Frequency in sentence comprehension

PDF

Download

Open document

Flip pages

Download a page range

Download transcript

Copy asset link

Request this asset

Request accessible transcript

Transcript (if available)

Content INFORMATION TO USERS This manuscript has been reproduced from the microfilm master. UMI 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 UMI 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. Each original is also photographed in one exposure and is included in reduced form at the back of the book. 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 UMI directly to order. UMI A Bell & Howell Information Company 300 North Zeeb Road, Ann Aibor MI 48106*1346 USA 313/761-4700 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. FREQUENCY IN SENTENCE COMPREHENSION LaQa Lalami A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL, UNIVERSITY OF SOUTHERN CALIFORNIA, In Partial Fulfillment of the Requirements for the Degree of DOCTOR OF PHILOSOPHY (Linguistics) May 1997 © 1997 Laïla Lalami Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. UMI Number: 9733083 UMI Microform 9733083 Copyright 1997, by UMI Company. All rights reserved. This microform edition is protected against unauthorized copying under Title 17, United States Code. UMI 300 North Zeeb Road Ann Arbor, MI 48103 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. UNIVERSITY OF SOUTHERN CALIFORNIA THE GRADUATE SCHOOL UNIVERSITY PARK LOS ANGELES. CALIFORNIA 90007 T T iis dissertatioK written by .................. under the direction of Dissertation Committee, and approved by all its members, has been presented to and accepted by The Graduate School, in partial fulfillment of re- cjuirements for the degree of DOCTOR OF PHILOSOPHY Date ..... DISSERTATION COMMnTEE ..... ...................... Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Acknowledgments Graduate study at USC has been a very exciting and wonderfully rew arding experience and there are m any people that I need to thank for that. First of these is my advisor, MaryeUen MacDonald, who introduced me to sentence processing and probabilistic constraints, tirelessly discussed ideas for experiments, and completely changed the way that I think about language. Many thanks also to Elaine Andersen for comments on earlier drafts of this dissertation, b u t especially for extremely relevant and always thought-provoking questions. Special thanks to Daniel Kempler for careful reading of this and m any other papers, as well as for insights into applications of constraint-based approaches to neuropsychology. I would also like to thank M ark Seidenberg and Jack Hawkins for rem arks on earlier drafts and for helpful discussions of connectionist and typology literatures, respectively. I am indebted to Joseph Aoun for teaching me to challenge assum ptions, those of others as well as my own. I am grateful to Jean-Roger Vergnaud for m any references and discussions on Chomskyan theory. Thanks are due to Barry Schein for sharing his views on the interface between syntax and semantics. My colleagues and friends at the language and cognitive neuroscience lab m ust also be thanked for fruitful exchanges on sentence processing and a great m any other topics: Am it Almor, Joe Allen, Lori Altm ann, M orten Christiansen, Tim Clausner, Joe Devlin, Laura Gonnerman, Mike H arm , Jim Hoeffner, Marc Joanisse, Donna LaVoie, Karen Marblestone, Sarah Schuster, Lynne Stallings, Karen Stevens and Robert Thornton. I am grateful to Sarah Schuster, Karen Stevens and Robert Thornton for helpful suggestions on data analyses. I would also like to thank m y fellow students in the linguistics departm ent here at USC: Pablo Albizu, Nancy Antrim, M aquela Brizuela, Jose Camacho, Lina Choueiri, Abdeslam Elomari, Cynthia Hagstrom , Roland Hinterholzl, Miao-Ling Hsieh, Sechang Lee, Ibtissam Kortobi, Karine Megerdoomian, Liliana Sanchez, Patricia Schneider-Zioga and Maki W atanabe. Their kindness, intelligence and generosity will always be remembered. I would especially like to acknowledge Pablo Albizu, Jose Camacho and Gorka Elordieta for help w ith Spanish and Basque data and Lina Choueiri for long debates on the nature of linguistic representations. The staff of the linguistics departm ent has always been there to support me, serve as guinea pigs for my surveys, and generally save my life in times of crises. Thanks to Laura Reiter, Kathy Stubaus, Linda W illiams-Culver, Don Bui, Zenobia Cann, Ayer Edwards, Jessica Hamilton and A udrey Messer. As always, I received the greatest support and encouragem ent from m y family, both here and in Morocco. Thanks to Alexander Yera, M adida El Fazzazzi, Ahmed Lalami, Gladys Yera, and Souâd, Abdelhak and Youness Lalami. 11 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table of Contents I. General introduction n. Chapter I: Models of sentence comprehension 4 Introduction 4 1. The garden-path m odel 4 1.1. The use of probabilistic information 10 1. 2. Separate stages of processing 11 1. 3. Ordering of principles 12 1. 4. Contradiction between principles 12 2. The linguistic tuning hypothesis 13 2.1. Do frequency counts relate to specific items or to general categories? 15 2. 2. W hat is the role of syntactic structure? 17 3. Gibson's m odel 18 3.1. Theoretical problems 23 3.2. Methodological problems 24 4. The com petition m odel 25 4.1. W hat constitutes a cue? 27 4.2. H ow are ambiguities handled? 28 4.3. Methodological problems 29 5. Crain, A ltm ann and Steedman's model 29 5.1. The notion of context 31 5.2. H ow are ambiguities other than M V/RR handled? 34 6. The constraint-based model 35 6.1. Late closure sentences 38 6.2. The role of non-lexical constraints 40 6.3. Long distance dependencies 40 7. Summary and Conclusions 41 m . Chapter II: Frequency as a constraint on sentence comprehension 46 Introduction 46 1. The issue of circularity 48 2. Frequency and voice 49 2.1. Structural fi-equency 49 2.2. Contingent frequency 53 3. The sources of frequency 55 3.1. Structural frequency 57 3.1.1. Pragmatic constraints 57 3.1.1.1. Natural salience 57 3.1.1.2. Length 58 3.1.2. Semantic constraints 59 31.2.1. Thematic structure 59 iii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3.2. Contingent frequency 60 3.2.1. Pragmatic constraints 60 3.2.1.1. N ature of the agent 60 3.2.2. Semantic constraints 61 3.2.21. Thematic structure 61 3.2.21. Transitivity 61 4. The representation of frequency 63 5. The effects of frequency 67 6. Summary and Conclusions 68 IV. Chapter HI: The com prehension of passive sentences 71 1. Studies of voice comprehension 71 2. Studies of voice production 81 3. Studies of voice in language acquisition 87 4. Studies of voice in aphasia 88 5. St. John and Gem sbacher's model 91 6. Summary and Conclusions 93 V. Chapter IV: Experiments 96 1. Experiment 1 98 2. Experiment 2 105 3. Experiment 3 116 4. Experiment 4 121 VI. Chapter V: Conclusions 130 1. The processing of unambiguous sentences 135 2. The processing of thematically ambiguous sentences 137 3. Representational issues 138 Vm. References 140 Vn. Appendixes 160 1. Hopper and Thompson's Parameters 160 2. Experiment 1 Verb Ratings 162 2. Stimuli for Experiments 2 and 3 166 3. Stimuli for Experiment 4 168 IV Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. List of Tables Chapter I Table 1: Experimental evidence for the garden-path m odel vs. the constraint-based model. 43 Chapter IV Table 1: Verb ratings for Experiment 1 100 Table 2: Verb rating as a function of thematic structure 102 Table 3: Example stim uli for Experiment 2 108 Table 4: Sentence regions for items in Experiment 2 110 Table 5: Example stimuli for Experiment 3 117 Table 6: Sentence regions for items in Experiment 3 117 Table 7: Example stimuli for Experiment 4 125 Table 8: Sentence regions for items in Experiment 4 126 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. List of Figures Chapter n Figure 1: Frequency hierarchy in English 51 Figure 2: Verb types: active-bias vs. passive-bias 54 Figure 3: A portion of Üie representation of a verb 66 Chapter IH Figure 1: Data from W right (1969) 75 Figure 2: Data from Frazier, Taft, Roeper, Clifton and Ehrlich (1984) 80 Figure 3: Data from Clark (1965) 83 Chapter IV Figure 1: Verb ratings and thematic structure 101 Figure 2: Correlation between percent of passive usage and verb rating 104 Figure 3: Experiment 2 reading times by voice type 111 Figure 4: Experiment 2 reading times in the passive voice 113 Figure 5: Experiment 2 reading times in the active voice 113 Figure 6: Experiment 3 reading times by voice type 118 Figure 7: Experiment 3 reading times in the passive voice 119 Figure 8: Experiment 3 reading times in the active voice 120 Figure 9: Experiment 4 reading times in the adjunct condition 127 Figure 10: Experiment 4 reading times in the argum ent condition 127 VI Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Abstract A lthough frequency effects have long been investigated in w ord recognition research, they have only recently been explored in sentence processing (Juliano & Tanenhaus, 1994; MacDonald, 1993; 1994; Trueswell, Tanenhaus & Kello, 1993). This dissertation investigates the w ay in which tw o types of frequency information, structural frequency (the rdative frequency of sentence types in the language) and contingent frequency (the relative frequency of lexical item s within sentence types), affect sentence com prehension. Proponents of current models of processing often disagree about the role of statistical properties of language. Based on available data, however, it is argued that probabilistic information is a crucial aspect of linguistic stimuli which com prehenders make use of during sentence processing. Accordingly, the role of frequency in the constraint-based model, its source, representation and effects are examined, using voice as a sample case. It is also argued that, while actives and passives have been extensively studied, the role of frequency in their processing has been somewhat ignored. Furtherm ore, because of the differences between actives and passives in term s of structural frequency and potentially in contingent frequency as well, gram m atical voice provides a unique opportunity for studying frequency effects. Four m ain issues are investigated in a series of experiments w ith a total of 300 participants: (1) the effect of contingent frequency on the com prehension of relatively unam biguous active and passive sentences containing active-bias verbs and passive-bias verbs; (2) the interaction of contingent frequency w ith structural frequency in parsing these sentences; (3) the role of contingent frequency in processing thematically am biguous passive sentences involving argum ent and adjunct complements; and (4) the interaction of contingent frequency w ith structural frequency in parsing such sentences. Results from these experiments suggest that contingent frequency directly affects the speed of processing in both thematically am biguous and unam biguous sentences. The interaction of structural and contingent frequency effects in on line sentence processing is also discussed. Thus it is argued that w hen the difference between structural and contingent frequencies is large, contingent frequency effects are masked in high-frequency structures (i.e. actives) but are m ade apparent with low-frequency structures (i.e. passives). This latter result is similar to the frequency-by-regularity interaction which has been docum ented in word recognition research (Seidenberg & McClelland, 1989). However, w hen the difference betw een structural émd contingent frequency is somewhat reduced, effects of contingent frequency appear in both high- frequency (i.e. argum ent complement) and low-frequency structures (i.e. adjunct complement). v ii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. GENERAL INTRODUCTION Frequency is an important aspect of linguistic stimuli that has long been recognized to affect the recognition and production of words, but has only recently draw n the interest of sentence comprehension researchers. This dissertation examines the role of frequency in sentence processing. I argue that statistical aspects of both lexical items and syntactic structures directly affect the speed and ease of parsing and that different types of frequency information can affect parsing in different ways. In a series of on-line experiments, I show that the frequency w ith which a verb is used in the passive voice affects its processing in unambiguous passive structures and that the frequency w ith which it appears w ith adjunct and argum ent complements determines the way that it is parsed in thematically am biguous structures. This dissertation is organized in five chapters. Chapter I provides a critical review of current models of sentence processing, starting w ith Frazier's garden-path model, through Mitchell's tuning hypothesis, Gibson's parsing principles. Bates and M acW hinney's competition model, Crain and Steedman's referential model and M acDonald's constraint-based m odel. This discussion of sentence processing makes a good case for the proposal that one of the best w ays of distinguishing betw een current models of sentence processing is to investigate relatively unam biguous structures. I also argue in favor of using the constraint-based m odel as a framework for the experimental w ork undertaken in this dissertation, based on its account of Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. current data, its unified explanation of lexical and syntactic parsing, and its applicability to other fields of human cognition such as language acquisition. Chapter H deals w ith the issue of firequency in sentence comprehension. Using the case of actives and passives as an example, I show how frequency differences that exist in the language are the result of several semantic and pragm atic factors such as thematic structure and transitivity and examine the way in which these frequency differences become represented in the language processing system. This is a crucial claim of the constraint-based model and it is important to support it with experimental evidence. Finally, I explore the kinds of effects that frequency has on syntactic parsing and investigate the way in which different sorts of frequency information interact. Chapter HI reviews the literature on passives. Passive sentences have been studied fi-om a psycholinguistic perspective since the 1960s, but fi-equency has rarely been a focus in this type of research. I argue that the results obtained in these studies can be at least partly attributed to probabilistic constraints. However, rather than comparing passives w ith actives as these studies typically do, I propose to use the structural fi-equency difference between these two structures in order to explore effects of structural frequency, contingent frequency, and their interaction. Chapter IV presents the experimental work undertaken to test some of the hypotheses presented in Chapters n and lU. Experiment 1 is a large scale study aimed at determining verb bias. Individual verb biases are then used for developing the materials for subsequent experiments. Experiment 2 2 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. investigates the processing of relatively unam biguous agentless passive sentences. Experiment 3 examines the processing of agented passives, in order to rule o u t an alternative explanation of the data from the previous experiment. Finally, Experiment 4 investigates thematically ambiguous passive sentences with argum ent and adjunct complements. Chapter V provides a critical review and sum m ary of previous chapters, as well as conclusions from the experimental work. I argue that contingent frequency directly affects the com prehension of somewhat unam biguous as well as thematically am biguous sentences. I also suggest that comprehenders keep track of contingent frequencies, and that this direct encoding guarantees that more general, structural frequencies are also encoded. In other words, sensitiveness to structural frequencies falls out of keeping track of contingent frequencies. Finally, I argue that the effects observed in Experiments 2, 3 and 4 w ould support the view that lexical and sentence processing are done via the same m echanisms and that the parsing of ambiguous and unambiguous structures can be done within the same fram ew ork. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER I - MODELS OF SENTENCE COMPREHENSION People parse utterances of their native language in a rapid, efficient way, with hardly any awareness of the large num ber of different sorts of ambiguities which perm eate these utterances. In processing sentences, people also develop m ental representations for the linguistic input they are hearing or reading. U nderstanding these two phenom ena is the goal of sentence comprehension research. Since the early 1970s, several accounts of sentence comprehension have been proposed. These accounts can typically be classified as serial and m odular or parallel and interactive in nature. Proponents of serial models claim that comprehenders handle possible sentence parses one at a time, and that only one kind of information (syntactic) leads the initial parse, without any input from other modules. In contrast, supporters of parallel models assum e that comprehenders handle all possible parses at once, and that different kinds of linguistic inform ation can inform the choice of correct alternatives. Mental representations for linguistic units postulated within each m odel thus turn out to be quite different. In this chapter, I review six models of sentence comprehension. The first two are serial; the last four are parallel. 1. The garden-path model Frazier's garden-path theory is perhaps the best-known model of sentence comprehension (Frazier, 1978; 1987; 1994; Rayner, Carlson & Frazier, 1983; Frazier & Rayner, 1987). It was initially form ulated in the late 1970s, appearing after a period of time in which psychologists were very skeptical about the behavioral study of language. Based on Chomsky's influential 4 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. theory. Miller (1962; Miller & Isard, 1963) had developed the 'Derivational Theory of Complexity’. The DTC addressed the question of how comprehenders recovered surface structure from deep structure. That is, it attem pted to find behavioral verification for the claim that transform ations were psychologically real, or at least plausible. Although the DTC was received w ith m uch enthusiasm , it soon became apparent that som e of the transformations had no efiects on real-time performance and that even the effects obtained w ere methodologically questionable. In the late 1970s, however, Frazier alm ost single-handedly revived the field of sentence comprehension by addressing the question of how surface structure was recovered (Frazier, 1978; Frazier & Rayner, 1982). The renewed interest in language that her work caused m ay be the reason why her m odel still enjoys so much popularity, despite an increasing num ber of questions about its ability to account for certain effects, such as the preference for non-minimal attachment interpretations in biasing contexts (Trueswell, Tanenhaus & Gamsey, 1994). Frazier's garden-path model assumes that comprehenders select only one interpretation of ambiguous structures. That interpretation m ay subsequently be confirmed or abandoned when enough evidence is m ade available. The theory is based on two major principles: minimal attachment and late closure. M inimal attachm ent assumes that the processing system favors structurally simpler structures. Thus, for each sentence encountered, the parser will opt for a structure that has fewer nodes in comparison with other possible structures. In contrast, late closure assumes that the processing system attaches new items into the clause or phrase postulated m ost recently, as long as this is grammatically permissible. 5 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Together, these two principles account for a large array of data. Thus minimal attachment accurately predicts that the main verb interpretation (rather than the reduced relative construal) is the one that first comes to m ind when unsuspecting readers are given reduced relative am biguous examples such as that in (la). Indeed, the parse in (lb), which assumes that the PP ends the clause, involves fewer nodes than that in (Ic), which assumes the existence of a reduced clause. (la) The horse raced past the bam fell.^ (Bever, 1970) (lb) [g In p The horse [vp raced [pp past the bam [y fell]]]]]. (Ic) [g [np The horse [g-bar [vP raced [pp past the bam [y fell]]]]]]. Frazier claims that minim al attachm ent also accounts for garden-path effects in NP-conjunction ambiguities such as (2a). As can be seen from the bracket diagrams below, (2b) involves the postulation of fewer nodes than (2c), and indeed the interpretation in (2b) is the one that first comes to mind. (2a) John tickled M arde and her sister laughed. (2b) [g [n p John [ y p tickled [n p M arde and her sister] [ y laughed]]]]]. (2c) [g [n p John [yp tickled [n p M arde [g [conj and [n p her sister [y laughed]]]]]]]]. ^ In a number of examples cited in this chapter, the ambiguity may only hold when the sentences are read rather than spoken. Ambiguity in spoken language has been extensively studied, particularly in terms of the role of prosodic cues. For example, Allbritton, McKoon and Ratcliff (1996) show that speakers do not consistently produce speech cues when reading ambiguous sentences in a disambiguating context. However, Warren, Grabe and Nolan (1995) argue that prosodic contrast is in fact used by comprehenders in the on-line interpretation of closure ambiguities. For detailed discussion on how prosodic cues may affect comprehension, see articles in the special issue of Language and Cognitive Processes (1996). 6 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In these kinds of examples, it is hypothesized that the N P her sister is initially parsed as p art of the direct object of the main verb. A ttaching the N P in this w ay requires fewer nodes tiian attaching it as a new clause, hence the preference for the NP construal. Finally, minimal attachm ent predicts a preference for VP (as in 3b) rather than NP modihcation (as in 3c) in PP- attachment ambiguities such as (3a), again because of structural simplicity. (3a) Jane saw the m an with the binoculars. (3b) [s [n p Jane [yp saw [n p the man [pp with the binoculars]]]]] (3c) [g [n p Jane [yp saw [n p [n p the man [pp with the binoculars]]]]]] In contrast, late closure handles ambiguities in w hich alternatives involve the sam e num ber of new nodes. The principle predicts garden-path effects in N P /S ambiguities as in (4a) where the NP will initially attach to the verb rather than to a new clause. It also accounts for tim e adverbial attachm ent ambiguities such as that in (5a), for which m ost native speakers prefer the modification of the adverbial to affect the em bedded verb. (4a) Since Jay always jogs a mile seems like light work. (4b) [g [gtar Since [n p Jay [Adv always [yp jogs [n p a mile]]]]] [yp seems..] (4c) [g [g-bar Since [n p Jay [Adv always [yp jogs]]]] [NP a mile [yp seems..]] (5a) Jane said Bill called yesterday. (5b) [g [n p Jane [yp said [g-bar [NP Bill [yp called Udv yesterday]]]]]]. (5c) [g [n p Jane [yp said [g-bar [NP Bill [yp called]]] [Adv yesterday]]]. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Since the original formulation of the garden-path m odel, Frazier has added a few more principles to the theory in order to accommodate the growing body of d ata that have identified, among other things, effects of probabilistic constraints on comprehension. For example, she proposed the existence of a thematic processor whose role is to evaluate initial analyses in term s of semantic plausibility and discourse appropriateness, and subsequently confirm or reject these analyses (Rayner, Carlson & Frazier, 1983; Frazier, 1987). W hen the thematic processor rejects an analysis, structural re-analysis becomes necessary. Reanalysis, Frazier (1994) contends, is associated with revision cost which depends on variables such as the length of the ambiguous phrase, the semantic plausibility of the first analysis (i.e. the structurally sim pler analysis) and the frequency of alternative subcategorization frames. However, the m ost im portant addition to the garden-path model is the principle of construal (Frazier & Clifton, 1996). This principle was incorporated into Frazier's theory in order to handle cases w here attachm ent preferences do not follow late closure. Recall that late closure predicts that these sorts of ambiguities will be initially disambiguated in favor of the lower N P (the colonel, in (6a)). (6a) The daughter of the colonel who was standing on the balcony.... (6b) [s [n p The daughter of the colonel] [S-bar who was standing on the balcony]].... (6c) [s [n p The daughter of [NP the colonel who was standing on the balcony]]].... 8 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Evidence for the preference of the low N P attachm ent in relative clauses in English was provided by Mitchell and Cuetos (1988) in a questionnaire study. However, the disambiguation of such relative clause attachments turns out to be at least partly determ ined by the semantics of the NPs involved, as dem onstrated by (7) and (8). (7) The woman in the room w ith red hair. (8) The woman in the room with big chandeliers. The principle of construal was developed in order to handle cases such as these, w here non-structural information determ ines the correct interpretation. Construal states that one m ust "associate a phrase XP that caimot be analyzed as instantiating a prim ary relation into the current thematic processing domain", and that "XP m ust be interpreted within that dom ain using structural and non-structural (interpretive) principles". The 'current thematic processing domain' is defined as the extended maximal projection of the last theta-assigner. Association is distinguished from attachm ent in that it is not governed by principles that will blindly favor certain structural attachments over others. It can leave certain syntactic relations to be determ ined by nonstructural as well as structural information (p.31). Frazier and Clifton have applied^ construal to relative clauses, as in (9) and adjunct predication, as in (10). (9) The daughter of the colonel who was standing on the balcony. (10) John ate the broccoli naked/John ate the broccoli raw. ^ Frazier and Clifton (1996) believe that construal could apply to conjoined structures such as George knew Dan voted for Sam and himself, though no analysis of such constructions has been formulated yet. 9 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Notice that construal will typically work w ith phrases that are adjuncts, or otherw ise optional sentence fragments, while late closure and m inimal attachm ent will be applied to arguments, or obligatory phrases. Frazier's m odel has thus come a long way in terms of specifying how various types of structures are interpreted by comprehenders, but several argum ents can be raised against this approach. 1. 1. The use of probabilistic information In its current formulation, the garden-path model cannot explain early effects of non-syntactic variables on the resolution of ambiguities involving argum ents. Indeed, several studies have show n that frequency, animacy and tem poral and referential context can ease or hinder the resolution of argum ent ambiguities. For example, Trueswell, Tanenhaus and Gam sey (1994) showed that subjects had difficulty reading ambiguous reduced relative clauses which begin w ith animate nouns (e.g. the lawyer examined) but not w ith those that begin w ith inanimate nouns (e.g. the evidence examined). MacDonald (1993) found that the relative frequency of use of a noun as a head and as a modifier partly accounted for reading times on N /V ambiguous sentences such as (11) and (12). (11) The warehouse fires m any workers each spring. (12) The corporation fires many workers each spring. Spivey-Knowlton, Trueswell and Tanenhaus (1993) show ed that prior contexts establishing the existence of one vs. m ultiple referents for am biguous fragments such as the actress selected have an early effect on 10 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. syntactic am biguity resolution because they help constrain the range of possible interpretations for that phrase. Similarly, Trueswell and Tanenhaus (1991) found that subjects had less difficulty reading reduced relative ambiguous clauses when they appeared in future contexts as com pared w ith past contexts. This is due to the fact that, in reduced relative clauses, the past participle reading is m ore readily available w hen the past tense interpretation is ruled out through the use of a future tense context. It should be pointed out that Frazier has come to include frequency as a factor which may influence parsing (as in Frazier & Clifton, 1996). However, firequency is not explicitly encoded in that m odel's representations, and it is difficult to see how one can use information that is not represented. That is to say, the garden-path m odel does not predict or explain how comprehenders w ould keep track or m ake use of firequency inform ation because the representations that it assumes do not include fi-equency or, indeed, any kind of probabilistic information. 1. 