University of Southern California Dissertations and Theses

Morphology and the lexicon: Exploring the semantics-phonology interface

(USC Thesis Other)

Morphology and the lexicon: Exploring the semantics-phonology interface

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content INFORMATION TO USERS This manuscript has been reproduced from the microfilm master. U M I films the text directly from the original or copy submitted. Thus, some thesis and dissertation copies are in typewriter face, while others may be from any type of computer printer. The quality of this reproduction is dependent upon the quality of the copy submitted. Broken or indistinct print, colored or poor quality illustrations and photographs, print bleedthrough, substandard margins, and improper alignment can adversely affect reproduction. In the unlikely event that the author did not send U M I a complete manuscript and there are missing pages, these will be noted. Also, if unauthorized copyright material had to be removed, a note will indicate the deletion. Oversize materials (e.g., maps, drawings, charts) are reproduced by sectioning the original, beginning at the upper left-hand comer and continuing from left to right in equal sections with small overlaps. Photographs included in the original manuscript have been reproduced xerographically in this copy. Higher quality 6” x 9” black and white photographic prints are available for any photographs or illustrations appearing in this copy for an additional charge. Contact U M I directly to order. Bell & Howell Information and Learning 300 North Zeeb Road, Ann Arbor, M l 48106-1346 USA U M T 800-521-0600 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. MORPHOLOGY AND THE LEXICON: EXPLORING THE SEMANTICS-PHONOLOGY INTERFACE by Laura Michelle Gonnerman A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (Linguistics) May 1999 Copyright ©1999 Laura Michelle Gonnerman Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. U M I Number 9933782 Copyright 1999 by Gonnerman, Laura Michelle All rights reserved. __ ___ __® UMI UMI Microform 9933782 Copyright 1999 by Bell & Howell Information and Learning Company. All rights reserved. This microform edition is protected against unauthorized copying under Title 17, United States Code. Bell & Howell Information and Learning Company 300 North Zeeb Road P.O. Box 1346 Ann Arbor, Ml 48106-1346 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. UNIVERSITY O F SOUTHERN CALIFORNIA THE CRA0UATE SCHOOL u n iv e r s it y PARK LOS ANGELES. CALIFORNIA 9000? This dissertation. written by L aura M i c h e l l e G onneraan under the direction of h.5*_ _ Dissertation Committee, and approved by all its members, has been presented to and accepted by The Graduate School, in partial fulfillment of re quirements for the degree of DOCTOR OF PHILOSOPHY ^ d m or S n iiits Date _____ DISSERTATION COMMITTEE OtMf-enm > //% ■ + . dr- ffo. <J Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. For my Grandmother Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. ACKNOWLEDGMENTS I would like to thank the members of my committee for their contributions to my graduate education. Laura Baker taught a fun and stimulating class on Regression Analysis and was very helpful in assuring that my statistical analyses were on target. Bernard Comrie provided valuable feedback on earlier versions of the manuscript, pointing me to interesting cross-linguistic facts. Unfortunately for me, he left USC before I actually finished the dissertation. Jack Hawkins has been a great help and wonderful to work with. He has been more than willing to discuss my ideas, always posing provocative questions and forcing me to be more precise about my views. Mark Seidenberg’s interest in my work and my ideas has been a great motivation. He has encouraged me to pursue the questions explored in this thesis as well as other interesting issues in language and cognition. Dan Kempler, though not on my committee, deserves special thanks for his kindness and advice, and for sharing his expertise in cognitive neuropsychology, in particular his interest in language impairments in Alzheimer’s disease. Most importantly, I would like to thank my advisor, Elaine Andersen. Elaine is a woman of unparalleled energy and optimism. I count myself truly lucky to have been among her students. She is one of the rare individuals who cares deeply about her role as mentor, being equally demanding and supportive, both academically and personally. Her insistence on clarity of thought and exposition and her willingness to work through difficult ideas with me has improved all of my endeavors. Plus, she’s great fun to work with and just to be around. I would also like to thank the professors who taught me as an undergraduate at Boston University and got me interested and excited about this field. Carol Neidle taught my first Introduction to Linguistics course and was instrumental in helping me make the decision to pursue graduate studies in this area. Jim Gee taught engaging classes on language universals and discourse analysis that I really loved. I would also like to thank my fellow graduate students who helped make Coglab an exciting place to learn and do research. Lori Altmann has been there from day one, sharing office space and the daily rituals of graduate school. Maquela Brizuela, Sarah Schuster, Mike Harm, Robert Thornton, Joe Allen, Morten Christiansen, Marc Joanisse, and Suzanne Curtin have all provided useful suggestions and interesting discussions. I would also like to thank Amit Almor for his friendship. And finally, Joe Devlin has been Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. a great friend. From our first memorable conference in Madeira, through qualifying exams and actually finishing our dissertations, Joe has been willing both to argue and to agree. I have really enjoyed our collaboration. I never would have finished this degree without the support o f family and friends. I’d like to thank Clancy and Charlotte Imislund for hosting volleyball in their back yard every Saturday morning. Thanks too to those who played along and trudged the road with me, especially Jennifer Kelly, Thor Gold, Kim Reisman, and Susie Nathan. Cahuenga Begue kept things in perspective, reminding me o f the world beyond graduate school. My parents, Madelyn Gonnerman and Wayne Gonnerman, traveled this path before me and offered consistent encouragement. My cousin, Marilyn Kiss, cheered me along, as did my siblings, Michael and Kathryn. My relatives on the Central Coast, Buddy and Lyn Jones, opened their home to me as a quiet haven whenever I could find the time to venture up there. And my grandmother, Nora Gonnerman, always urged me to continue my education and especially just to finish. This dissertation is dedicated to her. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. V TABLE OF CONTENTS Dedication................................................................................................................................ii Acknowledgments................................................................................................................. iii List of Tables.......................................................................................................................viii List of Figures............................................................................................................. xi Abstract..................................................................................................................................xii 1. In trod u ction .......................................................................................................... 1 2. Theoretical Approaches to M orphology..............................................................8 2.1 The Nature of the Problem....................................................................................9 2.2 Previous Research Exploring the Decomposition Question........................... 10 2.3 Previous Research on Formal and Semantic Factors........................................16 2.4 Research Based on a Different Metaphor......................................................... 21 2.4.1 Distributed representations...................................................................... 24 2.4.2 Basins of attraction................................................................................... 26 2.5 Focus on Derivational Morphology, in Particular, Suffixation...................... 28 3. A Computational Model o f M orphological Prim ing.........................................35 3.1 Marslen-Wilson et al. (1994) Study.................................................................. 36 3.2 Reanalysis of Marslen-Wilson et al. Data......................................................... 42 3.3 Computational M odel......................................................................................... 43 3.3.1 Semantic output representations.............................................................. 43 3.3.2 Phonological input representations..........................................................47 3.3.3 Model architecture and training.............................................................. 52 3.3.4 Priming in the m odel................................................................................53 3.3.4.1 Procedure................................................................................... 53 3.3.4.2 Results and discussion.............................................................. 54 3.4 Discussion............................................................................................................ 59 4. Behavioral Experim ents..................................................................................... 61 4.1 Introduction .................................................................................................... 61 4.1.1 Overview of experiments.........................................................................61 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4.1.2 The cross-modal lexical decision task.................................................... 62 4.2 Experiment 1: Degrees o f Semantic Relatedness............................................. 63 4.2.1 Method ....................... 64 4.2.1.1 Subjects........................................................................................64 4.2.1.2 M aterials..................................................................................... 64 4.2.1.3 Procedure.................................................................................... 68 4.2.2 Results and discussion............................................................................. 69 4.3 Experiment 2: Degrees of Phonological Relatedness----------------------------- 77 4.3.1 Method .....................................................................................................77 4.3.1.1 Subjects....................................................................................... 77 4.3.1.2 M aterials..................................................................................... 78 4.3.1.3 Procedure.................................................................................... 83 4.3.2 Results and discussion............................................................................. 83 4.4 Experiment 3: Role of Morphological Type.....................................................91 4.4.1 Method .....................................................................................................91 4.4.1.1 Subjects....................................................................................... 91 4.4.1.2 M aterials.....................................................................................92 4.4.1.3 Procedure....................................................................................96 4.4.2 Results and discussion............................................................................. 96 4.5 Experiment 4: Historically Unrelated Pairs.................................................... 104 4.5.1 Method ................................................................................................... 105 4.5.1.1 Subjects......................................................................................105 4.5.1.2 M aterials................................................................................... 105 4.5.1.3 Procedure..................................................................................108 4.5.2 Results and discussion........................................................................... 109 4.6 Summary and General Discussion....................................................................123 5. General Discussion........................................................................................... 127 5.1 Summary of Results...........................................................................................127 5.2 Implications of Results.....................................................................................129 5.3 Future Directions............................................................................................... 132 5.3.1 Cross-linguistic extensions....................................................................132 5.3.2 Developmental extensions.....................................................................133 5.3.3 Computational extensions...................................................................... 134 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. vii 5.3.4 Neuropsychological extensions..............................................................135 References.............................................................................................................. 137 Appendices............................................................................................................. 147 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. viii LIST OF TABLES 3.1 Sample stimuli used in Marslen-Wilson et al. (1994) Experiment 1................... 38 3.2 Sample stimuli used in Marslen-Wilson et al.’s Experiment 3 with semantic relatedness ratings and priming effects...................................................................39 3.3 Suffix Classification for Semantic Output Representations..................................46 3.4 Priming results from Marslen-Wilson et al.’s human subjects and the computational model................................................................................................. 54 4.1 Experiment 1: Mean semantic similarity ratings for sample items from pretest (where a score of 1 is very unrelated, 7 is very related)...........................65 4.2 Experiment 1: Sample stimuli and mean relatedness ratings (where a score of 1 is very unrelated, 7 is very related) for each condition................................. 67 4.3 Experiment 1: Mean reaction times and priming effects by condition................ 71 4.4 Experiment 1: Mean error rates for lexical decision to real word targets by condition......................................................................................................................74 4.5 Experiment 1: Mean reaction times and error rates for lexical decision to real word and nonword targets................................................................................ 75 4.6 Experiment 1: Mean reaction times and error rates for lexical decision to non word targets by condition...................................................................................76 4.7 Experiment 2: Mean semantic similarity ratings for sample items from pretest (where a score of 1 is very unrelated, 9 is very related)...........................80 4.8 Experiment 2: Sample stimuli and mean relatedness ratings (where a score of 1 is very unrelated. 9 is very related) for each condition................................. 82 4.9 Experiment 2: Mean reaction times and priming effects by condition................ 84 4.10 Experiment 2: Mean error rates for lexical decision to real word targets by condition..................................................................................................................... 88 4.11 Experiment 2: Mean reaction times and error rates for lexical decision to real word and nonword targets................................................................................ 89 4.12 Experiment 2: Mean reaction times and error rates for lexical decision to non word targets by condition...................................................................................90 4.13 Experiment 3: Mean semantic similarity ratings for sample items from pretest (where a score of 1 is very unrelated, 9 is very related)...........................93 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4.14 Experiment 3: Sample stimuli and mean relatedness ratings (where a score of 1 is very unrelated, 9 is very related) for each condition................................. 94 4 .15 Experiment 3: Mean reaction times and priming effects by condition.................97 4.16 Experiment 3: Mean error rates for lexical decision to real word targets by condition...................................................................................................................101 4.17 Experiment 3: Mean reaction times and error rates for lexical decision to real word and nonword targets.............................................................................. 102 4.18 Experiment 3: Mean reaction times and error rates for lexical decision to nonword targets by condition................................................................................ 103 4.19 Experiment 4: Mean semantic similarity ratings for historically unrelated sample items from pretest (where a score of 1 is very unrelated, 9 is very related)...................................................................................................................... 106 4.20 Experiment 4: Sample stimuli and mean relatedness ratings (where a score of 1 is very unrelated, 9 is very related) for each condition................................107 4.21 Experiment 4: Mean reaction times and priming effects by condition for Psychology subject group....................................................................................... 110 4.22 Experiment 4: Mean reaction times and priming effects by condition for Honors college subject group................................................................................ I l l 4.23 Experiment 4: Mean reaction times for Pseudo-suffixed' and 'Non-suffixed' targets following control and test primes. Priming effects are also shown......114 4.24 Experiment 4: Mean semantic relatedness ratings (where a score of 1 is very unrelated, 9 is very related) from pre- and post-test survey administration, shown for each condition by subject group................................116 4.25 Experiment 4: Correlation Matrix of Semantic Relatedness Ratings by Group........................................................................................................................ 117 4.26 Experiment 4: Mean error rates by condition for the Psychology subject group........................................................................................................................ 118 4.27 Experiment 4: Mean error rates by condition for the Honors college subject group........................................................................................................................ 119 4.28 Experiment 4: Mean reaction times for lexical decision to real word and non word targets by subject group..........................................................................120 4.29 Experiment 4: Mean error rates for lexical decision to real word and nonword targets by subject group..........................................................................121 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. X 4.30 Experiment 4: Mean reaction times and error rates for lexical decision to nonword targets by condition for the Psychology subjects group and the Honors College group.............................................................................................123 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. xi LIST OF FIGURES 3.1 Marslen-Wilson et al. (1994) stem-affix model for semantically transparent suffixed words.............................................................................................................. 41 3.2 Dynamic programming grid........................................................................................ 48 3.3 Identity comparison of one word to itself................................................................. 50 3.4 Model architecture: An attractor network with four layers, phonology, semantics, a layer of hidden units, and a layer of clean-up units. The numbers in each box indicate the number of units in that layer, while arrows indicate full connectivity between groups................................................................ 53 3.5 Relationship between semantic similarity and priming effects...............................56 3.6 Relationship between phonological similarity and priming effects....................... 57 3.7 Relationship between semantic similarity and phonological similarity................ 58 4.1 Experiment 1: Priming effects for items with different degrees of semantic relatedness: Low (Conditions 1 and 2), Mid (Condition 3), High (Condition 4, phonologically related; & 5, phonologically unrelated........................................73 4.2 Experiment 2: Priming effects for items matched for semantic similarity but with different degrees of phonological relatedness: Condition 1) No change, Condition 2) Consonant change. Condition 3) Vowel change, Condition 4) Consonant plus vowel change. Also Condition 5) Semantically unrelated, and Condition 6) phonologically unrelated synonyms............................................ 86 4.3 Experiment 3: Priming effects for suffixed word pairs: Condition 1) Low semantic, high phonological relatedness; 2) High semantic and high phonological relatedness; Condition 3) High semantic, low phonological relatedness.....................................................................................................................99 4.4 Experiment 4: Priming effects for historically unrelated word pairs: Condition 1) High semantic relatedness; 2) Mid semantic relatedness; Condition 3) Low semantic relatedness; Condition 4) High semantic relatedness, no phonological overlap; Condition 5) Low semantic relatedness, high phonological overlap. Results shown are from Honors College subject group only.................................................................................................................... 113 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. ABSTRACT In contrast to traditional views of the mental lexicon in which morphological relationships are explicitly represented, I propose an account that suggests instead that ‘morphology’ reflects degrees of compositionality in complex words and arises predominantly from regularities in the mappings between semantic and phonological codes. Support for this alternative account comes both from results of a computational model and from a series of four cross-modal priming experiments. The computational results demonstrate that a connectionist model, without an explicit representation o f morphological structure, can produce priming results strikingly similar to those from behavioral studies, suggesting that a system which relies on semantics and phonology alone can simulate what appear to be effects of morphological structure. The behavioral results demonstrate that: 1) subjects are sensitive to fine gradations both in the semantic similarities between related lexical items (e.g., teacher- teach; dresser-dress; comer-com), and in their phonological similarities (e.g., deletion- delete; vanity-vain; introduction-introduce); 2) the degrees of semantic and phonological relatedness of word pairs, when considered in conjunction, predict priming effects, and; 3) the nature of the morphological relationship between primes and targets does not predict priming effects: derived-stem pairs (e.g., teacher-teach), derived-derived pairs (e.g., saintly-sainthood), and pairs with no historical morphological relationship (e.g., trivial-trifle) all prime, if the words are highly semantically and phonologically related. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The results from the model and the behavioral experiments suggest that morphological structure is an emergent, interlevel representation that mediates computations between semantics and phonology, not an independent component of language. Moreover, I argue that this approach is superior to other approaches, because it accounts for intermediate, graded effects that present a challenge to more traditional, decompositional views of the lexicon. In addition, this account sheds new light on the longstanding debate within linguistic theory about the appropriate place for morphology, by changing its role to one of convenient description of phenomena that actually lie at the interface of semantics and phonology. Finally, the account provides interesting predictions about acquisition, cross-linguistic studies, and impairment, for which there is currently emerging support. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 1 Introduction Morphology has been an arena for major controversy within both linguistics and psycholinguistics. Linguists cannot agree as to whether it constitutes an independent sub component of language (Chomsky, 1970; Di Sciullo & Williams, 1987), is a part of syntax (Lieber, 1992) or phonology (Beard, 1995; Aronoff, 1976, 1994; Anderson, 1982, 1992), or is distributed among several different components o f language (Halle & Marantz, 1993). Similarly, within psycholinguistics, some researchers claim that inflected and/or derivationally complex words are stored as single "gestalt" forms (e.g., Butterworth, 1983), others suggest that they are stored according to their stems, with affixes added via rules of word formation (e.g., Taft and Forster, 1975; Taft, 1988); and still others propose hybrid systems that store some types of words (irregulars or semantically opaque forms) as gestalts and other types (regulars and semantically transparent forms) as separate stems and affixes. For example, in the case o f inflectional morphology, Prasada & Pinker (1993) posit a memory store for irregular instances of the past tense (e.g., go-went) but a rule that adds the affix -ed to a stem to form the past tense of regular verbs (e.g., walk-walked). Similar hybrid theories have been developed for derivational morphology (Cole, Beauvillain, & Segui, 1989; Marslen-Wilson, Tyler, Waksler, and Older, 1994). In their hybrid account, Marslen-Wilson et al. (1994) treat derivationally complex words either as stems plus affixes or as unanalyzed gestalt forms, based on whether the word is semantically transparent (i.e., government = govern + ment) or opaque (i.e., in department, there is no clear semantic relationship between depart and department) (Tyler, Behrens, Cobb, & Marslen-Wilson, 1990). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In this dissertation I propose an alternative account of the representation and processing of derivationally complex words which allows me to treat all forms in a uniform manner by exploiting the non-arbitrary relationship between sound and meaning that exists for these words. This account draws on many of the same connectionist principles that underlie much of the recent work on inflectional morphology (e.g., Plunkett & Marchman, 1991; Daugherty & Seidenberg, 1992; MacWhinney & Leinbach, 1991), and is compatible with other recent studies that do not view morphological structure as a primary determinant of lexical representation (e.g., Bybee, 199S; Rueckl, Mikolinski, Raveh, Miner, & Mars, 1997). For example, Bybee (1995) examines several languages to assess the role of type and token frequency in determining productivity of morphological patterns, explaining her findings within the 'network model' framework of Langacker (1987, 1988), which shares some ideas with the connectionist approach. In contrast, Rueckl et al. (1997) describe their experimental results in connectionist terms, making their ideas most obviously related to mine. Central to my approach is the role of semantic and phonological relatedness in the processing of morphologically complex words. The particular question addressed is whether effects that have been attributed to morphological structure actually derive solely from sound-meaning regularities. I explore this question in two different ways. First, I develop a computational model and show that without explicit morphological representations it can simulate the results of previously reported behavioral experiments that have looked at priming of derived-stem and derived-derived word pairs (Marslen-Wilson et al., 1994). Then I carry out a series of four behavioral experiments that examine more closely both the independent contributions of semantic and phonological factors to priming and their interactions. By comparing the results presented in this thesis with previously reported experimental results, I show that priming effects which seemingly support a Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3 decompositional lexicon with independent morpheme representations, in fact, can be accounted for by semantics and phonology alone. The dissertation includes four additional chapters, whose contents are briefly described below. In Chapter 2 ,1 outline the controversies surrounding morphology and some of the traditional approaches taken to explaining the role of morphological structure in the mental lexicon. I begin Section 2.1 by describing the relevant phenomena. I then provide a summary of previous important psycholinguistic investigations of morphological processing. Researchers have asked two major questions: 1) are complex words decomposed in lexical access or in storage; and 2) what is the role of semantic and formal (i.e., orthographic or phonological) factors in morphological processing? In Section 2.2,1 consider several experimental investigations that have addressed the former question. In Section 2 .3,1 look at the latter question, reviewing a variety of psycholinguistic studies that have considered one or more of a range of factors potentially involved in morphological access and storage, including phonological, orthographic, and semantic similarity, but that have never considered the interactions between these factors. I discuss why looking at each of these factors separately is misleading, causing researchers to conclude that morphological structure must play a role in the mental lexicon. In Section 2.4,1 suggest an alternative metaphor for the mental lexicon. Unlike the traditional dictionary metaphor, where words are stored as lists, I advocate a lexicon based on connectionist principles. I describe the primary principles that give rise to a different view of the lexicon, namely distributed representations and basins of attraction, explaining how these concepts apply to the representation and processing of derivational morphology. I argue that the connectionist metaphor is better suited to explain the nature of the mental lexicon and to account for the available data on morphological processing. Finally, in Section 2.S, I briefly discuss different morphological processes, concentrating Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4 on the distinction between inflectional and derivational morphology. The discussion of these morphological types motivates the decision to focus on stems and derived suffixed forms alone, to the exclusion of inflected, prefixed, or compound forms, for both the computational model described in Chapter 3, and the behavioral experiments described in Chapter 4 Chapter 3 describes a connectionist model o f comprehension (mapping from a sound pattern to the corresponding meaning pattern) that incorporates semantic and phonological information, but no explicit morphological structure. This model, a recurrent backpropagation model, was trained with the 208 prime and target words used in a study by Marslen-Wilson et al. (1994), where they tested priming of related stems and suffixed words (e.g., govem-govemment) as well as pairs of related derived suffixed words (e.g., govemor-govemment). In the model, words are represented as distributed patterns of activation that capture the phonological and semantic relatedness between prime-target pairs. A supervised learning algorithm is employed to adjust the weights between units until the model correctly produces the semantic representation of a word given its phonological form. A priming experiment is carried out by comparing the settling time of the model for a stem (e.g., govern) following presentation of a related suffixed word (e.g., government) compared to an unrelated control (e.g., brightness). The modeling results are strikingly similar to the results reported by Marslen-Wilson et al. for human subjects. While Marslen-Wilson et al. interpreted their results as indicating that the mental lexicon contains distinct representations for stems and affixes, the model reproduces their pattern of results without an explicit representation o f morphological structure. In the model, semantic and phonological relatedness alone account for the priming effects between morphologically related word pairs. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 5 The results from the model described in Chapter 3 lead to specific predictions about the nature of morphological priming: for example, degrees of relatedness between primes and targets should be reflected in a range of priming magnitudes; and highly related words should prime, regardless of the nature of their morphological relationship (i.e., derived-stem pairs, derived-derived suffixed pairs, or historically unrelated pairs). In Chapter 4 ,1 present a series o f four cross-modal lexical decision experiments designed to investigate these predictions. The first experiment examines the role of semantics, addressing the question of whether semantic relatedness can predict priming effects when phonological relatedness is held constant. The results indicated no priming between semantically unrelated words (e.g., comer-com). However, semantically related pairs did show reliable priming effects, with the magnitude of these effects increasing as a function of semantic relatedness (e.g., baker-bake primes more than dresser-dress). The second experiment was also designed to explore effects o f the degree of compositionality of derived words and stems, specifically to determine whether the degree of phonological relatedness could predict the amount of priming when all the forms were highly semantically related. The results reveal significant differences in priming effects tied to four different levels of phonological relatedness. The priming effects generally increased with increasing phonological relatedness. The third experiment looks at suffixed-suffixed pairs (e.g., observation- observant) to re-explore specific claims of Marslen-Wilson et al. (1994) who found that pairs of suffixed words did not prime each other, even when they were semantically related. I argue that the lack of priming in the Marslen-Wilson et al. experiment arises because the derived-derived pairs they used were slightly less semantically and phonologically related than the derived-stem pairs that did show priming effects. Therefore, in Experiment 3 ,1 use highly semantically related pairs where there is no Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 6 sound change between the derived form and the stem (e.g., saintly-sainthood). I also increase the number of items tested to increase the power of the experiment. Results show that the highly related derived-derived pairs do produce significant priming effects, supporting the view that suffixed forms are not stored or processed differently than either their stems or related prefixed forms. The fourth experiment questioned the assumption, implied in some work and explicit in others (e.g., Taft & Forster, 197S), that a morphological relationship between word pairs is necessary for priming. Instead, it investigates the hypothesis that semantic and phonological similarity alone can account for these effects. Thus, 51 word pairs that are not morphologically related but are similar in both sound and meaning are used in this experiment (e.g., trivial-trifle). Not surprisingly, I find that the historical morphological relationship of words is of no relevance in processing, rather, the effects can be explained by degree of phonological and semantic similarity alone. Finally, in Chapter 5 ,1 summarize the findings of the computational and behavioral studies and discuss how these findings support an account in which morphological structure is an emergent, interlevel representation that mediates computations between semantics and phonology. In so doing, I emphasize the importance of examining the interaction of relevant factors if we are to hope for an adequate, realistic account that matches human processing. For example, while I was able to find graded priming effects for different levels of phonological relatedness between primes and targets, this effect was only obtained because the word pairs were all highly semantically related. At low levels of semantic relatedness, the degree of phonological overlap no longer plays a role. Because phonologically transparent, but semantically distant word pairs do not prime, any approach that explores the effects of Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 7 one factor, such as phonology, in isolation, without at the same time considering other factors, is going to produce misleading results. While semantic and phonological similarity are the two most important factors contributing to apparent morphological effects in English, there are clearly additional factors which also impact the representation and processing of complex words. Indeed, to achieve a truly comprehensive story, one would want to include additional factors. I discuss some of these factors, among them productivity, type and token frequency, and orthographic similarity. The implications of this research for linguistic theories, psycholinguistic accounts of processing and storage, and methodology are evaluated in Section 5.2. In Section 5.3, I lay out the kinds of questions this type o f account will eventually need to address, for example: 1) cross-linguistic, by looking at languages with typologically different morphological structure; 2) developmental, to account for patterns in language acquisition; 3) computational, to push the limits of the connectionist implementation of a system of derivational morphology and thus provide further insights into the nature of morphological processing and storage; and 4) neuropsychological, to account for patterns of dissolution in language impairment. