . . . . . . . . . . 131 vi Contents 5.1.3 Two implementations: LSA and HAL . . . . . . . . . . . 132 5.1.4 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . 133 5.2 Reddy, McCarthy & Manandhar (2011) . . . . . . . . . . . . . . . 133 5.2.1 Selection procedure . . . . . . . . . . . . . . . . . . . . . 134 5.2.2 Reddy et al.’s human judgment data . . . . . . . . . . . . 136 5.2.3 Reddy et al.’s distributional semantics models . . . . . . 138 5.3 Pham and Baayen (2013) . . . . . . . . . . . . . . . . . . . . . . 142 5.3.1 Informativity based measures . . . . . . . . . . . . . . . 143 5.3.2 Pham and Baayen: compound selection and variable coding 149 5.3.3 Study 3: transparency rating experiment . . . . . . . . . 152 5.4 Marelli et al. (2015) . . . . . . . . . . . . . . . . . . . . . . . . . 154 5.4.1 Experiment 1: connotations . . . . . . . . . . . . . . . . 156 5.4.2 Experiment 2: semantic processing . . . . . . . . . . . . 157 5.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 6 Piloting semantic factors 161 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 6.2 The Reddy et al. data: a descriptive overview . . . . . . . . . . . 162 6.2.1 Linguistic characterization of the selected compounds 162 6.2.2 Descriptive overview of the rating data . . . . . . . . . . 166 6.3 Bell & Schäfer (2013) . . . . . . . . . . . . . . . . . . . . . . . . 170 6.3.1 Subsetting the Reddy et al. dataset . . . . . . . . . . . . 170 6.3.2 Semantic annotation of the compounds . . . . . . . . . . 171 6.3.3 Annotation results . . . . . . . . . . . . . . . . . . . . . 173 6.3.4 Bell and Schäfer (2013): the models . . . . . . . . . . . . 178 6.4 Bell & Schäfer (2013) revisited . . . . . . . . . . . . . . . . . . . 186 6.4.1 Classic model criticism . . . . . . . . . . . . . . . . . . . 187 6.4.2 Linear mixed effects modeling . . . . . . . . . . . . . . . 193 6.4.3 The role of the meaning shifts . . . . . . . . . . . . . . . 206 6.5 Conclusion and consequences . . . . . . . . . . . . . . . . . . . 213 7 Compound family based models 215 7.1 Semantic relations relative to constituent families . . . . . . . . 216 7.1.1 Gagné and Shoben . . . . . . . . . . . . . . . . . . . . . 217 7.1.2 Criticism and a corpus-based re-implementation . . . . 218 7.1.3 Relational distributions in other studies . . . . . . . . . 222 7.1.4 Conclusion: relations relative to families . . . . . . . . . 222 7.2 Assessing the role of constituent meanings . . . . . . . . . . . . 223 vii Contents 7.3 A database of compound families . . . . . . . . . . . . . . . . . 226 7.3.1 Initial families from the BNC . . . . . . . . . . . . . . . 226 7.3.2 Adding items from CELEX . . . . . . . . . . . . . . . . . 227 7.3.3 Usenet frequencies . . . . . . . . . . . . . . . . . . . . . 228 7.3.4 Further post-processing . . . . . . . . . . . . . . . . . . 229 7.4 Semantic coding . . . . . . . . . . . . . . . . . . . . . . . . . . . 229 7.4.1 Coding the semantic relations . . . . . . . . . . . . . . . 230 7.4.2 Coding the constituent senses . . . . . . . . . . . . . . . 231 7.5 Variables and predictions . . . . . . . . . . . . . . . . . . . . . . 233 7.5.1 Variables derived from the semantic coding . . . . . . . 233 7.5.2 Further explanatory variables . . . . . . . . . . . . . . . 235 7.5.3 Tabular overview of the explanatory variables . . . . . . 236 7.5.4 Restricting the target dataset . . . . . . . . . . . . . . . 236 7.5.5 Predicting semantic transparency . . . . . . . . . . . . . 237 7.6 The models from Bell & Schäfer 2016 . . . . . . . . . . . . . . . 240 7.6.1 N1 transparency . . . . . . . . . . . . . . . . . . . . . . . 241 7.6.2 N2 transparency . . . . . . . . . . . . . . . . . . . . . . 244 7.6.3 Whole compound transparency . . . . . . . . . . . . . . 246 7.6.4 The 2016 models: discussion and conclusion . . . . . . . 251 7.7 Re-modeling Bell & Schäfer (2016) . . . . . . . . . . . . . . . . . 252 7.7.1 New models for constituent transparency . . . . . . . . 258 7.7.2 New models for compound transparency . . . . . . . . . 261 7.7.3 Conclusion: re-modeling Bell & Schäfer (2016) . . . . . . 262 7.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263 8 Summary and outlook 265 8.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265 8.2 Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266 Appendix A: Semantic coding for Bell & Schäfer (2013) 269 1 Relations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269 2 Shifts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271 Appendix B: Semantic coding for Bell & Schäfer (2016) 273 1 N1 families . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274 2 N2 families . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314 Appendix C: Multiple readings and the 2016 coding 363 viii Contents Appendix D: Corpus and dictionary sources 369 1 Corpus identifiers . . . . . . . . . . . . . . . . . . . . . . . . . . 369 2 Online dictionaries . . . . . . . . . . . . . . . . . . . . . . . . . 374 References 375 Index 391 Name index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 391 Language index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395 Subject index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397 ix Acknowledgments My first thanks go to Barbara Schlücker. She initiated my work on compounds by suggesting, quite insistently, that I should submit an abstract to her and Matthias Hüning’s Naming Strategies workshop in 2008. Likewise, it was her initiative which brought me to a workshop on Meaning and Lexicalization of Word For- mation at the 14th International Morphology Meeting, Budapest, where I first met Sabine Arndt-Lappe. Both have been the best of colleagues, providing not only linguistic feedback, but also all-purpose advice and motivation. In 2011 I first met Melanie Bell when we both gave talks on English compounds at the 4th International Conference on the Linguistics of Contemporary English at the Universität Osnabrück. That was the starting point of a still ongoing collab- oration between the two of us, and our discussions and work together crucially shaped my thinking about semantic transparency and compounds. Apart from that, it was also a lot of fun, and overall a surprisingly and overwhelmingly fulfill- ing experience in a world of academia that I had almost come to see exclusively as a cynical caricature of its original purpose. Thank you Melanie! Preliminary versions of the material in this book were presented at numerous conferences and talks, and I thank all the audiences for their feedback. Special thanks go to Ingo Plag, Carla Umbach, and Thomas Weskott. Turning to my actual place of work, the English department of the University Jena, I would like to thank all my colleagues there, especially Volker Gast, Flo- rian Haas, Karsten Schmidtke-Bode and Holger Dießel, who witnessed the whole developmental progress of this work and provided feedback and encouragment throughout. Very special thanks go to my office mate Christoph Rzymski. He was my main statistics and R advisor, and also carefully read and helpfully com- mented on the manuscript before I submitted it. Quite over and above that, he also provided the office with much-needed Supertee, and generally made office life most enjoyable. This work is the revised version of my Habilitationsschrift, which was accepted in 2017 by the Philosophische Fakultät of the Friedrich-Schiller-Universität Jena. I thank the original reviewers of the Habilitationsschrift, Sabine Arnd-Lappe, Hol- ger Dießel, and Volker Gast as well as the anonymous referee for Language Sci- ence Press for their many helpful comments and suggestions. Acknowledgments Speaking of Language Science Press: many thanks to Sebastian Nordhoff, who made working with them a very pleasant experience. The work by Melanie Bell and me presented in this book was partially sup- ported by three short visit grants from the European Science Foundation through NetWordS—The European Network on Word Structure (grants 4677, 6520 and 7027). The corpus frequencies for our analyses presented in Chapter 7 were grate- fully provided by Cyrus Shaoul and Gero Kunter. xii Abbreviations The following conventions were used to reference examples taken from corpora and online dictionaries: BNC Example sentence from the British National Corpus. All BNC examples are followed by their unique BNC identifier. COCA Example sentence from the Corpus of Contemporary Ameri- can English. The exact references are given in Appendix D. DeReKo Example sentence from the Deutsche Referenzkorpus. The exact references are given in Appendix D. OED Example from the online edition of the Oxford English Dic- tionary. The exact references are given in Appendix D. WEB Example from the internet. The exact references are given in Appendix D. 1 Introduction This work is concerned with the notion of semantic transparency and its rela- tion to the semantics of compound nouns. On the one hand, my aim is to give a comprehensive overview of the phenomenon of semantic transparency in com- pound nouns, discussing its role in models of morphological processing, giving an overview of existing theories of compound semantics and discussing previous models of the semantic transparency of compounds. On the other hand, I will discuss in detail new empirical investigations into the nature of semantic trans- parency and the factors that make compounds appear more or less transparent. This part focuses on English noun noun combinations. 