2. Separate stages of processing Faced by the growing body of evidence supporting the early influence of lexical information on syntactic ambiguity resolution, Frazier has come to argue that, in her model, there is immediate checking of semantic plausibility after the initial syntactic parse is completed. It is interesting to note that since thematic checking was originally formulated, it m oved from being done at the end of the sentence (Frazier & Rayner, 1982; Frazier, 1987) to being done 'immediately' after parsing (Frazier, 1994). This no doubt reflects an attem pt at accommodating data that show that processing and lexical access occur simultaneously. However, immediate checking after syntactic processing may not work, simply because this entails a serial parser, and a serial parser m ust, by definition, have distinct stages. If the difference in processing time between 11 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. these stages is not m easurable, then one cannot reasonably continue to claim that these separate stages do exist. 1.3. Ordering of principles The principles of m inim al attachment and late closure, w hich handle ambiguities involving prim ary phrases (or arguments) have to be ordered in order to handle cases such as (13a). In this example, there are two possible attachment sites for the PP. Late closure would predict that the PP should modify reading, because that is the most recent site. The correct interpretation, however, is the one provided by minimal attachm ent, where the PP is an argum ent of put. (13a) Jay put the book Jane was reading in the library. (13b) [g [np Jay [yp put [n p the book [g-bar [n p Jane [vP was reading [pp in the library.]]]]]]]] (13c) [g [n p Jay [yp put [n p the book [g-bar (n p Jane [vP was reading]]] [pp in the library.]]]]] To resolve this conflict, Frazier and Fodor (1980) proposed that m inim al attachment should take precedence over late closure. The problem w ith this view is that there is no independent evidence for this ordering. The latter seems to follow solely from inadequate predictions of the theory. 1.4 Contradiction between principles A serious criticism that can be leveled against the garden-path theory is that its different principles clash. On the one hand, Frazier claims that there is 12 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. early use of syntactic information in resolving argum ent ambiguities such as the M V/RR, N P /S and PP-attachment ambiguities. On the other hand, she claims that there is early use of semantic inform ation in resolving adjunct ambiguities such as relative clause attachment and adjunct predication structures. Furthermore, both of these types of processing (i.e. syntactic for arguments, sem antic for adjuncts) are assumed to function within a serial model of processing. These are inconsistencies that have yet to be addressed. 2. The linguistic tuning hypothesis Another principle-based model of processing was put forth by Don Mitchell and his colleagues (Cuetos & Mitchell, 1988; Mitchell & Cuetos, 1991; Mitchell, Cuetos, Corley & c Brysbaert, 1995). This model, the linguistic tuning hypothesis, assumes that individual strategies are applied on the basis of comprehenders' previous contact with language. Unlike Frazier's model, Mitchell's m odel relies on w hat he calls 'records' of previous exposure to language. However, these records are seen as referring to categorial heads and modifiers, i.e. to general syntactic categories, rather than to the occurrence or co-occurrence of specific lexical items. According to Mitchell, his model is coarser-grained w hereas models that rely on the accum ulated frequencies of individual item s are mixed-grained or finer-grained models. Mitchell's linguistic tuning hypothesis is based on the premise that detailed lexical information is excluded from all records of the prevalence of different readings of ambiguous structures. In other words, the model relies on frequency of occurrence, but frequency only refers to sentence frames and not to lexical items. For example, in reference to the relative clause attachment ambiguity, Mitchell et al. (1995) write: 13 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Put simply, [our] proposal is that in sentences of this type, readers and listeners initially attadi the relative clause to whichever of the competing host positions has been modified most frequently by relative clauses in the past. By hypothesis, the statistical records that are consulted in this process are held to exclude all details of the nouns (or NPs) that occupied these positions in past encounters with a sentence-frame (p.l3). The existence of lexical influences is re-interpreted by Mitchell as being the result of later filtering phenom ena rather than first-pass parsing. Evidence for filtering processes was provided by Mitchell (1987), w ho used a self-paced reading task to investigate the processing of N P/S ambiguities such as (14) and (15). (14) After the child had sneezed the doctor prescribed a course of injections. (15) After the child had visited the doctor prescribed a course of injections. Mitchell found that there were garden-path effects for sentences such as (15), but m ore interestingly, that sentences such as (14) were m ore difficult than those such as (15) in spite of the fact that they contained intransitive verbs. If lexical information had been readily available, then there should have been no garden-path effects with intransitive verbs like sneezed. The presence of such effects was taken as evidence that the parser has no im m ediate access to lexical information and that it is taken care of by a second stage in processing which filters out incorrect interpretations. Notice, how ever, that Mitchell’ s experim ent measured reading times at three points in the sentence: over the entire first clause up to the verb, over the NP itself, and over the last phrase. This type of measurement may have missed subtle lexical effects. 14 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In sum m ary, M itchell's model includes serial processing, but allows frequency information to be used, though this w ould not come into play for spedfrc items. The evidence that he presents for his m odel centers around the finding that com prehenders do not utilize the frequencies of specific lexical biases. However, as will be shown in the next section, these findings are debatable. 2.1. Do frequency counts relate to specific items or to general categories? The evidence presented in support of the claim that frequency counts refer to categories rather than to individual items comes from studies of NP- PP attachm ent in English and French. For example, Mitchell et al. report the results of a study by Corley and Corley (submitted), in which they examined clause-attachment biases such as (16) vs. (17). (16) The satirist ridiculed the lawyer of the firm wh... (17) The satirist ridiculed the firm of the lawyer w h ... Corley and Corley found that relative clauses invariably attached to the second noun, which w ould indicate that it is som ething about the structural position of the noun that is attracting a modifier and not, or not only, some lexical property of the noun itself. This result provides very strong evidence for category-based parsing; however, modification by a relative clause may also be dependent on other factors besides structural position. For example, there are properties of NPs, such as definiteness, that can influence attachment. It seems that w hen one of the two NPs is indefinite and the other is definite, the additional information provided by the relative clause will tend to m odify the indefinite rather than the definite NP. Thus, in both the 15 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. satirist ridiculed a lawyer of the firm and the satirist ridiculed a firm of the lawyer, native speakers seem to prefer attachm ent to the first NP, which is the structurally disfavored site of attachment. Nevertheless, to the extent that preference for the second NP holds w hen both NPs are definite, Corley and Corley's argum ent is valid and provides good experimental support for Mitchell's theory. Other evidence adduced to the claim that processing refers to frequencies of sentence frames and not to individual lexical items comes from a study by Traxler and Pickering (1995). These researchers used pre testing procedures to identify verbs (e.g. realized) which strongly prefer the reduced com plem ent (e.g. realized this example is useful) rather than the direct object attachm ent (e.g. realized the goal of writing the paper). Eye- tracking m easurements revealed that w hen given examples such as (18) and (19) subjects have m ore trouble reading realized her shoes than realized her goals. (18) She realized her goals w ould be out of reach. (19) She realized her shoes w ould be out of reach. This is taken as evidence by Mitchell et al. (1995) that the N P following the verb is initially interpreted as a direct object, a fact which w ould be explained if coarse-grained rather than fine-grained tabulations were m ade, because one would indiscriminately opt for the m ore frequent direct object interpretation. However, other explanations of the data are also possible. For example, one constraint in processing verbs that are ambiguous between a direct object reading and a sentence complement reading is the relative frequencies of 16 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. these biases for each individual verb. Thus, Juliano and Tanenhaus (1994) showed that comprehenders experience processing difficulty when verbs such as insisted, which are m ore h ’ equently used w ith an embedded sentence, appear w ith an NP. Conversely, a verb such as visited, which is more frequently used with a direct object, was harder to process when it appeared w ith an embedded sentence complement (see also Gamsey, Pearlmutter, Myers & Lotocky, 1996). Thus, Mitchell's data could be reanalyzed as reflecting individual verb biases rather than categorial preferences. 2. 2. W hat is the role of syntactic structure? Mitchell and his colleagues did several experiments on PP-attachment in Spanish, in the hope of showing that late closure attachment is not universal and the type of syntactic structure that a noun is embedded in influences the way in which it is attached to a relative clause. For example, Mitchell and Cuetos (1991b) claim that the noun la chica does not attach to the relative clause w hen it appears in a prepositional phrase as in (20) but that the same noun does attach to the relative clause w hen it is itself em bedded in a relative clause as in (21). (20) Pedro miraba los libros de la chica qu ... Pedro was looking at the books of the girl who/which... (21) Pedro miraba los libros que pertenedan a la chica que... Pedro was looking at the books which belonged to the girl w ho/w hich... Notice that these tendencies are probably also subject to other constraints. In the sentence containing the embedded relative clause, the noun los libros is 17 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. m odified by the VP pertenedan a la chica so it is very likely that incoming inform ation will preferably m odify the as yet unspecified NP la chica, rather than the N P that has already been specified. By contrast, in the sentence Pedro miraba los libros de la chica que, both of the nouns los libros and la chica are open for modification. Furtherm ore, Mitchell and Cuetos' examples are also constrained by length. The longer the second NP is, the less likely it is for an incoming fragm ent to attach to the first NP, presum ably because of the higher dem and that such an attachm ent would place on w orking m emory capacities. As has been shown above, Mitchell's model relies on evidence that m ay be open to re-interpretation. 3. Gibson's Model Gibson (1991) proposed a parallel model of processing which assumes that all locally possible readings of an ambiguous string can be activated and m aintained by the parser simultaneously. He further argues that there are costs attached to retaining different interpretations, and that these costs are continually com pared with every new piece of stimuli. Due to m em ory limitations on the hum an processing system, only those interpretations that are deem ed 'least expensive' will be maintained. Thus, Gibson's m odel is a parallel m odel of processing, one that is limited by working memory capacities. Indeed, a sentence parse that is relatively less costly to m aintain will be pursued, leading to fast and accurate disambiguation. On the other hand, it is possible that, during on-line processing, a sentence parse that is too costly to m aintain will be abandoned, even if it ultimately turns out to be the correct one, leading to slow and difficult disambiguation (as is the case for garden-path sentences). The computation of relative cost in Gibson's m odel is construed as depending on a num ber of linguistic and non-linguistic 18 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. constraints such as the 0-criterion and recency preference. The 0-criterion is a linguistic principle which requires that each argum ent bear only one 0-role, and that each 0-role be assigned to only one argum ent. Recency preference is an attachment principle which captures the fact that constituents will preferably attach to the most recent site (see Kimball, 1973; and Frazier, 1978, for earlier and similar proposals). To see how processing cost is computed in Gibson's model, consider example (22), in which the constituent the cotton is initially analyzed as a prénommai modifier, but turns out to be a noun phrase that precedes a reduced relative clause. (22) The cotton clothing is m ade of grows in Mississippi. The prenom inal m odifier analysis {the cotton clothing) involves only one local 0-violation, because the parser has not yet received the head verb that should assign it a subject 0-role. On the other hand, the reduced relative clause interpretation (the cotton which clothing is made of....) would be associated with three local 0-violations, because the parser has not identified a 0-assigner for the NP the cotton (i.e. the verb grows), a 0-assigner for the N P the clothing (i.e. the verb is), and a 0-assigner for the embedded relative clause (i.e. the w ord which, deleted in the example). Thus the prenominal modifier analysis involves one violation while the reduced relative clause involves three. A difference of two 0-violations is considered by Gibson as being too costly to maintain, and so the reduced relative clause interpretation is not pursued any further, hence the garden-path effect at the end of the sentence. 19 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Gibson, Hickok and Schütze (1994) have provided some evidence for the parser's reliance on the 0-criterion in their investigation of the processing of empty categories. In English, sentences such as (23) are argued to contain a linguistically defined gap while sim ilar examples such as (24) do not. While processing the example in (23) an em pty category is postulated after the verb bring, because it is assum ed that the w ord who originated from the direct object position (which ultim ately turns out to be incorrect). However, while processing the example in (24), a direct object is expected after the verb, rather than a gap. (23) My brother w anted to know w ho Ruth will bring us hom e to at Christmas. (24) My brother w anted to know if Ruth will bring us hom e to Mom at Christm as. Gibson et al. claim that the processing difficulty that is clearly experienced with example (23) as compared w ith example (24) is due to this difference in number of 6-violations. The sentence in (23) has two possible parses, one which contains an em pty category (as in (25)) and one w hich doesn't (as in (26)). 1 (25) My brother wanted to know w ho Ruth will bring t us. (26) My brother wanted to know who Ruth will bring us. 20 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In (25), tile wh-N P who receives a 0-role by virtue of being linked to the object position of the verb bring. In (26), however, the wh-NP who requires a 0-role but lacks one. As predicted by Gibson's theory, the interpretation that posits a gap is the least costly to maintain and it is also the one that is first assum ed, hence the garden-path effect w hen the w ord us is encountered. Thus Gibson's m odel provides an account for both structural ambiguities and for long-distance dependencies by using the 0-criterion. Evidence for the parser's reliance on recency preference comes from a careful study of PP-attachment in Spanish and English. Indeed, Gibson, Pearlmutter, Canseco-Gonzalez and Hickok (1995) noted that English time adverbials are typically attached to the second of two potential sites, as shown in (27). Similarly, relative clauses seem to be generally attached to the second NP, as shown in (28). (27) John said that Bill left yesterday. I I (28) The reporter interviewed the daughter of the general who left for the Gulf. However, Gibson et al. noticed that while time adverbials in Spanish behave the same way as in English, in that they attach to the second NP, as shown in (29), relative clauses seem to prefer attachm ent to the first NP, as shown in (30). 21 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (29) Juan dijo que Maria se m uriô ayer. Juan said that Maria died yesterday. I I (30) El periodista entrevisto a la hija del coronel que tuvo el accidente. The journalist interviewed the daughter of the colonel who had the accident. These empirical facts led Gibson to claim that both recenq/ preference and predicate proximity govern attachm ent ambiguities. Recency preference states that attachm ent to the m ost recent sites m ust be favored while predicate proximity states that the parser attaches "as close as possible to the head of a predicate phrase" (p.23). In their Spanish study, Gibson et al. (1995) investigated constructions such as (31) in which the relative clause can potentially attach to either of the three NPs. (31) La(s) lâmpara(s) cerca de la(s) pintura(s) de la(s) casa(s) que fue(ron) dahada en la inundadon. The lamp(s) near the painting(s) of the house(s) that w as/w ere dam aged in the flood. Gibson et al. disam biguated the sentences by m anipulating the grammatical num ber of the verb fue vs. fueron {was vs. were). They found that subjects' latendes were faster for low attachments relative to both high or m iddle attachm ents and faster on high attachments relative to m iddle attachments. In other words, there was a hierarchy such that low attachments were preferred over high attachm ents which were in turn favored over m iddle attachments. These results were interpreted as suggesting that Spanish- 22 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. speaking subjects use both a recency preference (which makes them favor the lowest attachm ent site) and a predicate proximity preference (which makes them favor the attachment site closest to the predicate, i.e. the high attachm ent site). In the English study, the same pattern of behavior was observed: preference for the lowest N P over the other two, and preference for the highest over the middle NP. Using recency preference predicts attachm ent to the lower rather than the m iddle N P while predicate proximity favors higher over middle attachments. In general, Gibson's model is very promising, and offers well worked- out m echanisms for processing both local and long distance structural ambiguities as well as syntactic dependencies such as gap-filling. However, the model suffers from a few shortcomings which are discussed below. 3. 1. Theoretical problems One of the predictions of Gibson et al.’ s model is that in languages with rigid SVG order (e.g. English), the average distance from verbal heads to argum ents will tend to be short, hence the dominance of recency preference, and preference for last (or low) NP attachm ent in two-site attachment ambiguities. Conversely, in languages whose canonical word orders include VOS, VSO, SOV or OSV, the predicate proximity factor should be relatively strong, which w ould result in a preference for the first (or high) NP in relative clause attachm ent ambiguities. However, although Moroccan Arabic is a VSO language, there seems to be a preference for attaching the relative clause to the second (or low) NP in sentences such as (32). 23 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (32) tteswira hda ttem tal Ui kanbghi. the-picture near the-sculpture that Tense-I-like The picture near the sculpture that I like. In Basque, a language traditionally classified as SOV, the relative clause attaches to the closest N P the painting, in cases w here there are two potential attachment sites, as show n in (33a). W hen there are three potential attachment sites, the relative clause attaches to the m ost recent NP the house, or to the m iddle one the painting, but not to the highest one, as shown in (33b). These facts stand as a counterexample to Gibson's predictions. (33a) gustatzen zaidan koadrotik hurbil dagoen argia like-imper. aux-com p painting-from near is-comp lam p The lamp near the painting that I like. (33b) uholean izorratu zen etxearen koadrotik hurbil dagoen argia flood dam age aux-comp house-of painting-from near is-comp lamp The lamp near the painting of the house that was dam aged in the flood. 3. 2. Methodological problems One of the problem s with the Gibson, Pearlm utter, Canseco-Gonzalez and Hickok (1995) study is that it utilized subjects who spoke different varieties of Spanish (e.g. Mexican, Puerto-Rican, Spanish, Argentinean etc.). Such differences am ong participants m ay well have affected the behavioral data. For example, tw o of Gibson's subjects reported not knowing the word papalote ('kite'). A lthough this variation am ong subjects is merely lexical, 24 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. there is probably reason to be concerned about possible attachm ent preferences between subjects; yet dialect was not included as a factor in Gibson et al.'s analyses. 4. The competition model Bates and MacWhinney's competition m odel assumes that cues such as inflectional morphology and word order are m ade available by natural languages and that these cues compete during comprehension in such a way that the m ost valid cue and the least costly is the one that is attended to and m ade use of by comprehenders (Bates and MacWhinney, 1987; 1989; Bates, Friederid & Wulfeck, 1987, Bates & Devescovi, 1989). The competition model derives its appeal from the fact that the researchers involved in this approach have studied a num ber of languages, induding Chinese, Dutch, English, French, Germ an, Hebrew, Hungarian, Italian, Japanese, Serbo-Croatian, Spanish, Turkish and W arlpiri (see artides in the MacWhinney & Bates, 1989 volum e). The competition model relies on two major principles. One is cue validity, w hich refers to the degree of informativeness that a given linguistic cue carries. Three components of cue validity are distinguished: availability, reliability and conflict validity. The availability of one cue may be high in one language, but low in another. For instance, the cue of preverbal position is highly available in English because the verb in dedarative sentences is always preceded by its subject, but the availability of this cue is low in Italian because the subject is often dropped. The reliability of a cue determines the extent to which speakers can use the information it provides for processing sentences. 25 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. For example, the cue of preverbal position is very reliable in English, but it is not in Italian, because the absence of a preverbal element could signal several potential word orders. Conflict validity is a m easure of the validity of a cue w hen it is in competition w ith others. According to Bates and MacW hinney, this is a relatively infrequent' situation, which is handled by calculations of the validity of a cue w hen in conflict with others. The conflict validity estim ate of a cue is the num ber of competition situations in which that cue leads to a correct interpretation, divided by the num ber of competition situations in which that cue participates. These calculations of cue validity are done by reference to the frequency with which a given cue leads to correct interpretations. The second major principle is cue strength. This is a "quintessencially connectionist notion" because it refers to the probability attached to alternative interpretations of stim uli (Bates & MacWhinney, 1989, p.42). These probabilities have values w ith respect to the goal or m eaning w ith which they are associated. The stronger a cue, the more likely it is to be used during processing. Cue strength is intimately related to cue validity in that, ideally, the value of cue strength converges on the value of cue validity, because stronger cues should also be more informative. The order of strength of different cues should therefore reflect cue védidity estimates. This prediction is borne out in English, for example. In this language, the cue of pre-verbal position is stronger than that of animacy in determining the agent role in a sentence and this order in cue strength corresponds to the order in cue validity, with word order being more valid than animacy in English (McDonald and MacWhinney, 1989). Another example comes from comparisons of languages w here the same cues bear different degrees of 26 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. strength. Thus, given a sentence such as Peter eats cake, English speakers w ould pick Peter as the agent of the verb while Italian speakers w ould pick cake as the agent. Word order is indeed a stronger cue in English because this language has a more rigid w ord order than Italian. W ord order is also a more valid cue in English than in Italian because w ord order information is m ore reliable in English than in Italian. Thus, cue strength and cue validity are ultim ately related.^ The notion of competition between cues is very similar to the parallelism that is the basis of constraint-based lexicalist models such as the one described in section 6 below. The competition m odel is a lexicalist theory, because "it attempts to account for both lexical and grammatical phenom ena in terms of a single "dictionary" of content words and grammatical m orphem es, annotated to include information about elements that typically occur to the left and right of each item" (Bates & Wulfeck, 1989, p. 334). It is also a probabilistic theory, in which alternative possibilities compete for activation. The model leaves open the possibilities that cognitive costs are involved in acquiring or using new mappings, that a cue may be left unused because it is not perceivable, and that memory load differences will adjudicate in cases in which two cues have comparable validity levels. However, the model can be criticized on three fronts. 4. 1. W hat constitutes a cue? According to Gibson (1992), Bates and MacW hinney do not explicitly define w hat constitutes a cue in their model. Much of the work done w ithin ^ There are domains in which they do not correlate as closely. In language acquisition, the first cues acquired are often the lüghest in validity, but when others are acquired their strength values begin to reflect their validity in cases where they compete (McDonald, 1986). 27 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. the com petition Framework has concentrated on three major cues: word order, m orphological marking (including agreem ent and verb morphology) and animacy. W hile it is true that these are im portant features of sentences which m ake significant contributions to the understanding of sentence m eaning. Bates and MacWhinney do not m ake it clear how other irrelevant properties of the input stream (such as w hether a w ord contains a sibilant sound) m ay not qualify as possible cues. Thus, Gibson sees that a fundam ental problem with this model is that the notion of linguistic cues has rem ained largely unspecified. 4. 2. How are ambiguities handled? The competition model has explored issues that are different from those traditionally investigated in sentence com prehension research, hence it is som ew hat difficult to compare it to other models of sentence comprehension. This is especially evident in questions relating to ambiguity resolution, which has been the focus of sentence processing research for the past decade. W hen the notion of cues is taken to apply to cases of ambiguity, their adequacy appears to be challenged. Take the M V/RR ambiguity in English, in a sentence such as (34). (34) The judge sentenced for the bribery case was taken to jail. (35) The defendant sentenced for the bribery case was taken to jail. In English, Bates and MacWhinney claim, the cue of preverbal position is reliable because w ord order is fairly constrained by comparison with other languages. Thus the judge should be interpreted as the subject of the verb sentenced. This, of course, is not the correct interpretation for this sentence, 28 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. and a garden-path should result. The problem is that the m odel w ould not distinguish this case from one such as (35), which does not yield the same garden-path efrects. Both of these examples would be liable to the same analysis, and predicted to be handled in the same way by the competition m odel. Notice, however, th at plausibility considerations can easily be incorporated in this m odel because of its use of probabilistic measures. 