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 2 Theoretical Approaches to Morphology In this chapter I review some of the traditional psycholinguistic approaches taken to morphology, situating the studies described in the thesis within this broader research context. I begin in Section 2.1 by briefly reviewing the nature of the problem presented by derivational morphology for theories of lexical representation and processing. In Section 2 .2,1 summarize some of the main findings of previous research that addressed the question of whether speakers or readers process complex words by separating them into their components. I continue the review o f previous research in Section 2.3, focussing here on several experiments that have examined the roles of various factors in the processing of morphologically complex words. Included among these factors are semantic, phonological, and orthographic overlap between morphologically related words. The majority of these studies found that the additional contribution of morphological structure was necessary to account for the experimental data. I will argue that these accounts miss a crucial point; they consider each factor separately and fail to consider the interaction of factors. It is only by looking at the conjunction of these separate sources of information that a coherent, realistic picture emerges. In Section 2.4, I then describe an alternative approach to derivational morphology, which looks to the brain for its character, rather than the traditional dictionary metaphor of the lexicon. This alternative approach appeals to some of the general tenets of connectionism. I first discuss some recent work on inflectional morphology within the connectionist framework, before turning to a brief description of two aspects of connectionist models that are the most important for my account. Finally, in Section 2.5,1 provide an Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 9 explanation of the choice to focus on derivational suffixation in both the model described in Chapter 3 and the behavioral experiments described in Chapter 4. I also provide the rationale for focussing in particular on the process of suffixation within derivational morphology, highlighting dimensions on which suffixation and prefixation differ. 2.1 The Nature o f the Problem One of the fundamental problems in the study of language is how to characterize fluent speakers' knowledge of words and the ways in which this knowledge is used. It is often observed that the correspondences between sound and meaning are essentially arbitrary, such that the meaning 'four legged animal that barks' can be represented equally well by the sound patterns o f dog, chien, or Hund. However, there are also clear systematic relationships among some words. For example, talk, talker, talkative, and backtalk all include a common element, talk, that contributes systematically to the pronunciations and meanings of these words. Similarly, teacher, baker, and runner all share a final syllable that has the same phonological structure and contributes the same element of meaning: 'a person who performs an action'. It seems intuitively obvious that fluent speakers would take advantage of these systematic relationships to simplify the representation and processing of lexical items. The way some linguists have dealt with these phenomena traditionally has been to consider the mental lexicon as a store of all and only the information that is necessary to represent the inventory of words a speaker knows, that is, only the information that cannot be generated by a rule (Aronoff, 1976; Chomsky, 1970). To avoid redundancy, in a system of this sort, the stem talk would be represented, and all the related forms of the word (talk + -er, talk + -ative, back + talk) would be generated by rule. The intuitive Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 10 appeal of this account is that it provides an economy of storage and it easily accommodates the fact that speakers can comprehend and produce novel complex words. Many other researchers have argued, however, that it is more reasonable to assume that all the forms of known words are stored and directly accessed for use in comprehension or production (e.g., Butterworth, 1983; Manelis & Tharp, 1977). Such a system avoids the computational cost associated with the application of rules, as well as the problems that occur when the application o f the general word formation rules fails. For example, a system that applied rules to all forms stored as stems and affixes would have difficulty with words such as sister and raspberry, or comer, for which stripping the affix results in either a non-existent stem, (sist or rasp), or an unrelated whole word (com). Although the whole word model avoids the problems caused by applying rules to known forms, it still must posit rules to account for novel derived forms, since even very young children can produce or understand the form ricker given the nonsense form to rick (Berko, 1958). Both of these traditional approaches to the lexicon, decompositional and whole word, were based on the metaphor of a serial computer. Given this metaphor, the trade off between economy of access versus economy of storage was a central theoretical issue (see Sandra, 1994, for a discussion of these issues). In the next section, I review some of the most important empirical studies that have dealt with these issues. 2.2 Previous Research Exploring the Decomposition Question Much of the previous research attempting to characterize the mental lexicon has been devoted to determining whether morphologically complex words are decomposed in lexical access and/or in storage. In the following section I discuss some o f the research Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 11 that has been used to argue for either a decompositional, non-decompositional, or hybrid (either decompositional or whole word depending upon the nature of the word) mental lexicon. One of the earliest studies proposing a decompositional approach to morphological processing was that of Taft and Forster (1975). Using a lexical decision task, they found that subjects were slower to reject pseudo-affixed nonwords containing real morphological stems, such as dejuvenate, than they were to reject pseudo-af fixed nonwords without real morphological stems, such as depertoire. They interpreted the findings as evidence for a processing strategy that strips affixes from complex words and then searches for a stored stem. In their Affix-stripping model, lexical decisions to nonwords which are morphological stems (e.g.,juvenate) are slower than decisions to nonwords which are not stems (e.g., pertoire) because the stem juvenate actually accesses a stored representation and the subject must then determine that the form juvenate cannot stand alone, whereas lexical access with a stem like pertoire simply fails. An alternative explanation of the Taft and Forster (1975) results is that it takes longer to reject nonwords that are similar in form to real words (Coltheart, Davelaar, Jonasson, & Besner, 1977). Thus, subjects are slower to reject dejuvenate because it bears a closer phonological resemblance to rejuvenate than depertoire does to repertoire. Subjects may be more likely to pronounce depertoire without reducing the initial syllable, since there are many words in English that can begin with a full de syllable, as in department. Similarly, Schaeffer and Wallace (1970) found that semantically related items are more difficult to reject than semantically distinct items: it takes longer to determine that hemlock is different from daisy, than that hemlock is different from parrot. Although nonwords cannot be said to have semantic representations per se, they may activate more semantic information to the extent to which they resemble existing words. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 12 Thus, the results may arise from similarity to existing words (either semantic or phonological, or both) rather than from "morphological" aspects of the stimuli. Several subsequent experiments have provided similar experimental evidence that has been used to argue for the notion of Affix-stripping in lexical access (Taft, 1981, 1988; Taft & Forster, 1976), but these studies also suffer from similar methodological problems. While a major component o f Taft's approach is that Affix-stripping applies to all word forms, whether they are morphologically transparent or not, a second important aspect of this approach is the nature o f the units used for lexical access. Taft (1979) argues for an access unit that he calls a BOSS, or Basic Orthographic Syllabic Structure. This access unit is different from those proposed by other researchers, which have included units such as the syllable (e.g., Spoehr & Smith, 1973), morpheme (e.g., Libben, 1993), or whole word (e.g., Butterworth, 1983). Taft (1987) defines the BOSS as, "the first part of the stem morpheme of a word, up to and including all consonants following the first vowel, but without creating an illegal consonant cluster in its final position (p. 265).” The BOSS for lantern is therefore lant, for concentrate it is cent, for boycott, boy, and for spade, spad. The primary motivation for the BOSS is to avoid problems related to a syllabic access unit. For example, faster, contains the syllables fas and ter. Unfortunately, these syllabic access units do not overlap at all with the syllable in fast, which is clearly related to faster. However, using the BOSS as an access unit avoids this problem, since the BOSS for faster is fast. Thus, on this view, visual word recognition relies on a separate store of BOSSes which, when accessed, allow retrieval of all the information in the lexicon available for a particular entry. Seidenberg (1987, 1989) argues against the BOSS as an access unit, claiming that "perceptual groupings of letters in visual word recognition are due to orthographic redundancy" (1987: p. 259). He presents evidence from three experiments that support Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 13 the importance o f the role of frequency of orthographic patterns over the syllable, the BOSS, or the morpheme. He argues that orthographic properties of prefixed words might explain effects that appear to arise from morphological structure, pointing out that prefixes constitute common spelling patterns that occur across many words. It is the frequency of the orthographic pattern that makes the prefix highly salient, not its morphological properties. Thus, Seidenberg’s argument is similar to the approach I take in the thesis, in that he ascribes morphological effects in reading to factors other than the morphological make-up of the items. While Seidenberg’s argument is focussed on orthography and processing of visual letter patterns, in this thesis, I argue that the principles are more general, applying to semantic and phonological systems as well as orthographic, and to language comprehension and production, as well as to reading. More recently, Taft (1994) has dispensed with the notion of Affix-stripping for lexical access and proposes instead an interactive activation model where prefixed words are stored in decomposed form, but accessed directly, without the extra step of prefix- stripping. This change in theoretical approach is motivated primarily by the fact that the interactive activation model can account for the experimental results described above, but without positing an independent store of access representations for prefixes. The interactive activation model is seen as a more parsimonious account of the phenomena. In contrast, Manelis and Tharp (1977) developed one of the earliest psycholinguistic accounts that proposed a lexicon with whole word representations, based on findings from two different experiments that failed to show evidence for lexical decomposition. The researchers found no differences in lexical decision times for suffixed words (e.g., dusty) and pseudosuffixed words (e.g., nasty) and therefore concluded that the lexicon includes separate representations for all words. The authors argued that whole word representations save on computational costs for lexical Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 14 processing. The extreme version of this position, namely that the mental lexicon does not take advantage o f morphological structure in processing or storage, is not widely held. However, Bradley (1980) also argues for a view of the mental lexicon which capitalizes on efficiency of processing, noting that the lexicon, unlike syntax, is composed of a finite store of words and thus storing all words in the interest of efficient access to the information array is reasonable. There are a few other studies that take this general approach, sometimes arguing that the stored whole word forms are related in the lexicon by links between words that share common morphological elements (Butterworth, 1983; Lukatela, Gligorijevic, Kostic, & Turvey, 1980; Segui & Zubizarreta, 1985). A much more widely held view of the mental lexicon is a hybrid one, where morphological decomposition is variably applied, based either on the type of word or the type of lexical process. For example, Andrews (1986) argues that compound words are decomposed in lexical access, suffixed words are optionally decomposed, and all types of complex words are represented as whole forms. Another hybrid theory proposes that prefixed words are processed as whole forms and suffixed words are decomposed (Cole etal., 1989). One of the hybrid models that has been extensively explored is the Augmented Addressed Morphology (AAM) of Caramazza and his colleagues. Studies investigating the AAM in normal processing have been carried out on several different languages, including English, Italian, and Dutch (Burani & Caramazza, 1987; Burani, Salmaso, & Caramazza, 1984; Caramazza, Laudanna, & Romani, 1988; Laudanna & Burani, 1985; Laudanna, Badecker, & Caramazza, 1989, 1992; Schreuder, Grendel, Pouliss, Roelofs, & van de Voort, 1990). Investigations of morphological impairments have also been conducted within this theoretical approach (Caramazza, Miceli, Silveri, & Laudanna, 1985; Chialant & Caramazza, 1993). The AAM is a model where two types of lexical Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 15 access procedures work in parallel. Whole word access units are activated at the same time as access units for the individual morphemes that make up a complex word. The whole word access procedure is generally faster than the decompositional one for known words, because there is an extra computational cost associated with decomposing a complex word into its morphological constituents. The decompositional procedure is more important for processing of novel words, nonwords, or known words that are relatively unfamiliar and may have less stable whole word access units. The first of these two access procedures to reach a preset threshold eventually activates a lexicon in which words are stored as separate stems and affixes. An example o f the sort of experimental evidence used to argue for this account comes from a study of Italian by Burani et al. (1984), who used a lexical decision task and manipulated frequency of roots (i.e., the combined frequencies of all words that share a common stem, such as talk, talker, talkative, etc.) and whole words (the simple surface frequency of a single lexical item, e.g., talkative). Their results showed faster reaction times for words with higher root frequency, but not higher whole word frequency. For example, sentito ‘heard’ has higher root frequency than chiamavi ‘called’, but equal surface frequency, and subjects responded an average of 41 msec faster to the sentito type words. These results suggest that root frequency is more important than whole word frequency in determining priming effects. In order to explain their results, they argue that accessing a word form lowers the threshold for that form, as well as the thresholds for all the morphologically related forms. For English, this would mean, for example, that accessing walker augments the frequency not only of walker but also of walked, walking, walks, etc. Studies with morphologically structured pseudowords have also yielded results consistent with the AAM model of the lexicon, using both lexical decision tasks (Caramazza et al., 1988), and more recently, naming tasks (Laudanna, Cermele, & Caramazza, 1997). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 16 In the alternative approach that I am taking to morphology, questions of access versus storage are not central. Thus, I will not take a position either for or against decomposition in the mental lexicon, but will argue instead that the debate is uninformative, given a different framework from which to view the problem. Rather than dealing with all the experimental evidence supporting decomposition models separately and trying to account for it within this alternative framework, my approach will be to focus on one recent, well designed and executed set o f experiments, and to show how the results are interpretable without positing lexical decomposition. This is the focus of the model described in Chapter 3. In the next section I review some of the work that has asked a different question of the mental lexicon, namely whether semantic and formal factors alone, without morphology, can account for experimental evidence bearing on the mental lexicon. 2.3 Previous Research on Formal and Semantic Factors The majority of studies described above did not systematically explore the contribution of semantic or formal (orthographic or phonological) similarity to morphological processing. For instance, Rubin, Becker, and Freeman (1979) consider items such as remark to be prefixed words, and Taft (1981) includes words like intrigue and retard in his stimulus set as morphologically complex items, although synchronically the words are not analyzable into stems and affixes. In the case of items such as remark, there is no clear relationship between the meanings of remark and re- or mark, and for intrigue and retard the stems are nonwords and thus have no meanings. It is unclear how to interpret results from experiments that include in the same condition both unanalyzable forms, such as remark, and transparently analyzable complex words, such as rewrite. For Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 17 researchers interested in the decomposition question, the semantic relationships between complex words and their stems were often not considered because it was assumed that semantics played no role in the mechanisms of lexical access. However, other researchers have explored the extent to which semantic and formal factors influence morphological processing. Unfortunately, the majority of these studies involve what might be considered a ‘divide and conquer’ strategy, which I will argue leads them to the wrong conclusions. Thus, typically, studies are conducted where the researchers examine the role of one factor alone, while ignoring the other. For example, in a classic experiment, Murrell and Morten (1974) tested priming of words that were either morphologically related (both semantically and orthographically as in cars-car) or simply related in form, as in card-car. They found priming for the former, but not for the latter. They concluded that since formal relatedness did not influence priming, morphology had to be the cause of the priming effect and consequently the central organizing principle of the mental lexicon. However, they failed to consider that card-car pairs, while not morphologically related, are also not semantically related. It is certainly possible that the lack of semantic relatedness led to the lack of significant priming effects for these words, not the differences in morphological structure. The same problem is inherent in studies by other researchers who examined the role of orthographic similarity in processing complex words (e.g., Grainger, Cole, & Segui, 1991; Jarvella & Snodgrass, 1974; Laudanna et al., 1989; Stanners, Neiser, Hemon, & Hall, 1979). For example, Jarvella and Snodgrass (1974) presented subjects with two words simultaneously and asked the subjects to judge whether the words contained the same morpheme. Reaction times were faster for words that were orthographically similar. They concluded that complex words shared a lexical entry if there were no differences in the spelling of the base morpheme in the two words. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 18 Otherwise, separate lexical entries were created. Because they found that visual overlap alone could not account for priming, they concluded that morphology is the driving force behind processing of complex words, without considering semantics. Similarly, Kempley and Morton (1982) used a delayed repetition priming paradigm to explore the role of morphological structure in lexical processing. They found that regularly inflected words (e.g., reflected-reflecting) produced significant facilitation, whereas irregularly inflected forms (e.g., held-holding) did not. They concluded that "the morphemic basis of word recognition must be defined in terms of the structural, and not the semantic, properties of words” (p. 450). However, they failed to take into account the fact that irregularly inflected forms are generally less phonologically similar than regularly inflected forms (e.g., held-holding differ in vowel quality, as well as the -ing suffix, while reflected and reflecting differ only in their suffixes). In a second experiment, they found that experiencing a prime with a similar phonological pattern did not facilitate later recognition of a target word (e.g., part-party or deflecting-reflecting). Based on these results, they dismiss the role o f phonological relatedness in processing o f complex words. Again, while part-party and deflecting- reflecting are equated on one dimension, here phonological similarity, they differ on the other dimension, namely semantic relatedness. I will argue that both factors need to be examined simultaneously, and that their effects are not merely additive, but nonlinear. A similar concern can be directed toward another series of studies using repetition priming conducted by Napps and her colleagues (Fowler, Napps, & Feldman, 1985; Napps, 1985, 1989; Napps & Fowler, 1987). While they were concerned with both semantic and formal (phonological and orthographic) factors, by treating each factor independently, and ignoring any possible interactions, they failed to find effects for either semantic or phonological factors. They therefore concluded that morphology must b e a major Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 19 organizational dimension of the mental lexicon: "morphemic priming is not the result of the convergence o f semantic, orthographic, and phonological relationships but rather that morphemic relationships are represented explicitly in the lexicon" (Napps, 1989,729). Their conclusions may have been different had they incorporated both properties simultaneously into their experiments. This same general approach has been applied to testing the effects o f semantic and formal relationships crosslinguistically. For example, previous research in Hebrew has shown priming effects for morphologically related words (e.g., Bentin & Feldman, 1990; Berent & Shimron, 1997; Feldman & Bentin, 1994; Frost, Forster, & Deutsch, 1997). Bentin and Feldman (1990) used 24 different targets (e.g., miktav ‘letter’), each with a prime that was both semantically and morphologically related to the target (e.g., ktovet ‘address’), and another prime that was morphologically related, but had a distant meaning (e.g., katava ‘article’). Interestingly, Bentin & Feldman’s results showed significant priming effects for the morphologically related, but semantically distant forms. Based on these results, they argue for a lexicon that represents morphological structure. However, the priming effects for the semantically distant forms were smaller in magnitude than the semantically related forms. The results of Bentin & Feldman’s study suggest that both semantic and phonological similarity play important roles in the processing of Hebrew words. However, because they did not control for the degree of relatedness on either dimension, it is impossible to assess the relative contribution of each factor. There is one experiment that actually shows effects of phonological relatedness. Although he was testing decomposition versus whole word models, Mac Kay (1978) found different reaction times based on phonological similarity in a lexical production task, where subjects were faster to produce conclusion given conclude (consonant Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 20 change) than to produce decision given decide (consonant plus vowel change). It may be that Mac Kay's stimuli were more balanced in semantic relatedness and therefore provide a more accurate indication o f the role of different degrees o f phonological relatedness in processing complex words. Even when researchers have been more careful about the semantic transparency of their stimulus items, they often employ a relatively arbitrary measure for categorizing items as semantically transparent or opaque. For example, Marslen-Wilson et al. (1994), using subjects’ ratings on a 1 to 9 scale, select cut-off points of >7 (semantically transparent) and <4 (semantically opaque) to divide their stimuli on this dimension. But clearly most of the items in each of the two categories vary systematically from other items in the same category along this dimension. If semantic relatedness of words is an important aspect of our lexical knowledge, then no account will be adequate that does not take this factor into account in a serious way. The connectionist framework that I will employ allows one to account for gradations in similarity, while it is not clear how models that either represent separate morphemes or do not can account for gradations. The studies discussed above were concerned with semantic, phonological, or orthographic properties of complex words to see if these factors can account for morphological effects. The general consensus from these studies is that the additional contribution of morphological structure is necessary to account for the experimental results. But none of the studies looks at the conjunction of formal and semantic properties. Since it is clear that the complex interaction of semantic and formal factors is what underlies the acquisition o f morphological systems (Andersen, 1992; Clark & Berman, 1984, 1987; Slobin, 1973,1985), it seems only natural that adult morphological processing will be based on the interaction of these same properties. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 21 Morphology is not the only domain where it is important to look at the conjunction of multiple cues. Based predominantly on behavioral research, Bates and MacWhinney (1982, 1987, 1989) have discussed multiple cues in acquisition and processing of a range of linguistic phenomena for a number of languages. Similarly, using a modeling approach, Christiansen, Allen, & Seidenberg (1998) found that simple, single cues, such as prosody, were not sufficient to explain young children’s ability to segment the speech stream into words. They demonstrated that the conjunction of several factors within the model produced a better result in learning than any of the factors alone, or than the sum of all the factors. If the general framework for investigating morphological processing were not the metaphor of a serial computer, interactions between the separate semantic and formal codes become more plausible. The alternative metaphor described in the following section lends itself naturally to explorations of exactly this sort of empirical question. 2.4 Research Based on a Different Metaphor Most of the studies of morphological processing discussed above have assumed the basic metaphor of a dictionary listing o f words or morphemes, and a search process to access stored forms. A number of recent studies have looked to a different metaphor, one in which morphologically complex as well as simple words are represented as patterns of activity across simple units encoding aspects of meaning, sound, or orthography. "Accessing" a word (complex or simple) involves activating the corresponding sound and meaning units which participate in the phonological and semantic patterns of the word. When this basic metaphor is changed, many of the underlying assumptions change, as do consequently the possible explanations and the important questions. The connectionist Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 22 metaphor, while not even coining close to an actual approximation of brain processing, looks to the brain, rather than a computer, for its character. Much of the language research tied to this connectionist metaphor has focussed on inflectional rather than derivational morphology, with a large number of papers devoted exclusively to the acquisition and processing of the past tense in English (e.g., Daugherty & Seidenberg, 1992; MacWhinney & Leinbach, 1991; Plunkett & Marchman, 1991, 1993; Rumelhart & McClelland, 1986). The questions they have asked deal predominantly with differences in the processing of regular (bake-baked) and irregular (go-went) forms, and the main issue is single versus dual mechanisms. The dual mechanism approach assumes that there is a set of rules to process regular items, and a separate store for items that do not follow the rules and therefore must be learned individually (e.g.. Pinker, 1991; Pinker & Prince, 1988). In the single mechanism approach, the rule-governed items and the exceptions are processed by the same system (e.g., Rumelhart & McClelland, 1986; Seidenberg & McClelland, 1989). An important notion in this approach is the quasiregular nature of many of the exceptions to the rules (Seidenberg & McClelland, 1989). For example, some irregular past tense items seem to pattern together in sub-rules (e.g., ring-rang, sing-sang). Subjects often produce ‘irregular’ past tense forms when given a novel item that resembles one of these ‘irregulars’ (e.g., producing splung given spling, rather than splinged (Bybee & Moder, 1983)). The connectionist approach has been shown to capture the quasiregular nature of inflectional morphology. Many of the same issues that have arisen in the debate on inflectional morphology are also relevant to the work presented in this thesis, as derivational morphology exhibits quasiregularity, perhaps to an even greater extent than inflectional. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 23 In addition to these studies, some researchers have also begun to apply some of the same principles to derivational morphology. For example, Li and MacWhinney (1996) created a connectionist model to simulate the processing of two related prefixes, un- and dis-. These simulations were unable to handle both prefixes simultaneously, primarily due to the choice of architecture. Li & MacWhinney used a simple backpropagation network, without incorporating any recurrence. With a model that incorporates recurrent connections, it is probable that the same general principles of the connectionist framework can in fact be applied successfully to derivational as well as inflectional morphology. Perhaps the most relevant work to what is proposed here is a recent paper by Rueckl et al. (1997), who used a connectionist approach to account for effects from a processing study of derivational morphology. The researchers examined long-term morphological priming in three experiments, using both masked and standard fragment completion tasks. They found that morphologically related words produce significant priming, but argue that orthographic and phonological similarity cannot account for the results. While formal factors alone could not explain the experimental results, they point out that the priming effect varies in magnitude as a function o f orthographic similarity. While they interpret these graded effects of orthographic relatedness within a connectionist framework similar to the one I am describing here, the results do not actually distinguish between the single and dual mechanism approaches, nor do the authors actually implement a computational model. The computational model described in Chapter 3, as well as the behavioral experiments described in Chapter 4, were created to test the validity of the connectionist approach to derivational morphology further. The connectionist approach raises questions about the nature of morphological processing, which are tested first by the Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 24 model, and then in the behavioral experiments. Based on the full range of phonological and semantic relatedness of the prime-target pairs, and with no explicit coding of morphological type, the model reproduces the priming effects from the behavioral experiments with human subjects. The behavioral experiments in Chapter 4 test predictions that the connectionist approach makes for morphological processing. In the next two sub-sections I describe the important aspects of connectionist models that are crucial components of the alternative approach to morphology that I argue for here. 2.4.1 Distributed representations In the traditional view of the lexicon, where lexical items are stored as a list and accessed via some access mechanism, storing related forms like talk, talker, talking, and talkative separately seems like a waste of resources. Thus, on this view of the lexicon, ‘economy of storage’ becomes a major motivation for using rules to construct complex items from their base forms, thereby obviating the need for storage of items that can simply be generated by a rule (Taft, 1979). Alternatively, some theorists argued that using rules caused too much of a slow-down in processing speed, and therefore advocated the separate storage of all complex forms, thus while recognizing the huge memory requirements, citing a compensatory gain in access speed (Manelis & Tharp, 1977). The issue of ‘economy of storage’ versus ‘speed of processing’ is an important one for theories that see the lexicon as a list of stored forms (either whole words or decomposed portions of words) with accompanying processing mechanisms. In contrast, connectionist systems do not face this problem of an access/storage trade-off, in part because they can make use of distributed representations. Of course, connectionist models can also use localist representations, where, for example, each word Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 25 is represented by a different unit (McClelland & Rumelhart, 1981). However, distributed representations have several different properties which can provide advantages in modeling higher order cognitive functions (Hinton, McClelland, & Rumelhart, 1986; van Gelder, 1991, 1992). One of these properties is that distributed representations allow the system to reuse the same low level features in many different words. For example, in modeling derivational morphology, the same phonological units can be used repeatedly in encoding different words, with individual items stored as the weights on connections between these processing units. While it is true that more resources (e.g., more hidden units) are needed in order to store more patterns in a connectionist network, the problem exists to a much smaller degree than in the dictionary listing metaphor. In fact, the notion of lexical access becomes almost an irrelevant concept for connectionist systems that employ distributed representations (Seidenberg, 1990). Rather than accessing stored word forms, either whole words or stems, words are computed on the basis of activations over a set of features that encode either semantic, phonological, or orthographic information. Recently, Devlin (1998) has actually demonstrated in a model of semantic processing how distributed representations obviate the need for distinctions between access and storage. In patients with semantic impairments, effects of consistency, frequency, and priming have traditionally used as evidence for an access/storage distinction. In his simulations, Devlin shows that these phenomena can be accounted for in a connectionist network that does not distinguish between mechanisms of access and storage. Devlin’s work also emphasizes the importance of the dynamic nature o f lexical representations within connectionist systems. The model that I develop in this thesis takes advantage of both the distributed and dynamic aspects of lexical representations in connectionist systems. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 26 2.4.2 Basins o f attraction A second important feature o f the connectionist system advocated here is basins of attraction. The network units define a high dimensional space and legal patterns are points in this space. Basins of attraction form around these stable points in the processing space of the network. A range of input patterns will settle over time to one legal attractor point. The patterns in this region o f space make up the basin surrounding the attractor. Although it is impossible to imagine a high dimensional space, if we consider a three dimensional space, one could imagine a bowl, with the attractor at the bottom. Any point that lands inside the bowl will settle to the bottom over time. These basins can be deeper or shallower, wider or narrower, depending on the dynamics of the network. An important aspect of basins of attraction is that they can be componential (Plaut 1995; Plaut & McClelland, 1993; Plaut, McClelland, Seidenberg, & Patterson, 1996). In training an attractor network to read monosyllabic English words, Plaut and his colleagues showed how the componential attractors that developed for regular words allowed the network to generalize and pronounce nonwords. They also showed that although componential attractors reflecting correspondences between spelling and sound develop for regular words, the attractors for exception words are " far less componential" (Plaut et al., 1996: 83). When applied to the task of mapping meaning to sound for complex words, this means that words that are semantically and phonologically transparent will create componential attractors, whereas opaque words will not. Thus in processing government componential attractors can be expected to form for govern and ment, whereas department will create a single attractor basin. Basins of attraction also allow us to look explicitly at interactivity in the system, in this case at interactivity between semantic and phonological processing. In some Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 27 processing and linguistic theories, semantics is seen as an unnecessary complication, with lexical entries simply containing pointers to a complex system o f semantic information which is stored elsewhere (Garnham, 1985). In the system I am proposing, the role of semantics is very different. In a normally functioning connectionist system, it would be difficult for word recognition to proceed without at least some activation within the semantic system. This different metaphor for processing highlights a number of issues. One question concerns the frequency of forms in a language: how do type and token frequency interact with other factors, such as semantic transparency? Bybee and Newman (1995) have shown that type frequency plays a role in determining whether a suffix will be productive or not. Token frequency also plays an important role, in that low frequency irregular forms tend to be regularized over time (Bybee, 1988; Hare & Elman, 1995). Both type and token frequency are important factors in neural networks (Daugherty & Seidenberg, 1992; Gonnerman, Harm, & Andersen, 1997). The frequency of forms in the input set helps to determine the solution the network arrives at for a particular set of words. Another question concerns the role of cue validity for different suffixes, and its relationship to frequency. To answer these questions, it is important to know how consistent a particular meaning to sound mapping is, for example, er->agent. How many words end in -er but don’ t mean 'agent' and conversely, how many words include the meaning 'agent' and don't end in -er? Clark and Berman (1984, 1987) have shown that this issue also arises in language acquisition. Children learn one-to-one mappings of meaning to sound earlier and with fewer errors than words or morphemes with either many-to-one or one-to-many mappings. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 28 The answers to these questions have important consequences for the nature of the mental lexicon, and may turn out to be very different for different languages. For example, in English, semantic similarity to known words is an important cue to determining the meaning of a novel word, but in Hebrew formal relationships may be more central and informative than they are in English. 