1.1 A first notion of semantic transparency Semantic transparency is a measure of the degree to which the meaning of a multimorphemic combination can be synchronically related to the meaning of its constituents and the typical way of combining the constituent meanings. Se- mantic transparency is a scalar notion. At the top end of the scale are combina- tions whose meaning is fully transparent, that is, combinations whose meaning is predictable. Conversely, at the bottom end are combinations whose meaning is opaque. Their meaning cannot be predicted, and a link between the meaning of the constituents and the meaning of the resulting combination can hardly be established. In between, there are combinations with varying degrees of relat- edness between the constituents’ meaning and the meaning of the whole, and with varying degrees of predictability based on typical ways of combining these constituents. 1 Examples of English compounds with different degrees of semantic transpar- ency are given in (1). 1 Note that this view combines 2 lines of thinking about semantic transparency. In particular, Plag (2003: 46), in discussing derivations, links semantic transparency to meaning predictabil- ity, whereas Zwitserlood (1994: 344) understands the semantic transparency of compounds in terms of the synchronic relatedness between the meaning of their constituents and the com- pound meaning. 1 Introduction (1) a. For example, in the letters between Lady Sabine Winn and her milli- ner, Ann Charlton, sets of samples were sent, divided between gauzes, ribbons and silk fabrics . COCA b. The bronze lion was placed in the palace’s foundations to please the gods. COCA c. His dad worked for John Deere, his mother was a school teacher COCA d. I am the proud son of a hardworking milkman . COCA e. The creeping buttercup and Virginia creeper weren’t as plentiful as she’d thought. COCA f. But experts call the hypothesis hogwash . COCA g. To stay on postcoital cloud nine , stick to no-brainer subjects that won’t make him think that this one night of passion has changed everything. COCA The meaning of silk fabric in (1-a) appears to be predictable based on the meaning of its parts and the typical, or standard way of combining the modifier silk with the head fabric . This standard way can in this case either be seen as simple set intersection (a silk fabric is a fabric and is silk) or as an instantiation of some relation between the 2 constituents, here the made of-relation (a fabric made of silk). The meanings of the following 3 items, bronze lion , school teacher , and milk man , are somewhat less predictable: a bronze lion might have the corresponding color, or might be made out of bronze. In the latter case, he would not be a real lion, but the image of one. School and teacher can be linked by a local relation (teacher at a school), but both are not restricted in their combinatorics to a local relation, cf. the occurrences of the 2 constituents in other compounds: geography teacher or school finances . Likewise, neither milk nor man seem to suggest an interpretation along the lines of ‘HEAD who goes from house to house delivering MODIFIER’, cf. milkmaid, milk-soup, woodman, sandman, snowman , and garbage- man . For buttercup , some people might see a synchronic relatedness between its constituents and the whole compound, pointing to the resemblance of the color of a buttercup’s petals to the color of butter and the resemblance of the petals’ arrangement to the shape of a cup. Only the 2 final items in (1), hogwash and cloud nine , show no synchronic relation between their constituents and the respective compound meanings. Note that for combinations like hogwash the qualification that the meanings of the compound and its constituents must be synchronically related becomes im- portant. Thus, it is not a coincidence that hogwash means nonsense, and neither 2 1.2 Compounds and complex nominals of its 2 constituents are arbitrarily chosen terms. Rather, the ‘nonsense’ meaning is etymologically well motivated: According to the OED, it was originally used to refer to kitchen refuse that was used as food for pigs, as illustrated by the following quote. (2) Cooks who were not thrifty put all the kitchen leavings into a bucket. The content was called ‘wash’, and the washman visited regularly to buy it: he then sold it as ‘ hog-wash ’, or pigswill. J. Flanders