4. 3. Methodological problems The studies conducted by Bates and MacWhinney and their colleagues suffer from two methodological limitations. First, although m any different languages were investigated in their studies (e.g. Bates and Devescovi, 1989; Bates, Friederid and Wulfeck, 1987; Bates, Marchman, Harris, Wulfeck and Kritchevsky, 1995; Kail, 1989), few utilized on-line paradigm s. This indicates that their results may refer to post-comprehension processes rather than real time processing. A second methodological problem for the competition model is that the studies typically deal with the use of cues in simple active sentences (see, for example, Kilbom, 1994). It is of course commendable that studies of such simple sentences have been conducted on m any different languages, but the competition model would be tested m ore adequately if m ore complex sentences are investigated. 5. Crain, A ltm ann and Steedm an's M odel Crain, Altmann and Steedman's model of discourse preferences is a parallel m odel of sentence comprehension, in the sense that the processor reviews several alternatives which are then evaluated by other components of the m odel (Crain & Steedman, 1985; Altmann & Steedman, 1988; Steedm an & Altmann, 1989). But the model favors a weakly interactive view 29 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. of processing in which semantics helps select alternatives offered by syntax, so that semantics is not allowed to inform the processor's choice from the start. It is also w orth noticing that the model depends on fine-grained distinctions, which help it m ake choices based on the features of each w ord of the sentence encountered. One interesting feature of this m odel is that it relies fairly heavily on the pragm atic notion of referential support. For example, Altmann and Steedm an (1988) argue for a principle of referential support, which states that "an NP analysis which is referentially supported will be favored over one that is not". They point out th at in experimental situations, sentences typically appear w ithout a preceding context that establishes reference, and that, in these cases, expressions still carry their presuppositions, so that comprehenders would tend to rely more on a principle of parsimony, which recommends that "a reading which carries fewer unsupported presuppositions will be favored over one that carries m ore". Crain, A ltm ann and Steedman found support for their model by showing that non-m inim ally attached elements yielded response times that were not any longer than minim ally attached targets, given appropriate discourse/referential contexts (Crain & Steedman, 1985). The role of context in inducing these kinds of results is made even clearer in a study by Altmann and Steedman (1988). They found that, provided the context is referentially supportive, non-minimally attached targets can yield faster RTs than minimally attached ones, a finding which stands in sharp contrast to those of Frazier and her colleagues. Thus with sentences such as (36) and (37), the principle of Minimal Attachm ent predicts attachm ent to the VP regardless of referential context. 30 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (36) The burglar blew open the safe w ith the dynamite. (37) The burglar blew open the safe w ith the lock. Altmann and Steedm an found that in contexts which established the existence of two safes, the NP attachm ent was preferred, while in a context which established that there was only one safe, the VP attachm ent was preferred. These results provide clear support for the idea that context can constrain sentence interpretation. However, Altm ann and Steedm an also found that RTs for NP-attached targets (with the lock) w ere faster than those for the VP-attached targets (with the dynamite) in respectively supporting contexts. This result is surprising because context w ould be expected to have the same quantitative effects on both attachments. Their explanation for this is that comprehenders resort to different contexts for each type of attachment. The discourse context for the VP-attached cases is w orld knowledge (knowing that dynamite can blow up safes) while that necessary for assessing the NP-attached cases is discourse knowledge (which is provided by the sentence), hence faster response times for the latter. It seems plausible that the accessing of different kinds of contexts for the resolution of ambiguities involves different, and possibly disparate, processing costs. 5. 1. The notion of context As should be clear from the previous discussion, the Crain, Altmann and Steedman model relies quite heavily on the notion of context. However, context can be defined in many ways. In this model, context is used to refer to 31 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. the larger discourse constraints which com prehenders make use of in routine processing. For example, in the experim ent w here latencies on PP-attachm ent ambiguities were collected, the supporting and unsupporting contexts consisted of several utterances, as show n below: N P-Supporting context (38) A burglar broke into a bank carrying some dynamite. He planned to blow open a safe. Once inside he saw that there was a safe with a new lock and a safe w ith an old lock. The burglar blew open the safe with the new lock and m ade off with the loot. U nsupporting context (39) A burglar broke into a bank carrying some dynamite. He planned to blow open a safe. Once inside he saw that there was a safe with a new lock and a strongbox with an old lock. The burglar blew open the safe with the dynamite and made off with the loot. Thus, the Crain et al. model relies on discourse context of the kind that w ould provide a universe of discourse in which it is possible to establish precise referents and disambiguate sentences. However, the notion of context can be (and has been) taken quite differently. According to MacDonald et al. (1994), for example, each word makes its ow n lim ited contribution to the interpretation of the sentence it appears in. Thus while Crain et al. are probably correct in assuming that discourse context is constraining, it is also possible that the individual lexical items that appear with ambiguous phrases can help favor one interpretation over the other. These individual lexical preferences will tend to play a bigger role w hen discourse contexts fail to provide specific referents. To deurify this point, it m ight be helpful to think of discourse context as a continuum. Some contexts are strongly constraining, such as the safe examples above. Others are weakly constraining, while still 32 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. others are not constraining at all. Thus, a m odel which uses context to refer only to larger, discourse constraints m ight be ignoring local semantic constraints. For example, the principles of plausibility and parsimony do not appear to explain the difference between the ways in which the following ambiguities are handled; Strongly constraining (40) There was an earthquake last year, and its epicenter was by the small research hospital. Two patients/doctors had to be evacuated. One of the patients/doctors was sent to the m etropolitan hospital and the other was taken to a nearby medical center. The patient/doctor treated at the medical center survived. Weakly constraining (41) The new s cam e as a good surprise. The patientjdoctor treated at the medical center survived. It should be clear that the Crain et al. model can account for the absence of garden-path for both the doctor treated and patient treated examples in strongly constraining contexts. However, in weakly constraining contexts, comprehenders cannot use the principle of referential support (because it refers to larger discourse constraints), and hence m ust rely solely on the principle of parsim ony. Unfortunately, this principle does not distinguish between the patient case and the doctor case in weakly constraining contexts. According to the Crain et al. model, both examples should be comprehended in the same w ay because neither gives rise to m ore referential presuppositions than the other. These examples are not on a par, however, and it should be clear that the patient case does not yield garden-path effects. Thus, in weakly constraining contexts, the processor needs to rely on more 33 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. specific lexical preferences to guide its choice of interpretation and that is not possible in the C rain et al. model. 5. 2. How are ambiguities other than M V/RR ones handled? Kurtzm an (1985) points out that in dem onstrative/sentential ambiguities such as (42) the demonstrative reading implies the existence of a TV set, while the sentential reading does not im ply the existence of a referent. Crain's model w ould therefore favor the sentential reading over the dem onstrative interpretation. (42) That TV is bad for your eyes is obvious However, the dem onstrative reading is the one that is first assum ed for this kind of example. To Crain et al.'s credit, however, one should point out that this is true only of cases that involve nouns that can have both a mass reading (TV as a m edium ) and a count reading (a TV set). Furthermore, the Crain et al. model does predict the right interpretation for sentences such as (43). Still, the principle of referential support does not have a w ay of distinguishing betw een an implausible reading (as in the sentential case) and a plausible reading (as in the demonstrative case). (43) Howard claimed that TV is bad for your eyes. As an alternative explanation for this, consider Juliano and Tanenhaus (1993) work. Using the Penn Treebank database (Marcus, Santorini & Marcinkiewicz, 1993), they have shown that w hen the w ord that is used at the beginning of a sentence, it is almost always a determiner. W hen it follows a 34 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. verb, however, the w ord that is almost always a complementizer. If frequency is a factor in comprehension, one would predict garden-path efrects in examples such as (42) but not in examples such as (43). Such differences are not predicted by the Crain et al. model. 6. The constraint-based m odel The constraint-based approach to language com prehension has grown from the observation that linguistic ambiguities generally involve two or more alternatives that m ay have very different semantic and pragmatic implications. For example, with the classic exam ple the horse raced past the barn fell, the two possible uses of the verb race, transitive (racing the horse) and intransitive (the horse racing) are not equally probable or appropriate in the linguistic context provided. The basic hypothesis put forth in constraint satisfaction approaches is that these differences in probabilities are exploited by people during the process of language comprehension. For example, MacDonald, Pearlm utter and Seidenberg (1994) claim that "languages are structured at m ultiple levels simultaneously, including lexical, phonological, morphological, syntactic and text or discourse levels". Given this kind of structuring, it is argued that comprehenders process language in a constraint- based m anner, m aking use of all available inform ation in comprehending each piece of incom ing stimuli. The constraint-based approach to sentence com prehension relies on significant earlier research. The idea that com prehenders use phonetic, lexical, syntactic and semantic information from the input was proposed by Marslen-Wilson and Tyler (1975). They developed an on-line interactive model which assum es that syntactic processing is not autonom ous, but rather 35 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. interacts w ith semantic processing at every stage of a given sentence. Evidence for this model was provided by a study of noun/adjective ambiguities. The ambiguous fragm ents were presented in biasing contexts, as in if you walk too near the runway, landing planes ... vs. if you’ve been trained as a pilot, landing planes ... in a cross-modal naming task. Tyler and Marlsen-W ilson (1977) found that nam ing latencies were faster for appropriate than for inappropriate continuations, indicating that the parser has access to semantic information from the start. Ford, Bresnan and Kaplan (1982) argued that syntactic disam biguation depended on the strength of lexical alternatives. They proposed a principle of lexical preference which takes advantage of the fact that "the various lexical forms of a given verb have different strengths or saliences and that the strongest form somehow determines the preferred syntactic analysis" (p.745). Ford, Bresnan and Kaplan also suggested that a possible interpretation of the concept of "strength" was frequency. Since the late 1980s, Tanenhaus and his colleagues have also argued that syntactic processing depends on lexical aspects of the input. For instance, Tanenhaus and Carlson (1989) proposed that semantic commitment to an interpretation could be m ade on the basis of thematic role information and that there is rapid interaction betw een syntactic, discourse and real-world knowledge (see also Carlson & Tanenhaus, 1988; Trueswell, Tanenhaus & Gamsey, 1994; Tanenhaus & Trueswell, 1995). Several researchers currently work w ithin a constraint-based approach, but MacDonald et al.'s is so far the most detailed proposal for this type of theory, and we turn to it presently. MacDonald et al. (1994) view com prehension as a process in which several linked representations of linguistic inform ation are derived. They use three levels of interconnected representations: lexical, syntactic, and 36 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. discourse. At the lexical level, various types of linguistic and probabilistic inform ation are accessed. Linguistic information such as tense m orphology (past tense vs. past participle), voice (active vs. passive), verb finiteness (finite vs. infinitive), and argument structure (e g. transitive vs. intransitive) is represented in such a way that all options available for a lexical item (e.g. transitive vs. intransitive for a verb such as sit) carry weights directly representing how frequently these options occur. Units in the lexical entry (e.g. active vs. passive) are linked through connections which can be either fadlitory or inhibitory. Fadlitory connections link types of inform ation that are interdependent, whereas inhibitory connections exist betw een relations that are ruled out by the syntax. For instance, an inhibitory connection w ould exist betw een the passive and the past tense nodes, and between the passive and the active nodes, while an excitatory connection would exist between the passive an d the transitive nodes. The evidence supporting the constraint-based model comes from studies th at have demonstrated the existence of (lexical, syntactic and pragmatic) context effects, lexical frequency effects, and the interaction of the two. Consider die case of the main verb/reduced relative ambiguity. There is now a large body of data suggesting that parsing this kind of ambiguity is dependent on several interacting factors. For example, the semantic characteristics of the noun that precedes the m ain verb can directly affect the processing of reduced relatives. In an eye-movement study, Trueswell, Tanenhaus and Gam sey (1994) found that there were clear signs of difficulty in reading reduced relative clauses only w hen they began w ith animate nouns and that the semantic fit of a noun to potential argum ent positions has an im m ediate effect on the disam biguation of the clauses. Information 37 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. such as temporal context has also been shown to affect the processing of reduced relative clauses (Trueswell & Tanenhaus, 1991). Similarly, M acDonald (1994) varied the argument structure frequencies of verbs in transitive and intransitive frames. She found that the degree of difficulty often experienced w ith ambiguous reduced relative clauses w as a function of the relative frequency w ith which a verb was used transitively: w hether it w as preferably transitive (e.g. the rancher knew that the nervous cattle pushed into the crowded pen...) or preferably intransitive (e.g. the rancher knew that the nervous cattle moved into the crowded pen...). O ther evidence for the constraint-based model has been provided by w ork on N P /S ambiguities (Trueswell, Tanenhaus & Kello, 1993), subject/object attachm ent preference (Juliano & Tanenhaus, 1994), PP-attachment ambiguities (Thornton, Gil & MacDonald, 1996) verb control structures (Boland, Tanenhaus & Gam sey, 1990) and implicit argum ent phenom ena (Mauner, Tanenhaus & Carlson, 1995). The constraint-based m odel suffers from several shortcomings, however, to which we turn to in the next section. 6.1. Late closure ambiguities. One im portant criticism that can be leveled against the constraint- based model is that it cannot account for late closure phenom ena (Frazier, 1994). These include PP-attachment ambiguities such as (44) and (45). (44) The student bought a book with lots of illustrations. (45) The student bought a book with his allowance money. The basic argum ent is that since there is no explicit representation of structure, there cannot be any explanation for how PPs w ould attach to 38 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. different NPs that appear earlier on in the sentence. Proponents of the constraint-based model have attem pted to address this criticism by show ing that certain lexical and discourse constraints can influence the processing of these types of ambiguities. For example, definiteness (arguably a discourse variable) seems to come into play in the sense that definite NPs typically require the PP to attach to the VP. Furthermore, it seems that the lexical nature of the VP (whether it is actional or perceptual) interacts with definiteness in predicting attachm ent. Spivey-Knowlton and Sedivy (1995) found that, when the main verb is actional and the object N P is definite in sentences such as (46), subjects will prefer the PP to m odify the VP. W hen the m ain verb is perceptual and the object NP is definite as in (47) subjects will prefer the PP to modify the VP. W hen the main verb is perceptual and the object NP is indefinite as in (48), then the PP modifies the NP. (46) The fireman smashed dow n the door with the heavy ax/w ith the new lock. (47) The salesman glanced at the customer with suspicion/w ith ripped jeans. (48) The salesman glanced at a customer with ripped jeans/w ith suspicion. However, to the extent that there are generalizations that can only be expressed by reference to structure (i.e. that PPs will generally attach to the m ost recent site in English, and to the primary site in Spanish), the constraint-based model’ s account of these phenomena appears insufficient. 39 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 6. 2. The role of non-lexical constraints According to MacDonald et al. (1994), the lexical entry of a given item contains m uch (perhaps even all) of the information that is needed for processing that item. It is argued that accessing lexical information also involves accessing syntactic representations which consist of linked X- structures. It is also claimed that contextual information can interact with information relating to the relative frequencies of use of w ords in discourse contexts.^ The problem is that although MacDonald et al. postulate the existence of a syntactic and discourse levels, little is said about them. It is unclear how these levels are represented and how they interact. In the case of syntactic factors, for instance, the m odel relies on m any notions that were first identified by syntactitians (e.g. thematic roles, argum ent structures) and uses syntactic notions in order to express generalizations at the lexical level,^ but MacDonald et al. have not taken a stance as to w hat the exact role of syntactic structure is in their model. 6. 3. Long distance dependencies A frequent criticism of the constraint-based model is that it does not specify how lexical items are p u t together into a well-formed structure. The argum ent is that this makes the constraint satisfaction m odel unable to account for late clause (or N P/S) ambiguity because, although there exists a structural representation for lexical items, there is no representation of the ^ Contextual information can affect the choices that comprehenders make Ijetween alternative interpretations of sentences or alternative structures for expressing thoughts. Context includes such information as whether a verb (e.g. released) that is ambiguous between the main verb or reduced relative meaning is preceded by an auxiliary (e.g. was or had), whether the subject NP is animate or inanimate etc. ^ Research within linguistics itself seems to suggest that lexical entries of individual items have to be rich enough to include information about the syntactic structures that items can appear in (see Levin, 1993, and Levin & Pinker, 1991 for discussions). 40 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. structural relation between the two clauses that make up am biguous sentences such as (49). (49) While M ary was mending the sock fell off her lap. For example, Frazier believes that a global syntactic representation is necessary for disambiguating the sentence (Frazier, 1995). On this view, since the constraint satisfaction model does not have a representation for general syntactic structure, it should not be able to explain effects that rely on global structure. Indeed, MacDonald et al. (1994) never make it explicit how X’ structures are p u t together in a well-formed structure. 7. Summary and conclusions Six models of sentence com prehension have been reviewed and discussed. Frazier's garden-path m odel, while capable of handling various types of argum ent and non-argum ent ambiguities, runs into problem s when dealing w ith ambiguities that involve lexical items which exhibit preferences for one disam biguation over another. Mitchell's linguistic tuning hypothesis, although serial, does allow frequency to play a role in sentence comprehension; however, this m odel only refers to structural frequencies, and not to contingent frequencies, so it appears to be inadequate for handling cases in which the frequencies of individual lexical items constrain the com prehension of the sentences they appear in. Gibson's m odel, on the other hand, relies on both structural principles and probabilistic notions of cost, but runs into problems for languages other than English and Spanish. Bates and MacW hinney's model is also parallel in nature, and refers to the validity and strength of cues provided by utterances. However, more detailed proposals 41 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. are needed on how ambiguities are handled w ithin this fram ework. Crain, Altmann and Steedman's m odel is able to accurately characterize discourse context effects but depends almost exclusively on the notion of referential support, and leaves little space for other types of information to play a role. The MacDonald m odel is a constraint satisfaction m odel which takes advantage of the statistical properties of language, b u t falls short of providing a detailed account of how phrases are combined into bigger structures or of generalizations that hold over structures rather than lexical items. As the discussion in this chapter makes clear, each of these models has advantages and disadvantages and each has provided its own experimental evidence. Consider Table 1 below, in which the garden-path m odel and the constraint-based m odel are compared (the comparison was restricted to these two models because they best represent the contrast between the serial and parallel approaches). A cursory look at this table shows that nearly all the evidence given in favor of cxirrent comprehension m odels has to do with ambiguity resolution. W hile it is true that all sentences are am biguous at one level or another, the kinds of ambiguities that have been studied have tended to be of the type that considerably slow down the processing system , for example, by causing garden-path effects. These types of structural ambiguities seem to be the exception rather than the rule in norm al sentence processing, so one might want to look at relatively less ambiguous sentences as a way of testing one's model. In order to better distinguish between the serial and parallel approaches, one m ay w ant to look at new research questions. Since each theory has its ow n account of ambiguity resolution and has provided experimental support for it, it m ight be beneficial to look at relatively unambiguous sentences in the hope of distinguishing these m odels further. 42 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. This question is the first major concern of this dissertation. I investigate relatively unam biguous structures and show that they are processed no difierently than ambiguous sentences. Using the constraint-based approach, I show that the parsing of unambiguous sentences is affected by both structural fi-equency and by contingent frequency. Table 1: Experimental evidence for the garden-path model vs. the constraint- based model. Type of structure The garden-path model The constraint-based model MV/RR Rayner, Carlson & Frazier (1983) Rayner, Garrod & Perfetti (1992) Tanenhaus, Carlson & Trueswell Ü989) Trueswell & Tanenhaus (1991) Spivey-Knowlton, Trueswell & Tanenhaus (1993) Trueswell, Tanenhaus & Gamsey (1994) MacDonald (1994) Pearlmutter & MacDonald (1995) Example: The horse raced past the barn fell. NP/S Example: The student forgot the solution was on the back of the handout Ferreira & Henderson (1993) Clifton (1993) Trueswell, Tanenhaus & Kello (1993) Gamsey, Pearlmutter, Myers & Lotocky (in press) PP-attachment/Relative clause Clifton, Speer & Abney (1991) Rayner, Garrod & Perfetti (1992) Carreiras & Clifton (1993) DeVincenzi & Job (1993) DeVincenzi & Job (1995) Gilboy, Sopena, Qifton & Frazier (1995) Taraban & McClelland (1988) Spivey-Knowlton & Sedivy (1995) attachment Example: The burglar blew open the safe with the new lock The spy saw the cop with the binoculars I really liked the preface of the book that I read yesterday Lexical category ambiguities Rayner & Frazier (1989) Frazier & Rayner (1987) MacDonald (1993) Example: The desert trains young people to be tough Çap-filling Clifton & Frazier (1986) Frazier & Clifton (1989) Frazier & Flores d’ Arcais (1989) Tanenhaus & Carlson (1989) Example: That's the girl that the woman had forced to sing at Christmas. 43 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In investigating unam biguous sentences, I will be w orking w ithin the framework of the constraint-based approach. This choice is m otivated by the fact that, in spite of the problems that have been identified in section 6, this m odel is the m ost prom ising of all those reviewed. In particular, the constraint-based approach can be applied to different fields of cognition. For example, in w ord recognition research, there is a large body of work suggesting that m ultiple constraints are made use of sim ultaneously in reading w ords (Seidenberg & McClelland, 1989; Plaut, McClelland, Seidenberg & Patterson, 1996; Seidenberg, Plaut, Petersen, McClelland, 1994). Research on w ord segmentation has also relied on statistical approaches to language (Morgan & Safiran, 1995). For instance, Saffran, N ew port and Aslin (1996) address the issue of how the infant leam s to distinguish a w ord from a continuous speech stream . They proposed a notion of transitional probability which takes advantage of the fact that word-internal pairs of syllables (e.g. bay-bi) tend to occur m ore frequently in a corpus than w ord external pairs (e.g. o-bay-mi). This m eans that a given syllable is more predictive of a following syllable w hen they belong in the same word than w hen they do not. Saffran, N ew port and Aslin (1996) found that adult subjects exposed to an artificial language in which transitional probabilities w ere the only cues to w ord boundaries learned the words of that language. This result was replicated w ith eight-m onth old infants by Saffran, Aslin and N ew port (1996). Finally, the constraint-based approach is also compatible w ith some recent research on language acquisition suggesting that children m ay be more sensitive to probabilistic patterns in their environment than had previously been assumed (Allen, MacDonald, Seidenberg & Christiansen, 1996). As a sentence comprehension model, the constraint-based approach is certainly valid, and the criticisms which I outlined in section 6 can be countered with 44 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. m ore specific proposals about the representation of (global) syntactic structure. In general, then, the constraint-based model is to be preferred because it allows for a parsim onious theory of cognition. One of the fundam ental claims of the constraint-based m odel is that frequency afiects parsing in significant ways. However, the m odel has been criticized for not explicitly stating which kinds of frequencies should affect processing and how these fiequencies are com puted (Frazier, 1995; Frazier & Clifton, 1996). Work within the constraint-based model has tended to focus on the role of subtle contingent frequency information in ambiguity resolution, but global sorts of frequencies could also be investigated. Furthermore, one could also exam ine the interaction between local and global frequencies in order to help clarify the w ay in which the hum an processor is structured and determ ine the m anner in which these types of probabilistic information are encoded. This is the second major research question of this dissertation. To investigate these two research questions, and since frequency figures prom inently in the constraint-based model, I will start by discussing fi-equency as it constrains sentence comprehension. 