2.5 Focus on Derivational Morphology, in Particular, Suffixation While many of the studies described above focussed either exclusively on inflectional morphology (e.g., Murrell & Morton, 1974) or used a combination of inflected and derived forms (e.g., Rueckl et al., 1997), I have chosen instead to focus on derivational morphology, and even more particularly on suffixed forms. This choice applies both for the computational model and for the behavioral experiments. In the section that follows, I detail some o f the main differences between inflectional and derivational morphology and explain the rationale behind the choice to focus on derivation alone. This choice is based on three major observations: 1) derivational morphology is a more common process than inflectional; 2) by focussing on derivational morphology alone, the factors that differ between inflection and derivation are controlled (e.g., productivity); and 3) there is a greater range o f semantic and formal relatedness between derived forms and stems as compared to inflected forms and their stems, allowing for a more complete exploration of the effects o f similarity on processing of complex words. The morphological processes that underlie word formation are traditionally divided into three different types: inflectional (e.g., trees), where words are modified to reflect properties such as gender, tense, or number; derivational (e.g., treeless), where the Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 29 syntactic class and often the meaning of the stem changes; and compounding (e.g., tree top), in which two separate words are combined to form a new word. A distinction is often made between inflectional morphology on the one hand, and derivational and compounding processes on the other. Two main types of evidence have been taken as support for this distinction: typological and neurolinguistic. From the typological perspective, it has long been recognized that some languages have very rich inflectional systems, while others have no inflectional system at all (Bloomfield, 1933). At one extreme are languages such as Chinese, a language without inflections; instead Chinese relies heavily on compounding. At the other extreme are languages such as Eskimo, that have rich inflectional systems; in Eskimo a single inflected form can express an idea that would have to be translated by an entire sentence in English. The typological facts reveal that derivational morphology is a more common process in human language: "There are a considerable number of languages without inflections, perhaps none without compounding and derivation" (Greenberg, 1966: 93). Therefore, by focussing my investigation on derivational morphology, the findings will be more widely applicable crosslinguistically. Similarly, within the neurolinguistic literature, there is evidence that derivational morphology is a more basic process. There are occasional reports of aphasic patients who present with impairments to inflectional morphology, but whose derivational morphology appears intact (e.g., Miceli & Caramazza, 1988). For those patients who show a deficit in derived forms, these deficits typically appear in conjunction with problems on inflected forms (Badecker & Caramazza, 1991). Miceli and Caramazza (1988) interpret the dissociation between inflection and derivation as reason to posit separate processing components for the two types of processes: an Inflectional Processing Component (IPC) and a Derivational Processing Component (DPC). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 30 While Miceli and Caramazza (1988) treat inflectional and derivational forms in separate processing components based on neuropsychological evidence, some linguists separate inflection and derivation, based on other factors (e.g., Aronoff, 1976; Anderson, 1982, 1992; Jensen & Stong-Jensen, 1984; Scalise, 1984, 1988; Zwicky, 1990). Another approach is to consider the two word types as ends of a continuum (e.g., Bybee, 1985). My approach is similar to that of By bee (1985,1995), in that inflectional and derivational processes are handled by the same general mechanisms. Differences in the behavior of these forms within the language, in psycholinguistic experiments o f normal processing, and in studies of brain damage, are attributable to differences in their underlying properties. In the following section, I discuss some of the main factors that differ between inflection and derivation and on which many linguists have based their view of these processes as distinct. This discussion also explains the second motivation for focussing on derivation; by limiting the stimuli to one end of the continuum, I control for the factors that vary between inflection and derivation. One difference between inflected and derived forms is their role in syntax. Typically words are inflected for concepts like number, gender, tense, or aspect. These are relational concepts, showing dependencies between items within a sentence, but without substantially changing the meaning of the individual words. Because inflectional processes convey information about relationships between words, they are often considered part of syntax and inflectional morphemes are sometimes termed "grammatical" morphemes. Derivations, on the other hand, usually change the syntactic category of the word to which they apply; for example, adding -er to teach changes the verb to a noun. Derivational morphemes are also often restricted in the grammatical class to which they apply. Thus, the derivational suffix -er applies only to verbs. For the approach that I am considering here, the different syntactic roles of inflectional and Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 31 derivational morphology can be seen as having consequences for their semantic representations. In part, because inflected forms share the same grammatical category as their stems, they will often be more semantically related to one another than derived words and their related stems. Thus, teach and teaches both describe the same action, with the only difference semantically being the number of the agent carrying out the action, teach can indicate plural, but teaches must be singular. In contrast, run describes an action, but runner carries the additional semantic information o f an agent Another difference between inflections and derivations is in their productivity; derivational processes are less productive than inflectional, and their productivity is related to both formal and semantic properties of the roots and affixes. Bybee explains the contrast by noting that the fact that inflections carry littie semantic content allows them to apply across many more lexical items; "The greater specificity in derivational meaning restricts the applicability of derivational processes. In the case of inflection, then, the lack of lexical restrictions coincides with extreme semantic generality" (Bybee, 1985: 86-87). The role of the productivity of affixes (both inflectional and derivational) in morphological representation and processing has been investigated using both psycholinguistic experiments (Baayen, 1994; Schreuder & Baayen, 1997) and corpus based approaches (Baayen, 1994; Baayen & Lieber, 1991). Baayen and colleagues have looked at both English and Dutch corpora. Their findings support a view in which the productivity of suffixes has effects on the lexicon that extend beyond either simple frequency effects, or aggregate frequency effects of a stem and all its morphological relatives. For example, Schreuder and Baayen (1997) found that visual lexical decision latencies were faster to monomorphemic words that were part of large 'morphological families' compared to responses to monomorphemic words that had few lexical relatives. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 32 This effect of family size was found to be more important than the sum m ed frequency of all the forms belonging to the lexical family. While I will not test productivity explicitly in the research reported in this thesis, it is a factor that needs to be explored in further work along these lines. Connectionist simulations of derivational morphology would allow one to test the mechanisms behind productivity and to test explicitly subtle effects of this factor on the nature of representations for complex words and their processing. An additional difference between inflections and derivations is in their distribution; there tend to be fewer inflectional morphemes compared to derivational ones. New derivational morphemes can be added to the language much more easily than inflections; consider, for example, in English the relatively new -aholic suffix, as in chocaholic or workaholic, and also the extension of the cran- morpheme, as in cranapple and crangrape. "The general rule, then, is that affixes which are members of large classes to which new items can be freely added are not inflectional; those which are members of small classes to which it is not possible to add extra members are inflectional." (Bauer, 1983: 23) A final point related to differences in productivity is that, "in derivation there are likely to be large numbers of unpredictable gaps in the system, whereas inflection is much less likely to have such unpredictable gaps" (Bauer, 1983: 27). This means that the correlations between sound and meaning are much stronger for derivational forms than inflectional. This will be important when considering the behavior of derivational forms within a connectionist system designed to exploit the regularities in the mappings between forms and meanings. Inflectional and derivational processes differ on at least two other important dimensions: semantic and phonological similarity. These differences underscore the final motivation for focussing on derivational morphology; it allows for an exploration of a wider range of both semantic and phonological similarity. It is often said that inflectional Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 33 processes do not have a semantic effect on the forms to which they apply, whereas derivational processes do. Or it can be said that inflectional processes have a weaker semantic effect, for example, adding the plural changes the number of cats, but we are still dealing with the same animal, while adding a suffix can change the meaning drastically, for example, harm compared to harmless. Also, derivations can cause a wider range o f meaning changes, and these changes are not always transparent (e.g., serenity is transparent because it is the quality o f being serene, but authority is opaque because its relationship to author is not obvious), whereas inflectional changes are synchronically transparent: "the products of inflectional morphology are semantically regular, whereas the products o f derivational morphology tend not to be" (Bauer, 1983: 28). Applying the general framework I am advocating to derivational morphology therefore permits a wider range of semantic relatedness to be explored. Similarly, inflections tend to cause less change in the phonological form of the stem, for example, in bake-baked the vowel stays the same, versus serene-serenity where the vowel changes. Just as the greater variation in semantic compositionality of derived complex words makes them more amenable to my study, their greater variation in phonological compositionality makes derived forms more suitable for exploring the effects of phonological similarity on processing. Finally, within derivational morphology, there are also different types of sub processes, the main ones being affixation, as in English, and nonconcatenative processes, as in Hebrew. Affixes include prefixes (added to the beginnings of words, e.g., re-write), suffixes (added to the ends of words, e.g., teach-er), infixes (inserted within a word, e.g., guaran-damn-tee), or circumfixes (e.g., German ge-stohl-en). Suffixes are by far the most common cross-linguistically, followed by prefixes. Infixes and circumfixes are much rarer phenomena, and more difficult to learn (Andersen, 1992; Slobin, 1982), Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 34 probably because it is more difficult to process discontinuous forms (e.g., center embeddings, Christiansen, 1994; verb particles, Hawkins, 1994). Systems relying on nonconcatenative morphology, where, for example, different vowels are intercalated within a stem rather than appended to it, are rarer still. Therefore, by focussing on derivational suffixes, I am investigating the most widely applicable form of derivational morphology. Because suffixation is more basic in all these ways, it seems appropriate to begin to explore the account proposed here by focussing on derived forms that are suffixed. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 3 A Computational Model of Morphological Priming In Chapter 2 1 described briefly several psycholinguistic experiments where researchers examined the relative contributions of semantic, phonological, and sometimes orthographic, overlap to the processing of morphologically complex words (either derivational or inflectional). The majority of these studies concluded that morphological structure plays a central role in the representation and processing of the mental lexicon, although they vary somewhat in the exact form this morphological structure takes. In this chapter I look in greater detail at one of the more recent empirical studies of morphological priming, arguing for a different interpretation of the results and consequently an alternative view of role of morphology in the mental lexicon. Marslen-Wilson et al. (1994) found a pattern of priming among morphologically related words that they interpreted as evidence for a mental lexicon organized according to stems and affixes. In their theory, a word is treated either as an unanalyzed gestalt form or as a stem plus affix, based on whether the word is semantically opaque or transparent (Tyler et al., 1990). A morphologically complex word is considered transparent if its meaning can be determined from the combination of its component parts. For example, the word government is transparent, whereas department is opaque because there is no clear semantic relationship between depart and department. In contrast, I propose a more unified account in which gestalt and stem-plus-affix forms can be processed by a single system, simply by exploiting the non-arbitrary relationship between sound and meaning that exists for some complex words. The model described in this chapter seeks to capture this systematicity using explicit representations for Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 36 semantics and phonology, but without an explicit representation of morphological structure. I demonstrate that this connectionist model can exhibit the critical findings of Marslen-Wilson et al., even though the priming effects arise solely from an interaction of semantic and phonological similarity. The computational results strongly support the view that morphology is actually captured in the regularities in the relationships between the phonological forms of words and their m eanings. In the following section I briefly review the experiments and findings of Marslen- Wilson et al. (1994). I then present a reanalysis o f critical data from one of their experiments. The statistical reanalysis is followed by a description o f a connectionist model that was developed and implemented using stimuli from the Marslen-Wilson et al. experiments, testing whether a system that treats all words equally as a product of semantic and phonological correspondence can account for the processing of complex words. I conclude the chapter by enumerating specific predictions made by the results from the model that will further test the alternative view I am advocating. 3.1 Marslen-Wilson et al. (1994) Study Marslen-Wilson and colleagues (Marslen-Wilson et al., 1994) conducted a series of experiments using a cross-modal priming paradigm to explore whether the lexical entries for affixed words are structured according to their morphological characteristics. In these lexical decision tasks, subjects heard a word, then saw a word on a computer screen. The subject was required to press a button as quickly as possible indicating whether the word on the screen was a real word in English or not. Experiment 1 was designed to determine 1) whether phonological similarity could account for priming, and 2) whether the degree of phonological relatedness played a role Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 37 in processing. Morphological and phonological relatedness between primes and targets covaried in this experiment. Words that met the following criteria were considered morphologically related ([+Morph]): they must have entered the English language at the same time historically, and they must have come from the same language originally. Words which did not meet these stipulations were classified as having no morphological association ([-Morph]). There were three degrees of phonological relatedness. Stimuli were classed as phonologically transparent ([+Phon]) if a derived word contained a stem in its entirety and the phonetic realization of the stem was identical within the derived word. An example of this type of similarity is shown by govemment-govem. The stem govern is contained entirely within government and it sounds exactly the same in both words. In contrast, the other two types (both represented as [-Phon]), differed in the nature o f the difference between stem and derived forms. Pairs such as delete and deletion, in which there is a change in the stem-final consonant (the stop consonant ending delete becomes a syllable initial fricative in deletion), represent one class of phonologically dissimilar words ([-Phon]). The second set of phonologically dissimilar words, such as vain and vanity, have different vowels. For these words the underlying stem is said to be different from the surface form of either word (Chomsky & Halle, 1968). Chomsky and Halle suggested that the word vain, for example, has a phonological representation which contains an underspecified vowel, v/En. The same stem underlies vanity. Thus, the vain- vanity pairs are less related than the delete-deletion pairs. By using these two types of phonologically dissimilar word pairs, Marslen-Wilson et al. could determine whether the degree of phonological relatedness was important in lexical processing. Table 3.1 below shows examples o f the different words in each condition. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 38 Table 3.11 Sample Stimuli Used in Marslen-Wilson et al. (1994) Experiment I Condition Example 1. [+Morph, +Phon] friendly-friend 2. [+Morph, -Phon] elusive-elude 3. [+Morph, -Phon] serenity-serene 4. f-Morph. +Phonl tinsel-tin Marslen-Wilson et al. found priming effects for all of the morphologically related conditions, even those which were classed as [-Phon]. The only condition for which no priming was obtained was the [-Morph, +Phon] condition (e.g., tinsel-tin). The researchers concluded from these results that morphological relatedness was the only important factor determining priming in this experiment. As there was no difference in priming effects between the deletion-delete and the vanity-vain pairs, they assumed that phonological similarity, no matter what the degree, did not contribute to the effects seen. This assumption may be problematic. Using their results from Experiment 1 to dismiss phonology as a factor, Marslen- Wilson and his colleagues designed the next two experiments to assess the contribution to the priming effects of semantic relatedness between complex words and stems. Because Experiments 2 and 3 were essentially identical, the following discussion concentrates on Experiment 3 alone.2 In Experiment 3 all the prime-target pairs were morphologically related. Morphological type (stem vs. derived forms) and semantic relatedness were covaried. 1 Table adapted from Marslen-Wilson et al. (1994, p.7). ^Experiment 3 was designed to replicate surprising findings from Experiment 2. It simply incorporates adjustments made to stimulus items which had proven problematic in Experiment 2. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 39 Morphological relatedness was assessed as described for Experiment 1. Semantic similarity was evaluated using an experimental norming technique, in which subjects were asked to judge, on a scale from 1 to 9, how close in meaning two words were (1 standing for “very unrelated” and 9 indicating “very related” meanings). Results from the rating task were used to classify prime-target pairs as either [+] or [-] in semantic similarity, with cutoffs established for each category; scores higher than 6.2 counted as semantically related and scores less than 4.5 were categorized as semantically unrelated. Examples of word pairs from each condition are shown in Table 3.2 below. This table also contains the mean semantic ratings for each condition and the mean priming effects. Table 3.23 Sample Stimuli Used in Marslen-Wilson et al. ’ s Experiment 3 with Semantic Relatedness Ratings and Priming Effects Condition Morph Type Example Semantic rating Priming effect (msec) 1. [-sem, +morph] derived-stem casualty-casual 2.6 -1 2. [+sem, +morph] derived-stem punishment-punish 7.8 41* 3. [-sem, +morph] derived-derived successful-successor 2.0 4 4. [+sem, +morph] derived-derived confession-confessor 7.3 11 5. [+sem, +morph] stem-derived friend-friendly 7.6 52* *p < .05. Priming effects were not obtained for the semantically unrelated conditions, 1 and 3, regardless of whether they were derived-stem or derived-derived pairs. Priming was found for Condition 2 as well as for Condition 5, indicating that the semantically and 3 Table adapted from Marslen-Wilson et al. (1994, p. 16). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 40 morphologically related pairs prime each other in either the derived-stem prime-target order, or the reverse stem-derived order. An unexpected finding is the lack of priming for prime-target pairs in Condition 4, semantically and morphologically related derived-derived pairs such as confession and confessor. Marslen-Wilson and his colleagues concluded from this finding that an additional factor beyond semantic relatedness contributed to the priming effects. They carried out a three-way Anova including the factors of Prime Type (test or control), Semantic Transparency ([±Sem]), and Morphological Type (derived-derived or derived- stem). They found significant main effects o f Prime Type and Morphological Type and no interactions. Based on these statistical tests, the researchers concluded: "Semantic relatedness between a prime and a target is a necessary but not sufficient condition for priming to occur" (p. 17) and proposed that Morphological Type, that is, the fact that both prime and target are derived forms, caused the lack of significant priming in Condition 4. Based on these findings, they developed a model of the mental lexicon where semantically transparent stems are stored with links to suffixes, as shown in Figure 3.1 below. Because the derived-derived pairs in Condition 4 did not show significant priming effects, the researchers added inhibitory links between the suffixes. Thus, following their model, priming is not obtained for suffixed pairs because accessing one suffixed form inhibits any other suffixed form with the same stem. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 41 ment govern -o r Figure 3.1. Marslen-Wilson et al. (1994) stem-affix model for semantically transparent suffixed words^". While I agree with Marslen-Wilson et al. that there is a factor beyond semantic relatedness involved in the priming results, I disagree about the nature of this additional factor. Instead of Morphological Type, I will argue that the conjunction of semantic and phonological relatedness alone can explain the observed priming effects. This possibility was obscured in the Marslen-Wilson et al. study because they prematurely dismissed phonological similarity as a factor based on their Experiment 1 results. Indeed, carrying out a regression analysis on their reaction time data makes the role of phonological relatedness in addition to semantic relatedness between primes and targets clear. ^Figure reproduced from Marslen-Wilson et al. (1994, p. 19). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 42 3.2 Reanalysis of Marslen-Wilson et al. Data Two separate regression equations were calculated to probe the interaction of semantic and phonological relatedness (a procedure suggested in Aiken & West, 1991). The regression equations tested the specific prediction that for semantically related words, phonological similarity would be an important factor, but for semantically unrelated words, phonological relatedness would not play a role. The first equation examined the effect of phonological relatedness as a predictor of priming results when semantic relatedness was high. High semantic relatedness was defined as the mean of the semantic ratings for the two semantically related Conditions, 2 and 4; a value of 7.55. Low semantic relatedness was defined as the mean of the semantic ratings in Conditions 1 and 3, or 2.25. The regression equations are shown in 1) and 2) below (Y = priming effect, S-hi = semantic rating - 7.55, S-lo = semantic rating - 2.25, P = phonological relatedness). 1) Y = -100.88 + 20.11 (P) - 21.35 (S-hi) + 4.07 (P) (S-hi) 2) Y = 12.26 - 1.47 (P) - 21.35 (S-lo) + 4.07 (P) (S-lo) At high levels of semantic relatedness, a test of the coefficient for phonological relatedness indicated that it was a significant predictor of priming effects, t (79) = 1.98, p = .05. However, at low levels of semantic relatedness, a test of the coefficient for phonological relatedness indicated that it was not a significant predictor of priming effects, t (79) = -0.16, p = .87. This analysis supports the position that when words are semantically unrelated, such as bulletin-bullet, then the degree o f phonological similarity Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 43 has no effect on priming, but when words are semantically related, the degree of phonological similarity is crucial. Words in Condition 2, such as flattery-flatter, that are highly phonologically related, yield larger priming effects than words in Condition 4, such as maturity-maturation, with less phonological similarity. The connectionist model described in the next section was created to test this view further. Based on the full range of phonological and semantic relatedness of the prime-target pairs, and with no explicit coding of morphological type, this model reproduces the priming effects from the behavioral experiments with human subjects. 3.3 Computational Model A model was developed using the 208 prime and target words from Marslen- Wilson et al.’s Experiment 3. Words were represented as distributed patterns of activation that captured the phonological and semantic relatedness between prime-target pairs. A supervised learning algorithm was employed to adjust the weights between units until the model correctly produced the semantic representation of a word as output when given its phonological form as input. The model was then tested on the actual prime- target pairs used in Marslen-Wilson et al.’s Experiment 3 and its performance was compared with their data from human participants. 3.3.1 Semantic output representations Marslen-Wilson and colleagues obtained semantic similarity values by asking subjects to rate the correspondence in meaning of the prime-target pairs on a 1 to 9 scale. These ratings became the basis for the model’s semantic representations. For each pair of words, a binary vector of 45 zeros was first generated and then 20 percent o f the units (9) Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 44 were randomly chosen and set to 1. This vector was assigned to one of the words in the pair. The vector representing the second word was created such that the overlap between the active features in the two words equaled the semantic similarity rating rounded to the nearest whole number5. If the words in the pair were given a semantic rating of 8, then the second word was assigned a vector which overlapped by 8 units out of the 9 that were on in the first word. For example, delightful and delight had a mean semantic relatedness judgment of 7.9, consequently 8 active units overlapped in this pair. The corresponding vectors are shown in 3) below: 3) delight: 000000 1000000 1 10000 11100001000000000000000 10 1 delightful: 000000100 010011000011100000000000000000000101 Thus, similarity between words was captured by the number of units shared in overlapping semantic representations. While individual units had no particular meaning, nonetheless these representations allow the model to take advantage of the full range of semantic similarity available in the data, rather than dichotomizing the variable into semantically related and unrelated word pairs. In addition, vectors representing the suffix semantics were appended to the 45 unit stem vectors. Suffixes were divided into twelve semantic classes, with an additional unit created to represent each. A first pass at creating the semantic representation o f the suffixes involved classifying them according to whether derived words with these suffixes described things (N), actions (V), manners (ADV). or attributes (ADJ). The nature of the stem was also taken into consideration. Thus, the first division separated suffixes according to changes 5In some cases one prime participated in two prime-target pairs, such as confessor-confess and cofessor- confession. For these triplets, the repeated word was assigned the initial random vector. Two other vectors were created as above for the remaining words. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 45 they make in the meaning o f the words to which they are added. For example, -ment belongs in the group which changes actions to things, as in govern - government. In a second step, further divisions were made for suffixes that, although in the same larger category, resulted in clearly different meanings. For example, the suffixes, -ful and -less both describe attributes, but give opposite meanings when combined with stems, as in harmful compared to harmless. Similarly, the suffixes -or and -ee both can change an action into a person or thing, but while -or means roughly ‘performer of action X’, -ee indicates that the noun is the theme of the verb X, the person or thing which undergoes or experiences X. Therefore, the suffixes -ee and -or were separated into two groups based on differences in their thematic role assignment properties. A further category distinction was made to accommodate differences in the suffix, -able, and the other suffixes that apply to verbs to form adjectives. This is because -able creates an adjective which is understood as the theme of the verb, unlike -ive which also applies to verbs to form adjectives but which is understood as the agent (e.g., abusive). Consider the sentences in 4) below: 4) The boy hugs the teddy bear. The teddy bear is huggable. The teddy bear is the theme of the verb to hug, meaning more or less that teddy bears can be hugged, not that teddy bears do a lot of hugging (Anderson, 1982). The classification of suffixes was carried out in this manner for all 28 suffixes present in the items from Experiment 3 (see Table 3.3 below). Twelve units were added to the semantic representation described above to encode the meaning contained in the suffixes. For words with no suffixes, all of these units were set to zero. Suffixes with the same meaning had that meaning encoded by activating the same unit. One bit was assigned to Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 46 represent each class of suffix. Thus all derived words ending in -ment, -tion, -ance, - ence, and -ure overlapped in their representations; the first unit of the 12 suffix units was set to 1 for these words. Table 3.3 Suffix Classification fo r Semantic Output Representations Sem antic U nit C h a n g e in m ea n in g S uffix es 1. V — > N -ment, -tion, -ance, -ence, -ure 2. ADJ - > ADV -ly 3. ADJ — > Noun -ness, -ity 4. V --> ADJ (theme) -able 5. V — > ADJ (agent) -ent, -ive 6. N — > N (mass) -ship, -age, -ery 7. V — > N (agent) -or 8. V - > N (theme) -ee 9. N — > ADJ -ful, -al, -ary, -ish, -ile, -ic, -ate, -y 10. N — > V -ise 11. N — > ADJ (without) -less 12. ADJ - > V -en The criticism could be raised that by using semantic representations with localist encodings of the different suffixes, we have simply put morphological structure in the semantic system. However, the representations do not use separate units for each suffix in the input set; classes of words with similar meanings are grouped together. There is experimental evidence that words with related meanings tend to become correlated with Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 47 one another (cf. Devlin, Gonnerman, Andersen, & Seidenberg, 1998; McRae, De Sa, & Seidenberg, 1997). The representations used here capture some o f this shared semantic structure. These procedures resulted in a semantic vector for each word that has 57 units, where stems of prime-target pairs share from one to nine units o f meaning and suffixes are encoded with one unit of m eaning. These semantic representations are only an approximation of a system in which a more detailed set of semantic features is used to encode the actual meanings of words (cf. Hinton & Shallice, 1991). In that case, semantic relatedness would result from the pattern of shared features between concepts. The main simplification in these semantic representations was that features in the stem had no fixed, interpretable meanings. For the simulation described here, the approximation generated is adequate, since the important factor is the degree of similarity between word meanings, not the actual meanings themselves. For both the semantic and phonological representations, the individual bits are meaningless, except for the suffix representations. 3.3.2 Phonological input representations First, each word from Experiment 3 was transcribed phonetically based on the surface form of the word and using an inventory that included a total of forty phonemes. A dynamic programming algorithm was then used to determine how phonetically related the words were (Bavelier & Jordan, 1993). This technique was used to provide a quantitative assessment of the degree of phonological similarity between pairs of words in the training corpus. The dynamic programming algorithm uses a graph search routine to compare each phoneme in one word to each phoneme in another. A cost is associated with each comparison that was determined by consulting a confusion matrix for English Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 48 phonemes (Maeda, 1981). The confusion matrix uses experimental data from human subjects to compute a level o f auditory discriminability. A brief run through a sample comparison will illustrate how the dynamic programming algorithm works. The program begins in the lower left comer of the grid as shown in Figure 3.2 below for the words depart and govern. The first comparison for this pair is necessarily D to G, or the initial phonemes o f each word.6 D — i id N R E V o G Figure 3.2. Dynamic programming grid. The program now has the option of moving vertically, horizontally, or diagonally, as indicated by the arrows. The vertical move entails a comparison of the D in depart to the O in govern. A horizontal move would require comparing the G in govern with the E in depart. The last possible move is on the diagonal, a comparison of the two vowels, E and O. The algorithm is constrained so that it cannot move either left or downwards. There is a cost associated with each of the comparisons, extracted from the confusion matrix culled from the study by Maeda (1981). The cost for the three comparisons illustrated above are listed in S) below: ^Uppercase letters are used for convenience here to represent phonemes. However, the algorithm works on phonetic (not orthographic) similarity for purposes of this simulation. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 49 5) D <-> O = 5.6 G <-> E = 6.7 E <-> O = 4.3 The cost of comparing any phoneme to itself was arbitrarily set at 1. This sets up a baseline of I for identical sounds, with all other comparisons having a cost greater than 1. The algorithm continues to move in this manner, up, to the right, or diagonally, computing the cost of each comparison. The algorithm then adds the costs of each comparison as it moves through the grid towards its ultimate goal, the upper right-hand comer, the point where every phoneme from both words has been matched. This position on the graph, marked with an X, represents the comparison of the final two phonemes. The program is designed to select the ''cheapest” possible route through the graph. By cheapest, we mean that it reports the smallest possible sum of the comparisons it chose in traversing the graph towards its destination. This number output by the dynamic programming is then assigned to the prime-target pair as its phonological similarity rating. One problem is the cost differential associated with comparing longer as opposed to shorter words. If two words are identical, say spam and spam, and they have four phonemes as these do, then they would have a phonological similarity rating of 4 according to the dynamic programming algorithm. Each comparison has a cost of one, as the algorithm moves diagonally across the grid. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 50 A M M A P S S <-> S = 1 P<->P= 1 A <-> A = 1 M<->M = 1 Figure 3.3. Identity comparison o f one word to itself. However, two words which are identical but longer will have a higher cost. A comparison of strange to strange will also entail a diagonal search across the graph, but now the cost will be 6 because of the additional phonemes. I clearly do not want to suggest that strange and strange are somehow less related than spam and spam. In order to avoid this result, the dynamic programming algorithm was modified to normalize over the length of the words being compared. The value given as output from the path through the graph is divided by the number o f phonemes in the longer word o f the pair. Thus in the spam case the result of 4 from the four comparisons is divided by 4 phonemes to yield 1. The same thing applies to the strange contrast, 6 phonemes divided into an initial result of 6, leaves a response of 1. Now both comparisons are in line and the algorithm generates a 1 for any two phonetically identical words. In the case of a comparison between two words of unequal length, the algorithm does not stop when it gets to the end o f the shorter word. It must continue its trajectory until reaching the end of the graph. Precisely because the algorithm does not stop at the end of the shorter word, morphologically complex words compared to phonetically identical stems were given unreasonably high values. For example, govern compared to government yields a value indicating a higher degree of dissimilarity than govern Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 51 compared to depart. To control for this unwanted result, the algorithm was once again modified. If at the end of the shorter word, the algorithm has produced a value of 1, indicating phonological equivalence up to that point, then a set cost o f 0.6 is added to the outcome. Thus govern compared to governm ent receives a value o f 1.6. In words where there is a vowel or consonant alternation between the stem and the derived form, the matching is completed in the regular fashion. For example, in the case of delete and deletion the value would be computed using the entire graph search. A drawback o f the dynamic programming approach is that it does not take stress into account. What the algorithm cannot do is recognize the salience o f stressed syllables in the speech stream and adjust the similarity accordingly. The algorithm might judge m issing and smelling to be highly similar because o f the number o f phonemes which they share. In contrast, it would judge spring and flin g as equally or even less similar, since this pair only shares two phonemes, whereas the former pair shares six. However, people would be much more likely to consider spring and flin g more similar because of the prominence of their stressed rhymes (versus the same syllable when it is unstressed as in m issing or smelling). A more realistic algorithm would deal with the differences produced by stress. The output from the dynamic programming algorithm ranged from 1 for a comparison of a word to itself, to 4.9 for the most dissimilar pair in the training set (i.e., relative-relation). Therefore a scale from 1 to 5 was used as the phonological rating parameter. In order to have higher numbers represent more similar sounding words, and lower numbers represent dissimilar words (following the pattern for the semantic ratings), the scale was reversed. Thus, the govem -govem m ent example, for which the dynamic programming algorithm generates a value of 1.6, would have a phonological similarity rating of 5 - 1.6, or 3.4. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 52 To create the input based on these ratings, unique random binary vectors were generated, each with 100 units, 50 of which were on. If a pair of words had a phonological similarity rating of 3.4, this was translated into an overlap of 34 units that are on in the vectors of both words. Thus govern and government share 34 units that are on, and each word has 16 other units that are on but do not overlap. 3.3.3 M odel architecture and training The model was implemented in Xerion7 as a four-layer, connectionist network with 100 input units, 60 hidden units, 57 output units, and 25 clean-up units recurrently connected to the output layer as shown in Figure 3.4. The model received a phonological representation of a word as input and over the course of thirty time steps computed the corresponding semantic representation as output. The model was trained using conjugate gradient descent in conjunction with a line search algorithm. Iterative presentations o f the 208 word corpus continued until each unit's activity was within 0.2 of its target value for each pattern. The model learned the entire training set perfectly. Training took 167 epochs. 7The Xerion simulator was developed by Tony Plate, Drew van Camp and Geoff Hinton at the University of Toronto. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 53 Clean-up 25 Semantics 57 Hidden 60 Phonology 100 Figure 3.4. Model architecture: An attractor network with four layers, phonology, semantics, a layer of hidden units, and a layer of clean-up units. The numbers in each oval indicate the number of units in that layer, while arrows indicate full connectivity between groups. 3.3.4 Priming in the model 3.3.4.1 Procedure Priming was carried out by presenting the model with a test prime, such as absorbent, and allowing the network to settle into its stable state. Then, rather than set the activations produced by absorbent back to zero, as would be done in training, the activity of the output units was allowed to remain. The target, absorb, was then presented to the model. Priming effects were measured by computing the number of time steps necessary for absorb to settle into a stable state when the activation for absorbent was still present in the network. This value was compared to the value for absorb to settle into a stable, correct state when it was presented to the net following a semantically and phonologically unrelated prime, such as w itty. I interpret the time it takes for the Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 54 model to settle as analogous to reaction time in humans (cf. Masson, 1995; McRae et al., 1997; Plaut, 1997). This recognition provides the basis for making a lexical decision in the actual experiment. Because the model was trained on the 208 words from the Marslen-Wilson et al. experiment, I was able to use the same prime-target pairs to test the model. 3.3.4.2 Results and discussion In each condition, the modeling results parallel the behavioral results. In Conditions 1 and 3, where there is little semantic and phonological similarity between prime-target pairs, the model did not show priming effects, as was the case in the behavioral data of Marslen-Wilson et al. (1994)(see Table 3.4 below). Table 3.4 Priming Results from M arslen-W ilson et al. ’ s Human Subjects and the Computational Model Condition Prime-Target Semantic Relatedness Phonological Relatedness Human (msec) Model (time steps) 1 .derived-stem casualty-casual 2.6 2.9 -1 0.3 2.derived-stem punishment-punish 7.8 3.3 41* 2.2* 3 .derived-derived successful-successor 2.0 2.6 4 0.3 4.derived-derived confession-confessor 7.3 2.5 11 1.2 * p < .0 5 . In Condition 2, where there are clear semantic and phonological relationships between primes and targets, I found robust priming effects for the model just as Marslen- Wilson et al. observed for their subjects. I expected priming to be present for examples Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 55 such as absorbent-absorb in Condition 2 because of the strong overlap between the semantic vectors of the words in the pair. Because o f the degree o f overlap in the semantic vectors, the activations left over from the prime were already near the required output activations for the target. This boosted the network towards the correct pattern in a shorter amount of time than needed when following an unrelated word. The results for Condition 4 are the most informative. As in the behavioral data, the modeling results showed non-significant priming between derived-derived pairs such as government and governor. For Marslen-Wilson et al. this was an unexpected result. If government primes govern, and government is also semantically and morphologically related to governor, why shouldn't government prime governor? As discussed earlier, Marslen-Wilson et al. concluded that semantic relatedness alone is not sufficient to produce priming, interpreting these findings as clear evidence for morphological decomposition in the mental lexicon. However, by examining the composition o f the set of items which makes up Condition 4, an alternative explanation becomes available for the lack of priming between derived-derived suffixed word pairs: the items in Condition 4 are simply less related both semantically and phonologically than those in Condition 2, where priming effects were found. Graphing the means for each condition from the Marslen-Wilson et al. study shows the discrepancies in Condition 4. Figure 3.5 shows that semantic relatedness tracks the priming effects very closely, except in the case of Condition 4, where the two lines diverge. The same pattern holds for the correlation between phonological similarity and priming effects. Figure 3.6 shows that phonological similarity parallels priming in all conditions except for the fourth one. Finally, examining the relationship between phonological similarity and semantic similarity as illustrated in Figure 3.7, shows clearly that the two factors diverge only in Condition 4. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Semantic Relatedness 56 Semantic Relatedness and Priming Effects 8 r50 Semantic Prim ing 7 -40 6 -30 5 -20 4 -10 3 -0 2 1 -10 four two one Condition Figure 3.5. Relationship between semantic similarity and priming effects. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Priming Effect i n Msec Phonological Relatedness 57 Phonological Relatedness and Priming Effects 3.6-1 r 50 Phonology P rim in g -40 3.4 - 3.2 - -30 3.0 - -20 2.8 - -10 2.6 - -0 2.4 -10 four two one Condition Figure 6. Relationship between phonological similarity and priming effects Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Priming Effect i n Msec Semantic Relatedness 58 Semantic vs. Phonological Relatedness r3 .6 semantic phonological 6.2 ' -3.2 3.8 • - 2.8 1.4 2.4 four two three one Condition Figure 3.7. Relationship between semantic similarity and phonological similarity. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Phonological Relatedness 59 The results thus indicate a complex interaction between semantic and phonological relatedness, suggesting that it is the degree o f semantic relatedness between the prime-target pairs, in conjunction with their phonological relatedness, that accounts for these findings. 3.4 Discussion For derived-derived prime-target pairs (e.g., govemm ent-govemor) in Condition 4, both the behavioral study and the model showed intermediate priming effects. The model actually predicts this result based on the differences in the semantic and phonological similarity of primes and targets of this type. These word pairs are a little less similar phonologically than the words in Condition 2 (derived-stem pairs such as govem m ent-govem ), and their meanings are also a little more distant than those in Condition 2 where both studies found significant priming. This decrease in the amount of overlap in both the phonological and semantic representations between primes and targets accounts for the decrease in facilitation between Conditions 2 and 4. Because the model results parallel those of human subjects without explicitly encoding morphological properties of the words, these findings suggest that morphological structure need not be directly represented in the mental lexicon. Moreover, the model does not need to posit different storage and access mechanisms for semantically transparent versus opaque complex words. Both types of words are represented as distributed encodings over a set of units and are processed within one mechanism, which argues against the necessity of a hybrid processing system. The model also provides evidence that separate morphological structure need not be explicitly represented in order to show correspondences between stems and suffixed Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 60 forms. Correlations between semantic similarity and phonological similarity are sufficient to cause relationships to be formed between pairs of words, such that pairs with a high degree of overlap on both dimensions will prime each other. Morphological relatedness is not a relevant factor and there is no need to posit two processing mechanisms to account for these results. The results from the connectionist model make several specific predictions that will be tested by the empirical studies presented in Chapter 4: 1) differences in semantic relatedness between derived words and stems form a continuum, therefore priming results should be graded, reflecting the full continuum of relatedness; 2) phonological relatedness is also finely graded in nature, and this too should be reflected in priming results; 3) the important relationship between primes and targets is the amount of overlap on the dimensions of semantic and phonological similarity, not their morphological structure, hence pairs of suffixed words should prime one another if they are closely related in meaning and sound, and finally; 4) words that overlap in both meaning and sound should prime, even if there is no morphological relationship between them. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 4 Behavioral Experiments 4.1 Introduction In this chapter I describe four experiments designed to test the view that morphological structure emerges from the conjunction of two important types o f linguistic information, semantic and phonological similarity. The cross-modal lexical decision task is used to maximize semantic processing of the target words, while minimizing potential effects from overlap in nonlexical auditory or visual input that otherwise arise in intra-modal priming tasks. Words can be related to different degrees on both semantic and phonological dimensions. This variation in the amount o f overlap between primes and targets is predicted to have graded effects on the magnitude of priming produced in the lexical decision experiments. Therefore, in each of the four experiments, the degree of semantic and phonological relatedness is carefully controlled, by being either systematically varied or held constant. 4.1.1 Overview o f experiments The first experiment examines the role of semantic similarity in processing suffixed primes and related stems, predicting larger priming effects for more highly related prime-target pairs. Because phonological overlap is believed to interact with semantic relatedness, the amount o f phonological similarity between primes and targets is controlled in this experiment. In the second experiment I examine the role of phonological overlap on processing suffixed words and their stems. Because the interaction between semantic and phonological similarity is nonlinear (with phonological Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 62 similarity only mattering when words are semantically related), in addition to controlling for semantic similarity, I also used only highly semantically related prime-target pairs. Experiment 3 tests the idea that the type o f derivational relationship between primes and targets is not important for priming to occur between semantically and phonologically related pairs; suffixed primes should facilitate lexical decision to suffixed targets, if the words are highly related. Therefore, Experiment 3 uses suffixed forms for primes as well as targets, unlike Experiments 1 and 2 where suffixed forms were used as primes and related stems as targets. Finally, in Experiment 4 , 1 examine pure semantic and phonological overlap by using primes and targets that have no historical morphological relationship, yet sound and mean alike. Highly related word pairs are expected to prime regardless of their morphological profiles. 4.1.2 The cross-modal lexical decision task The cross-modal lexical decision task was chosen for several reasons. First, in order to examine effects of semantic similarity on processing, it is important to ensure that subjects are accessing the meanings of the words. Simple naming tasks can be performed without accessing semantics, but it is generally assumed that some processing of the meaning of a word must occur in order to make the decision of whether the item is a real word or not (Balota & Chumbley, 1984). Thus lexical decision was used in order to ensure semantic processing. The cross-modal task was chosen to avoid pure form priming that can arise in intra-modal tasks. In intra-modal tasks, when both the prime and target are visual, or both are auditory, simple sensory overlap can cause priming that does not reflect lexical processing (Morton, 1979). In the cross-modal task I used, visual targets were presented immediately after the offset of the auditory prime. This immediate Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 63 stimulus onset asynchrony (SOA) was used to reflect the maximum influence of semantic similarity between primes and targets. Studies of the time course of semantic processing have shown that semantic effects dissipate at longer SOAs. A longer SOA may cause only very strong semantic effects (i.e., between the most highly related word pairs) to be seen, while more subtle effects may be missed. Finally, using cross-modal lexical decision allows for a direct comparison of the results from the experiments reported here with those of other researchers. Specifically, any findings of priming for derived-derived pairs in Experiment 3 can be directly compared to the results from Marslen-Wilson et al. (1994), since their studies used the same methodology. Differences in results here cannot then be ascribed to uninteresting differences in experimental technique. 4.2 Experiment 1: Degrees of Semantic Relatedness This experiment is designed to test the notion that semantic relatedness between stems and derived words forms a continuum, predicting that differences in meaning should have finely gradient effects on priming. More specifically, in this experiment I test the hypothesis that degrees of semantic similarity between primes and targets can predict priming effects, when phonological similarity is held constant: words that are highly semantically related should yield the largest priming effects, words that are somewhat semantically related should yield moderate priming effects, and words that are semantically unrelated should not prime at all, and may even show inhibitory effects. Inhibitory effects could arise if subjects are momentarily confused by the similarity in sound of a target that is actually unrelated in meaning to the target. For example, a thought such as “what does ‘to com’ mean?” may briefly enter the mind of a subject who reads com after having just heard the prime comer. This type of intrusion would Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 64 probably lead to slower reaction times and consequently inhibitory effects for word pairs such as com er-com . 4.2.1 Method 4.2.1.1 Subjects 58 USC undergraduate students participated in the study. They ranged in age from 18 to 42 years, with an average of 20 years. The majority of subjects participated for course credit, although a few were instead paid 5 dollars for their participation. 4.2.1.2 M aterials Semantic relatedness pretest A semantic relatedness pretest was used to determine the degree of overlap in meaning between pairs o f stems and suffixed words (e.g., bake and baker). 135 word pairs were chosen for the pretest. These pairs were all phonologically transparent, such that the derived words contained the entire stem with no consonant or vowel changes (e.g., baker contains bake). In addition, each derived word included a recognizable suffix (e.g., -er, -able, -ment). The word pairs were alphabetized according to the stems and divided evenly into two lists. Subjects were asked to rate the similarity in meaning of the word pairs, using a scale from 1 (very unrelated) to 7 (very related). Subjects were given examples of highly related as well as unrelated pairs. The instructions also reminded subjects that some words may sound alike, but nevertheless have quite different meanings (e.g., ponder-pond) and that these pairs should be given a low rating. 138 USC undergraduates participated in the study for course credit. None o f the subjects who filled out the relatedness pretest participated in the lexical decision experiment. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 65 Mean relatedness ratings were calculated for each pair of words. Results show that subjects are sensitive to degrees o f semantic similarity between related words. Sample ratings are presented in Table 4.1 below. It is clear from this table that subjects were using the complete range o f the rating scale. Some items were rated as very unrelated, for example, message and mess were given a mean rating o f 1.1 out of 7. Other items were seen as clearly highly semantically related, for example, darken and dark were rated 6.2 out of 7. Even more important are the intermediate ratings. Subjects are not just considering words to be related or unrelated, but are able to use the full scale and see some pairs, for example, lately and late, as somewhat related. Moreover, the items are fairly evenly distributed along the rating scale and there is strong cross-subject agreement as to where any particular word pair falls, as demonstrated by low standard deviations for the items. The means and standard deviations for all o f the items are given in Appendix A. Table 4.1 Experiment 1: M ean Semantic Sim ilarity Ratings fo r Sample Item s fro m Pretest (Where a Score o f 1 is Very Unrelated, 7 is Very Related). Sample Word Pair Mean Relatedness Rating message-mess 1.1 pillage-pill • 1.5 • • lately-late • 3.4 shortage-short • 4.1 • • darken-dark • 6.2 hunter-hunt 6.5 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 66 Stimuli The semantic relatedness ratings were then used to select 84 prime-target pairs, falling into three conditions of 28 items each. Items in the low relatedness condition (Condition 2) were rated less than 3; these were items such as com er-com . In the mid condition (Condition 3) the ratings were greater than or equal to 3 and less than 5 (e.g., dresser-dress), and in the high condition (Condition 4) the ratings were equal to or greater than 5 (e.g., teacher-teach). To allow for the examination of phonological similarity in the absence of semantic relatedness, and semantic similarity in the absence of phonological relatedness, two additional conditions were created. Condition 1 consisted of 28 prime-target pairs that were phonologically transparent but semantically unrelated; these pairs were similar to those in the low relatedness condition (e.g., com er-com ), except that the test prime words in this condition did not have recognizable suffixes (e.g., spinach-spin). Using items with no recognizable suffixes could provide evidence of any independent effect of morphology. For example, finding significant facilitation effects for Condition 2 (e.g., com er-com ) but not for Condition 1 (e.g., spinach-spin), would indicate that a factor other than semantic or phonological relatedness, perhaps morphological structure, contributed to the priming effects, since these conditions are matched on semantic and phonological relatedness measures. The absence of differential effects between these two conditions would be consistent with the hypothesis that semantic and phonological relatedness, but not morphological structure, underlie any facilitation o f targets following related primes. However, because sufficient semantic relatedness is hypothesized to be crucial for any priming to occur, it would be very surprising and difficult to account for if either of these conditions produced facilitation effects since both conditions are extremely low in mean semantic relatedness (see Table 4.2). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 67 In the last condition (Condition 5), the 28 word pairs were synonyms and thus highly semantically related, but had no phonological similarity (e.g., idea-notiori). The word pairs in Conditions 1 and 5 were also rated for semantic relatedness in a separate rating pretest described in section 4.3.1.2, which describes the materials for Experiment 3. Sample stimuli for each condition and the mean relatedness ratings are shown in Table 4.2 below. Table 4.2 Experiment 1: Sample Stimuli and M ean Relatedness Ratings (Where a Score o f I is Very Unrelated, 7 is Very Related) fo r Each Condition. Condition Prime-Target Example Mean semantic relatedness I. lo sem, no morph spinach-spin 1.2 2. lo sem com er-com 1.9 3. mid sem dresser-dress 3.9 4. hi sem teacher-teach 6.1 5. hi sem, no phon idea-notion 6.0 For each o f the 140 test primes (S conditions of 28 items each), a control prime was selected to match the test prime in frequency, number of syllables, and part of speech. Test and control primes were neither phonologically nor semantically related. In addition, to avoid experiment-specific response strategies, 140 varied nonword fillers were included, some phonologically related (e.g., slither-slith), others not (e.g., basil- grook). The phonologically related nonwords consisted of two types, words with pseudo- Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 68 affixes (e.g., slither-slith) and words without recognizable affixes (e.g., bishop-bish). This was done to ensure that subjects could not develop a strategy whereby all phonologically related words or all words that ended in suffixes could be correctly responded to as real words. The items were divided initially into two lists, one with the test-target pair (e.g., cowardly-coward), the other with the control-target pair (e.g., demented-coward), so that each subject saw each target only once, preceded either by the corresponding test or control prime. Two separate, pseudo-random orders of all the items were generated to create a total of four lists. All of the test and control primes were digitally recorded using a MacRecorder by a female native English speaker. 4.2.1.3 Procedure Subjects were tested individually in a sound-proofed room. They were seated in front of a Macintosh Performa 6100 computer with a 14 inch Sony Trinitron monitor. Subjects listened to primes played on a high quality Sony speaker placed next to the testing computer. In addition, a button box was placed in front of the subjects. Subjects were instructed to use the button box to answer as fast as possible, while minimizing mistakes. A typical trial proceeded as follows. A fixation point consisting of three asterisks (***) was displayed on the center of the computer screen for 1000 msec. This was followed by the presentation o f the auditory prime. Immediately at the offset of the prime word, the target was displayed on the screen for 200 msec. All targets were presented as lower case letters in 14 point type using a sans-serif font. Subjects responded to the target by pushing the green button on the button box if the target was a real word or the red button if the target was a nonword. The button push ended the trial. A 500 msec delay followed the response before presentation o f the fixation point signaled Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 69 the beginning o f the next trial. To ensure that subjects were attending to the recorded primes, after 15 percent of the trials the instruction "Please repeat the word you just heard" was presented on the computer screen. Subjects repeated aloud the word they had heard, which was written down by the experimenter. Subjects then pressed either of the response buttons to generate the next trial. Subjects were given 20 practice items, followed by 4 warm-up items before presentation of the 280 real and nonword test stimuli. Thus, each subject responded to a total of 304 items. 15 subjects were tested on List 1, 15 on List 2, 14 on List 3, and 14 on List 4. The experiment took approximately 25 minutes to complete, including practice, warm-up, and test trials. Reaction times and lexical decision responses were recorded automatically by the computer. 4.2.2 Results and discussion Data from one subject were excluded from the following analyses because his high error rate makes his data uninterpretable. In addition, 9 items were excluded because more than half of the subjects made errors on those items. Three of these items were from Condition 1, three from Condition 2, and three from Condition 3. This left a total of 57 subjects and 131 items. For the following analyses, all errors were removed from the data set (3.0%), as were extreme values. The reaction times were entered into an ANOVA with the factors of Prime Type (test or control) and Condition (1-5). Both subject and item analyses were calculated. First, there were no significant effects of Prime Type, either by subjects or items, due to facilitation o f the target in some conditions, and its inhibition in others. There was Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 70 a main effect of Condition, which was significant both by subjects, F(4,224) = 29.9, and by items, F (4,252) = 13.7, both p < .001. Although the stimuli were matched across conditions, subjects were slower to respond to targets (e.g., spin) in Condition 1 overall, in both test (e.g., spinach) and control (e.g., m uffler) conditions, indicating that these items may have been slightly less familiar to USC undergraduates. Finally, there were Prime Type by Condition interactions which were significant for both subjects, F (4,224) = 7.9, and items, F (4,252) = 2.5, both p < .05, indicating significant differences between targets preceded by unrelated words as compared to targets preceded by related primes in some conditions. The differences between test and control primes in individual conditions are shown in Table 4.3 below. To evaluate the effects for each type of prime- target pair, separate analyses were carried out for each condition. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 71 Table 4.3 Experiment I: Mean Reaction Times and Priming Effects by Condition Condition Prime-Target Example Mean Reaction Time Control Test Priming effect (msec) 1. lo sem, no morph spinach-spin 649 668 -19 2. Io sem com er-com 607 631 -24 3. mid sem dresser-dress 588 569 19* 4. hi sem teacher-teach 613 573 40* 5. hi sem, no phon idea-notion 593 580 13* *p < .05 In both Conditions 1 and 2, where the primes and targets were unrelated in meaning, there were inhibitory effects, -19 and -24 msec, respectively. Neither of these effects was significant, (r< 1) for both. Thus, words that are semantically unrelated do not prime, whether there is a recognizable suffix (e.g., com er-com ) or not (e.g., spinach- spin). This result calls into question the importance of the morpheme for processing in the absence of a clear semantic contribution of the morpheme to the complex word. It certainly presents a finding that would be difficult to account for in a framework such as the Affix stripping theory of Taft (Taft, 1975, 1981, 1994), where affixes are stripped from stems regardless of the semantic relationship between stem and affix. The results of Conditions 3 and 4 clearly demonstrate that words related in meaning do prime, and that the degree of relatedness affects the magnitude of the priming: moderately related words (e.g., dresser-dress) prime half as much (19 vs. 40 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 72 msec) as highly related words (e.g., teacher-teach). The effects in both of these conditions are significant: Condition 3, t (55) = 3.79, p < .01; Condition 4, t (55) = 4.08, pc.OOl. The results of Condition 5 show that in the absence o f phonological similarity, there is, as predicted, a reduction in priming magnitude even when words are highly semantically related (i.e., 13 msec for idea-notion pairs). This relatively small effect was significant, however, due in large part to the consistency o f response times in this condition, t (55) = 3.85, p < .01. These results provide strong evidence that the magnitude of priming effects increases with increasing semantic similarity. To explore the extent to which this holds true, a regression analysis was carried out with the factor of semantic relatedness predicting differences in target reaction times following controls versus following primes. This analysis takes advantage of the full range of semantic relatedness present in the test stimuli. The semantic relatedness ratings were used to predict the test-control difference scores for items from Conditions 1-4 only. Condition 5 was not included because the items were not phonologically related and since I argue that phonological and semantic relatedness interact, it is important to control the level of phonological relatedness for this analysis. Semantic relatedness was a significant predictor of priming effects, r = .41, p < .001, indicating that subjects are sensitive to subtle differences in the similarity of meanings of pairs of words, and these ratings can predict priming effects in a lexical decision task. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 73 60 40 3 20 £ £ e o • C 0. -20 -40 Figure 4.1. Experiment 1: Priming effects for items with different degrees o f semantic relatedness: Low (Conditions 1 and 2), Mid (Condition 3), High (Condition 4, phonologically related; & 5, phonologically unrelated). Error analysis In addition to analyzing the reaction time data, analyses of the error rates were also carried out. The error rates were entered into an ANOVA with the factors of Prime Type (test or control) and Condition (1-5). The error rates for each condition are shown in Table 4.4 below. First, there was a significant main effect of Prime Type, F(l, 56) = 3.93, p < .05, indicating that it was generally more difficult to respond correctly to the targets when they were preceded by test primes compared to unrelated control primes. A significant main effect of Condition, F(4, 224) = 34.6, p < .001, shows that it was especially difficult Condition 1: lo sem, -morph, spinach-spin Condition 2: lo som, comer-com Condition 3: mid sem, drsssor-dress Condition 4: M sem, tsacher-teach Condition 5: hi sem, -phon, idea-notion 1 2 3 4 5 Condition *p < .05 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 74 to respond correctly to targets in Condition 1. The test primes and targets in Condition 1 were related in sound, but not in meaning. The only difference between Conditions 1 and 2 was that Condition 2 primes ended in ‘suffixes’ (e.g., -er in com er), whereas Condition 1 primes did not (e.g., spinach). There was also a significant Prime Type by Condition interaction, F(4, 224) = 10.9, p < .001, reflecting the fact that the greatest proportion of errors occurred when a Condition 1 stem (e.g., spin) followed a related prime (e.g., spinach). Table 4.4 Experiment 1: M ean Error Rates fo r Lexical Decision to R eal Word Targets by Condition. Condition Prime-Target Example Mean Error Rate Control Test Overall 1. lo sem, no morph spinach-spin .05 .10 .08 2. Io sem com er-com .02 .02 .02 3. mid sem dresser-dress .03 .02 .02 4. hi sem teacher-teach .02 .02 .02 5. hi sem, no phon idea-notion .01 .01 .01 Real vs. nonword targets Responses to real word and nonword targets were compared, using measures of both reaction time and error rate. Nonword targets were responded to more slowly than real words, t (56) = 111.2, p < .001. Subjects also made significantly more errors to nonword targets than to real words, t (56) = 14.6, p < .001. Reaction times and error rates for real words compared to nonwords are shown in Table 4.5 below. These results are Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 75 similar to findings from a classic study of lexical decision effects, in which subjects were significantly slower to respond correctly to nonword compared to real word targets (Rubenstein, Garfield, & Millikan, 1970). The results reported here show that subjects' increased difficulty in responding to nonword targets is reflected both in higher error rates as well as in slower reaction times. Table 4.5 Experiment 1: Mean Reaction Times and Error Rates fo r Lexical Decision to Real Word and Nonword Targets. Mean Mean Error Word Type Reaction Time Rate Real Word Targets 611 .03 Nonword Targets 760 .06 Nonwords Further analyses of the different types of nonword stimuli were also carried out. Mean reaction times and error rates for the various conditions are shown in Table 4.6 below. An ANOVA was calculated with error rate as the dependent variable and Condition (1-5) as the independent variable, F(4, 224) = 12.8, p < .001. Results from this analysis indicate that subjects were more likely to make errors on certain types of nonwords compared to others. The highest error rates were in the two conditions where the nonword target looks like the suffixless stem of its real word prime (e.g., bishop-bish and slither-slith). While there was also a significant effect o f Condition (1-5) on reaction time, F(4,224) = 3.47, p < .01, the results of this analysis are not as easy to interpret. Responses to the slither-slith condition were slower than to the phonologically unrelated primes and targets in the hostess-dight condition. However, responses to the bishop-bish Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 76 stimuli were not particularly slow, while the basil-grook responses were. The bishop- bish stimuli may be distinct enough from real suffixed targets that subjects are not likely to be lead astray by thinking that bish could be a word without the -op ending. There certainly are not very many English words that end in -o p (63 forms listed in Kugera and Frances (1967)) and even fewer that end in -op where the sounds form a syllable that could reasonably be separated from the ‘stem’ (13 forms). Compare the Ku^era and Frances (1967) frequencies o f wallop (1), dollop (0), and trollop (1), to shop (63), stop (120), drop (59). English speakers are not used to seeing non-suffix endings on words and are less likely to erroneously believe the portion of the word without the ending, in this case bish, is actually a real word. As for the slower reaction times in the basil-grook condition, I can only suppose that some of the nonwords in that condition happened to be particularly difficult to distinguish rapidly from real words. Table 4.6 Experiment 1: Mean Reaction Times and E rror Rates fo r Lexical Decision to Nonword Targets by Condition. Condition Prime-Target Example Mean Reaction Time (msec) Mean Error Rate 1. phon rel, change in target computation-compuse 751 .03 2. phon rel, no 'suffix' bishop-bish 752 .07 3. phon rel, pseudo suffix slither-slith 767 .09 4. no phon, suffixed prime hostess-dight 747 .05 5. no phon, no suffix basil-grook 771 .06 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 77 4.3 Experiment 2: Degrees of Phonological Relatedness The results from Experiment 1 support the idea that the relationship in meaning between stems and derived words is an important factor in the mental representation and processing of words. In this experiment I investigate the role o f sound overlap between complex words and their related stems. Marslen-Wilson et al. (1994) argued that the degree of phonological compositionality of complex words did not matter for lexical processing, based on results from a priming experiment where they found no effects for phonological relatedness. However, the items in their experiment were not controlled for semantic relatedness. In contrast, results from the model described in Chapter 3 indicate that phonological overlap does indeed play a role in processing complex words, albeit a secondary role to semantic relatedness; phonological overlap only matters when semantic relatedness is high. In the experiment described below, I test the hypothesis that the amount of phonological overlap does indeed affect processing, as measured by magnitude of priming effects in a lexical decision task, when the degree of semantic relatedness is high and is held constant. 4.3.1 Method 4.3.1.1 Subjects 51 USC undergraduate students participated in the study, none of whom had participated in Experiment 1. They were given course credit or were paid 5 dollars for their participation. Subjects ranged in age from 17 to 28 years, with an average of 20 years. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 78 43.1.2 M aterials Phonological relatedness metric Clearly some pairs of words sound more similar than others, for example triumph and triumphant have a more transparent phonological relationship than sign and signal do. However, while subjects were able to give reasonable ratings for the relationship in meaning of related stems and suffixed forms, the ratings given by subjects for phonological relatedness, reported in Chapter 3, did not seem as reliable. The ratings were clustered around the midpoint of the scale, ranging from approximately 4 to 6 (on a 1-9 scale) for all the related items. Therefore, rather than collecting ratings of phonological similarity from subjects, for this experiment I created a phonological relatedness metric. This metric relies on two simple notions: 1) that a vowel change between a stem and a derived form creates more distance in phonological space than a consonant change; and 2) that changes in phonemes are additive, such that a consonant change accompanied by a vowel change creates more distance than either a consonant change or a vowel change alone. These two assumptions led to the following set of four relatedness conditions, ordered from most to least similar phonologically: 1) no change, where the derived form contains the complete stem, without phonological modification (e.g., acceptable-accept); 2) consonant change, where there is a change in a consonant of the stem in the derived form (e.g., absorption-absorb); 3) vowel change, where the derived form differs from the stem in vowel quality only (e.g., criminal-crime); and 4) consonant and vowel change, where the stem in the suffixed form differs both in a consonant and a vowel from the stem in its simple form (e.g., introduction-introduce). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 79 Semantic relatedness pretest To be sure that any effects obtained in the lexical decision experiment are due to differences in the degree of phonological similarity between primes and targets, it is necessary to ensure that the prime and target pairs are controlled for semantic similarity across conditions. To obtain a set o f prime and target pairs that were matched in the degree of semantic similarity across conditions, candidate word pairs of different levels of phonological overlap were subjected to a relatedness pretest. For each o f the first three phonological relatedness categories described above, 56 word pairs were chosen. Items in the last category, word pairs that differed both in a consonant and vowel, were more difficult to find. Therefore only 46 items o f that type were included. Four separate lists were created, one for each phonological relatedness category. In addition to the 56 (or 46) phonologically related word pairs, 60 additional items were also included, which fell into four classes of 15 pairs each: 1) morphologically related but semantically distant word pairs (e.g., succession-successful); 2) both morphologically and semantically unrelated pairs (e.g., violence-violin); 3) both morphologically and semantically related pairs (e.g., boyish-boyhood); and finally, 4) synonyms (e.g., porpoise-dolphin). The first three classes of word pairs were phonologically related (to slightly varying degrees), while the last set of word pairs (synonyms) were phonologically distant. The same 60 items appeared on each list Thus approximately 25 percent of the items on each list were semantically unrelated, while 12.5 percent were synonymous, or highly semantically related. The additional 60 items were included to ensure that subjects used the full range of the scale available. The test and filler items were alphabetized and given to subjects in booklet form. Subjects were asked to rate the similarity in meaning of the word pairs, using a scale from 1 (very unrelated) to 9 (very related). Subjects were given examples of highly related as well as Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 80 unrelated pairs along with sample ratings. The instructions also reminded subjects that some words may sound alike, but nevertheless have quite different meanings (e.g., ponder-pond) and that these pairs should be given low ratings. Subjects were also instructed to skip any items which they did not recognize. This gave us an idea of the familiarity of the items to the USC undergraduate population. 138 USC undergraduates participated in the study for course credit. None o f the subjects who filled out the surveys also participated in the on-line experiment. Mean relatedness ratings were calculated for each pair of words. Results show that subjects are sensitive to degrees of semantic similarity between related words. Sample ratings are presented in Table 4.7 below. It is clear from this table that subjects were using the complete range of the rating scale. The means and standard deviations for all the items used in this experiment are presented in Appendix B. Table 4.7 Experiment 2: Mean Semantic Sim ilarity Ratings fo r Sample Items from the Pretest (Where a Score o f 1 is Very Unrelated, 9 is Very Related). Sample Word Pair Mean Relatedness Rating violence-violin 1.2 capitalism-capital • 3.6 • • righteous-right • 4.9 jouraalism-joumal • 5.6 • • intelligent-smart • 7.8 decision-decide 8.2 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 81 Stimulus pairs The semantic relatedness ratings were used to select 112 prime-target pairs, falling into four conditions o f different phonological relatedness levels: Condition 1) no phonological change between the stem and derived words (e.g., acceptable-accept); Condition 2) consonant change only (e.g., absorption-absorby Condition 3) vowel change only (e.g., criminal-crime); Condition 4) both a consonant and vowel change (e.g., introduction-introduce). To balance the experimental design, two additional conditions were created: Condition 5) semantically unrelated word pairs (e.g., accordion- accord); and finally, Condition 6) semantically related but phonologically unrelated pairs (e.g., porpoise-dolphin). Items in Conditions 1 through 4 were chosen such that they were highly semantically related and there were no differences in the semantic relatedness ratings across conditions. There were 28 items in each of these six conditions. Sample stimuli for each condition and the mean semantic relatedness ratings are shown in Table 4.8 below. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 82 Table 4.8 Experiment 2: Sample Stim uli and M ean Relatedness Ratings (W here a Score o f 1 is Very Unrelated, 9 is Very Related) fo r Each Condition. Condition Prime-Target Example Mean semantic relatedness 1. no change acceptable-accept 7.4 2. consonant change absorption-absorb 7.6 3. vowel change criminal-crime 7.5 4. consonant and vowel change introduction-introduce 7.4 5. lo semantic relatedness accordion-accord 2.0 6. hi sem, no phon porpoise-dolphin 7.6 For each of the 168 test primes, a control prime was selected to match the test prime in frequency, number of syllables, and part of speech. Test and control primes were neither phonologically nor semantically related. In addition, to avoid experiment- specific response strategies, 168 varied non word fillers were included, some phonologically related and others not. The 70 phonologically related primes and nonword targets consisted of four types, paralleling the differences in phonological relatedness of Conditions 1 through 4, such that some items exhibited no change (e.g., territory-territ), some differed by a consonant change only (e.g.,foundation-foundate), some by a vowel change only (e.g., marital-marite), and some by both a consonant and a vowel change (e.g., struggle-struge). There were 98 phonologically unrelated nonwords (e.g., boomerang-jaulic). The primes for the nonwords were matched in grammatical category, frequency, and number of syllables to the real word primes, to minimize any strategies that subjects might develop based on those stimuli characteristics. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 83 Subjects were given 20 practice items, followed by five warm-up items before presentation o f the 336 real and nonword test stimuli. Thus, each subject responded to 361 items. The items were divided into two lists so that each subject saw each target only once, preceded either by the corresponding test or control prime. Two separate, pseudo random orders of all the items were generated to create four lists. 13 subjects were tested on Lists 1 to 3, and 12 on List 4. 4.3.1.3 Procedure The procedure was exactly the same as described in Experiment 1. The experiment took approximately 30 minutes to complete, including practice, warm-up, and test trials. Reaction times and lexical decision responses were recorded automatically by the computer. 4.3.2 Results and discussion For the following analyses, all errors were removed from the data set (2.5%), as were extreme values (greater than 2000 msec or less than 200 msec). The reaction times were entered into an ANOVA with the factors of Condition (1-6) and Prime Type (test or control). Both subject and item analyses were calculated. First, there were significant main effects of Prime Type, both by subjects, F(l, 50) = 78.9, and by items, F(l, 324) = 18.2, both p < .001, indicating that responses to the target items were faster overall following the test primes compared to the control primes. There were also main effects of Condition, significant both by subjects, F(5, 250) = 34.5, and by items, F (5,324) = 8.9, both p < .001. Although the stimuli were matched across conditions, subjects were slower to respond to targets in Condition 5 overall, in both test Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 84 and control conditions, indicating that these items may have been slightly less familiar to USC undergraduates. Finally, the Prime Type by Condition interaction was significant by subjects, F(5, 250) = 8.4, p < .001, but not by items, F (5, 324) = 1.8, p = .11, indicating significant differences between targets preceded by unrelated words as compared to targets preceded by related primes for the subject analysis. The items analysis may not have reached significance due to the facilitation in most conditions of the targets following the test primes (Conditions 1-4 and 6) and their inhibition in Condition 5. The mean reaction times, as well as the differences between test and control primes in individual conditions, are shown in Table 4.9 below. To evaluate the effects in each condition, separate analyses were carried out for each condition. Table 4.9 Experiment 2: M ean Reaction Times and Priming Effects by Condition Condition Prime-Target Example Mean Reaction Time Control Test Priming effect (msec) 1. no change acceptable-accept 623 576 47* 2. consonant change absorption-absorb 662 597 65* 3. vowel change criminal-crime 656 608 48* 4. consonant and vowel introduction-introduce 674 639 35* 5. lo semantic relatedness accordion-accord 677 691 -14 6. hi sem, no phon porpoise-dolphin 661 621 40* *p < .001 In Condition 5, where the primes and targets were unrelated in meaning, there was an inhibitory effect, -14 msec, that did not reach significance, t (50) = 1.6, p = .21. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 85 Thus, as seen in Experiment 1, words that are semantically unrelated (e.g., accordion- accord) do not prime. In Condition 6, where the test prime and target pairs were synonymous, there was significant facilitation, t (50) = 19.0, p < .001. For Conditions 1 through 4 the results clearly indicate that primes do facilitate lexical decision to targets that are related in meaning and sound, and furthermore that the degree of phonological relatedness affects the magnitude o f the priming, with the general trend being an increase in magnitude of the priming effect as phonological similarity increases, see Figure 4.2 below. The effects in all four of these conditions are significant: Condition 1, t (50) = 39.0, p < .001; Condition 2, t (50) = 38.5, p < .001; Condition 3, t (50) = 33.0, p < .001; Condition 4, t (50) = 13.5, p < .001. The prediction for this experiment was for decreases in the magnitude of the priming effects, with the greatest effect predicted for Condition 1, considered the most phonologically related, the smallest effect for Condition 4, the least phonologically related, and intermediate effects predicted for the moderately related Conditions 1 and 2. This prediction was upheld for Conditions 2, 3, and 4. However, Condition 1 primed less than Condition 2 (47 vs. 65 msec), although the difference between the conditions was not significant. The smaller effect in Condition 1 may be because many of the derived words in Condition 1, while changing neither the consonantal nor the vowel qualities of the stem, are sometimes resyllabified when a suffix is added, for example absorbent and absorb. Perhaps absorption, even though it entails a consonant change, should be considered more similar to absorb than absorbent, because absorption retains the syllable structure of absorb, with the stop in the final coda of the stem. In any case, the most important prediction of a general decrease in magnitude o f priming effects as stimulus pairs decrease in phonological relatedness is held up in this experiment, especially in that Condition 2 items (e.g., absorption-absorb) prime significantly more than Condition 4 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 86 items (e.g., decision-decide), F(l,50) = 6.44, p < .01. It may be that a more sensitive measure of phonological relatedness would be better able to predict subtle differences in priming magnitude across decreases in phonological relatedness, much as was done with the semantic relatedness ratings in Experiment 1. C ondi: acceptable-accept Cond 2: absorptlon-absorto Cond 3: criminal-crfme Cond 4: introduction-introduce Cond 5: accord ion-accord Cond 6: porpoisa-dolphin 1 2 3 4 5 6 Condition *p < .05 Figure 4.2. Experiment 2: Priming effects for items matched for semantic similarity but with different degrees o f phonological relatedness: Condition 1) No change, Condition 2) Consonant change, Condition 3) Vowel change. Condition 4) Consonant plus vowel change. Also Condition 5) Semantically unrelated, and Condition 6) phonologically unrelated synonyms. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 87 Error Rate Analysis In addition to analyzing the reaction time data, analyses of the error rates were also carried out. The error rates were entered into an ANOVA with the factors of Prime Type (test or control) and Condition (1-6). The error rates for each condition are shown in Table 4.10 below. First, there was a significant main effect o f Prime Type, F(l, 50) = 9.86, p < .005, indicating that it was generally more difficult to respond correctly to targets preceded by unrelated control primes than targets preceded by test primes, although the actual magnitude o f the difference was small (97% correct for targets following control primes, 98% correct for targets following test primes). There was also a main effect of Condition, F (5 ,250) = 23.42, p < .001, indicating that subjects were significantly more likely to decide incorrectly that real word targets were nonwords for items in Condition 5, the only condition with semantically unrelated primes and targets in this experiment (e.g., accordion-accord). There was also a significant Prime Type by Condition interaction, F(5,250) = 2.24, p < .05, reflecting the finding that the greatest proportion of errors occurred when a Condition 5 target (e.g., accord) followed a phonologically related, but semantically unrelated, prime (e.g., accordion). This last result is similar to the finding in Experiment 1 that phonologically related, but semantically unrelated, primes and targets produced the most errors (e.g., spinach-spin). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 88 Table 4.10 Experiment 2: Mean E rror Rates fo r Lexical Decision to Real Word Targets by Condition. Condition Prime-Target Example Mean Error Rate Control Test Overall 1. no change acceptable-accept .02 .01 .01 2. consonant change absorption-absorb .02 .01 .01 3. vowel change criminal-crime .02 .01 .01 4. consonant and vowel introduction-introduce .04 .02 .02 5. lo semantic relatedness accordion-accord .06 .07 .06 6. hi sem, no phon porpoise-dolphin .03 .01 .02 Real and nonword targets Reaction times and error rates to real word targets were compared to responses to nonword targets. Nonword targets were responded to more slowly than real words, t (50) = 95.73, p < .001. Subjects also made significantly more errors to nonword targets than to real words, t (50) = 9.07, p < .005. Reaction times and error rates for real words and nonwords are shown in Table 4.11 below. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 89 Table 4.11 Experiment 2: M ean Reaction Times and E rror Rates fo r Lexical Decision to R eal Word and Nonword Targets. Mean Mean Error Word Type Reaction Time Rate Real Word Targets 669 .02 Nonword Targets 811 .05 Nonwords Further analyses comparing phonologically related (e.g., territory-territ) and phonologically unrelated (e.g., boomerang-jaulic) nonword stimuli were also carried out. Mean reaction times and error rates for the two conditions are shown in Table 4.12 below. Results indicate that subjects were more likely to respond slowly, t (50) = 28.17, p < .001, as well as to make errors, t (50) = 15.53, p < .001, on nonword targets that were phonologically related to their primes. This result is similar to the finding in Experiment 1, for which the highest error rates were also found for the conditions where the nonword target looked like the sufGxless stem of its real word prime (e.g., bishop-bish and slither- slith). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 90 Table 4.12 Experiment 2: M ean Reaction Times and Error Rates fo r Lexical Decision to Nonword Targets by Condition. Prime-Target Mean Mean Error Condition Example Reaction Time Rate 1. phonologically related territory-territ 833 .07 2. phonologically unrelated boomerang-jaulic 795 .04 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 91 4.4 Experiment 3: Role o f Morphological Type The model of the mental lexicon proposed by Marslen-Wilson et al. (1994) and discussed in Chapter 3, relied heavily on the result of their priming experiment that showed significant facilitation for related derived and stem pairs, such as govemment- govem , but no significant priming for pairs of related suffixed words, such as govemm ent-govemor. This crucial result led Marslen-Wilson and colleagues to propose a model of the mental lexicon where stems are linked to suffixes, which themselves inhibit one another. In Chapter 3 ,1 challenged this conclusion, arguing that a combination of semantic and phonological relatedness could account for the lack of significant priming between suffixed primes and suffixed targets. This experiment is designed to test explicitly the idea that hearing a suffixed word will indeed facilitate lexical decision to another suffixed word if both words are sufficiently similar in meaning and in sound. 4.4.1 Method 4.4.1.1 Subjects 51 USC undergraduate students participated in the study. They were either given course credit or were paid 5 dollars for their participation. None of the subjects had participated in either Experiment 1 or 2. They ranged in age from 16 to 44 years, with an average of 20 years. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 92 4.4.1.2 Materials Semantic relatedness pretest To obtain a set of suffixed-suffixed prime and target pairs that were both highly semantically and phonologically related, a relatedness pretest was developed. 68 pairs of related suffixed words were chosen (e.g., sainthood-saintly). The survey also included 28 synonyms (e.g., sorcery-magic), 28 semantically unrelated but phonologically similar pairs (e.g., catacomb-catalog), and 78 morphologically unrelated filler items. Thus there were a total of 202 items, which were evenly divided into 2 lists of 101 items. The surveys were printed in booklet form and given to USC undergraduates who filled them out for course credit. Subjects were asked to rate the similarity in meaning of the word pairs, using a scale from 1 (very unrelated) to 9 (very related). Subjects were given examples of highly related as well as unrelated pairs along with sample ratings. The instructions also reminded subjects that some words may sound alike, but nevertheless have quite different meanings (e.g., successfitl-succession) and that these pairs should be given low ratings. Subjects were instructed to skip any items that they did not recognize. This gave an indication of the familiarity of the items to the USC undergraduate population. 27 subjects returned List 1, of whom 18 listed English as their first language, and 35 returned List 2, including 29 native English speakers. Mean relatedness ratings were calculated for each pair of words. Sample ratings are presented in Table 4.13 below. It is clear from this table that subjects were using the complete range of the rating scale. Means and standard deviations for each word pair are shown in Appendix C. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 93 Table 4.13 Experiment 3: M ean Sem antic Sim ilarity Ratings fo r Sample Items from the Pretest (Where a Score o f I is Very Unrelated, 9 is Very Related). Sample W ord Pair Mean Relatedness Rating properly-property 1.5 handsome-handful • 1.8 • • drinker-drinkable • 4.3 blindly-blindness • 5.6 • • tourism-tourist • 7.5 sainthood-saintly 7.9 Stimulus pairs The semantic relatedness ratings were used to select 60 prime-target pairs, falling into three conditions: Condition 1) word pairs with no phonological change between test primes and targets (e.g., useful-useless), but only moderate semantic relatedness; Condition 2) phonologically transparent, highly semantically related word pairs (e.g., scientific-scientist); Condition 3) phonologically dissimilar, highly semantically related pairs (e.g., observation-observant). There were 15 prime-target pairs included in Condition 1, 30 in Condition 2, and 15 in Condition 3. The number of items in Conditions 1 and 3 were limited to 15 each to avoid having too many related items in the design. While it would have been interesting to include more items, and would lend more power to any effects in those conditions, the main items of interest were in Condition 2, which made it more important to include 30 items there. Many of the items in Condition 1 were opposites, such as harmless-harmful, or exhibited differences in thematic role, for example drinker-drinkable. Subjects Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 94 consistently gave these types o f word pairs lower semantic relatedness ratings. It was not therefore clear whether test primes in this condition would facilitate reaction times to their related targets or not. Items in Condition 2 were expected to prime, even though both primes and targets were suffixed forms. Condition 3 consisted of suffixed pairs with characteristics more similar to those used by Marslen-Wilson et al. (1994). These were word pairs that were not phonologically transparent: there were changes in either vowel or consonant quality in the stems o f the two suffixed forms (e.g., observation-observant). Items in Condition 3 were also slightly less semantically related that items in Condition 2. Therefore, because Condition 3 items were less related both phonologically and semantically than pairs in Condition 2 ,1 expected them to yield only moderate facilitation effects, that most likely would not reach statistical significance. Thus, I expected essentially the same results for Condition 3 that Marslen-Wilson et al. obtained for similar word pairs. Sample stimuli for each condition and the mean relatedness ratings are shown in Table 4.14 below. Table 4.14 Experiment 3: Sample Stimuli and M ean Relatedness Ratings (W here a Score o f 1 is Very Unrelated, 9 is Very Related) fo r Each Condition. Condition Prime-Target Example N Mean semantic relatedness 1. low sem, hi phon useful-useless 15 4.7 2. hi sem, hi phon saintly-sainthood 30 7.7 3. hi sem, lo phon observation-observant 15 7.4 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 95 For each of the 60 test prunes, a control prime was selected to match the test prime in frequency, number o f syllables, and part of speech. Test and control primes were neither phonologically nor semantically related. In addition, 90 real word filler items were included: 15 filler pairs were semantically unrelated, but historically morphologically related (e.g., casually-casualty); 15 pairs were both semantically and historically unrelated (e.g., catacomb-catalog); and 15 were semantically related, but phonologically unrelated (e.g., sorcery-magic). Since there is strong evidence from Experiments 1 and 2 as to the nature of priming effects for words of this type, matched control primes were not included for any o f these items. Instead, 45 completely unrelated (i.e., morphologically, semantically, phonologically unrelated) prime-target pairs were included (e.g., admiration-exclusive). These last primes were matched for part of speech, frequency, and number of syllables with the primes in Conditions 1-3. Thus, there were a total of 150 real word pairs. 150 nonword pairs were also constructed. These were of three types: 30 pairs were phonologically overlapping and included real suffixes (e.g., respectful-respection), 30 were overlapping in sound, but the target did not end in a real suffix (e.g., stylist- styleem), and 90 were unrelated phonologically (e.g., optimal-brovian). The primes for the nonword targets were matched in grammatical category, frequency, and number of syllables to the primes for the real word targets, to minimize any strategies that subjects might develop based on these stimuli characteristics. Subjects were also given 20 practice items, followed by four warm-up items before presentation of the 300 real and nonword test stimuli. Thus, each subject responded to 324 items. The items were divided into two sets so that each subject saw each target only once, preceded either by the corresponding test or control prime. Two Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 96 separate, pseudorandom orders o f all the items were generated to create four lists. 13 subjects were tested on Lists 1 to 3, and 12 on List 4. 4.4.13 Procedure The procedure was the same as described in Experiment 1, except that the targets were displayed for 250 rather than 200 msec. The extra time was considered necessary for subjects to be able to read the somewhat longer suffixed targets. The experiment took approximately 30 minutes to complete, including practice, warm-up, and test trials. Reaction times and lexical decision responses were recorded automatically by the computer. 4.4.2 Results and discussion For the following analyses, all errors were removed from the data set (3.4%), as were extreme values (over 2000 msec and under 200 msec). The reaction times were entered into an ANOVA with the factors of Condition (1-3) and Prime Type (test or control). Both subject and item analyses were calculated. First, the main effect of Prime Type was significant by subjects, F(l, 50) = 4.60, p < .05, but not by items, F(l, 114) = 2.50, p = . 11, because responses to the target items were faster following the test primes in some conditions, but slower in others. The main effect of Condition was marginally significant by subjects, F(2, 100) = 2.76, p = .07, and not significant by items, F (2, 114) < 1. Finally, the Prime Type by Condition interaction was significant by subjects, F(2, 100) = 3.07, p < .05, but not by items, F (2, 114) < 1. Of greatest interest in this experiment, however, are the effects for each individual condition because rather than making predictions across all the conditions, there were separate Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 97 hypotheses for each condition. The mean reaction times, as well as the differences between reaction times to targets following test versus control primes in individual conditions, are shown in Table 4.15 below. To evaluate the effects in each condition, separate analyses were carried out for each condition. Table 4.15 Experiment 3: Mean Reaction Times and Priming Effects by Condition Condition Prime-Target Example Mean Reaction Time (msec) Control Test Priming effect (msec) 1. low sem, hi phon useful-useless 624 628 -4 2. hi sem, hi phon saintly-sainthood 655 621 34* 3. hi sem, lo phon observation-observant 652 638 14 *p < .05 In Condition 1, where the primes and targets were less related in meaning, there was a nonsignificant effect o f -4 msec, (t <1). In Condition 2, where items were highly semantically and phonologically related, there was a significant 34 msec facilitation effect, t (50) = 21.0, p < .001. Finally, in Condition 3, where the test primes and targets were less phonologically similar than those in Condition 2, there was a slight facilitation effect, 14 msec, that did not reach significance, t (50) = 1.3, p = 0.26. Interestingly, the result in Condition 3 is very similar to the result obtained by Marslen-Wilson et al. (1994) who found an 11 msec facilitation effect for their suffixed prime-target pairs that also failed to reach significance. There are thus two possible explanations, either: 1) the semantic relatedness of the items in Condition 3 was simply not high enough and it needs to reach a threshold before there is significant priming, or more likely; 2) the semantic Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 98 and phonological similarity of these items is low enough that a marginal facilitation effect is produced, which would reach significance if there were sufficient power to detect the result. Unfortunately, because there were only IS prime-target pairs in Condition 3, there may not have been enough items. Similarly, Marslen-Wilson et al. only had 20 items in their experiment. To determine whether more power would lead to a significant priming result for items of the observation-observant type, it would be necessary to test a greater number o f items o f this sort. Perhaps with 30 items in Condition 3 there would have been a small, yet significant, facilitation effect. Clearly, the results from this experiment show that hearing a suffixed word facilitates lexical decision to another suffixed word, when the words are similar enough in both meaning and sound. Furthermore, the results underscore the importance of the factors of semantic and phonological relatedness working in conjunction to produce facilitation between primes and targets: either factor alone does not produce priming. Thus, in Conditions 1 and 3 where only one factor was high (i.e., high phonology but low semantics in Condition 1, higher semantics, but lower phonology in Condition 3) there were no significant priming effects. Only in Condition 2, where primes and targets were both highly semantically and phonologically related was there significant facilitation. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 99 Condition 1: useful-useless Condition 2: saintly-sainthood Condition 3: observation-observant 34* Condition Figure 4.3. Experiment 3: Priming effects for suffixed word pairs: Condition 1) Low semantic, high phonological relatedness; 2) High semantic and high phonological relatedness; Condition 3) High semantic, low phonological relatedness. Error Rates Analyses of the error rates for real word targets in the various conditions were also carried out. The error rates were entered into an ANOVA with the factor of Condition (1-7). The error rates for targets following control primes versus following test primes were not examined separately because there were no control primes for Conditions 4 through 7 in this experiment. There were negligible differences between the Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 100 targets following control and test primes in Conditions 1 to 3. The error rates for each condition are shown in Table 4.16 below. There was a significant main effect of Condition, F(6,294) = 25.73, p < .001. The greatest proportion of errors occurred in Condition 5, where prime-target pairs were phonologically, but not semantically related, and Condition 7, where primes and targets were unrelated both semantically and phonologically. There were essentially no differences between the error rates for the three conditions with semantically and phonologically related suffixed prime-target pairs. This is essentially the same pattern of results that occurred in Experiments 1, where the most difficult real words were pairs such as spinach-spin, that were semantically unrelated, but phonologically related. It appears that subjects are thrown by the lack of relatedness in meaning, and forget momentarily that the target has an entirely separate, unrelated meaning. Thus, a subject sees catalog after hearing catacomb, and realizes only that catalog has no meaning related to catacomb, and only secondarily recognizes the actual meaning of catalog, after having quickly and erroneously pushed the nonword button in response. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 101 Table 4.16 Experiment 3: M ean Error Rates fo r Lexical Decision to Real Word Targets by Condition. Condition Prime-Target Example Mean Error Rate 1. low sem, hi phon useful-useless .02 2. hi sem, hi phon saintly-sainthood .02 3. hi sem, lo phon observation-observant .01 4. no sem, hi phon, suffixed casually-casualty .03 5. no sem, hi phon, no suffix catacomb-catalog .07 6. hi sem, no phon sorcery-magic .01 7. no sem, no phon admiration-exclusive .08 Real word vs. nonword items Responses to real word targets were compared to responses to nonword targets, using measures of both reaction time and error rate. Nonword targets were responded to more slowly than real words, t (49) = 93.0, p < .001. Subjects also made significantly more errors to nonword targets than to real words, t (49) = 6.07, p < .05. Reaction times and error rates for real words compared to nonwords are shown in Table 4.17 below. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 102 Table 4.17 Experiment 3: Mean Reaction Times and Error Rates fo r Lexical D ecision to Real Word and Nonword Targets. Mean Mean Error Word Type Reaction Time Rate (msec) Real Word Targets 651 .04 Nonword Targets 773 .07 Nonwords Further analyses of the different types of nonword stimuli were also carried out. Mean reaction times and error rates for the various conditions are shown in Table 4.18 below. An ANOVA was calculated with error rate as the dependent variable and Condition (1-3) as the independent variable, F (2,98) = 47.44, p < .001. Results from this analysis indicate that subjects were more likely to make errors on certain types of nonwords compared to others. The highest error rate was in the condition where the nonword target was composed of a real stem and real suffix, that happen not to form a real word in English when they are combined (e.g., respection). Because the form actually creates a reasonable English word, it is much more difficult to realize that this possible form is not attested in English, compared to judging a form that is composed of either a non-existent stem and a real suffix, or two non-existent components. Subjects with smaller vocabularies may be especially prone to falsely accepting such nonwords as real words because they are unsure of the words they know, and forms such as respection clearly are possible words, conforming to norms of English. It would be interesting to correlate vocabulary size with tendency to make errors on nonwords of this type. Unfortunately, no measure of vocabulary size was collected. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 103 The second most difficult condition consisted of nonword targets that were composed of non-occurring stems (e.g., brov-) combined with real suffixes (e.g., -ian, brovian). There was also a significant effect of Condition (1-3) on reaction time, F(2, 98) = 75.55, p < .001, the pattern of reaction times paralleling the error rate pattern, namely responses to nonwords which are formed from unattested combinations of real suffixes and real stems (e.g., respection) elicit the slowest responses, followed by targets with real suffixes (e.g., brovian) and finally targets without real suffixes (e.g., styleem). Table 4.18 Experiment 3: M ean Reaction Times and Error Rates fo r Lexical Decision to Nonword Targets by Condition. Condition Prime-Target Example Mean Reaction Time Mean Error Rate 1. phon rel, pseudo suffix respectful-respection 840 .12 2. phon rel, no 'suffix’ stylist-styleem 738 .02 5. no phon, pseudo suffix optimal-brovian 764 .06 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 104 4.5 Experiment 4: Historically Unrelated Pairs The experiments described thus far have all used stimuli that were related to varying degrees on phonological and semantic dimensions. In addition, these word pairs were all historically morphologically related. While it seems obvious that diachronic relationships between words should have no effect on the synchronic representation and processing of related words (for anyone except maybe a linguist or classics scholar), there are some instances where the importance of historical origins is implied. For example, Marslen-Wilson et al. (1994) were very careful to ensure that all the prime-target pairs in their experiments were morphologically related, where a pair of words was considered morphologically related only when they shared "the same historical source word (or etymon), as determined by the Oxford Dictionary of English Etymology (1983) or the Longman Dictionary of the English Language (1983)" (Marslen-Wilson et al., 1994: 7). In an earlier study, Taft and Forster (1975), found that subjects were slower to reject pseudo-affixed nonwords containing real morphological steins, such as dejuvenate, than they were to reject pseudo-affixed nonwords without real morphological stems, such as depertoire. They interpreted the findings as evidence for a processing strategy that strips affixes from complex words and then searches for a stored stem. I would argue instead that the result simply reflects subjects’ tendency to have more difficulty rejecting non words that more closely resemble real words (cf. Schaeffer & Wallace (1970) for a related example from the semantic priming literature). The logic of this study implies that subjects are sensitive to the historical status of the stem, since neither juvenate, nor pertoire are currently free standing words. It seems important then to answer the question of whether primes that are related in meaning and sound, but happen to come from different roots etymologically, will significantly facilitate recognition of target words. In this experiment I use the cross-modal lexical decision task to test priming of Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 105 semantically and phonologically related, but historically unrelated, word pairs, predicting that those pairs which are both highly semantically and phonologically related will yield significant facilitation effects. 4.5.1 M ethod 4.5.1.1 Subjects Two groups of subjects participated in the experiment: 29 participants were recruited from the University Honors College Program, and 24 participants from the Psychology subject pool. Subjects received course credit for their participation. The Honors College students were tested because the stimulus set included infrequent words which might not have been familiar to a general undergraduate population, represented here by students from the Psychology subject pool. 4.5.1.2 M aterials Semantic relatedness pretest As in the previous experiments, candidate word pairs were chosen and then rated for semantic similarity using a pretest. It is difficult to find word pairs that are related in meaning and sound but that have different origins in the language. Nonetheless, I was able to find 78 potential word pairs. These word pairs were combined with pairs of historically and semantically related suffixed words, synonyms, and morphologically related but semantically unrelated pairs for a total of 202 items. These items were divided into two lists of equal length and printed in booklets. The booklets were filled out by undergraduate students taking an Introduction to Psychology course. The procedure for the survey is described in the Materials section of Experiment 3 (Section Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 106 4.4.1.2). Sample ratings for the historically unrelated items are presented in Table 4.19 below. Table 4.19 Experiment 4: Mean Semantic Similarity Ratings fo r Historically Unrelated Sample Items from the Pretest (W here a Score o f 1 is Very Unrelated, 9 is Very Related). Sample Word Pair Mean Relatedness Rating forceps-force 1.9 scalpel-scalp • 2.5 • • rivulet-river • 4.2 dismal-dismay • 4.6 • • pulley-pull • 6.3 mortuary-morgue 7.4 The semantic relatedness ratings were then used to select 125 prime-target pairs, falling into five conditions: Condition 1) historically unrelated prime-target pairs with high semantic relatedness ratings (greater than 5) (e.g., trivial-trifle); Condition 2) historically unrelated prime-target pairs with moderate semantic relatedness ratings (greater than or equal to 3 and less than 5) (e.g., dismal-dismay)', Condition 3) historically unrelated prime-target pairs with low semantic relatedness ratings (less than 3) (e.g., rankle-rank); Condition 4) highly semantically related prime-target pairs with no phonological overlap (e.g., sorcery-magic); Condition 5) semantically unrelated prime- target pairs with high phonological overlap (e.g., pumpkin-pump). Because it was so difficult to find appropriate historically unrelated prime-target pairs, it was impossible to control for the amount of phonological overlap between test primes and targets in each condition. Thus, some of the test primes and controls are transparently phonologically Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 107 related (e.g., rankle-rank), some contain different vowels (e.g., glamour-gleam), others exhibit a consonant change (e.g., mortuary-morgue), and a very few include both vowel and consonant changes (e.g., trivial-trifle). Also, unlike in the previous three experiments, it was not possible to control for the "morphological" relationship between the test primes and the targets; while the majority of targets appeared to be stems related to "suffixed” test primes (e.g., lunatic-loon), some targets appeared to be "suffixed" forms related to "suffixed" primes (e.g., jubilee-jubilant). Sample stimuli for each condition, the number of items in each condition, and the mean semantic relatedness ratings are shown in Table 4.20 below. The complete list of mean relatedness ratings for all of the items, including the standard deviations, can be found in Appendix D. Table 4.20 Experiment 4: Sample Stim uli A nd Mean Relatedness R atings (Where a Score o f 1 is Very Unrelated, 9 is Very Related) fo r Each Condition. Condition Prime-Target Example N Mean semantic relatedness 1. hi semantic related trivial-trifle 26 6.1 2. mid semantic related dismal-dismay 25 4.0 3. lo semantic related rankle-rank 18 2.4 4. hi sem, no phon sorcery-magic 28 7.3 5. no sem, hi phon pumpkin-pump 28 1.6 For each of the 125 prime words, a control word was selected to match the prime in frequency, number of syllables, and part of speech. Test and control primes were neither phonologically nor semantically related. Thus, there was a total of 125 real word targets, 125 test primes, and 125 control primes. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 108 125 non word pairs were also constructed. These were o f four types: 20 pairs were made up o f a suffixed prime and a stem-like nonword target, where there was a change in the final consonant of the target (e.g., brackish-brap), 10 pairs were composed of a suffixed real word prime, and a nonword created from the stem of the prime combined with a real suffix to create a non-attested form (e.g., dipper-dippance), 20 were phonologically transparent pairs, in which removing the ending of the prime created a nonword target (e.g., bishop-bish or dervish-derv), and finally, 75 were unrelated phonologically (e.g., quiver-proof). The primes for the nonword targets were matched in grammatical category, frequency, and number o f syllables to the real word primes, to minimize any strategies that subjects might develop that would interfere with the word/nonword response. Subjects were given 20 practice items, followed by five warm-up items, before presentation of the 250 real and nonword test stimuli. Each subject therefore responded to a total o f 275 items. The items were divided into two sets so that each subject saw each target only once, preceded either by the corresponding test or control prime. Two separate, pseudorandom orders of all the items were generated to create four lists. 4.5.1.3 Procedure The procedure for the on-line cross-modal lexical decision task was the same as described for Experiment 1. However, at the conclusion of the priming experiment, participants were asked to fill out the semantic relatedness survey that was described earlier in Section 4.5.1.2 on Materials. The on-line experiment took approximately 25 minutes to complete, including practice, warm-up, and test trials. The survey took between five and ten minutes to complete. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 109 4.5.2 Results and discussion Two separate sets o f analyses were carried out: first the data from the Psychology subjects were analyzed, then the data from the Honors college subjects. For the following analyses, all errors were removed from the data set (8% for the Psychology group, 4% for the Honors college group), as were extreme values (greater than 2000 msec and less than 200 msec). In addition, three items were removed from Condition 1 and two from Condition 3 because further investigation showed that they were actually historically related. The reaction times were entered into two separate ANOVAs (one for Psychology subjects and one for Honors subjects) with the factors of Condition (1-5) and Prime Type (test or control). Both subject and item analyses were calculated for both groups. Results for the Psychology group are reported first. Psychology subject group First, the main effect of Prime Type was not significant by subjects or by items, F( 1,23) < 1 for both, indicating that overall there was no difference in reaction time to targets following test primes when compared to targets following control primes. The main effect of Condition was significant both by subjects, F (4 ,92) = 25.49, p < .001, and by items, F (4,230) = 9.76, p < .001. This effect arose from overall faster responses to targets in Condition 4 where the pairs were synonyms (e.g., sorcery-magic). Finally, the Prime Type by Condition interaction was not significant by subjects, F(4, 92) = 1.32, p = .27, or by items, F (4,230) < 1 . As for the priming effects in individual conditions, the only significant result was for the synonym pairs in Condition 4, where there was a 21 msec facilitation effect, t (23) = 4.13, p < .05. No other results reached significance. The Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 110 mean reaction times, as well as the differences between test and control primes in individual conditions, are shown in Table 4.21 below. Table 4.21 Experiment 4: Mean Reaction Times and Priming Effects by Condition fo r the Psychology Subject Group. Condition Prime-Target Example Mean Reaction Time Control Test Priming effect (msec) 1. hi sem related trivial-trifle 681 665 16 2. mid sem related dismal-dismay 637 625 12 3. lo sem related rankle-rank 604 592 12 4. hi sem, no phon sorcery-magic 579 558 21 5. no sem, hi phon pumpkin-pump 625 644 -19 Honors college subject group The results for this group of subjects were quite different from the results for the Psychology group. First, the main effect of Prime Type was significant by subjects, F (1, 28) = 8.12, p < .01, although not by items, F (l, 230) = 1.28, p > .05, indicating that reaction times to targets following test primes were generally faster than reaction times to targets following control primes. There was also a main effect of Condition, significant both by subjects, F(4, 112) = 22.30, p < .001, and by items, F(4, 230) = 5.82, p < .001. As was the case for the Psychology subject group, this effect arose because responses to targets in Condition 4 (sorcery-magic) were faster than in the other conditions. Finally, the Prime Type by Condition interaction was significant by subjects, F(4, 112) = 5.18, p Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. I l l < .001, and marginally significant by items, F (4,230) = 2.04, p = .09. The mean reaction times, as well as the differences between test and control primes in individual conditions, are shown in Table 4.22 below. The pattern of priming effects in individual conditions was much more interesting for this subject group than for the Psychology group. Statistical analyses o f the results in each condition are described in detail below. Table 4.22 Experiment 4: M ean Reaction Times and Priming Effects by Condition fo r the Honors College Subject Group. Condition Prime-Target Example Mean Reaction Time Control Test Priming effect (msec) 1. hi sem related trivial-trifle 657 633 24* 2. mid sem related dismal-dismay 631 636 -5 3. lo sem related rankle-rank 634 615 19 4. hi sem, no phon sorcery-magic 603 549 54* 5. no sem, hi phon pumpkin-pump 613 642 -39* *p < .05 As was the case with the Psychology subject group, the synonymous prime target pairs in Condition 4 yielded a significant facilitation effect, t (28) = 23.3, p < .001. More interestingly, whereas that was the only significant effect for the Psychology group, there were other effects that reached significance for the Honors College group. In Condition 1, where the primes and targets are historically unrelated yet highly semantically related, there was a significant effect facilitation effect of 24 msec, t (28) = 4.27, p < .05. The Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 112 effects for both Conditions 2 and 3, where the prime target pairs were at best moderately related in meaning, did not reach significance. Finally, there was a significant inhibition effect in Condition 5, where the primes and targets were phonologically transparent, but semantically unrelated, t (28) = 5.79, p < .05. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 113 o o M E o « * » o O) c a ^ H E * z Q. -50 Condition 1: trifle-trivial Condition 2: dismal-dismay Condition 3: rankle-rank Condition 4: sorcery-magic Condition 5: pumpkin-pump — r 2 — r 3 54 -39* — r 4 — r 5 Condition *p < .0 5 Figure 4.4. Experiment 4: Priming effects for historically unrelated word pairs: Condition 1) High semantic relatedness; 2) Mid semantic relatedness; Condition 3) Low semantic relatedness; Condition 4) High semantic relatedness, no phonological overlap; Condition 5) Low semantic relatedness, high phonological overlap. Results shown are for the Honors College subject group only. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 114 The priming effect obtained in Condition 1, the highly semantically related but historically unrelated items (e.g., trivial-trifle), was examined more closely. Stimuli within Condition 1 were of two basic types: either 1) pseudo-suffixed items (e.g., two ‘suffixed’ forms, jubilee-jubilant or a ‘suffixed’ prime and a ‘stem’ prudent-prude), or 2) items with no clear ‘stems’ or ‘suffixes’ (e.g., halo-holy). In order to examine the difference in priming for the two types o f prime-target pairs in this condition, the stimuli were divided into two groups. There were 14 ‘pseudo-suffixed’ pairs, and 9 unsuffixed pairs. The mean reaction times for these two stimuli groups are presented in Table 4.23 below. Table 4.23 Experiment 4: M ean Reaction Times fo r ’ Pseudo-suffixed’ and ‘ N on-suffixed’ Targets Following Control and Test Primes. Priming Effects are A lso Shown. Mean Reaction Time (msec) Priming Effect Suffix Status Prime-Target Example Control Test (msec) 1. Pseudo-suffixed jubilee-jubilant 656 626 30* 2. Non-suffixed halo-holy 655 642 13 *p < .05 Reaction times for Condition 1 targets were entered into an ANOVA with Prime Type (test or control) and Suffix (pseudo or non) as factors. There were no significant effects, indicating that whether or not the primes or targets appeared to include a suffix had no bearing on the priming results. However, analysis of the individual priming Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 115 effects for the pseudo-suffixed and the non-suffixed items did show differences. Whereas the pseudo-suffixed items yielded a significant 30 msec facilitation effect, the non-suffixed pairs produced a non significant 13 msec effect. The small number of items in the non-suffixed set (n = 9) may have contributed to the lack o f significance for this effect by lowering the power in the analysis. One might argue that these results do support the notion of a role for morphology in priming effects; however, there is an alternative explanation that is consistent with the results of the other three experiments and the approach I am advocating here. The explanation lies in the fact that the items that primed, while not significantly more related semantically, were much more related phonologically; 75% of the non-suffixed primes and targets changed either in a vowel or a consonant, or both, whereas only 14% of the suffixed pairs were not phonologically transparent. The role of phonological overlap was demonstrated clearly in Experiment 2, where suffixed pairs that were phonologically transparent yielded significant facilitation (e.g., saintly-sainthood), but pairs that were more phonologically different did not yield significant priming effects (e.g., observation-observant). The same principle is at work in the case of the pseudo- versus non-suffixed stimuli in this experiment. Psychology vs. Honors College semantic relatedness ratings An interesting question is why results from the Honors College group showed the expected facilitation effects in Condition 1, while the Psychology group did n o t One explanation is that the Honors college group was simply more familiar with many of the words. There can be no semantic priming if the subjects do not know the meanings of the primes and targets. The error analyses described below lend support to this hypothesis. Further support comes from a comparison of the semantic relatedness ratings given by the Psychology group in the exit survey and those provided by the Honors college group. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 116 Mean semantic relatedness ratings for the Psychology group and the Honors college group are shown below, along with the ratings from the Pretest group originally obtained to determine the division of prime-target pairs into high, mid, and low semantic relatedness conditions. Table 4.24 Experiment 4: Mean Semantic Relatedness Ratings (Where a Score of 1 is Very Unrelated, 9 is Very Related) From Pre- and Post-Test Survey Administration. Shown for Each Condition by Subject Group. Condition Prime-Target Example Mean Semantic Pretest Psyc Post Psyc Rating Post Honors 1. hi semantic related trivial-trifle 6.1 6.0 6.6 2. mid semantic related ruffian-rough 4.0 3.8 4.4 3. lo semantic related rankle-rank 2.4 2.4 2.6 While the ratings are all highly correlated with one another (see Table 4.25 below), the ratings from the Honors college students are significantly higher than those from the Psychology students in Condition 1, t (44) = 1.67, p < .05. The higher semantic relatedness ratings, which suggest richer semantic representations, are reflected in the facilitation found for the Honors subjects, but not the Psychology subjects. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 117 Table 4.25 Experiment 4: Correlation M atrix o f Semantic Relatedness Ratings by Group Name Pretest (Psyc) Post (Honors) Post (Psyc) Pretest (Psyc) 1 Post (Honors) .876 1 Post (Psyc) .844 .843 1 Error rate analyses In addition to analyzing the reaction time data, analyses of the error rates were also carried out. The error rates for the Psychology subject group and the Honors College group were entered into separate ANOVAs with the factors o f Prime Type (test or control) and Condition (1-5). Only subject analyses were calculated. Psychology subject group First, the effect of Prime Type was not significant, F (l, 23) = 1.22, p = .28, indicating that making a correct lexical decision to target words was equally difficult following test primes compared to following control primes. There was a significant main effect of Condition, F (4, 92) = 22.83, p < .001. This effect stems from the fact that subjects were more likely to respond incorrectly to items in Conditions 1 and 2, presumably because these items were less familiar. Finally, the Prime Type by Condition interaction was not significant, F(4,92) = 1.8, p = .19. The mean error rates for targets following test and control primes in individual conditions, as well as the rates for the conditions overall, are shown in Table 4.26 below. The results of the error analysis underscore the notion that the lack of significant facilitation for the highly semantically related prime-target pairs for the Psychology subject group is due to the subjects’ lack of familiarity with the items in Conditions 1 and 2. There is no priming for the semantically related word pairs for the Psychology students because many of these students may either Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 118 not know or have less rich meanings for many of the primes and targets. Without semantic knowledge, there can be no semantically based priming. Table 4.26 Experiment 4: M ean E rror Rates by Condition fo r the Psychology Subject Group. Condition Prime-Target Example Mean Error Rate Control Test Overall 1. hi sem related trivial-trifle .13 .09 .11 2. mid sem related ruffian-rough .11 .12 .11 3. lo sem related rankle-rank .04 .01 .03 4. hi sem, no phon sorcery-magic .02 .003 .01 5. no sem, hi phon pumpkin-pump .04 .06 .05 Honors college subject group The results for the group of Honors College subjects were similar in overall pattern to the results for the Psychology group. First, the main effect of Prime Type was not significant, F (1, 28) < 1, indicating no difference in error rate in responses to targets based on the type of prime. There was a significant main effect o f Condition, F(4, 112) = 10.62, p < .001, due to greater proportions of errors in Conditions 1 and 2. Finally, the Prime Type by Condition interaction was significant, F(4, 112) = 3.13, p < .02, reflecting the greater difficulty subjects had correctly responding to targets in Condition S (e.g., pump) when they followed test primes (e.g., pumpkin) compared to unrelated control primes (e.g., splendor). The mean error rates for targets following test and control primes Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 119 in individual conditions, as well as for each condition overall, are shown in Table 4.27 below. While the general pattern of errors was similar for both subject groups, the Honors college group made fewer errors. The differences in error rates for both subject groups is explored in greater detail below. Table 4.27 Experiment 4: M ean E rror Rates by Condition fo r the Honors College Subject Group. Condition Prime-Target Example Mean Error Rate Control Test Overall 1. hi sem related trivial-trifle .07 .06 .07 2. mid sem related ruffian-rough .08 .06 .07 3. lo sem related rankle-rank .05 .04 .05 4. hi sem, no phon sorcery-magic .01 .002 .005 5. no sem, hi phon pumpkin-pump .02 .06 .04 Psychology versus Honors College subject groups The reaction times of the Psychology subject group and the Honors College group to both real and nonword targets were compared. Mean reaction times were entered into an ANOVA with the factors of Group (Psychology vs. Honors College) and Target Type (real vs. nonword). There was no effect of Group, F (l, 51) < 1, indicating no difference between the reaction times of subjects from the two groups to either real or nonword targets. There was a significant main effect of Target Type, F(l, 51) = 109.0, p < .001, indicating that responses were significantly slower to nonword targets compared to real words. This effect was found in the first three experiments as well. There was no Group Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 120 by Target Type interaction, F (l, 51) < 1. Reaction times for both Psychology and Honors College subjects on real word and nonword targets are shown in Table 4.28 below. Table 4.28 Experiment 4: M ean Reaction Times fo r Lexical Decision to Real Word and Nonword Targets by Subject Group. Mean Reaction Time (msec) Word Type Psyc Honors Real Word Targets 620 619 Nonword Targets 773 747 The error rates of the Psychology subject group and the Honors College group to both real and nonword targets were also examined. An ANOVA with the factors of Group (Psychology vs. Honors College) and Target Type (real vs. nonword), using error rate as the dependent variable, was calculated. There was a significant main effect of Group, F(l, 51) = 13.74, p < .001, reflecting greater error rates for the Psychology subject group compared to the Honors College group. The effect of Target Type was not significant, F(l, 51) < 1, indicating relatively equal error rates to both real and nonword targets. There was no Group by Target Type interaction, F(l, 51) = 1.63, p = .21. Mean error rates for both Psychology and Honors College subjects on real word and nonword targets are shown in Table 4.29 below. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 121 Table 4.29 Experiment 4: M ean Error Rates fo r Lexical Decision to R eal Word and Nonword Targets by Subject Group. Mean Error Rate Word Type Psyc Honors Real Word Targets .06 .04 Nonword Targets .09 .04 Further analyses of the different types of nonword stimuli were also carried out. Mean reaction times and error rates by subject group for the four conditions are shown in Table 4.30 below. An ANOVA was calculated using reaction time as the dependent variable, with Condition (1-4) and Group (Psychology vs. Honors College) as factors. The effect of group was not significant, F (l, 51) < 1, reflecting overall similarity in the reaction times of both subject groups. There was a significant main effect of Condition (1-5), F(3, 153) = 26.33, p < .001, indicating slower mean reaction times to the more phonologically related targets in Conditions 2 (dipper-dippance) and 3 (dervish-derv). This result is very similar to the results for nonword targets in Experiments 1, 2 and 3. Finally, there was a significant Group by Condition interaction, F(3, 153) = 4.0, p < .01. The interaction arose because for some conditions there was no difference in the reaction times of the Psychology and Honors College subjects, and in other conditions (2 and 4), the Honors College subjects responded faster than the Psychology subjects. The mean error rates for the nonword targets were entered into an ANOVA with Condition (1-4) and Group (Psychology vs. Honors College) as factors. The main effect of Group was not significant, F (l, 51) = 2.84, p = .10, indicating that the two subject groups performed similarly on the nonword targets. Although the Group effect was only Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 122 marginal, the effect is still worth examining. The fact that the Psychology students made more errors in most of the conditions is probably due to a lack of confidence in their own vocabularies. The subjects are aware that many of the words are unfamiliar to them, and therefore assume that they simply are real words that they have not learned yet. On the other hand, the Honors college students are more confident in their vocabularies, and assume that a word they do not recognize is a nonword. There was a significant main effect o f Condition (1-4), F (l, 51) = 14.23, p < .001. Results from this analysis indicate that subjects were more likely to make errors on certain types of nonwords compared to others. As in the reaction time analysis, the most difficult nonword targets, marked by higher error rates, were in the two conditions where the nonword targets either look like the suffixless stem of their real word primes (e.g., bishop-bish o r dervish-derv), or the nonwords are composed o f real stems and real suffixes whose combinations are unattested (e.g., dippance). Finally, there was a significant Group by Condition interaction, F(3, 153) = 4.68, p < .005, because both groups had the most difficulty with the Condition 2 (dipper-dippance) nonwords, but the Honors College subjects made half as many errors or less than the Psychology subjects in the other three conditions. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 123 Table 4.30 Experiment 4: Mean Reaction Times and Error Rates fo r Lexical Decision to Nonword Targets by Condition fo r the Psychology Subject Group and the H onors College Group. Condition Prime-Target Example Mean Reaction Time fmseci Psyc Honors Mean Error Rat? Psyc Honors 1. phon change in target brackish-brap 745 729 .06 .02 2. unattested stem + suffix combination dipper-dippance 836 795 .10 .12 3. phon transparent bishop-bish dervish-derv 778 797 .11 .05 4. unrelated quiver-proot 111 733 .09 .03 4.6 Summary and General Discussion In the four behavioral experiments described in this chapter, I find evidence consistent with results from the modeling described in Chapter 3 that suggest semantic and phonological relatedness drive morphological priming, and not morphological structure. In the following section I summarize the major findings from the four studies and discuss the importance of these results for a theoretical account of morphological processing. The findings from Experiments I and 2 suggest that subjects are sensitive to gradations in both the semantic and phonological similarity of related words. Words that are closely related in meaning (e.g., baker-bake) produce larger facilitation effects than words that are more distant in meaning (e.g., dresser-dress), while semantically unrelated words do not prime at all (e.g., com er-com ). A similar pattern of results was found for Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 124 word pairs that varied in their phonological relatedness. Highly related pairs (e.g., deletion-delete) prime more than moderately related pairs (e.g., vanity-vain) which in turn prime more than even less related pairs (e.g., introduction-introduce). These findings are difficult to accommodate within theories of the mental lexicon where words are formed by applying rules to combine discreet suffix and stem units. Rather, the graded effects found here are easily explained by a lexical system that builds up representations through mappings between semantic and phonological patterns, and regularities in the mappings give rise to similar representations. Items that are more similar in meaning as well as in sound will tend to be closer in lexical space and more strongly connected than unrelated words. Results from Experiment 3 show that the morphological relationship between primes and targets does not drive priming effects. Whereas other researchers (Marslen- Wilson et al., 1994) found no priming for pairs of derived suffixed words (e.g., observation-observant) and argued based on these findings that the mental lexicon is organized according to stems and affixes, I found that suffixed pairs can prime one another. Crucially, this priming only occurs when the primes and targets are highly related both semantically and phonologically (e.g., saintly-sainthood). Furthermore, I obtained the same results as Marslen-Wilson et al. (1994) for less semantically and phonologically related pairs (e.g., observation-observant), which produced intermediate effects (14 msec) that failed to reach significance. The results of this experiment call into question the need to posit separate processing and storage mechanisms for suffixes and suffixed words. Finally, results from Experiment 4 showed that words whose only relationship was semantic and phonological, but not morphological (e.g., trivial-trifle), produce significant priming effects. These results again support the view that semantic and Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 125 phonological similarity are important dimensions in the mental lexicon, yielding results that mimic morphological structure. Even when there is a lack of “morphology”, morphological effects emerge. The effects for nonword targets were consistent across all four experiments, and reflect classic findings in the psycholinguistic literature (cf. Rubenstein et al., 1970), that subjects are generally slower to respond to nonword than real word targets. Furthermore, subjects had the most difficulty responding to nonwords that more closely resembled real words (cf. Schaeffer & Wallace, 1970). This difficulty was reflected by both longer reaction times and greater error rates for pairs such as respectful-respection compared to pairs such as boomerang-jaulic. The results for nonwords underscore the role of similarity in processing, since all nonwords did not produce the same effects, instead there were differences based on their similarity to existing words. While the discussion here has centered on the important role of semantic and phonological similarity, factors other than these will also affect the lexicon. For example, both type and token frequency play strong roles in lexical representation and processing. Thus, patterns that are presented repeatedly will have stronger representations (Gonnerman et al., 1997), and patterns that are repeated across many different lexical items will be more robust (Bybee, 1995). Even though I controlled for frequency rather than examining it as a separate factor in these experiments, some of the results from Experiment 4 speak to the issue of frequency. To create stimulus pairs that are related in meaning and sound but have different historical origins in the language, I was forced to use some low frequency words. The fact that the Honors College participants showed priming effects for these low frequency items but the Psychology students did not, can be explained by appealing to the role of frequency in lexical representation. The Honors College students are accepted into their program based in part on their scores on the Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 126 vocabulary portion of the Scholastic Aptitude Test (SAT). Students who score higher on this part o f the SAT typically have read more and more diversely than lower scoring students. I would argue that the difference in levels of exposure to words between the two subject groups in Experiment 4 is the primary reason for differences in the priming effects. Exposure to a greater number o f different words and more exposure to each of these words builds up their sound-meaning mappings such that Honors College students ultimately have somewhat different lexical representations than the Psychology students. The strong lexical relationships between related items produce priming in subjects who are familiar with the stimuli, but fail to show effects in subjects who are not. I would predict that Classics scholars who are familiar with the Latin and Greek roots of many English words would also show different results for lexical items whose relationships may be obscure to the general public. Their semantic representations include information that is not present in the lexicons of other non-Classics scholars. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 5 General Discussion In this final chapter, I briefly review the major findings from both the computational model and the behavioral experiments. This summary is followed by a discussion of the implications of these results for psycholinguistic theories. I end the chapter by proposing future directions for this research. 5.1 Summary o f Results The model described in Chapter 3 was developed to explore whether a system based on semantic and phonological relatedness between lexical items could reproduce results from a behavioral experiment that were claimed to support a decompositional model of the mental lexicon (Marslen-Wilson et al., 1994). Results from the priming experiment with the computational model were strikingly similar to the results from humans, suggesting that a system which relies on semantics and phonology alone can simulate what appeared to be effects o f morphological structure. The results from the four behavioral experiments presented in Chapter 4 demonstrate that: 1) subjects are sensitive to fine gradations both in the semantic similarities between related lexical items (e.g., teacher-teach; dresser-dress; comer- com ), and in their phonological similarities (e.g., deletion-delete; vanity-vain; introduction-introduce); 2) the degrees of semantic and phonological relatedness of word pairs, when considered in conjunction, predict priming effects, and; 3) the nature of the morphological relationship between primes and targets does not predict priming effects; thus, derived-stem pairs (e.g., teacher-teach), derived-derived pairs (e.g., saintly- Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 128 sainthood), and pairs with no historical morphological relationship (e.g., trivial-trifle) all prime, provided the word pairs are highly semantically and phonologically related. It is important to emphasize the graded nature of the priming effects found for both semantic and phonological relatedness. These effects are crucial, as they provide evidence that can differentiate between the connectionist approach to morphology that I am taking and more traditional, decomposition, or dual-route theories. How would a theory that proposes storing stems and affixes separately determine which complex words to decompose? For example, in Marslen-Wilson et al.’s approach, semantically transparent forms are decomposed, but opaque forms are not. This theory is easy to apply to forms that are at either end of the semantic similarity continuum, but it is unclear how to handle intermediate forms, such as dresser-dress, within such an approach. Are intermediate forms decomposed or not? If they are not, then why do they prime one another? And if they are decomposed, then why do they prime less than more related forms? Thus, the intermediate effects that I have reported here both capture the generalizations one can make about morphology better than more traditional views, and present a challenge to these views. Another important aspect of the results reported here is that semantic relatedness seems to play a more primary role than phonological relatedness. In Experiment 2, all of the words were highly semantically related. This was done because the results from the model and the reanalysis of the Marslen-Wilson et al. behavioral data in Chapter 3 showed that phonological relatedness only plays a role in priming when the prime-target pairs are highly semantically related. These effects are not simply additive, they interact. One explanation involves the nature of the space for each system. Language is first and foremost about conveying meaning. Moving slightly in phonological space, for example from bank to bang, causes a big change in meaning. However, moving a bit in semantic Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 129 space, for example from a cat with a tail to a cat without a tail, does not have as large an effect on meaning. An additional property of the semantic system that makes it more primary than phonology is the fact that it is a much richer system. Semantics includes information from various input modalities, including visual, tactile, auditory, and kinesthetic domains, while phonological information is based on only one sensory modality, namely auditory. These differences between the semantic and phonological systems contribute to the primacy of semantics in processing complex words in English. 5.2 Implications o f Results In this thesis I have taken a somewhat novel approach to examining the factors that contribute to the processing of morphologically complex words. While most psycholinguistic research in this area has looked at factors such as phonological, semantic, and orthographic similarity independently, in the experiments reported here I looked at the simultaneous effects of semantic and phonological overlap. The two factors of semantic and phonological similarity can account for a wide range-of behavioral data, but only when they are considered in conjunction. Examining complex interactions of factors constitutes a different approach to psycholinguistic research than traditional experimental techniques, such as that of Marslen-Wilson et al. (1994), who looked at individual factors in turn, but not together. A second major finding from the studies reported here that has implications for psycholinguistic research is the importance of the graded nature of phonological and semantic similarity and their effects on priming magnitude. Rather than dichotomizing variables— for example, dividing stimuli into transparent and opaque derived words, or derivational versus inflectional forms, or productive versus nonproductive affixes— the Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 130 use of items that span the full range of relatedness allows one to see important intermediate effects; considering only the ends of a continuum can provide misleading information. Several other factors also presumably play a role in morphological processing. For example, while I controlled for token frequency of test and control primes, I did not control for type frequency. It is certainly relevant for processing to know how many ‘friends’ and ‘enemies’ a given derived word has, for example, teacher is a ‘friend’ to runner because both forms have consistent meanings for the ending -er, but com er is an enemy, because the ending -er does not mean ‘agent’ in this case, as it does for the other two forms. In addition, the relative frequency of particular form-meaning pairings (e.g., -ee in employee) compared to cases where the form is not associated with the same meaning (e.g., -ee in frisbee) also surely effect the representation and processing of words that end in that particular phonological sequence. Another factor that would be important to investigate is the productivity of particular affixes. For example, while in- is an unproductive affix for contemporary English speakers, who typically form novel words using un-, there are still many attested forms with in- as a prefix (e.g., inconceivable). For most forms, productivity of the affix will tend to correlate with neighborhood size, but for some forms these two factors will diverge. It would be interesting to explore the effects of productivity and neighborhood size both separately and in conjunction, just as I have looked at semantics and phonology separately and in conjunction. The factors mentioned here, phonological, semantic, and orthographic overlap, frequency, neighborhood size, consistency, and productivity are some of the dimensions that define the highly complex space of the mental lexicon. While most of these factors are relatively easy to specify, for example orthographic overlap could be reasonably Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 131 quantified simply as the number of letters in common between two forms, semantic relatedness is very difficult to measure precisely. I have operationalized semantic similarity by using relatedness ratings from subjects, but it is not entirely clear what these relatedness ratings are actually measuring. Presumably the ratings reflect some real distance between semantically related items in a complex, highly dimensional semantic space. The exact dimensions which compose that space are not easy to delineate. The question of what constitutes semantics opens the door for more research exploring exactly what the nature of semantic similarity is. If the mental lexicon is neither decompositional, nor made up o f stored whole word forms, nor some combination of the two, as I suggested in Chapter 2, what then does the mental lexicon look like? The view of the mental lexicon I am advocating is one which takes its character from the brain, rather than the traditional metaphor of a serial computer, with stored forms and mechanisms for accessing those forms. In the view I am adopting, word forms are distributed across many simple, neurcn-iike processing units. Comprehending a word involves computing its semantic representation from a phonological pattern, and producing a word involves computing the phonological pattern based on an activated semantic pattern. Words are stored, not as independent, separate patterns, but in the weights on connections between these simple processing units. All the words that a speaker knows are superimposed on the same set of weighted connections. The differences between my account and more traditional accounts which reify morphology are subtle and concern the representational structure of morphological information. Both accounts accept that “morphological” effects exist, but they differ in terms of their explanation for the genesis of these effects. In the traditional account morphological effects arise as the system decomposes complex words into stems and Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 132 affixes in either storage or access, or both. In the account proposed here, morphological effects arise as the consequence of the interactions in a dynamic system that maps meanings onto forms and vice versa. Componential basins of attractions will behave like stems and morphemes. Although it may seem like this account has simply transferred morphemes into these basins of attraction, there are implications that differentiate my approach from those proposing morphological decomposition, specifically in the areas of acquisition, normal processing, and impairments. 