Victorian House (2004) iii. 87 OED Probably via the intermediate step of the second meaning reported in the OED, ‘Any liquid for drinking that is of very poor quality, as cheap beer, wine, etc.’, hogwash then came to be used with its now most frequent meaning, ‘nonsense’. Both of these 2 last steps, that is, from liquid waste for pigs to cheap alcohol and again from cheap alcohol to nonsense are metaphorical extensions that are easy to follow; its current meaning is therefore quite well motivated on the basis of its historical origin. For cloud nine , not even a good etymological explanation is available. In addi- tion, it is more restricted in typically appearing following the preposition on , and, perhaps bearing witness to its unclear etymology, an alternative, on cloud seven , is available, apparently with exactly the same meaning, compare the 2 earliest quotes from the OED in (3). (3) a. Oh, she’s off on Cloud Seven —doesn’t even know we exist. 1956 O. Duke Sideman ix. 120 OED b. I don’t like strange music, I’m not on Cloud Nine 1959 Down Beat 14 May 20 OED Even though neither cloud nine nor cloud seven have been attested for long, their etymology remains unclear; the best one can find are statements like the follow- ing attempt for cloud nine : “the number nine is said by some to come from a meteorologist’s classification of a very high type of cloud” (Walter 2014). 1.2 Compounds and complex nominals Compounds share many properties with other complex constructions having a nominal head. The term ‘complex nominal’ is used in this work to refer to con- structions of the general format MODIFIER HEAD, with the head always being a noun and the resulting construction likewise being substitutable in noun con- 3 1 Introduction texts. It is a cover term that subsumes constructions that are traditionally called compounds (e.g. blackbird, railway , and volcano ash ) as well as constructions that are traditionally considered as phrases (e.g. superconducting cable and brown hair ), extending on the usage of the term in Levi (1978: 1–2), where it was used to encompass nominal compounds as well as combinations of nonpredicating adjectives with nouns (e.g. electric clock or musical talent ). 2 For English, with no binding elements nor specific word forms as formal mark- ers of compoundhood, stress placement is often accepted as the only fail-safe criterion for compoundhood: if an X-N construction is stressed on the first con- stituent, then it is a compound (this has been most famously formalized by Chom- sky & Halle 1968: 17–18, who distinguish between a nuclear stress rule and a compound stress rule). However, as Plag et al. (2008: 761) point out after listing the many authors stating exceptions to this rule, there is a considerable number of constructions that are typically regarded as compounds but that do not show fore-stress, compare the examples in (4), drawn from (1) in Plag et al. (2008). (4) apple píe, Michigan hóspital, summer níght, aluminum fóil, spring bréak, silk tíe (the acute accent marks the vowel of the most prominent syllable) In this work, all these constructions are complex nominals and the term com- pound is also used with the wider, more general usage in mind. In the discussion of other criteria that have been introduced to diagnose compoundhood the main focus has been on noun noun constructions. Bauer (1998) shows that none of the criteria traditionally employed to distinguish between 2 constructions (listed- ness, orthography, stress, syntactic isolation of the first constituent, constituent coordination, one -substitution) yields strong evidence for a distinction between 2 types of noun noun constructions. Bell (2011) follows Bauer (1998) in that the cri- teria do not allow to distinguish between 2 different categories and argues for the analysis of all noun noun constructions as compounds. In a similar vein, Bauer, Lieber & Plag (2013: 434) acknowledge that “there seems to be no established set of trustworthy procedures that could tell us reliably and theory-neutrally for a given NN construction whether it is a noun or a phrase”, arguing for a maximally inclusive approach in assigning compound status. Note that the 2 major academic reference grammars of English both maintain a distinction between 2 different categorical types of noun noun combinations: 2 Levi (1978: 1–2) specifically mentions a third group of constructions where the head noun is a deverbal nominalization (e.g. presidential refusal or metal detection ). However, as far as I can tell these constructions are always a subset of either of the first 2 constructions. 4