45 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER H: FREQUENCY AS A CONSTRAINT ON SENTENCE COMPREHENSION Frequency is an im portant aspect of linguistic stimuli that has often been dem onstrated to influence the way in w hich w e process and understand language. For example, it takes longer to nam e a picture of an object when the object's name is low in frequency than w hen it is high in frequency (Forster & Chambers, 1973). Similarly, studies have show n that the speed and accuracy with which a w ord is recognized or named is a function of its frequency, with low frequency w ords taking longer to identify or pronounce than high frequency words (Broadbent & Gregory, 1968; Rice & Robinson, 1975; Rubenstein, Garfield & Millikan, 1970). However, although frequency effects have long been investigated in word recognition research, they have only recently began to draw the interest of sentence processing researchers (MacDonald, 1994; Juliano & Tanenhaus 1993). These researchers have reasoned that frequency m ight affect phenom ena such as syntactic ambiguity resolution because it could alter the level of activation of syntactic alternatives and hence determine the choice of interpretation of a structurally ambiguous sentence. For instance, MacDonald (1994, MacDonald, Pearlm utter & Seidenberg, 1994) showed that the frequency w ith which a verb is used in transitive or intransitive frames, the frequency w ith which a verb appears in the active or passive voice, and the frequency of a verb as a past tense or a past participle are all relevant in disambiguating M V /RR ambiguities (e.g. the horse raced past the barn fell). Furthermore, Juliano and Tanenhaus (1993) found that com prehenders are sensitive to the frequency with which the complementizer that appears in a sentence-initial compared to a post-verbal 46 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. position and m ake use of that inform ation in comprehending sentence- initial vs. post verbal that clauses. Other contingent frequency effects were identified by MacDonald (1993) for noun-verb ambiguities such as fires and judges. She found that the relative fi-equency of a word as a head or as a m odifying noun together with semantic context (supportive or unsupportive of disambiguation) constrain the way in w hich sentences containing hom ophone ambiguities are comprehended. In a reading time study, Trueswell, Tanenhaus and Kello (1993) investigated N P/S ambiguities such as the student forgot the solution was printed on the back of the handout. They found that syntactic misanalysis effects appeared at the em bedded verb when the matrix verb was more frequently used w ith noun-phrase than with sentence complements. Stowe, Tanenhaus and Carlson (1991) found that whether a wh-phrase could act as a plausible complement of a verb depended on the relative frequency of use of that verb in a transitive frame. The pattern in all of these studies is such that frequency helps activate certain alternatives, and w hen these alternatives are consistent with context, comprehension is greatly facilitated; if these alternatives are inconsistent with context, comprehension is severely taxed. Constraint-based models have thus sought to identify the effects of frequency and their interaction with other factors such as context. The results of these endeavors can then be used to address issues of representation. For instance, a processing system that takes frequency into account w ould have to be structured in such a way that it could keep track of frequency differences, and that it could serve for both word and sentence processing. Thus it is typically assum ed in constraint-based m odels that linguistic information is distributed over representational units, and that the level of activation of 47 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. these units is altered with exposure to linguistic stimuli. Although the constraint-based approach is prom ising, it has been severely criticized. The most im portant of these criticisms is discussed in the next section. 1. The issue of circularity One of the m ost serious criticisms of constraint-based models is that they are circular. Current research w ould seem to suggest that the frequency of individual items, the frequency of structures, and the frequency of items within structures can cause certain patterns of behavior to appear. How ever, by acknowledging the existence of these effects, one is faced with the following problem: frequency cannot reasonably be assumed to be pre-existent in the hum an processing system. Frequency m ust itself be created by (the conjunction of) other factors. One crucial aspect of constraint satisfaction models is that they provide a way of encoding the frequency differences which cause effects of greater accuracy and speed of processing to appear. In the constraint-based model of M acDonald et al. (1994), for example, the weights of w ords are strengthened or weakened through exposure to linguistic stimuli. In this manner, frequency differences between individual lexical items start building up over time. Beginning in childhood and continuing through the life span, linguistic stimuli are processed, and weights are accordingly altered, to reflect the frequency with which w ords and structures are heard. Each time a w ord or structure is encountered, the processing system adjusts its weights, and so more frequent words and structures come to have stronger weights than less frequent words and structures. Therefore, frequency encoding is a consequence of every day processing: it is done systematically and involuntarily. In sum, frequency is both a CAUSE (in the sense that it is responsible for faster latencies on higher 48 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. frequency w ords and structures) and a CONSEQUENCE (in the sense that it is a by-product of processing words and sentences). This fact m ay have given the impression th at constraint satisfaction m odels are circular. This is an extremely serious criticism, because, if true, it would m ean that the constraint-based m odel is unfalsifîable, and hence that it no longer qualifies as a possible theory of processing.^ However, the impression that the constraint-based m odel is circular is probably due to an imderspedfication of the w ay in which fi"equency difierences come to exist, how they are represented, and how they influence comprehension. In w hat follows, I investigate this issue by taking actives and passives as a sam ple case. The idea is to explore existing frequency differences (section 2), exam ine their possible sources (section 3), the way in w hich they become represented (section 4), and the w ay in which they affect comprehension (section 5). I will then present the fundamental research question of this thesis and show why it is im portant to sentence com prehension research. 2. Frequency an d voice 2.1. Structural frequency In English, as in many Indo-European languages, the active voice is canonical (Givôn, 1979). The most plausible explanation for this fact is that sentences tend to be more about agents than about patients. In English, agents tend to occupy the TOPIC rather than the COMMENT position so that m ost sentences will be in the active voice. The reverse situation, one in w hich the ^ Here, I am assuming that falsifiability is a condition on theories, not just a desirable aspect of models. 49 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. patient is the topic, occurs less frequently. Givôn (1979) found that passive structures occur, on average, 20% of the time. Furthermore, the use of the passive voice seems to be a function of the type of discourse adopted by speakers. Thus, in Svartvik's (1966) scientific texts, the percentage of passive sentences was as high as 32% while in Givôn's (1979) sports reports, arguably an informal register, the average frequency of passives was 4%. Evidence from other available frequency counts supports the idea that active sentences are more frequent than passive sentences.^ A search for the frequency of appearance of verbs in the active and passive voice was perform ed using the Perm Treebank database (Marcus, Santorini & Marcinkiewicz, 1993). This corpus consists of one million words from the Wall Street Journal, whose syntactic structure has been tagged and hand- checked. The search showed that 1235 verbs appeared in the active voice, while only 654 appeared in the passive voice.® The total num ber of tokens was 23215 for actives and 2275 for passives. This means that 91% of the sentences in the corpus w ere in the active, while approximately 9% appeared in the passive. Finally, within passive structures, "agentless" passives (The patient was told that she needed to have surgery) are m ore frequent than "agented" ones (The patient was told by the doctor that she needed to have surgery). In Thompson's (1987: 498) corpus from radio news broadcasts, 82% of all passives were agentless. This tendency was also supported in the Penn ^ Data from the Penn Treebank database are compared with traditional frequency counts in this section, in section 12., and in Chapter IV, section 1, as a means of assessing the validity of relying on this database in developing materials for on-line experiments. ® The frequency of active verbs ranged from 1 to 5764, while the frequency of passive verbs ranged from 1 to 138. 50 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Treebank corpus. Indeed, 81% of passive sentences in the Penn Treebank corpus were agentless, while 19% were agented. These findings suggest that active and passive sentences follow a certain frequency hierarchy, schematized in Figure 1 below. Frequency Hierarchy in English Voice Passives Actives Agented Agentless Figure 1 This kind of hierarchy, w ith actives more frequent than passives, does not necessarily hold cross-linguistically. If frequency differences between actives and passives were random , one would expect to find three groups of languages: (1) languages in which the passive is less frequent, (2) those in which the passive and active are equally frequent, and (3) those in which the 51 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. passive is m ore frequent.^ Most languages that have been studied so far seem to belong in the first group. For example, Samoan does not have a passive construction at all (Siewierska, 1984). Furthermore, w hen the passive construction does exist, it is often the case that the agentive noun, which in English appears in the by-phrase, is left unexpressed. This is the case for Classical Arabic, Classical Hebrew, H ungarian and Latvian. Other languages shift attention aw ay from the agentive phrase by disallow ing object agreement in the passive, as is the case for Basque an d Fijian (Keenan, 1987). In the second group of languages, where actives and passives are approximately equally frequent, one finds Malagasy (Keenan, personal communication). Furtherm ore, the frequency of the passive voice seems to vary with the tense used, so that passives are slightly m ore frequent in the past tense while actives are more frequent in the present and future tense (Polinsky, personal communication). The third group of languages seems to be much m ore restricted. To my knowledge, there are no languages in which the passive is overall more frequent than the active.^° In general, then, actives seem to be m uch more frequent than passives in a num ber of hum an languages, as is the case in English. This type of frequency difference is structural: it holds at the sentential rather than at the lexical level. However, there is at least one other type of frequency inform ation which can influence processing: contingent frequency. This refers to the frequency of individual lexical items w ithin given structures. ^ Ergative languages are not included in this discussion. In ergative languages (e.g. Eskimo) the theme of transitive verbs and the agent of intransitive verbs display similar characteristics. There are a few languages in which the passive is very frequent (Cebuano and other Philippine languages) but it is unclear if the passive is in fact the canonical voice. 52 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 2. 2. Contingent frequency The generalizations about relative frequencies of structures outlined above do not make any predictions about the process of com prehension at specific sentence regions. The type of verb used in one or the other voice may also be a factor in their processing. While the passive voice is used, on average, 20% of the time (Givôn, 1979), not all English verbs are equally represented in that percentage. Some verbs are quite frequently used in the passive.iï Others are rarely ever used in the passive. For the purposes of this study, these two sets of verbs will be designated passive-bias and active-bias, respectively. Passive-bias verbs are defined as verbs that are m ore frequently used in the passive voice than in the active. For example, appointed and sentenced are used more frequently in the passive voice than verbs such as phoned and resisted.^^ Active-bias verbs are defined as verbs that are almost never used in the passive voice. For example, slapped and wanted are rarely used in the passive, com pared w ith verbs such as elected and released. These differences are schematized in Figure 2 below. In fact, there exist verbs (e.g. rumor, reincarnate, repute) that can only be used in the passive. (See Levin, 1993). These frequency data are taken from the Penn Treebank corpus. The proportion of passive voice use for the verbs cited in this paragraph is: appointed: 78% and sentenced: 88% vs. phoned: 0% and resisted: 9%. Also, slapped: 0% and wanted: 1% vs. elected: 77% and released: 60%. 53 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Verb Types: Active-bias vs. Passive-bias Act Act Appointed Resisted Pas Pas Passive-bias Verb Active-bias Verb Figurez To test the idea that English verbs can be classified into active-bias and passive-bias groups, the frequency of use of verbs in the active and passive voice was compared, again using the Penn Treebank corpus. However, there was some difficulty in comparing these two sets in that not all the verbs that appeared in the passive corpus were present in the active corpus, and vice- versa. To resolve this problem, a strategy was adopted such that w hen a verb appeared in the passive but not in the active (which was the case for 218 verbs), it was given an active frequency of 0. Conversely, if a verb appeared in the active but not in the passive (which was the case for 799 verbs), it was given a passive frequency of 0. In this manner, the two groups, now totaling 1453 verbs each, were amenable to statistical analysis.A correlation analysis Here, I am referring to verbs as types, not as tokens. 54 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. was perform ed, in which the overall (active + passive) frequency of use of verbs was correlated with a verb’s relative frequency of use in the passive. This was m otivated by the need to determine whether passive-bias verbs were sim ply verbs that were very frequent, and so had a higher frequency in the passive voice, or whether the frequency of use of a verb in the passive was som ewhat independent of its overall frequency. The two factors did not correlate (r = -.044), a result w hich demonstrates that overall frequency does not necessarily relate to the frequency of use of a verb in the passive voice. For example, the verb added has an overall frequency of 579 in the Penn Treebank, w ith an active frequency of 551 and a passive frequency of 28, while the verb expected has an overall frequency of 43, with an active frequency of 19 and a passive frequency of 24. This confirms the view that some verbs are more frequently used in the passive and that this is not necessarily related to their overall frequency. To sum m arize, while actives are far m ore frequent than passives in English, there exist verbs which are m ore often used in the passive than in the active. The form er tendency is structural while the second is contingent. The question that now arises is how these voice frequencies come about. The following section deals with this issue. 3. The sources of frequency In order to understand how frequency differences come about, we will shift attention from comprehension to production. For frequency differences to arise, some words or structures have to be encountered more frequently than others. Accordingly, it is necessary to understand mechanisms of production that w ould increase or decrease the frequency of certain words or structures. An im portant point m ust be made about production, however. It is a very different process than comprehension. Sentence comprehension 55 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. occurs 'on-line'/ with the linguistic signal being interpreted as soon as it is perceived. This means that comprehenders process sentences by m aking certain hypotheses, which m ay or may not turn out to be the correct ones. By contrast/ in production/ the speaker is fully aw are of the content of the message that is being produced. Choices that face the speaker have to do with the kind of syntactic structure to use for communicating the message and the kinds of lexical items to use w ith that structure. The discourse situations in which the speaker finds herself often constrain the choices that she m akes in term s of the structures that she uses. If the sam e situations end up occurring m ore often than others because of extraneous factors, then one w ould expect the speaker to produce certain structures m ore often than others. To illustrate this point, the choice betw een active and passive voice in sentence production is examined, w ith a hope of show ing how actives can become m ore frequent than passives. The basic question that arises is this: how do actives come to be more ft’ equent than passives? This is taken as a sam ple case of the larger question that faces the constraint-based model: how do frequencies of argum ent structures (word co-occurrences, verb frames etc..) come about? There is no sim ple answer to these sorts of questions, as it is probably a conjunction of semantic and pragmatic factors that conspire in m aking words or structures m ore frequent than others.^'* However, by investigating the case of actives The terms pragmatics and semantics are taken in their conservative senses here. In current linguistic research, both semantidsts and pragmatidsts concern themselves with explaining how a sequence of words is used to convey a linguistic message. However, while semantics deals witt\ the more restricted, literal meaning of utterances, pragmatics deals with the way in which these utterances can be used by speakers to convey different sorts of information. The classic example that is given to illustrate this is the sentence it's cold in here, which can be used to provide information but can also be used as an indirect request to turn on the heater. Thus, when utterances are used in different contexts, their literal meanings are interpreted against a background of world knowledge and constraints. In this dissertation, 1 will adopt 56 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. and passives in English, it is hoped that the relation between sources of frequency differences and the comprehension of sentences will become clearer. In w hat follows, several pragmatic and semantic constraints are argued to influence the use of actives and passives in sentence p r o d u c t i o n . ^ ^ 3.1. Structural Frequency 3.1.1. Pragmatic Constraints 3.1.1.1. N atural salience Discourse interests usually focus on agents and experiencers— which are typically hum an— rather than on patients, goals or locations— which usually are not (Givôn, 1979). This is probably due to the fact that hum an beings are naturally m ore interested in other human beings than they are in inanim ate objects (Berlin, Boster & O ’ Neill, 1981). Animate entities are capable of mobility and interaction; they can be threatening or comforting; they can create and communicate. Inanim ate objects cannot. The interest that hum an beings m anifest towards other hum an beings is illustrated by the fact that the hum an visual system leam s to attend to animate entities very early on. For example, infants seem to attend m ore to moving than to static objects in their visual field (Hartlep, 1983; Johnson, Dziurawiec, Ellis & Morton, 1991). It is also a common finding in face perception research that people attend m ore to hum an faces than to other objects (Yu & Blake, 1992). Given that hum ans focus their attention on anim ate objects m ore than they do on inanim ate objects, it seems plausible that the sentences they use to this conservative definition of the terms pragmatics and semantics. Pragmatics will l>e taken to refer to the larger discourse situations and world knowledge which are distinct from the literal meanings of lexical items. Although some of the constraints discussed below apply to many languages, the discussion is specific to English. 57 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. communicate would reflect that difference in saliency.^^ The active word order does just this: it places the agent in a prom inent, earlier position in the sentence, and relegates the patient to the end of the clause. Thus the active word order fits m ost discourse situations in which a speaker is likely to find herself in, hence its higher frequency of use by comparison with the passive word order. The less frequent voice, the passive, seems to depend for its use on very specific discourse constraints. Typically, w hen speakers need to focus on the patient rather than on the agent, the passive w ord order is preferred.^ ^ 3.1.1.2. Length Speakers typically tailor their speech to suit the communicative needs of their audience. In most situations, however, unnecessarily long sentences may not be welcomed. Grice (1975) expressed this generalization in his maxim of M anner "Be brief. The preference for shorter over longer items is also particularly clear for individual words. Zipf (1935) showed that the longer a w ord is in length, the less likely it is to be used. This tendency generalized to English, German, Chinese and Latin, which led Zipf to argue that the relationship between high frequency and shortness in length was valid for language in general. Although the length of passives differs by only two w ords from that of actives, this difference should become more im portant w hen other things are equal, i.e. in situations in which topicalizing the agent or the theme is equally felicitous. In those cases, the shorter structure should presumably be preferred. Arguably, length is a weak This point seems to be supported by the typological facts cited earlier. It has been shown that actives are more frequent than passives in manylanguages and that they are equally frequent in Malagasy. There is yet no evidence of languages in which the passive is more frequent than passives. Notice, however, that natural salience is only one factor and that it may be overruled in other languages by the conjunction of other semantic and pragmatic factors. For experimental evidence of this, see Chapter III, section 2. 58 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. factor, but it is one whose strength depends on other variables that m ay be involved (see McDonald, Bock and Kelly, 1993). 3. 1.2. Semantic constraints 3. 1.2.1 Thematic structure The argum ent structure of predicates can be an im portant factor in the choice of w ord order and hence the choice of active vs. passive structure. It is relevant to note here that m ost English verbs place the agentive elem ent in subject position. For example, w ith agent-theme verbs such as hit and drink, the anim ate entity appears first. This is also true of experiencer-theme verbs such as fear and feel, benefactive-theme verbs such as receive and lose and theme-location verbs such as live and come. Thus the thematic structure of m ost predicates in English is such that the active w ord order is m ost appropriate. The pragmatic and semantic factors outlined above conspire to m ake actives more frequent than passives in English, a tendency which is found in m any other languages. Notice that, because m ost sentences in English are actives, it follows that m ost verbs are used in the active voice m ore often than they are used in the passive voice. Thus, structural frequency and contingent frequency coincide for most predicates. However, there are verbs that are m ore likely to be passivized than others given appropriate circumstances. The following section reviews some of the semantic and pragm atic factors that might m ake passive-bias verbs more frequent in the passive than other verbs. 59 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3.2. Contingent Frequency 3.2.1. Pragmatic constraints 3.2.1.1. N ature of the agent The agent of a verb can have a variety of features. It can be known to the speaker, in which case it can be m ade explicit or left implicit, or it can be unknown to the speaker. W hen the agent is entirely unknow n (the speaker doesn’t know who perform ed the action as in M ariam 's car was stolen); implicit in the context used (The breeder had several Dalmatian puppies. They were sold this morning); or dedudble from w orld know ledge (The criminal was sentenced to life in prison), then the passive voice is preferred. This is due to the fact that the passive voice allows the agent to remain unexpressed whereas the active voice does not. In English, an agent can be left unexpressed in passive sentences (The puppies were sold) but not in active sentences (*Sold the p u p p ie s ).Givôn (1979) found that in a 1974 English text^9, the agent of short (agentless) passives w as dedudble from pragmatic knowledge in 46% of the cases and from discourse context in 31% of the cases (i.e. a total of 77% of cases).^° Notice that leaving out the agent does not increase the processing load for the comprehender. Thus, Manner, Tanenhaus and Carlson (1995) investigated the idea that semantic information assodated w ith certain unexpressed argum ents of verbs is routinely encoded as p art of understanding a sentence. They found that it is no m ore difficult to process a verbal passive when the agent is present (e.g. the game show’ s wheel was spun by the contestant to win a prize and lots of cash) than w hen it is not (e.g. the game show's wheel was spun to win a prize The agent can be left unexpressed in the active voice in PRO-drop languages. The text is Trout, 1974, pp 7-47 The remaining passives were lexical passives (23%), as in Mike was disappointed. In these cases, it is not at all clear that the action had an agent. 60 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. and lots of cash/the contestant spun the game show's wheel to win a prize and lots of cash), which indicates that implicit arguments m ay not be any m ore difficult to encode than explicit arguments. 3.2.2. Semantic constraints 3.2.2.I. Thematic structure As was argued above, the thematic structure of m ost verbs reflects the general tendency that actives are more frequent than passives. However, there are verbs (e.g. theme-experiencer verbs such as frighten, amaze, challenge and surprise) that do not follow this tendency. W hen theme- experiencer verbs are used in the active voice, the theme is placed in topic position, and this m ay not be felicitous, because in most discourse situations the agent is topicalized. Therefore, when the thematic structure of the verb is such that the theme appears first, the passive voice is usually preferred. This is corroborated in a study by Ferreira (1994), who has shown that significantly m ore passives are produced with theme-experiencer verbs than w ith agent- them e or experiencer-theme verbs. She contends that this is due to speakers' compliance with a them atic hierarchy in which experiencers are ranked higher than themes (Grimshaw, 1990). Thus, the thematic structure of predicates can play an im portant role in determining their contingent fi-equency. 