5.3 Future Directions While the research described in this thesis is based primarily on processing in young normal adults, the results make clear predictions for research in other domains, such as crosslinguistic research, acquisition, and impairment. Implications of this research in other domains are described in the following sections. 5.3.1 Cross-linguistic extensions Hebrew provides a challenge to a connectionist system of morphology. Because the language is nonconcatenative, using consonantal roots with vocalic patterns inserted between the consonants, the system is very different than the English system. Morphological processes seem to play a much more central role in this language. This being the case, how then can a system that relies only on phonology and semantics and the relationships between them capture the phenomena of Hebrew? It would seem that surely morphology must be explicitly represented here. It is important to extend this work crosslinguistically to demonstrate that morphological phenomena can be captured by correlations between semantics and Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 133 phonology for languages such as Hebrew, which have less conventional morphological systems than English. These kinds of languages have traditionally been very difficult to integrate into theories of morphology and many theorists have developed separate constructs to deal with them (see, for example, the Morphological Rule Constraint of McCarthy (1981)). The status of the root in Hebrew is also debated among linguists, with some researchers opting for roots (e.g., Ephrat, 1997) and others arguing against the psychological reality of roots (e.g., Bat-El, 1994). Previous research in Hebrew has shown priming effects for morphologically related words (e.g., Bentin & Feldman, 1990; Berent & Shimron, 1997; Feldman & Bentin, 1994; Frost et al., 1997). If the account developed here for English is correct, it should also extend to Hebrew, predicting that priming effects in Hebrew arise from the phonological and/or semantic similarity of the primes and targets in morphologically related Hebrew words. It will be important to systematically investigate phonological and semantic overlap in Hebrew and to look at how tlicsc factors mientcr with one another in this typologically different language. 5.3.2 Developm ental extensions There is already a substantial body of work on acquisition of English verb morphology in connectionist networks (Plunkett & Marchman, 1993; Rumelhart and McClelland, 1986). This work has concentrated on inflectional processes. While researchers such as Clark and Berman (1984) have described the developmental trajectory for derivational morphology in English and in Hebrew, there have been no attempts to model the phenomena. Modeling acquisition data from Hebrew as well as English will allow an examination of exactly how the principles of semantic and Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 134 phonological transparency, frequency, and productivity interact to produce the type of errors children show. Simulating acquisition data would also allow one to examine the effects of differences in formal similarity and semantic relatedness on the ease with which complex words are learned. This can then be compared to differences in adult processing of complex forms revealed by the priming experiments. Results from such a study would show whether properties which affect adult processing of derived words also affect children’s acquisition of the same forms. If it could be shown that the same principles underlie both acquisition and skilled processing in two typologically different languages, English and Hebrew, this would lend support to this alternative approach to morphology based on connectionist principles that unites phenomena traditionally considered disparate. 53.3 Computational extensions The model described in this thesis was based closely on Experiment 3 of Marslen- Wilson et al. (1994). It used the 208 prime and target words from that experiment as the entire training and testing corpus. Words were based on random patterns designed to capture the phonological and semantic relatedness between prime-target pairs. The model was required to learn a mapping from sound to meaning. While this model was highly succesful in simulating the Marslen-Wilson et al. priming results, it could be expanded in several ways. For example, in this model, the semantic representations were based on random vector patterns that capture overlap between pairs of words, such that govern and government overlapped quite a bit, but depart and department did not. These representations incorporated sufficient regularities to capture the priming results Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 135 demonstrated, but they would not be adequate for comparisons among neighborhoods of words, for example a large set of words that include ‘agent’ as part of their meaning. The phonological representations used in the model suffer from similar drawbacks as the semantic representations, namely they were random bit patterns that only captured similarities between pairs of words, and sometimes triples in the case of sets like govern, governor, and government. In order for the model to capture adequately a broader range of behavioral phenomena, the phonological representations will need to encode relationships among the entire training set, rather than just pairwise correlations. The model could be further improved by modifying the architecture to allow simulation of both production and comprehension tasks. The model described here was only able to model comprehension, as it was trained solely on mappings from sound to meaning. To examine production data, a model should also be trained on semantic to phonological mappings. To do this, new sets of connections would need to be added from the semantic layer to the hidden units and from the hidden units to the phonological layer. An expanded model, capable o f more realistically representing a larger set of words would significantly advance our understanding o f the roles of various factors in morphological representation and processing. 5.3.4 Neuropsychological extensions Finally, this work makes clear predictions about the consequences of brain damage on morphological processing. There is some evidence that aphasic patients can present with different patterns of impairment for opaque (e.g., department) and transparent (e.g., departure) forms (Tyler et al., 1990). If morphology reflects the conjunction of semantics and phonology, then a ‘morphological’ impairment should always be accompanied by impairments to either the semantic or phonological systems, Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 136 or both. In future work I will explore systematically the role of phonological and semantic relatedness o f derived words and their steins in on-line studies with impaired subjects, investigating the relationship between morphological processing impairments and semantic or phonological deficits. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 137 REFERENCES Aiken, L., & West, S. (1991). M ultiple Regression: Testing and Interpreting Interactions. Newbury Park, CA: Sage Publications. Andersen, E. S. (1992). Complexity and language acquisition: Influences on the development of morphological systems in children. In J. A. Hawkins & M. Gell- Mann (Eds.), The Evolution o f Human Languages, (pp. 241-271). Redwood City, CA: Addison-Wesley. Anderson, S. (1982). Where’ s morphology? Linguistic Inquiry, 75,571-612. Anderson, S. (1992). A-morphous Morphology. Cambridge: Cambridge University Press. Andrews, S. (1986). Morphological influences on lexical access: Lexical or nonlexical effects. Journal o f M emory and Language, 2 5 ,726-740. Aronoff, M. (1976). Word Formation in Generative Grammar. Cambridge, MA: MIT Press. Aronoff, M. (1994). M orphology by Itself: Stem s and Inflectional Classes. Cambridge, MA: MIT Press. Baayen, H. (1994). Productivity in language production. Language and Cognitive Processes, 9 ,447-469. Baayen, H., & Lieber, R. (1991). Productivity and English derivation: A corpus-based study. Linguistics, 29, 801-843. Badecker, W., & Caramazza, A. (1991). Morphological composition in the lexical output system. Cognitive Neuropsychology, 8, 335-367. Balota, D., & Chumbley, J. (1984). Are lexical decisions a good measure of lexical access? The role of word frequency in a neglected decision stage. Journal o f Experimental Psychology, 10, 340-357. Bat-El, O. (1994). Stem modification and cluster transfer in modem Hebrew. Natural Language and Linguistic Theory, 1 2 ,571-596. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 138 Bates, E. & MacWhinney, B. (1982). Functionalist approaches to grammar. In E. Wanner, & L. Gleitman (Eds.), Language Acquisition: The State o f the A rt (pp. 173-218). New York: Cambridge University Press. Bates, E. & MacWhinney, B. (1987). Competition, variation and language learning. In B. MacWhinney (Ed.), Mechanisms o f Language Learning (pp. 157-193). Hillsdale, NJ: Erlbaum. Bates, E. & MacWhinney, B. (1989). Functionalism and the competition model. InB. MacWhinney, & E. Bates (Eds.), The Crosslinguistic Study o f Sentence Processing (pp. 3-73). New York: Cambridge University Press. Bauer, L. (1983). English Word-Formation. Cambridge: Cambridge University Press. Bavelier, D., & Jordan, M. I. (1993, November). Representing words in connectionist networks. Paper presented at the Psychonomics Society, Washington, DC. Beard, R. (1995). Lexeme-morpheme Base Morphology: A General Theory o f Inflection and Word Formation. Albany: State University of New York Press. Bentin, S., & Feldman, L. B. (1990). The contribution of morphological and semantic relatedness to repetition priming at short and long lags: Evidence from Hebrew. Quarterly Journal o f Experimental Psychology, 42A, 693-711. Berent, I., & Shimron, J. (1997). The representation of Hebrew words: Evidence from the obligatory contour principle. Cognition, 64, 39-72. Berko, J. (1958). The child’s learning of English morphology. Word, 1 4 ,150-177. Bloomfield, (1933). Language. New York: Holt, Rinehart, & Winston. Bradley, D. (1980). Lexical representation of derivational relation. In M. Aronoff & M- L. Kean (Eds.), Juncture (pp. 37-55). Saratoga, CA: Amma Libri. Burani, C., & Caramazza, A. (1987). Representation and processing o f derived words. Language and Cognitive Processes, 2 ,217-227. Burani, C., Salmaso, D., & Caramazza, A. (1984). Morphological structure and lexical access. Visible Language, 75,419-427. Butterworth, B. (1983). Lexical representation. In B. Butterworth (Ed.), Language Production (pp. 257-294). San Diego, CA: Academic Press. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 139 By bee, J. (1985). M orphology: A Study o f the Relation Between M eaning and Form. Philadelphia: Benjamins. Bybee, J. (1988). Morphology as lexical organization. In M. Hammond & M. Noonan (Eds.), Theoretical M orphology, (pp. 119-141). San Diego, CA: Academic Press. Bybee, J. (1995). Regular morphology and the lexicon. Language and Cognitive Processes, 1 0 ,425-455. Bybee, J., & Moder, C. (1983). Morphological classes as natural categories. Language, 59, 251-271. Bybee, J., & Newman, J. (1995). Are stem changes as natural as affixes? Linguistics, 33, 633-654. Caramazza, A., Laudanna, A., & Romani, C. (1988). Lexical access and inflectional morphology. Cognition, 2 8 ,297-332. Caramazza, A., Miceli, G., Silveri, M.-C., & Laudanna, A. (1985). Reading mechanisms and the organisation of the lexicon: Evidence from acquired dyslexia. Cognitive Neuropsychology, 2, 81-114. Chialant, D., & Caramazza, A. (1995). Where is morphology and how is it processed? The case of written word recognition. In L.B. Feldman (Ed.), M orphological Aspects o f Language Processing (pp. 55-76). Hillsdale, NJ: Erlbaum. Chomsky, N. (1970). Remarks on nominalization. In R. A. Jacobs & P. S. Rosenbaum (Eds.), Readings in English Transformational Grammar, (pp. 184-221). Waltham, MA: Ginn. Chomsky, N., & Halle, M. (1968). The sound pattern o f English. New York: Harper & Row. Christiansen, M. (1994). Infinite languages, finite minds: Connectionism, learning, and linguistic structure. Ph.D. dissertation, University of Edinburgh. Christiansen, M., Allen, J., & Seidenberg, M. S. (1998). Learning to segment speech using multiple cues: A connectionist model. Language and Cognitive Processes, 13, 221-268. Clark, E., & Berman, R. (1984). Structure and use in the acquisition of word formation. Language, 60, 542-590. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 140 Clark, E., & Berman, R. (1987). Types o f linguistic knowledge: interpreting and producing compound nouns. Journal o f Child Language, 14,547-568. Cole, P., Beauvillain, C. & Segui, J. (1989). On the representation and processing of prefixed and suffixed words: A differential frequency effect. Journal o f Memory and Language, 28, 1-13. Coltheart, M., Davelaar, E. Jonasson, J., & Besner, D. (1977). Access to the internal lexicon. In S. Domic (Ed.), Attention and Performance VI (pp. 535-555). Hillsdale, NJ: Erlbaum. Daugherty, K., & Seidenberg, M. S. (1992). Rules or connections? The past tense revisited. In Proceedings o f the 14th Annual Conference o f the Cognitive Science Society (pp. 259-264). Hillsdale, NJ: Erlbaum. Devlin, J. T. (1998). The role o f dynamic representations in understanding semantic impairments: Investigations from connectionist neuropsychology. Ph.D. dissertation, University of Southern California. Devlin, J. T., Gonnerman, L. M., Andersen, E. S., & Seidenberg, M. S. (1998). Category-specific semantic deficits in focal and widespread brain damage: A computational account. Journal o f Cognitive Neuroscience, 1 0 ,77-94. Di Sciullo, A-M., & Williams, E. (1987). On the Definition o f Word. Cambridge, MA: MIT Press. Ephrat, M. (1997). The psycholinguistic status of the root in modem Hebrew. Folia Linguistica, 3 1 ,77-103. Feldman, L., & Bentin, S. (1994). Morphological analysis o f disrupted morphemes: Evidence from Hebrew. Quarterly Journal o f Experimental Psychology, 4 7 ,407- 435. Fowler, C. A., Napps, S. E., & Feldman, L. (1985). Relations among regular and irregular morphologically related words in the lexicon as revealed by repetition priming. Memory and Cognition, 13, 241-255. Frost, R., Forster, K., & Deutsch, A. (1997). What can we learn from the morphology of Hebrew? A masked-priming investigation of morphological representation. Journal o f Experimental Psychology: Learning, Memory, and Cognition, 2 3 ,829- 856. Gamham, A. (1985). Psycholinguistics: Central Topics. London: Methuen. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 141 Gonnerman, L. M., Harm, M. W., & Andersen, E. S. (1997). The role o f frequency in modeling double dissociations. In Proceedings o f the 19h Annual Conference o f the Cognitive Science Society (p. 934). Mahwah, NJ: Erlbaum. Grainger, J., Cole, P., & Segui, J. (1991). Masked morphological priming in visual word recognition. Journal o f Memory and Language, 30, 370-384. Greenberg, J. (1966). Language Universals. The Hague: Mouton. Halle, M., & Marantz, A. (1993). Distributed morphology and the pieces of inflection. In K. Hale & S. J. Keyser (Eds.), The View from Building 20: Essays in Linguistics in Honor o f Sylvain Bromberger (pp. 111-176). Cambridge, MA: MIT Press. Hare, M., & Elman, J. (1995). Learning and morphological change. Cognition, 5 6 ,61- 98. Harm, M. W., & Seidenberg, M. S. (in press). Phonology, reading acquisition, and dyslexia: Insights fromconnectionist models. Psychological Review. Hawkins, J. A. (1994). A Performance Theory o f Order and Constituency. Cambridge: Cambridge University Press. Hinton, G. E., McClelland, J. L., & Rumelhart, D. E. (1986). Distributed representations. In D. E. Rumelhart, J. L. McClelland, & the PDP Research Group (Eds.), Parallel Distributed Processing: Explorations in the M icrostructure o f Cognition. Volume 1: Foundations (pp. 77-109). Cambridge, MA: MIT Press. Jarvella, R. J., & Snodgrass, J. G. (1974). Seeing ring in rang and retain in retention: On recognizing stem morphemes in printed words. Journal o f Verbal Learning and Verbal Behavior, 1 3 ,590-598. Jensen, J. T., & Stong-Jensen, M. (1984). Morphology is in the lexicon! Linguistic Inquiry, 1 5 ,474-498. Kempley, S. T., & Morton, J. (1982). The effects of priming with regularly and irregularly related words in auditory word recognition. British Journal o f Psychology, 73,441-454. Langacker, R. (1987). Foundations o f cognitive grammar. Vol. 1. Theoretical prerequisites. Stanford, CA: Stanford University Press. Langacker, R. (1988). A usage-based model. In B. Rudzka-Ostyn (Ed.), Topics in Cognitive Linguistics (pp. 127-161). Amsterdam: John Benjamins Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 142 Laudanna, A., & Burani, C. (1985). Address mechanisms to decomposed lexical entries. Linguistics, 23,115-192. Laudanna, A., Badecker, W., & Caramazza, A. (1989). Priming homographic stems. Journal o f M emory and Language, 28, 531-546. Laudanna, A., Badecker, W., & Caramazza, A. (1992). Processing inflectional and derivational morphology. Journal o f Memory and Language, 31, 333-348. Laudanna, A., Cermele, A., & Caramazza, A. (1997). Morpho-lexical representations in naming. Language and Cognitive Processes, 1 2 ,49-66. Li, P., & MacWhinney, B. (1996). Cryptotype, overgeneralization and competition: A connectionist model o f the learning of English reversive prefixes. Connection Science: Journal o f Neural Computing, Artificial Intelligence & Cognitive Research, 8, 3- 30. Libben, G. (1993). Are morphological structures computed during word recognition? Journal ofPsycholinguistic Research, 22, 535-544. Lieber, R. (1992). Deconstructing Morphology: Word Formation in Syntactic Theory. Chicago: University of Chicago Press. Lukatela, G., Gligorijevic, B., Kostic, A., & Turvey, M.T. (1980). Representation of inflected nouns in the internal lexicon. Memory and Cognition, 8 ,415-423. Maeda, A. (1981). A multidimensional map of the phonemes of English: A perceptual study. Ph.D. dissertation. University of Hawaii. MacKay, D. (1978). On the retrieval and lexical structure of verbs. Journal o f Verbal Learning and Verbal Behavior, 15, 169-182. MacWhinney, B., & Leinbach, A. J. (1991). Implementations are not conceptualizations: Revising the verb learning model. Cognition, 40, 121-157. Manelis, L., & Tharp, D. A. (1977). The processing of affixed words. Memory and Cognition, 5 ,690-695. Marslen-Wilson, W., Tyler, L., Waksler, R. & Older, L. (1994). Morphology and meaning in the English mental lexicon. Psychological Review, 101, 3-33. Masson, M. (1995). A distributed memory model of semantic priming. Journal o f Experimental Psychology: Learning, Memory, and Cognition, 21, 3-23. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 143 McCarthy, J. (1981). A prosodic theory of nonconcatenative morphology. Linguistic Inquiry, 12, 373-418. McClelland, J. L., & Rumelhart, D. E. (1981). An interactive activation model o f context effects in letter perception: Part 1. An account of basic findings. Psychological Review, 88, 375-407. McRae, K., de Sa, V., & Seidenberg, M. (1997). On the nature and scope of featural representations o f word meaning. Journal o f Experimental Psychology: General, 1 2 6 ,99-130. Miceli, G., & Caramazza, A. (1988). Dissociation o f inflectional and derivational morphology. Brain and Language, 3 5 ,24-65. Morton, J. (1979). Facilitation in word recognition: Experiments causing change in the logogen model. In P. A. Kolers, M. Wrolstad, & H. Bouma (Eds.), Processing o f visual language (pp. 259-268). New York: Plenum Press. Murrell, G. A., & Morton, J. (1974). Word recognition and morphemic structure. Journal o f Experimental Psychology, 102,963-968. Napps, S. E. (1985). Morphological, Semantic, and Formal Relationships in the Organization o f the "Mental Lexicon". Ph.D. dissertation, Dartmouth College. Napps, S. E. (1989). Morphemic relationships in the lexicon: Are they distinct from semantic and formal relationships? Memory and Cognition, 17, 729-739. Napps, S. E., & Fowler, C. A. (1987). Formal relationships among words and the organization o f the mental lexicon. Journal o f Psycholinguistic Research, 16, 257- 272. Pinker, S. (1991). Rules of language. Science, 253, 530-535. Pinker, S., & Prince, A. (1988). On language and connectionism: Analysis of a parallel distributed processing model o f language acquisition. Cognition, 2 8 ,73-193. Plaut, D. (1995). Semantic and associative priming in a distributed attractor network. In Proceedings o f the 17th Annual Conference o f the Cognitive Science Society (pp. 37-42). Hillsdale, NJ: Erlbaum. Plaut, D. (1997). Structure and function in the lexical system: Insights from distributed models of word reading and lexical decision. Language and Cognitive Processes, 12, 765-805. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 144 Plaut, D., & McClelland, J. (1993). Generalization with componential attractors: Word and nonword reading in an attractor network, hi Proceedings o f the 15th Annual Conference o f the Cognitive Science Society (pp. 824-829). Hillsdale, NJ: Erlbaum. Plaut, D., McClelland, J. L., Seidenberg, M. S., & Patterson, K. (1996). Understanding normal and impaired word reading: Computational principles in a quasi-regular domain. Psychological Review, 703,56-115. Plunkett, K., & Marchman, V. (1991). U-shaped learning and frequency effects in a multi-layered perceptron: Implications for child language acquisition. Cognition, 3 8 ,43-102. Plunkett, K., & Marchman, V. (1993). From rote learning to system building: Acquiring verb morphology in children and connectionist nets. Cognition, 48, 1-49. Prasada, S., & Pinker, S. (1993). Generalizations o f regular and irregular morphological patterns. Language and Cognitive Processes, 8, 1-56. Rubenstein, H., Garfield, L., & Millikan, J. A. (1970). Homographic entries in the internal lexicon. Journal o f Verbal Learning and Verbal Behavior, 9 ,487-494. Rubin, G. S., Becker, C. A., & Freeman, R. H. (1979). Morphological structure and its effect on visual word recognition. Journal o f Verbal Learning and Verbal Behavior, 18, 757-767. Rueckl, J., Mikolinski, M., Raveh, M., Miner, C., & Mars, F. (1997). Morphological priming, fragment completion, and connectionist networks. Journal o f M emory and Language, 36, 382-405. Rumelhart, D, & McClelland, J. (1986). On learning the past tenses of English verbs. In J. McClelland, D. Rumelhart, & the PDP Research Group (Eds.), Parallel Distributed Processing: Explorations in the M icrostructure o f Cognition, Vol. 2: Psychological and Biological M odels (pp. 216-271). Cambridge, MA: MIT Press. Sandra, D. (1994). The morphology o f the mental lexicon: Internal word structure viewed from a psycholinguistic perspective. Language and Cognitive Processes, 9, 227-269. Scalise, S. (1984). Generative M orphology. Dordrecht: Foris. Scalise, S. (1988). Inflection and derivation. Linguistics, 26, 561-581. Schaeffer, B., & Wallace, R. (1970). The comparison of word meanings. Journal o f Experimental Psychology, 86, 144-152. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 145 Schreuder, R. & Baayen, H. (1997). How complex simplex words can be. Journal o f Memory and Language, 37, 118-139. Schreuder, R., Grendel, M., Poulisse, N., Roelofs, A., & van de Voort, M. (1990). Lexical processing, morphological complexity and reading. In G. B. Flores d'Arcais, & D. A. Balota (Eds.), Comprehension Processes in Reading (pp. 125- 141). Hillsdale, NJ: Erlbaum. Segui, J., & Zubizarieta, M.-L. (1985). Mental representation of morphologically complex words and lexical access. Linguistics, 2 3 ,759-774. Seidenberg, M. S. (1987). Sublexical structures in visual word recognition: Access units or orthographic redundancy? In M. Coltheart (Ed.), Attention and Performance Vol. XII: Reading (pp. 245-263). Hillsdale, NJ: Erlbaum. Seidenberg, M. S. (1989). Reading complex words. In G. N. Carlson, & M. K. Tannenhaus (Eds.), Linguistic Structure in Language Processing (pp. 53-105). Dordrecht: Kluwer Academic Publishers. Seidenberg, M.S. (1990). Lexical access: Another theoretical soupstone? In G. B. Flores d'Arcais, & D. A. Balota (Eds.), Comprehension Processes in Reading (pp. 33- 71). Hillsdale, NJ: Erlbaum. Seidenberg, M. S., & McClelland, J. L. (1989). A distributed, developmental model of word recognition and naming. Psychological Review, 96, 523-568. Slobin, D. I. (1973). Cognitive prerequisites for the development of grammar. InC . Ferguson, & D. I. Slobin (Eds.), Studies o f Child Language Development (pp. 175- 208). New York: Holt, Rinehart, & Winston. Slobin, D. I. (1982). Universal and particular in the acquisition of language. InE . Wanner, & L. Gleitman (Eds.), Language Acquisition: The State o f the A rt (pp. 128-170). New York: Cambridge University Press. Slobin, D. I. (1985). Cross-linguistic evidence for the language-making capacity. InD. I. Slobin (Ed.), The Cross-linguistic Study o f Language Acquisition. Volume 2: Theoretical Issues. Hillsdale, NJ: Erlbaum. Spoehr, K. T., & Smith, E. E. (1973). The role o f syllables in perceptual processing. Cognitive Psychology, 5 ,71-89. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 146 Stanners, R. F., Neiser, J. J., Hemon, W. P., & Hall, R. (1979). Memory representation for morphologically related words. Journal o f Verbal Learning and Verbal Behavior, 18, 399-412. Taft, M. (1979). Recognition of affixed words and the frequency effect. M emory and Cognition, 7,263-272. Taft, M. (1981). Prefix stripping revisited. Journal o f Verbal Learning and Verbal Behavior, 20, 289-297. Taft, M. (1988). A morphological-decomposition model of lexical access. Linguistics, 26, 657-667. Taft, M. (1994). Interactive-activation as a framework for understanding morphological processing. Language and Cognitive Processes, 9, 271-294. Taft, M., & Forster, K. (1975). Lexical storage and retrieval of prefixed words. Journal o f Verbal Learning and Verbal Behavior, 1 4 ,638-647. Taft, M., & Forster, K. (1976). Lexical storage and retrieval of polymorphemic and polysyllabic words. Journal o f Verbal Learning and Verbal Behavior, 1 5 ,607-620. Tyler, L., Behrens, S., Cobb, H„ & Marslen-Wilson, W. (1990). Processing distinctions between stems and affixes: Evidence from a non-fluent aphasic patient CognitionA 36, 129-153. van Gelder, T. (1991). What is the “D” in “PD F’? A survey of the concept of distribution. In S. Stitch, D. Rumelhart, & W. Ramsey (Eds.), Philosophy and Connectionist Theory (pp. 33-59). Hillsdale, NJ: Erlbaum. van Gelder, T. (1992). Defining ‘distributed representation’. Connection Science, 4, 175-191. Zwicky, A. (1990). Inflectional morphology as a (sub)component of grammar. In W. Dressier, H. Luschiitzky, O. Pfeiffer, & J. Rennison (Eds.), Contemporary M orphology (pp. 217-236). Berlin: Mouton de Gruyter. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 147 Appendix A Stimuli from Experiment 1: Primes, targets, and semantic relatedness ratings Condition 1 Prime Target Semantic Relatedness Rating Mean Std. Dev. lettuce let 1.053 0.229 dragon drag 1.056 0.236 rugby rug 1.056 0.236 wallow wall 1.056 0.236 catalog cat 1.103 0.310 wiggle wig 1.111 0.323 pumpkin pump 1.158 0.501 cartoon cart 1.167 0.514 ribbon rib 1.167 0.514 candid can 1.176 0.728 spinach spin 1.222 0.732 bucket buck 1.241 0.786 litmus lit 1.278 0.575 kidney kid 1.379 0.622 pocket pock 1.389 1.037 bulletin bullet 1.500 1.249 palace pal 1.517 1.503 matador mat 1.724 1.360 riddle rid 1.793 1.398 lentil lent 1.862 1.642 pickle pick 1.897 1.676 captain cap 1.897 1.398 rapid rap 1.897 1.345 stubborn stub 1.966 2.079 musket musk 2.000 1.711 perturb pert 2.056 1.589 sandal sand 3.138 2.475 tinsel tin 3.483 2.747 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 148 v^unuiuuKi ^ Prime Target Semantic Relatedness Rating Mean Std. Dev. comer com 1.069 0.306 message mess 1.091 0.420 penance pen 1.131 0.465 figment fig 1.143 0.490 pigment Pig 1.169 0.575 manage man 1.348 0.734 pillage pill 1.484 1.112 department depart 1.583 1.123 aspen asp 1.692 1.462 customer custom 1.697 1.301 fasten fast 1.708 1.261 tenable ten 1.845 1.335 barber barb 1.986 1.510 fabrication fabric 2 1.499 apartment apart 2 1.914 slipper slip 2.108 1.371 hardly hard 2.139 1.638 appliance apply 2.175 1.487 missive miss 2.2 1.366 passive pass 2.262 1.450 comely come 2.339 1.469 foundation found 2.364 1.399 portable port 2.423 1.704 gingerly ginger 2.464 1.990 winsome win 2.655 1.743 after aft 2.679 1.964 craven crave 2.87 2.019 homely home 2.889 2.080 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 149 ^ U l I U l U U I l J Prime Target Semantic Relatedness Rating Mean Std. Dev. allowance allow 3.061 1.744 backer back 3.206 1.696 stately state 3.227 1.690 molten molt 3.25 2.047 postage post 3.258 1.748 publication public 3.258 1.658 basement base 3.348 1.957 lately late 3.394 1.944 likely like 3.415 1.870 suitable suit 3.424 2.083 tiresome tire 3.621 2.286 jabber jab 3.65 2.378 barker bark 3.672 2.275 dresser dress 3.681 1.830 active act 3.879 1.785 lovely love 3.909 1.821 shipment ship 3.985 1.917 shortage short 4.121 1.949 warden ward 4.188 1.790 personable person 4.236 1.570 plantation plant 4.444 1.677 favorable favor 4.576 1.857 ashen ash 4.644 1.910 grossly gross 4.662 2.056 placement place 4.773 1.634 storage store 4.833 2.012 equipment equip 4.944 1.528 notation note 4.969 1.741 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 150 Condition 4 Prime Target Semantic Relatedness Rating Mean Std. Dev. reasonable reason 5.028 1.601 emergence emerge 5.234 1.974 treatment treat 5.262 1.881 taxable tax 5.722 1.355 chatter chat 5.75 1.340 sicken sick 5.771 1.321 toughen tough 5.819 1.314 deadly dead 5.833 1.463 spillage spill 5.845 1.283 endurance endure 5.889 1.295 effective effect 5.955 1.841 payment pay 6.028 1.210 deafen deaf 6.069 0.969 consideration consider 6.091 1.367 golden gold 6.167 1.171 darken dark 6.194 0.973 government govern 6.211 0.893 chilly chill 6.227 1.187 wreckage wreck 6.292 1.119 honorable honor 6.292 0.879 awaken awake 6.323 1.239 kindly kind 6.324 0.938 hunter hunt 6.525 0.931 acceptance accept 6.556 0.748 boldly bold 6.597 0.781 annoyance annoy 6.606 0.839 teacher teach 6.621 0.760 cowardly coward 6.639 0.756 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 151 Condition 5 Prime Target Semantic Relatedness Rating Mean Std. Dev. demand ask 4.667 2.567 agile nimble 6.647 3.181 veranda porch 6.722 2.539 pocketbook purse 6.793 1.859 ravenous starving 7.000 2.062 crooked bent 7.034 1.822 blemish spot 7.222 1.700 barbecue grill 7.379 1.916 celebrity star 7.500 1.505 idea notion 7.684 1.600 marionette puppet 7.684 1.493 ancient old 7.724 1.131 injure hurt 7.737 1.851 detonate explode 7.778 1.665 conceal hide 7.897 1.633 happy cheerful 7.947 1.508 intelligent smart 8.000 1.333 imitate copy 8.000 1.414 quickly fast 8.053 1.545 immense huge 8.069 1.307 profit gain 8.069 1.033 forest woods 8.103 0.939 narcotic drug 8.158 1.344 location place 8.345 0.814 garbage trash 8.483 0.738 sorcery magic 8.483 0.871 sofa couch 8.655 0.614 automobile car 8.667 0.686 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 152 Appendix B Stimuli from Experiment 2: Primes, targets, and sem antic relatedness ratings by Condition Condition 1 Prime Target Semantic Relatedness Rating Mean Std. Dev. performance perform 8.286 0.845 arrival arrive 8.238 1.446 inventive invent 8.200 0.894 maturity mature 8.143 1.195 cultural culture 8.000 1.673 correspondence correspond 8.000 1.304 alcoholic alcohol 8.000 1.449 ceremonial ceremony 8.000 1.140 talkative talk 7.952 1.284 humorous humor 7.952 1.244 conformity conform 7.952 1.359 existence exist 7.905 1.786 temptation tempt 7.857 1.621 acceptable accept 7.700 1.809 falsify false 7.650 1.725 symbolize symbol 7.381 1.987 attractive attract 7.238 1.480 festival festive 7.143 1.236 clinical clinic 7.095 1.786 creative create 6.952 1.658 critical critic 6.950 2.373 notation note 6.905 1.868 alteration alter 6.762 2.567 formation form 6.619 2.500 solidly solid 6.524 2.228 scenic scene 6.429 2.420 guardian guard 6.143 2.414 adorable adore 5.900 2.490 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 153 Condition 2 Prime Target Semantic Relatedness Rating Mean Std. Dev. absorption absorb 8.296 1.815 assertion assert 7.577 2.176 completion complete 7.885 2.100 conclusion conclude 8.148 1.695 deletion delete 8.231 1.023 departure depart 8.185 1.360 elevation elevate 6.519 2.748 ethnicity ethnic 7.185 2.063 evasive evade 6.926 2.455 exclusion exclude 7.889 1.618 explosive explode 7.333 1.341 exposure expose 7.074 2.148 facial face 7.074 1.414 factual fact 7.667 1.271 financial finance 7.346 1.379 lyricist lyric 6.962 1.808 migration migrate 8.185 1.219 moisture moist 8.000 1.091 musician music 7.815 1.570 narration narrate 7.963 1.126 operation operate 7.741 1.852 persuasive persuade 7.926 1.400 pollution pollute 7.815 1.460 promotion promote 7.926 1.285 racial race 6.444 1.728 responsive respond 7.741 1.325 spiritual spirit 7.148 1.586 thievery thief 7.259 1.683 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 154 Condition 3 Prime Target abdominal abdomen abolition abolish abstinence abstain accusation accuse acidic acid adaptation adapt admirable admire angelic angel athletic athlete biblical bible cleanliness clean comedian comedy competition compete conspiracy conspire courageous courage criminal crime curiosity curious dramatic drama heroism hero idiotic idiot intervention intervene magnetic magnet maintenance maintain metallic metal national nation poetic poet preference prefer traumatic trauma Semantic Relatedness Rating Mean Std. Dev. 7.667 1.494 6.952 1.884 7.333 2.436 7.571 1.568 7.810 1.327 7.667 1.653 7.048 1.802 7.150 1.785 8.286 1.007 7.429 1.805 7.333 1.528 7.952 1.117 8.048 1.161 7.150 1.387 7.429 1.535 7.619 1.499 7.619 1.322 7.000 2.049 7.619 1.431 7.190 1.167 7.381 1.499 7.857 1.236 6.850 2.231 7.476 1.289 6.857 1.931 7.571 0.978 7.571 1.630 7.429 1.434 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 155 V ^ U I I U I L I U I I Prime Target Semantic Mean Relatedness Rating Std. Dev. allegation allege 7.741 1.289 assumption assume 8.333 0.877 brevity brief 6.519 2.190 collision collide 8.000 1.687 consumption consume 8.000 1.177 deception deceive 7.741 1.723 decision decide 8.148 1.099 descriptive describe 7.519 1.369 despicable despise 6.556 2.242 division divide 7.889 1.281 habitual habit 7.000 1.754 inquisitive inquire 6.926 1.542 introductory introduce 6.815 1.469 logician logic 7.000 1.523 magician magic 7.778 1.672 negligence neglect 6.667 2.094 perceptive perceive 6.704 1.958 precision precise 7.630 1.149 prescription prescribe 7.741 1.347 redemption redeem 6.593 2.099 reduction reduce 7.667 1.240 revision revise 7.963 1.192 sacrificial sacrifice 6.852 1.955 seductive seduce 7.556 1.649 statistician statistic 7.037 2.278 subscription subscribe 7.185 1.962 vocal voice 7.000 1.961 wisdom wise 7.667 1.387 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 156 Appendix C Stimuli from Experiment 3: Prim es, targets, and sem antic relatedness ratings by Condition Condition 1 Prime Target Semantic Relatedness Rating Mean Std. Dev. harmless harmful useful useless hugger huggable 6.000 2.236 deafen deafness 6.000 1.982 brighten brightly 5.897 2.425 childish childhood 5.889 2.272 employee employer 5.778 2.691 boyish boyhood 5.722 1.873 dusty duster 5.690 2.537 blindly blindness 5.586 2.292 mover moveable 5.421 2.168 containment container 5 2.249 guiltless guilty 4.526 3.044 cloudless cloudy 4.517 3.439 drinker drinkable 4.345 2.553 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 157 Condition 2 Prime Target Semantic Relatedness R ating Mean Std. Dev. actor actress 8.722 0.575 alcoholic alcoholism 8.655 0.897 bakery baker 8.345 1.143 zoology zoologist 8.276 0.922 specialize specialty 8.138 1.125 joyous joyful 8.069 1.132 continual continuous 7.893 1.595 sainthood saintly 7.862 1.060 quickly quickness 7.862 1.545 awkwardly awkwardness 7.759 1.300 heroism heroic 7.737 1.558 loudly loudness 7.69 1.168 scientific scientist 7.667 1.534 terrorism terrorist 7.611 1.577 waiter waitress 7.517 2.181 tourism tourist 7.5 1.618 friendship friendly 7.474 1.349 rudeness rudely 7.444 1.504 cleverness cleverly 7.444 1.381 drunken drunkard 7.389 1.650 crispy crispness 7.389 1.787 bitterly bitterness 7.379 1.449 mastery masterful 7.368 1.770 kindly kindness 7.368 1.571 sadness sadly 7.333 1.645 boldness boldly 7.333 2.196 calmness calmly 7.241 1.704 cheery cheerful 7.241 2.166 respectful respectable 7.111 1.711 sharpness sharpen 7.103 1.896 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 158 V ^ U i l U l U U U J Prime Target Semantic Mean Relatedness Rating Std. Dev. absorbent absorption 7.139 1.944 accusation accuser admission admittance 7.167 1.890 adoration adorable attraction attractive 6.829 2.176 confessor confession 7.767 1.549 creation creative 6.8 1.972 criticism critical generosity generously magnetize magnetic 7.417 2.062 maturity maturation 7.833 0.558 narration narrator 7.778 1.807 observation observant 7.833 1.117 relative relation 7.114 1.922 symbolic symbolize 7.629 1.352 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 159 Appendix D Stimuli from Experiment 4: Primes, targets, and semantic relatedness ratings by Condition Condition 1 Prime Target Semantic Relatedness Rating Mean Std. Dev. mildew mold 7.526 1.645 mortuary morgue 7.448 1.404 jubilee jubilant 7.167 1.790 jaunty jaunt 6.882 2.147 pester pest 6.632 2.216 ritzy glitzy 6.500 2.149 island isle 6.483 2.734 sheen shine 6.389 2.146 pulley pull 6.345 2.023 sorrow sorry 6.056 2.578 sprightly spry 6.000 2.478 halo holy 6.000 2.560 lunatic loon 5.842 2.243 trivial trifle 5.833 2.307 sacrilegious religious 5.690 2.855 mattress mat 5.684 2.110 pustule pus 5.586 2.706 haggard hag 5.474 2.590 fidget fiddle 5.389 2.831 prudent prude 5.211 2.463 allegiance alliance 5.167 2.282 lad lass 5.138 3.056 hurtle hurl 5.069 2.477 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 160 Condition 2 Prime Target Semantic Relatedness Rating Mean Std. Dev. enervate energy 4.931 2.520 ruffian rough 4.862 2.924 gourmand gourmet 4.778 2.547 gardenia garden 4.737 2.469 caricature character 4.655 2.075 sentinel sentry 4.571 2.795 gastric gas 4.552 2.229 dismal dismay 4.552 2.458 pantry pan 4.263 2.766 rivulet river 4.222 2.708 mastiff massive 4.214 2.644 glamour gleam 4.158 2.292 regatta regal 3.944 2.555 lingerie linen 3.862 2.341 harass harry 3.857 2.877 reindeer reins 3.828 2.633 sideburns side 3.667 2.142 massacre mass 3.655 2.497 lazy lay 3.526 2.435 tornado turn 3.379 2.456 pundit pun 3.357 2.198 bandanna band 3.222 2.756 tippler tipsy 3.111 2.324 rhapsody rapt 3.000 2.228 pedestrian pedantic 3.000 2.625 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 161 Condition 3 Prime Target Semantic Relatedness Rating Mean Std. Dev. rankle rank 2.964 2.151 rampart ramp 2.889 2.374 mysterious mist 2.759 2.198 adultery adult 2.759 2.047 hangnail hang 2.690 1.671 guttural gutter 2.690 2.301 clammy clam 2.655 2.143 greyhound gray 2.586 2.180 scalpel scalp 2.517 2.198 piggyback Pig 2.211 2.275 teetotaler tea 2.207 1.953 parchment parch 2.172 1.965 forceps force 1.947 2.013 coleslaw cold 1.722 1.904 salmonella salmon 1.556 1.294 boomerang boom 1.552 0.985 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Evaluating heaviness: Relative weight in the spoken production of heavy-NP shift