3.2.2 2 Transitivity Since verbs m ust be transitive in order to be passivizable, one could ask the question of w hether there is a relationship between the two properties. Transitivity is a property of clauses which reflects the degree to which the action expressed by the verb affects its theme. This property is often 61 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. viewed as a continuum, which H opper and Thompson (1980) have tried to circumscribe by distinguishing several features (e g. ASPECT and MODE) which m ay contribute to m aking clauses transitive. As a simple test of the idea that the degree of transitivity of individual verbs may correlate with their ability to be passivized, 100 verbs were selected from the Penn Treebank database and rated using H opper and Thompson's parameters. However, since the aim is to rate lexical items, only those parameters relating to verbs (not clauses) were used to rate the verbs from the Penn Treebank. These param eters were PARTICIPANTS, KINESIS, PUNCTUALITY, and VOLrnONALITY (see Appendix l).2i A verb was given a score of 1 when it had a feature described by one of the parameters and 0 when it did not. The composite transitivity score thus varied between 0 and 4. Once the set of 100 verbs was scored, a correlation was run between the verbs' transitivity score and their percentages of passive use in the Penn Treebank corpus. These two variables correlated significantly (r= .26; p < .01), indicating that the more strongly transitive a verb is, the more likely it is to be used in the passive voice. This result might be due to the fact that highly transitive verbs are m ore capable of handling the loss of transitivity that is associated with passivization. Many researchers, including Hopper and Thompson, have argued that passive clauses undergo w hat is traditionally referred to as detransitivization' (Givôn, 1994). During this process, the agent is often deleted, which results in a decrease of the transitivity of the action. Thus, The parameter of PARTICIPANTS refers to the number of entities involved in the action. No transitivity can take place unless two or more participants are involved. KINESIS simply refers to wheÂer the verb is stative or actional. Actions can be transferred from one participant to another, but states cannot. PUNCTUALITY refers to whether the verb denotes an action carried out in transitional or immediate stages. A strongly transitive action is one whose effects are immediately observed. VOLmONALITY refers to whether the action was purposeful or involuntary. Purposeful actions are seen as being more transitive than involuntary actions. 62 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. highly transitive verbs (e.g. issued) w ould be m ore likely to be passivized w ithout m uch harm to the overall message of the sentence. In sum, the semantic properties of the predicate which contribute to transitivity play a direct role in predicting the contingent frequency of this predicate in the passive voice. H aving established (1) that actives are far m ore frequent than passives in English, as is the case in many different languages, (2) that various semantic and pragmatic factors can contribute to this state of affairs and (3) that some verbs may not follow this structural tendency, the question that arises is this: are these differences in structural frequency between actives and passives, and in contingent frequency between active-bias and passive-bias verbs, reflected in the human processing system? To p u t it another way, does the processing system represent frequency differences that exist in the language? The answer would seem to be 'yes'. 4. The representation of frequency As w as noted in Chapter I, the constraint-based model assumes that the frequencies with which words are heard are directly encoded in the lexical representations of these items. It is unclear how one would go about proving this assum ption, which is perhaps the weakest point of the constraint-based model. How ever, it is no m ore a weakness of this theory than the construction of syntactic nodes is for the garden-path model, or the postulation of a theta-criterion is for Gibson's model. Theories often have to m ake assum ptions because there is no other w ay of explaining available data. Still, to the credit of the constraint-based m odel, notice that this assum ption about the language processing system is in fact plausible, given that 63 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. researchers from very different fields have found that the hum an brain keeps track of the frequencies of a num ber of events, even those that m ay sometimes seem trivial (see Kelly & Martin, 1994, for a review). For example, native speakers of a language seem to be sensitive to the frequency w ith which they encounter letters of the alphabet (Attneave, 1953). As readers, they are also aware of the relative frequency with which letters (such as w and h) appear in tandem i.e. in bigrams (Broadbent & Gregory, 1968; Rice & Robinson, 1975). O n a higher level, it has been dem onstrated that people are attentive to the frequency of syllables in their language (Rubin, 1974). Finally, it is well-known that native speakers have clear intuitions about the frequencies of occurrence of w ords as well as about their degree of familiarity (Shapiro, 1969). It is true that these intuitions are sometimes unreliable: speakers m ay overestimate the frequencies of low-frequency items or underestimate the frequencies of high-frequency items (Baron, 1988; Gemsbacher, 1984). However, the intuitions of native speakers are, by and large, consistent w ith available frequency counts (Gordon, 1985). Outside of the realm of language, hum an beings have been shown to be sensitive to the relative lethality of events (Lichenstein, Slovic, Fischoff, Layman & Combs, 1978); the relative frequencies of right- and left-handers (Coren & Porac, 1977); the comparative frequency of surnam es (Zechmeister, King, G ude & Opera- N adi, 1975; Valentine & Moore, 1995); the periodicity of encounters with different people (Saegert, Swap & Zajonc, 1973); the frequency of occurrence of novel shapes such as ideograms (Wiggs, 1993); the frequency of traumatic exchanges in adolescence (Vemberg, Ewell, Beery, Freeman et al., 1995) and even the commonness of fast-food restaurant chains (Schedler, Jonides & Manis, 1985). Evidence from different fields of research thus favors the view that the hum an information processing system "seems to support the 64 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. inevitable encoding into m em ory of certain fundamental aspects of experience" (Hasher and 2^cks, 1984, p. 1372). The assumption that frequency is encoded by the hum an processing mechanism leads to interesting hypotheses about the kind of language processor that is capable of handling continuous encoding of information. MacDonald, Pearlm utter and Seidenberg (1994) provided a detailed proposal about the structure of the parsing mechanism. They suggested that distributed representations of individual item s include information about syntactic identity, thematic properties, contingent frequency and other aspects relevant for the production and processing of these items. Distributed representations, they argue, are necessary in order for the processor to be able to encode different pieces of information simultaneously. Figure 3 represents a portion of the lexical representation of a verb, which includes inform ation about tense morphology, voice and argum ent structure (reproduced from MacDonald, Pearlmutter and Seidenberg, 1994). These various types of interconnected information become activated in parallel w hen an item is encountered, which leads to competition between alternatives that are m utually exclusive (excitatory connections are shown in plain lines while inhibitory connections are show n in bold lines). For instance, if the w ord raced is encountered in an am biguous context, both the past tense and the past participle alternative w ould presum ably be activated and competition is assum ed to be won by whichever alternative is higher in frequency, all other things being equal. The m ore frequent a particular alternative is, the higher its degree of activation is likely to be. In comprehension, higher frequency m ost often results in greater speed and accuracy; in production, higher frequency could result in greater likelihood of use. 65 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. A Portion of the Lexical Representation of a Verb Tense M orphology Past Tense P ast Faitidple Active Transitive Voice Argument Structure Passive Figures The MacDonald et al. proposal about lexical representation is clearly an appealing one, but it needs to be elaborated further in order to match the claim that syntactic, semantic, and discourse information interact during the processing of sentences. Indeed, MacDonald et al. need to be specific about how various aspects of a w ord’s identity are represented. For a verb, that w ould presum ably include information about its selectional restrictions (whether it takes animate or inanim ate argum ents or both), its thematic structure (whether it requires agent, patient, goal or benefactive roles), its grammatical number (e.g. walk vs. walks), its grammatical person (e.g. is vs. 66 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. are), its co-occurrence inform ation ( whether it tends to occur w ith certain nouns to its left or right), its discourse implications (e g. w hether it requires prior establishing of reference, causality etc.) and perhaps even its length relative to other elem ents of a sentence. The difficulty comes from having to specify all the possible interactions between these types of information in order to validate the constraint-based theory's claim about the richness of lexical representations and about the simultaneous activation of m ultiple types of information. In sum , the assumption that frequency is encoded seems to be supported by data from different fields of research, b u t current proposals still have a long w ay to go in order to be explanatorily adequate. Assuming that frequency is encoded, we can now move on to look at its effects. 5. The effects of frequency The list of pragm atic and semantic factors presented in section 3 may not be exhaustive, but it helps m ake the following point: Frequency is the currency over which tendencies created by semantic and pragm atic factors are expressed. It is possible to think of contingent frequency as a log of previous situations in which using the active voice was more appropriate than using the passive voice (and vice-versa for certain verbs). That log is constantly updated and is thus responsible for the activation of active and passive alternatives. However, this approach raises the question of w hether frequency can have effects of its own. To put it differently, if frequency is the result of semantic and pragm atic factors, then it may seem reasonable to attribute frequency effects to semantics and pragmatics. 67 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Whether effects of frequency can be attributed to semantic and pragmatic factors is in fact beside the point. Clearly, h-equency differences are the result of a num ber of variables including non-linguistic factors. The point, however, is that the hum an processing system is able to utilize this frequency information in order to rapidly and efficiently process language. The evidence reviewed in section 4 suggests that hum ans effortlessly keep track of a number of probabilistic patterns in their environm ent, so the processor seems able to exploit these statistical properties. 6. Summary and C onclusions To recapitulate, frequency differences are created during production by the interaction of several semantic and pragm atic factors. While exposed to linguistic stimuli in which these differences exist, the processing system alters the weights of relevant nodes in a way that reflects the periodicity with which items and structures are encountered. This allows the parser to take advantage of statistical properties of the language and exploit them in order to m ake processing m ore efficient. It should be clear that the path taken by frequency is not circular. The frequency of individual items plays a direct role in the comprehension of sentences in which these items appear, but it plays a m ore limited role in production. Indeed, one does not produce utterances in order to conform w ith frequency, but in order to convey a specific message. W hile it is true that the level of activation of high-frequency items makes them more accessible and hence more likely to be used in production, there are other linguistic factors such as register variation that can quite noticeably affect the choice of lexical items during production.^ ^ For example, in formal social situations, one is more likely to use more complex structures, or less frequent items. This would not necessarily be predicted by a frequency-driven production mechanism. 68 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Current research seems to be increasingly interested in probabilistic constraints on comprehension. Sometimes, how ever, probabilistic values attached to certain types of linguistic information are missed. For example, arguing that the doctor treated the old man is sem antically more plausible than the patient treated the old man is itself a probabilistic (a frequency) difrerence: the first case has a higher likelihood of occurrence than the second case. Yet, the notion of semantic plausibility is generally understood as denoting semantic context effects, so that the probabilistic notions that it relates to are overshadowed. The point to be m ade here is that few people w ould question the idea that semantic plausibility is a factor that can influence processing and yet there is still some resistance to the idea that frequency is a viable alternative for explaining data (e.g. Frazier & Clifton, 1996). Being sensitive to frequencies is merely an efficient way for hum ans to simultaneously keep track of different sorts of linguistically relevant inform ation. By acknowledging the fact that frequency can have effects, one can investigate research questions which can be inform ative about the way in which hum ans use linguistic information. For exam ple, one can explore the extent to which com prehenders rely on frequency vs. semantic and pragm atic context. Thus Gamsey, Myers, Pearlmutter & Lotocky (1996) did several experiments in which they m anipulated the bias of verbs that preferably take direct objects (e.g. protested) or sentence-complements (e.g worried). They also m anipulated the plausibility of the continuation. They found rather robust effects of verb bias b u t that plausibility had an effect only when the verbs were equi-biased. This pattern of behavior, so similar to findings in 69 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. word recognition, (Ehiffy et al., 1988) suggests that syntactic and lexical ambiguity resolution follow the sam e mechanisms. More interestingly, one can also investigate the way in which different sorts of frequency information interact in on-line processing. This is a fundamental concern of the research reported in Chapter IV. It has been shown that actives are far more frequent than passives in English. However, there are some verbs that run contrary to that structural trend and are thus biased towards the less ti-equent voice type. This situation allows one to explore the interaction of contingent and structural frequencies. If structural frequency turns out to affect parsing of passives relative to actives regardless of the subtle contingent frequencies of verbs such as sentenced, then one w ould have to conclude that structural frequency overrides contingent frequency, which w ould indicate that speakers keep track of only the m ore general kinds of frequencies and otherwise rely on context to guide parsing. However, should contingent frequency turn out to affect parsing, then one would have to assume that the hum an processing system keeps track of subtle kinds of frequencies, and that structural frequency is Itself merely a reflection of the conjunction of many contingent frequencies. Keeping track of structural frequency would then naturally fall out of keeping track of contingent frequencies. This research question will be examined in this dissertation in the light of passives and actives in English. 70 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER n i; THE COMPREHENSION OF PASSIVE SENTENCES Passive sentences have been studied continually in psycholinguistics since the 1960s. Early studies focused on the relationship between deep and surface structure in passives and were geared towards the verification of hypotheses form ulated in linguistics (Gough, 1965,1966; Slobin, 1966, Clark, 1965; W right, 1969; Greene, 1970). By the 1980s, however, those sorts of research questions were all but abandoned, and people looked at passives for resolving traditional psycholinguistic issues such as prim ing phenom ena (Frazier, Taft, Roeper & Ehrlich, 1984). More recently, researchers in sentence production have explored pragm atic factors that affect the choice of voice (Bates & Devescovi, 1989; M archman, Harris, Juarez & Bates, 1992). The fact that the passive voice is used under specific conditions has prom pted language acquisition researchers to investigate the temporal course of its appearance in children (Fraser, Bellugi & Brown, 1963; Mills, 1985; Berman, 1985; DeMuth, 1989). Aphasia researchers, however, have tended to focus on the working m em ory dem ands of passives compared with actives (Caramazza & Miceli, 1991; Caplan, Baker & Dehaut, 1985; Ratcliff & McKoon, 1989). This chapter provides a critical review of this literature and proposes a re-interpretation of some of these results which will enlighten the experimental research undertaken in Chapter IV. 1. Studies of voice comprehension In early sentence comprehension research, a substantial am ount of evidence w as produced in support of the general assum ption that passive sentences are m ore difficult to process than active sentences. Theories of 71 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. linguistics, such as Chom sky's (1957) initial form ulation of Transformational Grammar, suggested that passive constructions could be derived from active structures through a series of simple transform ations. Psycholinguists reasoned that sentences w hich require m ore transform ations from deep- to surface-structure should take longer to com prehend than those which required fewer or no transformations. In an attem pt to provide behavioral verification for this claim, several off-line sentence com prehension studies examined the processing of actives and passives. Miller (1962) used a sentence matching test in which subjects had to pair transform ed sentences with the corresponding kernel (i.e. source) sentences. Estim ated transform ation times w ere longer for m atching passive to kernel sentences, than for matching kernel sentences to themselves. The task of m atching passive sentences to active counterparts required computing the passive transform ation in reverse, and since it took longer for subjects to do this task than when there w as no transformation required. Miller concluded that the results supported a transformational view of grammar. These findings were replicated by Gough (1965), using a sentence- picture matching paradigm (see also Clifton, Kurcz and Jenkins, 1965; and Smith, 1965). G ough found that the m ean verification time for passives was longer than for actives w hen they matched the presented pictures as well as w hen they did not. G ough was well aware of the fact that, while his results were consistent w ith the transformational hypothesis, they did not establish it. For example, Gough noticed that passive sentences were both longer and less frequent than active sentences, facts which could have been responsible for the effect observed. Yet, he did not consider the possibility of a frequency effect in later work (Gough, 1966). Rather, he exam ined the effects of sentence 72 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. length and delay of picture presentation on the task of matching sentences to pictures. The former m anipulation was m otivated by the fact that full passive sentences are longer than active sentences, so Gough compared shorter passives to full length actives. The latter m anipulation was aim ed at obtaining verification RTs which reflected only the time required for matching sentences to pictures, unconfotm ded w ith the time necessary to perform a transformation. Gough found that shorter passive sentences were still verified w ith longer RTs than active sentences, and that delay had no effect on verification time. The transformational explanation of the findings suffered a setback, however, in Slobin's (1966) investigation of reversible and non-reversible passives. In reversible passives such as those in (1) and (2), both the agent and the theme are capable of accomplishing the action denoted by the verb, while in non-reversible passives such as (3) and (4), the agent is the only argum ent capable of assum ing such a function. (1) Julia was criticized by Tom during the presentation. (2) Tom was criticized by Julia during the presentation. (3) Julia was tempted by the cheesecake on the dining table. (4) *The cheesecake was tempted by Julia on the dining table. Slobin noticed that although the two passive types were derived through the same kind of transformation, they appeared to be verified with different RTs. He found that, for both children and adults, RTs for passive sentences were longer than for active sentences, but that non-reversible passives did not take any longer to comprehend than actives. Slobin argued that this latter result 73 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. was due to the fact that only one noun could be the agent of the action in non-reversible passives, and proposed that semantic factors such as reversibility alter the perform ance predictions m ade on the basis of the transformational theory. In other words, both syntactic (i.e. grammatical rules) and semantic factors were thought to contribute to the comprehension of passives. Further invalidation of Gough's results came from a study by W right (1969). She presented active and passive sentences auditorily to subjects. After a 5-second delay time, subjects were asked a comprehension question which was itself posed in either the active or the passive voice. W right found that there were fewer com prehension errors in the 'same-voice' conditions than in the different-voice' conditions. Furthermore, there was no difference in performance on actives relative to passives when each appeared in the matched conditions. She concluded from this that passive sentences were handled as passives, and w ere not transformed into kernel sentences, and she attributed results such as those of Gough (1966) to the difference in voice between the test sentences and the comprehension questions. In other words, previous results were due to a mismatch problem, not to a voice problem. A very interesting finding by W right, and one that is replicated in Tannenbaum and Williams's (1968b) production study, was an interaction between sentence type and question type such that differences between error rates for comprehension questions w ere m uch larger in the passive comprehension question condition. This interaction is shown in Figure 1 below. 74 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Data from W right (1969) 140 Active sentence Passive Sentence 120 100 I S 8 0 ' I M 60 - 40 Active question Passive question Type of Comprehension Question Figure 1 Thus subjects had more trouble answ ering a comprehension question when it was presented in a different voice type from the one in which the test sentence was formulated. This difficulty was exacerbated in the passive question condition.23 The transformational hypothesis was thus seriously challenged. Using a sentence-sorting task, Greene (1970) investigated the hypothesis that passives w ould be in fact easier to com prehend when they were used for their natural' f u n c t i o n . ^ 4 Greene m ade the assum ption that "the passive can be This type of interaction, with larger effects for passives relative to actives, will be observed in several other studies described in section 2. Her assumptions sound surprisingly similar to constraint satisfaction views: "each type of syntactic construction has a specific semantic function (...) the difficulty of handling a particular syntactic form will depend on whether it is being used in a context in which it is performing this natural semantic function" (Greene, 1970, p. 17). 75 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. considered as being used unnaturally w hen it is used to reverse meaning" (p.l8). Subjects w ere asked to determ ine if the two m em bers of each of the pairs given below had the same or a diHerent meaning. Same m eaning pair (5) X exceeds y, (6) y is exceeded by x. Different m eaning pair (7) X exceeds y. (8) X is exceeded by y. Greene found that it took subjects an average of 1.17 seconds to decide that pairs such as (5) and (6) have the same m eaning but 1.53 seconds to decide that (7) and (8) have a different m e a n i n g . 2 5 in both cases, the second sentence was in the passive voice, yet the second case yielded m ore errors and longer sorting times. This was interpreted as suggesting that com prehending the passive construction was dependent on prior linguistic context and not merely on the num ber of transformations required to switch it to its kernel form. Although the DTC was completely rejected by the 1970s, the experimental findings of greater difficulty with passive than with active sentences w ere very consistent.^^ However, these results w ere challenged by ^ Notice that there is a confound here. In general, it takes longer to make negative verifications than positive ones, and this is exactly what Greene found. ^ The difficulty experienced with passive sentences seems to extend to other languages where the passive construction is less frequent than the active. Roman, Pavard and Asselah (1985) ran a simple off-line study, in which subjects were asked to read out loud Arabic sentences 76 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Carrithers and her colleagues in a series of on-line experiments. In an eye- movement study, Carrithers and Bever (1984) found that readers had longer fixation durations o n the last words of canonical sectors (active clauses of the type NP1-V-NP2, w here NP% is the agent of V) than for final w ords of non- canonical sectors, including passives. This result was replicated in another study which com pared final-phrase reading times of actives and passives (Carrithers, 1986). It is possible, however, that these diverging results m ay have to do with the nature of the stim uli. Full passive sentences have more predictable endings than do actives sentences, because of their agentive by phrases. This fact alone could explain faster reading times in the final phrase of passive clauses and points to the need for controlling for predictability in reading time studies. Moreover, a com parison of final-phrase RTs does not fully reflect the level of difficulty of passive or active clauses, because the entire construction m ust be taken into account when doing comparisons. Accordingly, Carrithers (1989) attem pted to generalize the result of faster RTs in the final phrases of passives to the entire passive structure. She used a moving w indow experiment in which subjects read sentences off a computer screen one w ord at a time. Subjects controlled the pace of presentation by pressing the spacebar on the computer's keyboard. Carrithers found that passive clauses and subject clefts were read 20 ms per w ord faster than active clauses and object clefts. She also found that subjects w ere faster which were either vocalized or non-vocalized. The Arabic writing system represents consonants only, which creates an ambiguity as to whether the main verb of a sentence is in the active or in the passive voice. Roman, Pavard and Asselah found that subjects made more errors reading and recalling passive sentences when the main verbs were not voice-marked. In a second experiment, they recorded the eye-movements of subjects reading the same materials. Results showed longer regressions with passive than active sentences, whether the verbs were vocalized or n o t Longer regressions are often taken to signal difficulty with, and re interpretation of material. Thus, comprehenders appeared to have more trouble reading passive relative to active sentences. 77 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. on the function w ords in passives in com parison with their m ean overall reading rate. Notice that large differences in speed of reading times at a given w ord can affect reading times on subsequent words, a phenomenon know n in the reading literature as spillover effect*. If subjects beneHted ffom increases in speed on the function w ord by', spillover effects could appear in the rem ainder of the agentive phrase. This could explain why the per-w ord reading time of passive sentences was overall faster than that for active sentences. Furthermore, Carrithers' stim uli were not controlled for length, so her passive sentences tended to be longer than her active sentences. In self- paced reading, subjects tend to read faster towards the middle of the sentence and slow er at the end of the sentence (Mitchell & Green, 1978). For longer sentences, the speed in the middle region increases, so that the average reading time per w ord for longer sentences turns out to be shorter than for brief sentences. Thus, Carrithers' results are open to re-interpretation. Frazier, Taft, Roeper, Clifton and Ehrlich (1984) obtained results sim ilar to those of Carrithers'. In an on-line experiment, they m easured reading time on the second clause of conjoined sentences such as (9) through (12): (9) The tall gangster hit John and the short thug hit Sam (active-active). (10) John was hit by the tall gangster and Sam was hit by the short thug (passive-passive). (11) The tall gangster hit John and Sam was hit by the short thug (active- passive). (12) John was hit by the tall gangster and the short thug hit Sam (passive- active). 78 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Frazier et al. found that subjects read second segments faster when the two segments w ere parallel in structure than when they w ere n o t However, the parallelism effect was significant for actives but n ot for passives. Frazier and her colleagues claim that this is because passive sentences are used in discourse contexts in which the theme of the verb is in fact the topic of discourse, and Üiat this requirem ent is not fulfilled w hen passives are preceded by actives. Thus, because passives require a context that establishes the theme as the topic of discourse, and because this requirem ent fails to obtain, RTs on the second fragm ent are longer than w ould normally be expected, so that differences between the active-passive condition and the passive-passive condition are reduced. However, a rather surprising result of this study was that the mean reading time per w ord for the second fragment was faster when it was passive than when it was active, regardless of the syntactic form of the first fragment, as shown in Figure 2. 79 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Data from Frazier, Taft, Roeper, Clifton and Ehrlich (1984) 3 2 0 Faim o f f in t K gm ent 300 I a 280- I g ) 260. 1 240- 220 A ctive P assiv e Form of second segment Figurez This finding, which goes against predictions made by the garden-path model that simpler structures (such as actives) should be processed w ith greater ease than complex structures (such as passives), was not addressed in Frazier's study. However, it is interesting to note here that both Frazier et al.'s and Carrithers' study used full passives. As mentioned earlier, full passives use a by-phrase as argument. This by-phrase is expected to occur after the verb, an expectation which is reflected in greater speed in the post-verbal region. This difference could be reflected in the per-w ord reading times. In other words, w hen comparing full passives w ith actives, it is possible to obtain results in which the per-word reading time for passives is shorter, but this m ay not necessarily be interpreted as indicating greater ease of com prehension of passive sentences. 80 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Frazier et al.'s (1984) results contradict the garden-path m odel, but this was ignored, probably because Frazier's model, like other m odels of sentence processing, has mainly concerned itself with structurally am biguous sentences, rather than w ith relatively unambiguous ones. C hapter IV deals w ith this issue in depth and shows that parsing models need to m ove beyond issues of ambiguity resolution in order to characterize people's sentence processing abilities. It is interesting, in this respect, to note that production and aphasia studies have long dealt w ith unambiguous sentences. It is to these that w e turn next. 2. Studies of voice production Production research has concerned itself with the lexical and syntactic choices which face speakers during spontaneous speech. Active and passive sentences generally convey the same linguistic meaning, but they tend to communicate different pragm atic content. It is therefore unsurprising that m anipulations of pragmatic situations in which the speaker is likely to find herself have a strong effect on the type of voice used to express com m unicative intentions. Clark (1965) implemented a large scale production study in which subjects were asked to form sentences by filling in the blanks of fixed active and passive sentence frames such as (13) and (14). (13) The _________ __________ ed the __________ . (14) The _________ w as ed by the 81 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The purpose of the experim ent was to derive and compare common properties of agents, patients and verbs of active and passive sentences. Clark did this by computing, for each noun or verb used by subjects, an informational uncertainty factor, which is a m easure of the likelihood of appearance of a given noun or verb in a given position (e.g. the higher the uncertainty factor for a w ord in one position, the higher the num ber of different w ord types in that position). Results show ed that the agent had lower uncertainty m easures than the verb or object in the active frame, whereas they did not differ in the passive frame; that the agent in the active sentence fram e had lower uncertainty measures than the agent of the passive sentence frame; and that the patient of the active sentence fram e had higher uncertainty m easures than the patient of the passive sentence frame. In other words, the agents of active sentences tended to be better defined than their verbs or patients; die agents of actives tended to be better defined that those of passives; and the patients of passives tended to be better defined than those of actives. Furtherm ore, Clark analyzed the properties of the w ords used by subjects as agents and patients. He found that 81% of the agents in actives were anim ate (compared w ith 68% of agents in passives) and that only 26% of the patients in actives were anim ate (compared with 45% of objects in passives). 82 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. D ata from Clark (1965) 1 0 0 S 5 I s o g a & 3 E w u A g e n ts Patients A ctive Passive V o ice Figures Clark's study show ed that there are asymmetries betw een thematic arguments in actives and passives and that these sentence types differed in both animacy and uncertainty (see also Johnson, 1967, for matching experiments and conclusions). Tannenbaum and Williams (1968a) used a sentence description task in order to investigate the interaction of sentence focus w ith sentence voice. They gave subjects short pream bles in one of three conditions of conceptual focus: neutral, subject-focus or object-focus. The subject-focus preambles consisted of active sentences, while the object-focus pream bles consisted of passive sentences. Subjects were then asked to describe a picture by using the active voice or the passive voice. The picture depicted an event involving the 83 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. subject or object previously p u t in focus. Tannenbaum and W illiams found that passive sentences were produced w ith significantly shorter latencies in the object-focus than in the subject-focus condition whereas the reverse obtained for active sentences. In a subsequent study, Tannenbaum and Williams (1968b) used simple active sentences such as that in (15), then used either the subject, verb or the object of that sentence as a prom pt for recalling deleted sentence parts, as in (16) to (18). (15) The cat took the mat. (16) The ______ took the m at. (17) The cat ______ the m at. (18) The cat took the ______ . Tannenbaum and Williams found that, across all conditions, subjects displayed better overall recall in actives com pared with passives, but that the subject/agent was a better cue for rem em bering active sentences and that the object/them e was a better cue for rem em bering passive sentences. Furtiiermore, this difference in recall was exacerbated for passives. In general, then, it seems that the agent is m ore prom inent for actives while the patient is more prominent for passives, which w ould explain why it is better defined and is a better cue for recall. Tannenbaum and W illiams' results are consistent with those of Prentice (1966). She trained subjects to leam active and passive sentences in response to nouns. Thus subjects w ould be given a sentence such as (15) and trained to recall it when given a noun such as (20), that was a semantic associate of one of the nouns in (19). 84 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (19) A tiger frightened the wom an. (20) Lion She found that overall, it was easier to associate a noun w ith a sentence w hen that noun was semantically associated w ith the grammatical subject of the sentence (the first noun, in both actives and passives), than when it was associated to the grammatical object of the sentence (the second noun in both sentence types). It is interesting to note that in Prentice’s study, the difference in recall performance between subject and object was greatly enhanced for passive sentences. Again, this is an asym m etry between actives and passives, perhaps stemming from the difference in frequency between the two voice types. As will be shown in Chapter IV, certain effects may be more m arked in passives than in actives because actives are higher in frequency (see Seidenberg & McClelland, 1989). More general discourse factors w ere investigated by Bates and Devescovi (1989). In a free description task, they found that English speakers produced very few passives (around 1%) however, when the experimenter focused on a 'nonagentive element' of the scenes presented, production of the passive rose to 12%.27 This dem onstrates the importance of discourse context on the production of passives. Clearly, the choice of active or passive structures in sentence production partly depends on pragmatic factors such as w hether the discourse focus is the agent of the action or its patient, w hether the speaker is being formal or informal, etc. A nd it is clear how these Passives seem to be more frequent in spoken language in Italian than in English. In the free description task, 8% of Italian speakers' utterances were passives. In the probed condition, the percentage rose to 20%. 85 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. pragmatic constraints could influence structural h-equency: most discourse contexts revolve around agents rather than themes, and since the active word order places the agent in the TOPIC position, it follows that discourse contexts probably help the active structure become more frequent than the passive construction. Bates and her colleagues investigated the production of passives in im paired populations as well. In a description task. Bates, Harris, Marchman, and W ulfeck (1995) found that Alzheimer’s patients produce fewer passives than elderly controls who, in turn, produced fewer passives than college-age controls. Another result was that Alzheimer patients produced considerably m ore gef-passives as in (21) than bc-passives as in (22), a tendency which also seems to obtain in children's spoken language (Marchman, Harris, Juarez & Bates, 1992). (21) The goat got bitten by the horse. (22) The goat was bitten by the horse. Bates, Harris, Marchman, and Wulfeck (1995) argue that the first finding is a direct result of the low frequency of the passive structure relative to the active construction in English. They go on to claim that access to such low-frequency constructions is made more difficult by the process of normal aging and that this difficulty becomes particularly clear in dementia. However, Bates et al. adm it that they do not have any explanation for the result that demented patients produce more get- than be-passives. They simply suggest that get- passives are more frequent because get denotes a change of state, whereas be simply denotes a state, so be is less likely to fit the semantic intention of 86 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. speakers wishing to passivize a sentence. In other words, in a competition between get and be, the winner is get.28 Bates and her colleagues thus recognize the role of frequency in comprehension, b u t there is yet no explanation for how it can afreet the processing of specific sentence types. In general, the production studies described here show that the thematic properties of arguments in actives and passives are different. Agents are more prom inent for actives while patients are m ore prom inent for passives. This explains why focusing on a nonagentive elem ent of discourse results in the production of a greater num ber of passives. The influence of pragmatic factors on the production of passives has been reinterpreted by language acquisition researchers w ho have sought to determ ine the time- course of the acquisition of the passive. 3. Studies of voice in language acquisition Passives have been studied extensively by language acquisition researchers. Fraser, Bellugi and Brown (1963) show that three year old children acquiring English are unable to com prehend passive sentences if the participants in the action cannot be determ ined on non-linguistic grounds. It is not until age five that children are able to understand full passives in English (Slobin, 1966). This tendency also holds for Germ an and for Hebrew (Mills, 1985; Berman, 1985). However, children acquiring certain Bantu languages seem to leam the passive construction very early on. Sesotho is a case in point. For example, DeMuth (1989) shows that children learning Sesotho as their native language spontaneously produce passives before the ^ The curious absence of such competition in written language is taken by Bates and colleagues to result from the power of convention. 87 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. age of 3. It is im portant to keep in m ind, however, that in DeM uth's child corpus, the rate of passives produced was fairly small, as she herself admits (139 passives for 14218 utterances). The Sesotho data show early acquisition of the passive, but the view that the passive construction is learned after the active construction rem ains unchallenged, since there were m ore active than passive sentences in the child corpus. It m ight well be the case that the reason w hy the passive construction is acquired so early in Sesotho is that the frequency of Üie passive in this language is fairly high. DeM uth (1989) reports that in her adult caregivers’ corpus, 6% of utterances were passives, a rather high rate for spoken l a n g u a g e . 2 9 Thus the point to be m ade here is that the age at which the passive construction is acquired is a function of its frequency in the language in question. Some researchers might argue that it is necessary to distinguish adjectival from verbal passives, because there are languages (e.g. Hebrew), where children acquire adjectival passives long before verbal passives (Berman, 1985). It is im portant to realize that this pattern of acquisition is closely related to frequency of occurrence. In Hebrew , verbal passives rarely occur, and, in fact, seem to be restricted to w ritten discourse. The language acquisition data available are thus consistent w ith a frequency- driven acquisition process in which higher frequency structures w ould be acquired earlier than lower frequency ones. 4. Studies of voice in aphasia The classic finding in sentence processing research that passives are m ore difficult to com prehend than actives is reflected in aphasia by the greater likelihood for passive sentences than active sentences to be affected in Recall that, in their free description task. Bates and Devescovi (1989) found that approximately 1% of English speakers' utterances were passives. 88 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. brain-dam aged patients. A large body of evidence from language breakdown supports this view. For example, Caramazza and Miceli (1991) described a stroke patient w ho displayed an impairment in processing them atic roles but retained tiie ability to process the morphological structure of sentences. Specifically, Caram azza and Miceli's patient w as relatively m ore impaired in the com prehension and production of passive reversible sentences than active sentences. This type of selective im pairm ent has also been documented in other aphasie patients (Schwartz, Saffran & M arin, 1980). Caplan, Baker and D ehaut (1985) explored the nature of syntactic im pairm ents in aphasie patients using an object-manipulation task, in w hich subjects enacted thematic roles of sentences. Caplan et al. found that syntactic structure influenced sentence comprehension in aphasia such that sentences with non- canonical them atic role orderings, greater num ber of thematic roles, or larger num ber of verbs w ere difficult to comprehend for aphasies. Thus, aphasie subjects perform ed m ore poorly on passive sentences (which use a non- canonical them atic role ordering) than on active sentences; they also had more difficulty w ith dative passives such as (23) which use m ore thematic roles than w ith sim ple passives such as (24). (23) The elephant was given to the monkey by the rabbit. (24) The elephant w as hit by the monkey. One reason that passive sentences tend to break dow n m ore easily under brain dam age m ight be that they place higher working memory dem ands on com prehenders than do active sentences. This view is developed in work by Miyake, Just and Carpenter (1994) who argued that aphasie patients m ay have reduced working m em ory capacities by 89 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. comparison w ith norm al adults. They reasoned that, by severely increasing the com putational dem ands on norm al adults, their perform ance w ould show some of the characteristics of aphasies' performance. In a "simulation" experiment, Miyake et al. presented subjects w ith sentences which they had to read one w ord at a time, using the RSVP (Rapid Serial Visual Presentation) technique. Miyake et al. varied the presentation speed of the stimuli, from slow to fast. They found that, overall, normal adults perform ed more poorly on passive sentences than on active sentences, and that subjects' performance under stressful conditions (faster pace of presentation) correlated directly with their w orking memory capacities, w ith high working m em ory individuals being less affected, and low working memory individuals being more affected, by increases in computational demands. Thus, it appears that passive sentences w ould require more working memory for their comprehension, hence the finding that they are more affected in disorders that severely constrain w orking memory capacities. Further evidence that passives require more working memory resources is provided by the greater difficulty in rem embering passives relative to actives. Using a response signal procedure, Ratcliff and McKoon (1989) exam ined the time-course of retrieval in a sentence m atching procedure. Subjects had to leam lists of active and passive sentences and were then presented with test items which either matched the learned stimuli or approxim ated them, but w ithout being correct reproductions of the memorized sentences. For example, if (25) was a learned sentence, (26) would be a correct active version, but (27) would be an incorrect passive version. 90 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (25) John was hit by Bill. (26) Bill h it John. (27) Bill w as hit by John. Ratcliff and McKoon found that subjects tended to take longer in answ ering identification questions about passive sentences than about active sentences.^® The studies described so far, w hether in comprehension, production, acquisition or aphasia, have looked at various kinds of factors such as complexity, length, discourse contexts etc. in their investigations. How ever, there is at least one study that has investigated actives and passives w ithin the context of a connectionist model. 5. S t John and Gem sbacher's model St. John and Gemsbacher (1995) investigated w hether the difficulty reported in the processing of passive sentences in im paired populations w as a question of complexity or if it was a m atter of frequency. They built a recurrent parallel distributed processing model which they trained on a corpus of m ore than 22, 000 occurrences of four sentence types, including the active and the passive. The model was given the task of determining the agent and patient of the action described by the verb. This kind of task was plausible, they argued, because the questions "represent the sort of information a reader or listener is likely to be expected to obtain from a ^ In the same vein, Pleszeswka (1985) investigated the relationship between sentence structure of college lectures and students' retention of the lecture content. Her results show that the content of lectures consisting of active voice sentences is retained better than the content of lectures consisting of nominal or passive voice sentences. 91 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. sentence" (p3). St. John and Gernsbacher found that the m odel perform ed better on active than passive sentences, when actives w ere eight times m ore frequent than passives in the training corpus. More interestingly, they found that when passive sentences were m ore frequent than active sentences in the training set, the m odel's performance was instead better on passive than on active sentences. Their conclusion was that the mere frequency of active and passive structures in English could cause passive structures to be harder, regardless of any structural complexity inherent in these structures. While m odeling work such as St. John and Gem sbacher's cannot prove that frequency, rather than linguistic factors such as complexity, influences the processing of sentences, it does show that a theory based on frequency differences is at least plausible, and can produce the same kinds of results that are observed in behavioral studies, w ithout encoding complexity. St. John and G em sbacher’ s model makes interesting points about two other relevant differences betw een active and passive sentences. First, the model performed better on shorter sentences (actives) than on longer sentences (passives, or passives with a modifier) when it was trained on both constructions with an equal frequency. Notice, however, that length no longer mattered w hen the passive construction was m ore frequent: this time, the model perform ed better on longer sentences. Thus, length effects were observed only w hen no frequency differences existed. Second, St. John and Gemsbacher m odeled brain damage by using McClelland’ s (1993) noisy neuron procedure. The activation value of each unit was subjected to a small am ount of random noise, which resulted in poorer perform ance on lower frequency structures (actives or passives, depending on the simulation). This 92 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. result m irrors behavioral findings that brain-dam aged subjects are more im paired on the low frequency voice type (passive sentences). 6. Summary and conclusions The results of the studies reviewed here tend to support the contention that full, reversible passive sentences are m ore difficult to comprehend than active sentences in on- and off-line tasks. A common problem with studying the com prehension of active and passive sentences is that they differ along m ore than one variable. First, full passive sentences are longer than active sentences, w hich can create a confound, as it did in Slobin's (1966) study. But length does not appear to make passive sentences more difficult, since shorter passive sentences are harder to match to pictures than corresponding active sentences (Gough, 1966). Second, thematic role allocation in passive sentences differs from that in active sentences. A rgum ent positions in reversible and non-reversible passive sentences are not assigned canonical thematic roles: the subject of a passive sentence is the them e of the verb, rather than its agent. It is true that the animacy of the them e in non-reversible passives (e.g. the stone was carried hy the boy) gives an indication to the comprehender about the them atic relation between the verb and the theme. However, reversible passives do not have this option available, and are, in this sense, m ore complex' than active sentences. Third, the processing of passive sentences seem s to place greater working m em ory dem ands on the comprehension system. Decreases in working m em ory capabilities, which are common in cases of brain damage, m ight m ake passive sentences more vulnerable to language breakdown. Fourth, passive sentences are less frequent than active sentences, a confounding factor in all the experimental studies m entioned. 93 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The m ore im portant concern, however, is that these four factors are intimately related. Thus w hen sentences increase in length, they require more working m em ory for their comprehension and production. Furthermore, it is a fact of English, and languages in general, that m ore complex sentences tend to be less h-equent than sim pler sentences. In sum , increases in length alm ost inevitably entail increases in sentence complexity, which lead to greater working memory dem ands. In the case at hand, these converging tendencies will lead to lower frequencies of use for passive sentences relative to active sentences. In other w ords, the variables studied above contribute to creating frequency differences, and these frequency differences are exploited by the sentence processing mechanism during com prehension. Thus it seems probable that structural frequency differences could be at least partly responsible for the behavioral results of greater difficulty with passives than w ith actives found in the studies reviewed here. Miller's (1962) results of greater difficulty w ith passives in a sentence-matching task could be attributed to structural frequency. The same goes for Gough's (1965) sentence- picture m atching task. Furthermore, the effect of structural frequency seems to persist over and above effects of length, as show n by Gough's (1966) investigation of short passives. Structural frequency can interact w ith context, however. In W right's (1970) off-line comprehension task, there was no difference in perform ance on active and passive comprehension questions when they were presented after a sentence w ith the identical voice type. Thus context can all b u t eliminate the effect of structural frequency, though it can be seen in the interaction between sentence type and comprehension 94 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. question type (for the different-voice condition, error rates w ere m uch larger w hen the passive comprehension question was in the passive). The interaction of context and hrequency can also be seen in the production experiments. Typically, subjects produce few passives, but w hen prior context is focused on the theme rather tium the agent, die rate of passives produced increases, though it never surpasses that of actives (e.g. in Bates & Devescovi, 1989). The lower frequency of passives can also be used to explain their disruption in aphasia (Caramazza & Miceli, 1991; Caplan, Baker & Dehaut, 1985). Less frequently used structures would be expected to be affected first when the processing mechanism is impaired. In general then, structural frequency can be argued to be at the source of many of the results reviewed above. This view is supported by St. John and Gemsbacher's (1995) neural network sim ulations. The research reported below aims to incorporate structural frequency différences, but also extends its views to contingent frequency. Rather than focusing on actives and passives to determine which is harder to process, I propose to use their difference in structural frequency to investigate the role of statistical properties of lexical items and syntactic structures in parsing. I argue that the comprehension of active and passive sentences depends on the frequency of these structures in the language, but more im portantly, that the relative frequency of use of m ain verbs in each voice type affects the w ay in which active and passive sentences are comprehended. Furtherm ore, I investigate the interaction of structural and contingent frequency, in the context of both unam biguous and ambiguous passive sentences. 