PDF

The phonemics and morphology of Hokkaido Ainu

PDF

Frequency in sentence comprehension

PDF

Intra-lexical noun-verb dissociations: Evidence from Chinese aphasia

PDF

Correspondence and faithfulness constraints in optimality theory: A study of Korean consonantal phonology

PDF

On the nature of particles in Japanese and its theoretical implications

PDF

Syntactic structures in nominals: A comparative study of Spanish and Southern Quechua

PDF

Metaphor and metonymy: A comparative study of Chinese and Western poetics

PDF

Effects of working memory and semantic impairments on speech in Alzheimer's disease.

PDF

On the representation and licensing of Q and Q-dependents

PDF

Investigating Stages Of Information Processing and Attentional Capacity: Contributions Of Startle Eyeblink Modification and Measures Of Signal Detection.

PDF

Definiteness types in Spanish: A study of natural discourse

PDF

In the event of focus

PDF

Studies in the psychopathology of time: 1. an exploration of temporal behavior in schizophrenia

PDF

Morpheme recognition in prosodic morphology

PDF

Language and context: toward a model of psycholinguistic competence

PDF

Perceptions of parental behaviors by a sample of gifted and no gifted women: an exploratory study

PDF

Women and the unspeakable: Rape in Ovid's "Metamorphoses"

PDF

New American fiction and the aesthetics of Camus and Robbe-Grillet

PDF

Phonology and syntax of Korean morphology

Asset Metadata

Creator Gonnerman, Laura Michelle (author)

Core Title Morphology and the lexicon: Exploring the semantics-phonology interface

Degree Doctor of Philosophy

Degree Program Linguistics

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag language, linguistics,OAI-PMH Harvest,psychology, cognitive,psychology, experimental

Language English

Contributor Digitized by ProQuest (provenance)

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c17-369777

Unique identifier UC11350791

Identifier 9933782.pdf (filename),usctheses-c17-369777 (legacy record id)

Legacy Identifier 9933782.pdf

Dmrecord 369777

Document Type Dissertation

Rights Gonnerman, Laura Michelle

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the au...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus, Los Angeles, California 90089, USA