95 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER IV: EXPERIMENTS A large part of current psycholinguistic research is devoted to the question of on-line ambiguity resolution: how com prehenders deal with alternative m eanings or interpretations. This interest in ambiguity resolution is due to two m ain reasons. One is that any valid theory of hum an language processing will need to explain how com prehenders consistently choose the correct interpretation of ambiguous sentences with hardly any awareness of the existence of alternatives. Indeed, languages are full of ambiguities, and since linguistic information becomes available over time, the processing system m ust have a mechanism that enables it to deal w ith ambiguous material in a rapid and efficient way. The second reason for the focus on ambiguity resolution is that ambiguities tend to slow dow n the process of parsing, thus giving researchers a w indow on the m anner in which language processing operates. It is not surprising, therefore, that ambiguity resolution has been at the center of sentence comprehension research for over 20 years. While this focus on ambiguity resolution has contributed to the development of the field, it leaves open the question of w hether current models of sentence processing can account for phenomena that may not necessarily qualify as structural am biguity resolution. Indeed, since these aspects of processing have hitherto been ignored, it is difficult to determ ine whether they can be handled by current theories of processing. The experiments presented here seek to answ er this concern. They investigate the processing of relatively unam biguous sentences such as the competent lawyer was asked/elected to represent the new firm. 96 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. C urrent m odels of sentence processing have p u t forth different proposals about how ambiguities are resolved (Mitchell & Cuetos, 1991; Cuetos & Mitchell; 1988; Gibson, 1991, Gibson, Pearlm utter, Canseco-Gonzalez & Hickok, 1995; Crain & Steedman, 1985; Altmann & Steedm an, 1988). Serial models assume that only one structural interpretation is constructed at a time, with no input other than syntactic (Frazier, 1978; 1987; Frazier & Clifton, 1996; Rayner, Carlson & Frazier, 1983). Constraint-based m odels, on the other hand, assume that m ultiple sorts of information are com puted at the same time and that the level of activation of alternatives is determ ined by their frequency (MacDonald, Pearlm utter & Seidenberg, 1994; Tanenhaus & Trueswell, 1995). Several sentence comprehension studies have identified frequency effects, b u t few have addressed the way in w hich different kinds of frequency inform ation interact during on-line processing (Juliano & Tanenhaus, 1994). The experiments presented below deal w ith this issue by investigating cases in which structural frequency and contingent frequency clash. Indeed, while the structural frequency of voice in English is such that actives are overwhelm ingly more frequent than passives, the contingent frequencies of individual predicates can run in reverse direction. Thus verbs such as sentence and elect are more frequently used in the passive than in the active while verbs such as sue and ask conform w ith the general tendency and are more often used in the active than in the passive. According to the constraint-based m odel adopted here, the latter type, active-bias verbs, would be predicted to be easier to process when in the active voice, while passive- bias verbs w ould be predicted to be easier to process w hen in the passive voice. Serial models of processing, such as the garden-path m odel, predict that structurally sim pler sentences (actives) wiU be easier to process than 97 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. more complex sentences (such as passives), but do not make any predictions about structures that differ only in term s of the lexical items that they contain (i.e. whether they involve active-bias o r passive-bias verbs). The experiments presented below investigate the interaction of structural and contingent frequencies in relatively unambiguous sentences, thus providing a new w ay of testing the serial and parallel approaches to processing. 1. Experiment 1 Experiment 1 was a large-scale survey designed to collect ratings of the preference for the active vs. passive version of verbs. The purpose of these ratings was to help classify a large num ber of verbs as being relatively active- bias or relatively passive-bias, thus obtaining a database which would be used for developing the materials for Experiments 2 and 3. A second aim was to investigate possible correlations betw een thematic structure and verb bias. Indeed, Ferreira (1994) proposed that the choice of active vs. passive voice in sentence production is parüy dependent on the thematic structure of the m ain verb. She found that passives occur m ore frequently w ith theme- experiencer verbs than with agent-theme or experiencer-theme verbs. This finding was tested in Experiment 1 by com paring verb ratings as a function of their thematic structure. The third goal was to investigate the relationship between verb bias and proportional voice fi-equency in the Penn Treebank database (Marcus, Santorini & Marcinkiewicz, 1993) in order to determine whether independent frequency corpora provide reliable evidence about verb bias. 98 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Subjects 184 u s e undergraduates participated in this survey in exchange for course credit Each subject rated one of 5 lists, each containing approximately 34 items, for a total of 175 verbs. Each list was seen by an average of 36 subjects (range: 33 to 43), Survey materials All the verbs selected were transitive. Each verb was presented in both the active and the passive voice. Procedure Stimulus item s were presented as pairs of verbs, w ith the active version on the left side and the passive version on the right side,^^ as shown below: the X reminded the y 1 2 3 4 5 6 7 the y was rem inded by the x the X phoned the y 1 2 3 4 5 6 7 the y was phoned by the x the X sentenced the y 1 2 3 4 5 6 7 the y was sentenced by the x The two fragments w ere separated by the numbers 1 to 7. Subjects circled a number closest to the version which they preferred. The verbs were presented with the generic arguments the x' and the y' in order to ensure that subjects visualized active and passive versions of verbs in sentences, but without using specific words. The definite article was used w ith the generic arguments so as to encourage subjects to think of an argum ent rather than of a name (as would probably be the case if 'X' and Y' had been used). This order of presentation was chosen so as to make subjects more sensitive to the passive structure. 99 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Results The mean rating for the m aterials of Experiment 1 was 3.60 (SD = .62). As is typical in such surveys, subjects did not utilize the full rating scale (1 to 7), so the actual ratings ranged from 1.76 to 4.97. Rather than grouping item means that were above the m edian (= 3.62) as passive-bias, and those that were below the median as active-bias, a m ore stringent criterion w as used. The scale was divided into three parts, w ith means under 3.30 indicating active-bias verbs and means above 3.80 indicating passive-bias verbs.^2 This allowed for greater discrimination betw een active-bias and passive-bias verbs. The ratings results are shown in Table 1 below. Table 1: Verb ratiniÇ S for Experiment 1 Verb Type Active bias Equibiased Passive bias Rating Range 1.76-3.30 3.31-3.80 3.81-4.97 C ount 52 53 70 Mean Rating (SD) 2.85 (.38) 3.54 (.15) 4.19 (.29) Example 'asked' 'praised' 'elected' Although these ratings are divided into three distinct groups, it is im portant to bear in m ind that verb bias is a continuum and that the observations made here are not to be interpreted in an absolute, but rather in a relative sense. Notice also that the intuitions of native speakers about the verbs in Experiment 1 give us an idea about the biases of these individual verbs in the absence of any context. Naturally, the predictions that one could make about the way in which these verbs will be comprehended or produced will also depend on contextual and situational constraints. For the purpose of this study, equibiased items are defined as verbs that are equally good in active and in passive fragments. 100 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In order to address the issue of the relationship betw een thematic structure and verb bias, the verbs used in Experiment 1 w ere divided into two types of thematic structure: agent-them e and theme-experiencer. An impaired t-test revealed that theme-experiencer verbs had significantly higher verb ratings than agent-them e verbs (t (173) = 4.91, p < .0001). This difierence means that theme-experiencer verbs tended to m ake better passive- bias verbs than active-bias verbs, as shown in Figure 1 below. V erb R atings and Them atic Structure I § (0 o Î Active-bias 0 Equibiased □ Passive-bias A gent-T hem e Them e-Experiencer V erb T ype Figure 1 This result is compatible w ith Ferreira’ s (1994) findings that subjects tend to produce passive sentences m ore often when the m ain verb is theme- experiencer than w hen it is not. The likely reason for this behavior, Ferreira 101 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. contends, is that speakers comply w ith a universal thematic role hierarchy which places agents above experiencers, and experiencers above themes. While this them atic ordering is m et for agent-them e verbs w hen they are in the active voice, speakers have to resort to the passive voice for theme- experiencer verbs. Notice, however, that theme-experiencer verbs differ from agent-theme verbs along two dimensions: w ord order and animacy. O n the one hand, theme-experiencer verbs require a non-canonical w ord ordering that easily m aps onto the passive voice, w hile agent-theme verbs fit in the canonical w ord order of English. On the other hand, the theme in theme- experiencer verbs is typically inanimate, w hereas it can be anim ate in agent- theme verbs. In order to clarify which of these two variables is prim arily responsible for the difrerence in ratings, the agent-theme verbs w ere divided into two sub-groups: agent-animate them e and agent-inanimate theme, as shown in Table 2. Table 2: Verb rating as a function of thematic structure Verb Type A gent-them e A gent-them e Theme-Experiencer Animacy A nim -A nim . A nim -Inanim . Inanim .-A nim . Count 77 52 46 Mean Rating 3.46 (.63) 3.47 (.58) 3.96 (.51) Example sue distribute impress A factorial ANOVA was then run w ith verb type and verb rating as factors (F (2,172) = 12.02, p <.0001). Post-hoc analyses showed that theme- experiencer verbs had significantly higher ratings than either of the agent- theme verb types {Scheffe F = 10.35 for anim ates and Sch^e F = 8.46 for inanimates, p < .05). Both types of agent-theme verbs had similar ratings, however, regardless of whether the theme was animate or inanim ate. This finding indicates that the difference in rating between theme-experiencer and 102 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. agent-theme verbs is not due to differences in animacy between the them atic roles of a verb, b u t rather to their ordering. The third goal of Experiment 1 was to examine how the ratings obtained relate to the relative frequency of use of verbs in the active vs. passive voice in the Penn Treebank corpus. In order to do this, the occurrences of each of the 175 verbs in the database was pulled from the Wall Street Journal Penn Treebank corpus and coded as active or passive. The percentage of passive usage for each verb was then entered as a predictor of its verb rating in a regression analysis. The regression was positive and significant (r = .32, p < .0001), suggesting that the higher the passive frequency of a verb in the Penn Treebank corpus, the m ore likely it was to also have a high (passive) verb rating. Thus the relative frequency of use of a verb in the passive voice is a good indicator of whether that verb is a passive-bias verb or not. Frequency corpora such as the Penn Treebank m ay therefore be relied on for determ ining verb bias. 103 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. # V eib R atin g I > # # l« 0 -20 20 0 40 60 80 100 120 % P assiv e U sag e Figure 2 Conclusions In sum m ary. Experiment 1 demonstrated the existence of different voice biases for verbs in English. These biases correlated with thematic structure in such a w ay that theme-experiencer verbs tended to m ake better passive-bias verbs than agent-theme verbs. One should be careful, however, not to confuse thematic structure with bias. While a verb's thematic structure does play an im portant role in its use in the active or passive voice, and hence in its voice bias, them atidty is only one of several factors that interact in various ways in determining the biases of a verb (see Chapter H). Indeed, context, transitivity, animacy, and general pragmatic constraints can all influence the relative voice bias of a verb. The end product of these interactions is expressed in the relative bias of a verb and in its relative 104 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. frequency of use in the active or passive. In this respect, note that Experiment 1 dem onstrated that verb bias as determ ined by native speakers' ratings tended to correlate significantly with the relative frequency of use of a verb in the active vs. passive voice in the Penn Treebank corpus, suggesting that such a corpus is indeed reliable. This finding validates the reliance on the Penn Treebank in developing the stimuli for subsequent experiments. 2. Experiment 2 Since Experiment 1 demonstrated the existence of voice biases for verbs in English, it seems reasonable to ask w hether these biases can influence sentence comprehension. Thus the aim of Experiment 2 was to test the hypothesis that passive sentences would be easier to com prehend when they contained passive-bias rather than active-bias verbs and conversely that active sentences w ould be easier to parse w hen they contained active-bias rather than passive-bias verbs. If contingent frequency (i.e. verb bias) effects are obtained, then this would constitute strong evidence for the constraint- based model and for statistical approaches to language. If, however, there is no effect of contingent frequency, then principle-based parsing would be supported. The second purpose of Experiment 2 was to compare overall reading times on active sentences with reading times on passive sentences. Indeed, the literature reviewed in Chapter n has given a picture that is not entirely consistent. While some studies found clear advantages for actives over passives in production, processing and m em ory tasks, others found no evidence for such differences. If structural firequency effects are obtained along with contingent fi-equency effects then one w ould have evidence for the idea that people keep track of various statistical aspects of linguistic stimuli 105 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. simultaneously. Experiment 2 utilized a self-paced reading paradigm in order to examine subjects' comprehension processes, thus providing data that are relevant for both of the issues under discussion. Subjects 40 u s e undergraduates participated in this experiment in exchange for course credit. All subjects w ere native speakers of American English and had norm al or corrected-to-normal vision. Selection of materials 24 verbs were selected from the active-bias and the passive-bias groups identified in Experiment 1, and placed in active and passive sentence contexts. Pilot work on Experiment 2 showed that, for comparisons of reading times at verb regions, it w as im portant to control for the appropriateness of the agent with respect to the verb, a notion that has come to be referred to in am biguity resolution studies as thematic fit. In order to take this factor into account in the experiment, the stimuli were divided up into two lists (one for each type of sentence context), and given to 20 USC undergraduates^^ who were asked to rate the appropriateness of the agents for the verbs selected. Stimulus items were presented as pairs of agents and verbs, w ith the agent on the left side and the verb on the right side, as shown below. Subjects were asked to decide on a scale of 1 to 7, how appropriate the agent w as for that verb, with 1 indicating a very good agent, and 7 indicating a very bad agent. All verbs were presented in the past tense. the entire committee asked 1 2 3 4 5 6 7 the company chairman appointed 1 2 3 4 5 6 7 ^ The subjects received course credit for their participation. 106 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Planned comparisons revealed that sentences w ith active-bias verbs did not differ in terms of thematic fit from sentences with passive-bias verbs (t(22) = .73). In addition to ratings for thematic fit, the stimuli were also rated by another set of 20 subjects for appropriateness of the patient in relation to the verb in passive sentences. The same m ode of presentation was used. the new representative was asked 1 2 3 4 5 6 7 the prom inent banker was appointed 1 2 3 4 5 6 7 U npaired t-tests confirmed that the active-bias and passive-bias groups did not differ in terms of the appropriateness of the patient to the verb (t (22) = 1.11). The stimulus items were also m atched w ith respect to their overall frequency, as measured by both the Wall Street Journal corpus (t (22) = .37) and the Kucera & Francis corpus (t (22) = 1.60). As expected, the stimulus items did, however, differ in terms of their relative frequency of use in the active vs. passive voice in the Penn Treebank corpus, w ith passive-bias verbs appearing more frequently in the passive voice. Furthermore, active-bias and passive-bias verbs differed significantly in verb rating. Materials The stimuli consisted of 24 experim ental sentences, each appearing in 4 versions. The experiment m anipulated two variables: sentence type (active vs. passive) and verb type (active-bias vs. passive-bias). Each sentence began w ith a prepositional phrase, followed by a subject noun phrase, the verb, an object noun phrase and an ending complem ent phrase or adjunct. The introductory phrase was three words in length for all the stimulus items. All sentential arguments were definite, animate, and equal in length. Table 3 107 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. presents example stimuli, and the full list of sentences is given in A ppendix m . The experimental sentences w ere intermixed with 45 filler items, which w ere sentences for an unrelated experiment. Subjects were also given 10 practice items so as to practice reading using the spacebar. Table 3: Example stimuli for Experiment 2 Active After several discussions, the entire conunittee asked the competent lawyer to represent the new firm, (active-bias) After several discussions, the entire committee elected the competent lawyer to represent the new firm, (passive-bias) Passive After several discussions, the competent lawyer was asked to represent the new firm, (active- bias) After several discussions, the competent lawyer was elected to represent the new firm. (passive-bias) Design The experiment had a 2x2 design: sentence type and verb type w ere manipulated. Each subject received all the experimental items, but in only one of the four conditions. Procedure The experiment consisted of a self-paced reading task. Subjects w ere instructed to press the space bar on a keyboard to see each word of the sentence on a computer screen. At the beginning of each trial, a series of hyphens appeared on the screen. Subjects pressed the spacebar and the first series of hyphens w ould be replaced by the first word of the sentence. Each subsequent keypress revealed a new word, and replaced the previous w ord w ith hyphens. Reading times represented the time in milliseconds that subjects spent between two spacebar presses. Comprehension questions for all 108 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. experimental sentences appeared at the end of each trial. Subjects answered "Yes" or "No" by pressing keys on the keyboard, and were given feedback on the accuracy of their response. Each subject saw a different random ization of the experimental and filler item s and did the experiment in less than 25 m inutes. Results The analyses w ere conducted on adjusted rather than actual reading times (Ferreira & Clifton, 1986). Regression equations predicting reading time from word length were com puted on the basis of each subject's reading times in both the experimental and the filler items. The predicted reading times w ere then subtracted from the actual reading times. Thus, a negative reading tim e on a given sentence region means that subjects w ere faster in that region than would be predicted by w ord length alone. The purpose of this m anipulation was to reduce baseline subject differences in reading rate and to adjust for subjects' sensitivity to length. For the purposes of data analyses and comparisons, the w ords of each stim ulus sentence were grouped in 5 sentence regions, show n in Table 4. Region 1 was comprised of the three words in the introductory prepositional phrase. Region 2 represented the subject of the sentence (that is, the agent phrase in active sentences and the theme in passive sentences). Region 3 consisted of the verb. This was w ord 7 for the active sentences, and w ords 7 and 8 for the passive sentences. Region 4 grouped together the object of the verb for active sentences (words 8 ,9 and 10) and the complement prepositional phrase for passive sentences (words 9,10 and 11). Region 5 was composed of the rem aining w ords of the sentence. 109 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 4 - Sentence regions for item s in Experiment 2 Region 1 Region 2 Regions Region 4 RegiimS After several discussions the entire committee asked the competent law yer to represent the new firm. After several discussions the entire committee elected the competent law yer to represent the new firm. After several discussions the competent lawyer was asked to represent the new firm. After several discussions the competent lawyer was elected to represent the new firm. Notice that the same w ords (e.g. to represent the’ ) are placed in different regions, depending on whether the sentence is active or passive. This makes overall comparisons of active sentences with passive sentences a little difficult. However, it does not affect comparisons w ithin passive sentences or within active sentences, contrasts which are crucial. N ote also that the active and passive sentences differed in length by two words. This difference was systematic across all experimental items. Thus, active sentences had a mean length of 15.16 (SD = 1.30) while the passive sentences had a m ean length of 13.16 (SD = 1.30), a non-significant difference. The length-adjusted data w ere trimmed by disregarding RTs that were 2 SD above or below the appropriate condition m ean (voice/ verb / region). All trials on which the comprehension question was incorrectly answ ered were rem oved prior to doing the analyses. A 2x2x5 ANOVA was then ru n with voice, verb type and sentence region as factors. Subjects' overall performance on active and passive sentences is shown in Figure 3. There was no m ain effect of voice (Fi (1,39) = 2.42, F2 (1,21) = 1.75). However, there was a reliable Voice X Region interaction (F % (4,156) = 8.38, p < .0001, F % (4, 84) = 7.41, p < .0001), suggesting that differences in reading times in active vs. passive 110 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. conditions appeared in certain regions only. Pairwise comparisons at each sentence region confirmed that there were significant differences^^ between active sentences and passive sentences in reading times at the end of the sentence (Fi (1,39) = 18.67, p <.0001, Fz (1,21) = 8.26, p < .OD.^s Experiment 2 - Reading Times by Voice Type A ctive sentence P assive sen ten ce I ' I 40- f I ■ ? n- P < -2 0- W ) 5 -60 NP1 VP Region NP2/PP End Figures There was also a marginally reliable Verb x Voice interaction (Fi (1, 39) = 2.65, p = 11), indicating that differences in reading times between active-bias and passive-bias verbs appeared in only one voice type. Subjects' performance ^ Arrows in Figures 3-10 indicate significant differences. ^ Two items had to be removed from the F2 analyses because they contained eleven words only, which yielded empty cells at region 5. I l l Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. in the passive voice is presented in Figure 4. Pairwise comparisons confirmed this interaction, w ith reading times in the passive voice being significantly faster for passive-bias verbs than active-bias verbs at the verb region (Fi (1,39) = 3.55, p =.06), the postverbal region (Fi (1,39) = 3.00, p =.09) and at the end of the sentence (F% (1,39) = 6.49, p < .01), although the items analyses were not significant. Subjects' performance in the active voice is show n in Figure 5. There was no overall trend, and reading times for active sentences did not differ reliably at any sentence region. At first sight, this lack of effect might be surprising; one w ould expect active-bias verbs to be easier w hen they appear in the active voice. However, such an effect was not observed. One explanation for this fact is that, being exposed to active sentences more often than to passive sentences, comprehenders of English develop far better reading skills w ith the active than with the passive sentence frame. It is possible that this eliminates any differences between passive-bias and active- bias verbs in the active frame. 112 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Experiment 2 - Reading Times in the Passive Voice A ctive-bias v e ib P assiv e-b ias v e ib I H . r I 13 f -20 bp J -40 -60 AficrMvaldôaaiioM theœmpetait lawyer was aaked/efeded loiepmmtthe Region Figure 4 Experiment 2 - Reading Times in the Active Voice A ctive-bias v eib P assive-bias v erb 1 3 I 1 I -20 -60 Aâeraevenldtfoiasions the entire cominttee asked/elected thecompeterth%vyer torepresodthenewfinn Region Figures 113 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Discussion Although there was no main effect of voice in Experiment 2, active sentences were relatively easier to read than passive sentences at the end of the sentence. Notice that this result cannot be attributed to other confounding variables. First, the difference in length between active and passive sentences in this experiment w as systematic and minimal (in fact, it was not significant). Second, the result of greater difficulty w ith passives could not be due to complexity alone. Recall that passive sentences in this study w ere agentless, so that there were fewer thematic roles to assign in passive than in active sentences. If complexity played an im portant role in the case at hand, then passive sentences should have been easier than active sentences; they were not. In other words, results of slightly greater difficulty with passive sentences could not be due to length, complexity or discourse factors. A second result of Experiment 2 is that passive sentences w ith passive- bias verbs tended to be read faster than sentences w ith active-bias verbs. Since thematic fit and overall verb frequency were controlled for, it seems reasonable to conclude that contingent frequency alone was at the source of this effect. The more frequently a verb is used in the passive voice, the more activation its passive node receives, and the easier it becomes to process it in the passive voice. This result indicates that com prehenders are sensitive to subtle co-occurrence frequencies, which suggests that they som ehow keep track of the different uses of individual verbs in the active and passive voices. Finally, notice that active-bias verbs appear alm ost exclusively in the active voice, so that one would expect them to m ake easier-to-comprehend active sentences than passive-bias verbs. However, there appears to be no 114 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. difference betw een the ways in which these two types of sentence are read. Differences between active-bias and passive-bias verbs in the active voice seem to be eliminated, probably because of the effect of structural frequency. Active sentences are far m ore frequent than passive sentences, and English speakers can be said to no longer be sensitive to differences affecting actives. Research on w ord recognition suggests similar patterns of effects: orthographic regularity effects are only observed for low frequency words, not for high frequency ones (Seidenberg & McClelland, 1989; Taraban & McClelland, 1987). This kind of result is compatible w ith a constraint-based model such as the one assum ed in this study. The pattern of activation over units for high frequency words or structures allows them to reach conscious awareness long before low frequency words or structures, thereby elim inating any differences that m ight have existed between high frequency items. In sentence processing research, Juliano and Tanenhaus (1994) also identified a frequency-by-regularity interaction. They found that, while reading times for the determiner the were significantly slower for low frequency verbs that take sentential complements (e.g. hinted the...) than for low frequency verbs that take direct object complements (e.g. invited the...), there were no differences in reading times when the verbs were high in frequency (e.g. studied the... vs. realized the...). The absence of a contingent frequency effect in active sentences in this study therefore signals a high degree of reading skill for these kinds of sentences (Seidenberg & McClelland, 1989). In summary. Experiment 2 dem onstrated that both structural and contingent frequency effects appear in parsing, but that the effects of contingent frequency may be modulated by structural frequency. 115 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 3. Experiment 3 Experiment 2 provided some evidence that voice bias affected the comprehension of passive sentences. However, an alternative explanation for these data may be that agentless passive sentences are infelicitous when presented in isolation, as they were in Experiment 2. By chance, it m ay be that the active-bias verbs w ere adversely affected by this lack of previous context for agentless passives and that this contributed to slower RTs on active-bias verbs. In order to address this concern. Experiment 3 uses full rather than short passives. The aim of this experim ent was thus to see w hether voice effects would hold with agented passives. Subjects 40 u s e undergraduates participated in this experiment in exchange for course credit. All subjects were native speakers of American English and had norm al or corrected-to-normal vision. Materials The stimuli consisted of the 24 experimental sentences of Experiment 2. Passive sentences, however, appeared in their agented form. Example stimuli are given in Table 5 below. 116 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 5: Example stimuli for Experiment 3 Active After several discussions, the entire conunittee asked the competent lawyer to represent the new firm, (active-bias) After several discussions, the entire committee elected the competent lawyer to represent the new firm (passive-bias) Passive After several discussions, the competent lawyer was asked by the entire committee to represent the new firm, (active-bias) After several discussions, the competent lawyer was elected by the entire committee to represent the new firm (passive-bias) Design and Procedure The design and procedure were identical to those of Experiment 2. Results The w ords of each stimulus sentence in Experiment 3 were grouped in 5 sentence regions, as was done for Experiment 2. In the present experiment, however, region 4 grouped together the theme in active sentences (words 8, 9 and 10 ) and the by-phrase for passive sentences (words 9,10,11 and 12). Region 5 consisted of the remaining words in the sentences. Table 6 - Sentence regions for items in Experiment 3 Region 1 Region 2 Regions Region 4 Regions After several discussions the entire committee asked the competent law yer to represent the new firm. After several discussions the entire committee elected the competent law yer to represent the new firm. After several discussions the competent law yer was asked by the entire committee to represent the new firm. After several discussions the competent law yer was elected by the entire committee to represent the new firm. The length-adjusted data were trimmed by disregarding RTs that were 2 SD above or below the appropriate condition m ean (voice/verb/region). All 117 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. trials on which the com prehension question was incorrectly answ ered were rem oved prior to doing the analyses. A 2x2x5 ANOVA was then ru n with voice, verb type and sentence region as factors. Subjects' overall performance on active and passive sentences is shown in Figure 6. There was no effect of voice (Fi (1,39) = ,50, F2 (1,23) = .10). There was, however, a reliable Voice x Region interaction (Fi (4,156) = 3.97, p < .005, F2 (4,92) = 4.16, p < .005), suggesting that differences in reading times in active vs. passive conditions appeared in certain regions only. Pairwise comparisons at each sentence region confirmed this trend: active sentences w ere read m arginally faster at the verb region (Fi (1,39) = 2.47, p =.12, Fz (1, 23) = 3.34, p = .08) while passive sentences were read significantly faster at the N P 2/P P region (Fi (1,39) = 5.96, p <.01, F2 (1,23) = 13.01, p < .001). E xperim ent 3 - R eading T im es b y Voice Type A ctive sen ten ce P assiv e sen ten ce H T3 I 1 -20 bp g -40 N - J -60 NPl NP2/PP End Region Figure 6 118 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. There was also a reliable Verb x Region interaction (Fi (4,156) = 2.79, p < ,05 ), indicating that there were differences in reading times between active- bias and passive-bias verbs. Subjects' performance in the passive voice is presented in Figure 7. Pairwise comparisons confirmed this interaction, with reading times in the passive voice being reliably faster for passive-bias verbs than active-bias verbs at the verb region (Fi (1,39) = 3.35, p =.07, F% (1,23) = 2.05, p = 16) and at the end of the sentence (Fl (1,39) = 6.34, p < .05, F% (1,23) = 4.76, p < .05). Subjects' performance in the active voice is shown in Figure 8 below. There was no overall trend, and reading times for active sentences did not differ reliably at any sentence region. Experiment 3 - Reading Times in the Passive Voice 60 I I 40 ■ a < 20 0 - -20 -40- -60 A ctive-bias v eib Passive-bias v e ib I j - - r AftersevenldÎKUflsions the competent lawyer was asked/elected by the entire committee to teptesent the new firm Region Figure 7 119 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Experiment 3 - Reading Times in the Active Voice 6 0 I I 40 « ^ 20 I " 3 0 Î -40 A dive4> ias v e ib P assiv e-b ias v e ib After Kvenl dacimkna theenU i to om m iH ee T ----------- r ■ I askcd/decled thecompdentlaivyer loicpiaentlhcwwfinn Region Figures Discussion Experiment 3 displayed no effect of voice; however, the reliable interaction of voice and verb typewas such that active sentences were easier to process at the verb region while passive sentences benefited from the predictability of the by-phrase. This result, taken together with that of Experiment 2, suggests that structural frequency effects m ay be observed but that they are m odulated by semantic constraints such as predictability. More im portantly, the effect of voice bias obtained in Experiment 2 was replicated in Experiment 3 as well. Passive sentences w ith passive-bias verbs tended to be read faster than sentences with active-bias verbs at the critical regions. The robustness of this effect provides strong evidence that contingent frequency affects sentence comprehension, a finding which 120 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. strongly supports the constraint-based approach, but is not predicted by principle-based theories such as the garden-path model. Given that the contingent frequency effect is robust, one is forced to conclude that the lexical representations of individual verbs include information about the structures that they appear in. The null result in the active sentences was again observed in Experiment 3 and lends further support to the hypothesis that structural and contingent frequencies interact in such a w ay that the effect of contingent frequency is m asked in the high frequency voice. If this claim is correct, then given a situation in which the disparity between the two types of frequency is reduced, the effect of verb bias should be clearly observed. This is what Experiment 4 attem pts to accomplish. 4. Experiment 4 Experiments 2 and 3 validated the idea that voice biases influence sentence processing in relatively unam biguous active and passive sentences. In Experiment 4 , 1 investigate these biases as they relate to thematic ambiguity resolution. Passive sentences contain a tem porary ambiguity as to the interpretation of the post-verbal prepositional phrase. Thus, the fragment the shrubs were planted could end with inform ation about the agent of the planting e.g. by the head gardener or about its location, e.g. by the greenhouse. These ambiguities contrast an argum ent interpretation, the agentive one, w ith an adjunct one, the locative one.^^ ^ Arguments are usually defined as constituents that bear a thematic relationship with the predicate while adjuncts are constituents that may be removed without affecting the structural identity of the construction. 121 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Argument-adjunct ambiguities have been the subject of m uch debate in recent sentence com prehension literature. Thus, Pickering, Liversedge, Branigan and van Compel (1996) found a preference to interpret the ambiguous phrase as an agentive when the sentences were presented in isolation. This preference is predictable, given that the agentive interpretation seems to be m uch m ore frequent than the locative one. However, when a context supporting the locative interpretation was introduced, the locative and agentive interpretations were equally easy to interpret This result led Pickering et al. to postulate that the parser always proposes an optional argum ent analysis rather than an adjunct analysis. In contrast, Spivey-Knowlton, H anna and Tanenhaus (1996) found an effect of context for both the agentive and the locative targets, when each was preceded by a supportive context. Both Pickering et al. and Spivey-Knowlton et al. used eye-tracking m easurem ents in virtually identical experiments, yet obtained contradictory results. It is possible that the differences in the results obtained in the two studies could be due to the items used. Given the results of Experiments 2 and 3, one m ight venture that the items in the two experiments differed in terms of their voice biases. An analysis was thus conducted in which the relative voice bias of the verbs from each experiment was compared. Results showed that in both Pickering et al.'s and in Spivey-Knowlton et al.'s experiments, the verbs tended to be active-bias verbs. The verbs in Liversedge et al.'s study had a m ean percentage of passive use of 29.24% while those in Spivey-Knowlton et al.'s study had a mean percentage of passive use of 18.02%. The two sets of verbs did not differ reliably in terms of voice bias (t (25) = 1.46). However, one 122 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. m ight also want to compare these verbs in terms of any possible argum ent vs. adjunct biases. Indeed, since voice biases do exist for individual verbs, it is conceivable that some verbs could be biased towards an argum ent interpretation while others could be biased towards an adjunct interpretation. Using the Penn Treebank database, the passive usages of the verbs in each experim ent were coded as involving either a locative or an agentive interpretation. Results revealed that Pickering et al.'s items had a lower locative frequency (7.69%) than Spivey-Knowlton et al.'s items (26.14%), a statistically signiHcant difference (t (25) = 2.63, p < .05). Thus it is quite possible that the lack of difference in RTs between the locative and the agentive interpretations in the presence of a locative supportive context in Pickering et al.'s study might be due to the weak biases of the verbs used in that study. A third point on which one m ight w ant to compare the two studies is that they both restricted themselves to two possible continuations for the am biguous by-phrase, i.e. the locative and the agentive interpretation. Notice, however, that the by-phrase is in fact multiply ambiguous. Indeed, it can also be interpreted as a tim e-adverbial as in the transaction was completed by eleven thirty a.m. or as a m anner-adverbial as in the report was illustrated by adding nice figures. Both of these interpretations are adjunct ones. Thus, we have a situation in which the argum ent interpretation (i.e. the agentive) is the m ost frequent interpretation while the adjunct interpretation (i.e. the locative, time or manner adverbial) is overall less frequent.^^ However, there are also verbs that are biased towards one or the other of these interpretations. For example, the passive version of the verb park often Hanna, Spivey-Knowlton and Tanenhaus (1996) report that "out of over 300 passive sentences containing by-phrases, not a single one introduced a location, while agents were frequent" (see also Hanna, Barker & Tanenhaus, 1995). 123 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. involves a location adjunct while the verb drive frequently involves an agent argum ent, hence the clear contrast betw een the car was parked hy the building entrance and the car was driven by the building entrance. In sum m ary, the argum ent vs. adjunct am biguity in passive sentences is subject to constraints that are very similar to those examined in Experiments 2 and 3. Structural frequency (the overwhelming frequency of the argum ent interpretation) contrasts w ith contingent frequency (the individual biases of certain verbs). This time, however, the difference in structural frequency is not quite as large. Indeed, Experiments 2 and 3 compared actives with passives, while Experiment 3 compares a low frequency structure, the passive, in two of its forms (argument vs. adjunct). It is thus conceivable that one could observe effects that were masked in the previous experiments. The fact that Experiment 4 deals with ambiguous sentences only increases the chances of finding such effects. Subjects 36 native speakers were paid $5 for their participation in this experiment. All w ere native speakers of American English. Materials 20 verbs w ere selected from the Penn Treebank database and coded for whether their passive usages contained argum ent or adjunct complements. The 10 argum ent-bias verbs had significantly higher argument-frequency than the 10 adjunct-bias verbs. However the two groups did not differ in overall frequency as measured by both the Wall Street Journal (t (18) = .61) and the Kucera and Francis (t (18) = 1.23) counts. The stimuli consisted of a 124 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. noun phrase and a passive verb followed by an agentive, locative, time, or manner by-phrase. Sample materials are presented in Table 7 below. Table 7- Sample stim uli for Experiment 4. Argument Continuation The car was driven by the parking attendant, (argument bias verb) The car was parked by the parking attendant, (adjunct bias verb) Adjunct Continuation The car was driven by the building entrance, (argument bias verb) The car was parked by the building entrance, (adjunct bias verb) Design The experim ent employed a 2x2 design w ith continuation (argum ent vs. adjunct) and bias (argument vs. adjunct) as factors. Procedure Same as in Experiment 2. Since the 20 experimental sentences were all exactly 9 words in length, and fit on one line on the computer screen, all of the 70 filler sentences w ere less than 80 characters long (i.e. they fit in one line on the screen as well). Results The words of each stimulus sentence w ere grouped in 4 experimental regions. Regions 1, 2 and 3 consisted of 2 w ords each (e.g. the car, was parked, by the). Region 4 com prised the adjective and the final noun (e.g. parking attendant). Region 4 w as thus the critical region. 125 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 8 - Sentence regions for item s in Ex]périment 4 Region 1 Regi<Hi2 Regions Region 4 The car was parked by the parking attendant The car was parked by the parking attendant The car was driven by the building entrance The car was driven by the building entrance All analyses were conducted on length-adjusted reading time data, w hich were then trim m ed by rem oving any reading times that w ere 2 SD above or below the appropriate condition m ean (continuation/bias/region). Item s on which participants gave an incorrect answer were rem oved prior to the analysis. A 2x2x4 factorial ANOVA was run with continuation, verb bias and sentence region as factors. There were no main effects of continuation or verb bias. There was, however, a reliable interaction between continuation and sentence region (Fi (3,105) = 3.93, p < .01; P2 (3,57) = 3.57, p < .05). In addition, there was a continuation by verb bias interaction (F % (1,35) = 5.01, p < .05; F2 (1,19) = 2.85; p = .12) and a three-way interaction between continuation, verb bias, and sentence region (F % (3,105) = 4.37, p < .01; f 2 (3, 57) = 3.02; p < .05). Reading times in the adjunct condition are show n in Figure 9 below. Pairwise comparisons at each sentence region revealed that sentences containing adjunct-bias verbs (e.g. parked) were significantly easier to disambiguate in the adjunct continuation condition, as evidenced by faster reading times at the end of the sentence (Fi (1,35) = 4.97, p < .05; F2 (1,19) = 6.13; p < 05). Reading times in the argum ent condition are shown in Figure 10 below. Comparisons between reading times at each sentence region showed th at sentences containing argum ent-bias verbs (e.g. driven) were reliably easier to disambiguate in the argum ent continuation condition, with faster RTs at the end of the sentence (F % (1,35) = 4.91, p < .05; F2 (1,19) = .93; p > .10). 126 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Experiment 4 - Reading Times in the Adjunct Condition I T3 < S I I Î A ig u m en t-b iaa veib A djunct-biaa v eib I -20 ^ -60 -80 by the Region Figure 9 Experiment 4 - Reading Times in the Argument Condition 80 60 40 20 0 -20 -40 -60 -80 A ig u m en t-b ias veib A djunct-bias v eib I The car wasdriven/puiaed — '1 by the Region puting attendant Figure 10 127 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Discussion The results obtained in Experiment 4 can be summarized succinctly. Adjunct-bias verbs allowed for a faster disam biguation than argument-bias verbs in the adjunct-continuation condition. The reverse obtained in the argum ent-continuation condition. Furtherm ore, sentences disam biguated in favor of the adjunct interpretation took significantly longer to read than those ending with the argum ent interpretation, regardless of the verb used. These findings seem to be compatible w ith Spivey-Knowlton et al.'s (1996) but not w ith Pickering et al.'s (1996). Indeed, Spivey-Knowlton et al. found an effect of context for both the argum ent and the adjunct conditions, while Pickering found this effect only for the argum ent condition. Given that Spivey-Knowlton's items had an overall higher frequency of use in the passive w ith locative complements, the difference in results between the two studies follows directly. The results obtained in Experiment 4 also indicate that contingent frequency has an impact on the disam biguation of passive argum ent vs. adjunct ambiguities. Structural ft-equency effects are also seen in the increased length of time (regardless of verb bias) that it takes to disambiguate the adjunct interpretation relative to the argum ent interpretation. Notice that contrary to w hat was the case for Experiments 2 and 3, contingent frequency effects are not masked by global frequency here. The lack of effects for the active sentences in those experiments could have been construed as signaling that com prehenders do not in fact keep track of contingent frequencies. How ever, the presence of contingent frequency effects in Experiment 4 in the high-frequency (i.e. the agentive) as well as in the low-frequency structure 128 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (i.e. the adjunct) dispels this doubt. These findings confirm the idea that comprehenders automatically keep track of contingent fi-equency. The data obtained therefore indicate that com prehenders are sensitive not only to global or structural sorts of frequencies, b u t even to subtle co-occurrence fi'equendes. This lends support to the view that the hum an sentence processing m echanism is capable of encoding various statistical patterns in the language and exploit them during on-line sentence comprehension. 129 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER V: GENERAL CONCLUSIONS Since the 1970s, sentence com prehension research has tended to focus on the manner in which structural ambiguities are resolved. For instance, m any studies have investigated the w ay in which the correct interpretations of sentences such as (la) and (lb) are eventually recovered (Rayner, Carlson & Frazier, 1983; Ferreira & Henderson, 1991; 1993; Spivey-Knowlton, Trueswell & Tanenhaus, 1993; Pearlm utter & MacDonald, 1995; Trueswell, Tanenhaus & Gamsey, 1994; MacDonald, 1994). W hile (la) is ambiguous between a main verb and a reduced relative reading, (lb) is ambiguous between a direct object and a sentential com plem ent construal. The M V/RR and the N P /S ambiguities represent well the kinds of structures that are typically investigated in current sentence processing studies. (la) The horse raced past the bam fell. (lb) Jane knew the answ er to the problem was on the back of the handout. This focus on am biguity resolution is motivated by two im portant reasons. Firstly, any theory of processing m ust be able to account for the way in which the correct interpretation of sentences such as (la) and (lb) is ultim ately recovered w ith great efficiency. Secondly, ambiguity resolution tends to slow down the language parser, thus providing researchers with a w indow on the m echanisms that underlie syntactic processing. Thus ambiguity resolution is one of the best paradigm s for sentence com prehension research. 130 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. However, it is im portant to note that sentences such as (la) and (lb) are so strongly ambiguous that they tend to halt the processing system. Com prehenders usually become conscious of the existence of the ambiguity and of the possibility of alternative readings as they reach the second verb. Thus, sentences such as (la) and (lb) are at the extreme end of a continuum of ambiguity. Indeed, sentences can be ambiguous to different degrees, and sentences (la) and (lb) w ould probably be considered by m ost readers as being severely ambiguous. Clearly, there exist many other types of sentences (e g. the competent lawyer was elected by the entire committee to serve as representative) which are not as ambiguous as (la) and (lb). Unfortunately, these relatively unam biguous sentences have not received as much attention in sentence comprehension research. The point to be m ade here is that comprehenders are hardly ever aware of syntactic ambiguities precisely because the sentences that they encounter are not as ambiguous as those traditionally studied in sentence processing studies. Therefore, a complete theory of sentence processing m ust account for the entire continuum of ambiguity, i.e. for both am biguous and relatively unam biguous sentences, in order to be explanatorily adequate. If a given theory can provide a complete account of how ambiguities are resolved, as well as characterize relatively unam biguous sentences, it will demonstrate that it is more than just a set of clever mechanisms for disambiguation, but rather a complete account of the nature of hum an language processing. Current theories of sentence processing m ake different claims about how syntactic ambiguities are resolved. These theories can be classified under two broad headings, serial or parallel. Proponents of serial models claim that only one reading of an am biguous fragment is accessed at a time, in 131 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. accordance with general syntactic principles (Frazier, 1978; 1987; 1995; Rayner, Carlson & Frazier, 1983; Frazier & Rayner, 1987; Frazier & Clifton,1996). On the other hand, supporters of parallel models argue that all possible readings of an ambiguous fragment are activated to a degree that is comm ensurate w ith their frequency in the language (Trueswell & Tanenhaus, 1991; Trueswell, Tanenhaus & G am sey, 1994; Tanenhaus & Trueswell, 1995; MacDonald, 1993; 1994; M acDonald, Pearlmutter & Seidenberg, 1994). If unam biguous sentences constitute an important part of the array of data that theories of processing m ust account for, the question that arises is how serial and parallel models w ould deal w ith unambiguous sentences. The best-known serial m odel of processing is the garden-path model of Frazier and her colleagues (Frazier, 1978; 1987; 1995; Rayner, Carlson & Frazier, 1983; Frazier & Clifton, 1996). In this model, syntactic principles such as m inim al attachment and late closure govern the processing of ambiguous sentences. These principles w ere developed specifically to account for am biguous structures such as those in (1), so it is som ew hat unclear how they can apply to unambiguous structures. In processing these latter types of sentences, one must decide w hether to use the general attachm ent principles of late closure and minimal attachment or the association principle of construal. Primary phrases (in general, predicates and their argum ents) should be processed in accordance with attachment principles while non prim ary phrases (adjuncts and other optional elements) should be processed through association rules. In neither case, however, is contingent frequency allowed to inform the parser’ s choice until after the first analysis has been m ade. 132 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. On the other hand, parallel m odels make the prediction that unambiguous and ambiguous sentences are processed via the same mechanisms. Indeed, one of the fundam ental claims of the constraint-based model is that the traditional separation between lexical ambiguities and syntactic ambiguities has no basis. Proponents of this approach have presented strong evidence suggesting that lexical and syntactic ambiguities are resolved through the same mechanisms (see MacDonald, Pearlm utter & Seidenberg, 1994, for a review). There is no reason to assum e that there would be separate mechanisms for am biguous and unam biguous sentences in the constraint-based approach either. Unam biguous sentences would be argued to be processed by reference to the same probabilistic constraints as ambiguous sentences. The experim ental research undertaken in this dissertation provides data that are relevant for these issues. The passive construction was selected for the self-paced reading experiments because it is a relatively unambiguous structure and because it provides a unique opportunity for studying the kinds of frequency effects that constitute the bone of contention between serial and parallel models. Frazier (1995) writes: According to the view held by those working in a constraint-satisfaction framework, all the action in sentence processing— the heart of the problem-concems the processor's use of frequency and contextual probability information. Therefore, it would seem incumbent on those pursuing this approach to explicitly specify in advance of an experimental test of the theory which frequencies should matter and why. This matter is two-sided. On the one hand, proponents of constraint-satisfaction models, and others (see Gibson, Schütze & Salomon, 1995; Merlo, 1994; Mitchell, Cuetos & Corley, 1995), deserve credit for empirical attempts to sort out which frequencies matter. On the other hand, the resulting evidence, though not predicted in advance, has often been taken as evidence favoring a constraint-satisfaction model. Surely, such claims are premature in the absence of an articulated account of which frequencies should matter. If it is contingent frequency, then the model should explicitly define contingent frequency. Is it computed over a two-word window, a five-word window, a 25-word window? Is it contingent frequency given a string of syntactic categories, or does the higher syntactic 133 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. structure matter? Or is it computed over semantic categories and, if so, which ones? Frazier (1995), pp. 447-448. Frazier's criticism can be sum m arized as follows. Proponents of the constraint-based m odel should determ ine w hich frequencies m atter in comprehension. Specifically, they need to show whether structural frequency, or contingent frequency, or both, affect processing, and also define how these frequencies are com puted. The passive construction is an excellent paradigm for answering these questions. Indeed, passive sentences are m uch less frequent in English than active sentences (Svartvik, 1966; Givon, 1979). The difference in structural frequency between actives and passives is therefore a good area for the investigation of w hether general sorts of frequencies affect sentence comprehension. However, although actives are far m ore frequent in English than passives, there exists a class of verbs that are m ore often used in the passive than in the active (e.g. sentenced and elected). These verbs, referred to as passive-bias verbs, are contrasted with the vast m ajority of verbs, which are active-bias verbs (e.g. sued and asked). The contingent frequencies of individual verbs can thus coincide with structural frequency (as is the case for active-bias verbs) or run counter to it (as is the case for passive-bias verbs). To summarize, current models of processing m ust be able to account for unambiguous structures as well as am biguous ones. If frequency turns out to play a role in their processing, then this w ould constitute additional evidence for the constraint-based model. If, however, frequency does not affect the processing of unambiguous sentences, then principle-based models such as the garden-path theory will be supported. 134 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 1. The processing of unambiguous sentences Experiments 2 and 3 investigated processing in relatively unam biguous active and passive sentences. Experim ent 2 contrasted short passives w ith active sentences while Experiment 3 compared full passives w ith active sentences. The garden-path model predicts that actives should be easier to process than passives because the form er are structurally simpler than the latter. However, it does not predict that passive sentences such as the competent lawyer was asked by the entire committee to represent the firm should be read any differently from the competent lawyer was elected by the entire committee to represent the firm because the difference between these tw o sentences is m erely lexical. Neither does the garden-path model predict contingent frequency differences for the passive sentences. On the other hand, the constraint-based model predicts that passive-bias verbs should be easier to process in the passive and that active-bias verbs should be similarly easier in the active. As was shown in Chapter IV, contingent frequency effects were identiffed in both Experiment 2 and 3, with passive-bias verbs having a signiffcant advantage over active-bias verbs in passive sentences. Structural fi’ equency did not seem to have a very strong effect: Active sentences did not seem to be easier to process, except at the sentence-final region. However, structural and contingent frequency interacted in both experiments in such a w ay that the effect of contingent frequency (i.e. the effect of verb bias) was significantly larger in the low-frequency voice type (i.e. the passive) than in the high-ft'equency voice type (i.e. the active). This interaction is strikingly similar to the frequency-by-regularity interaction observed in word processing (Seidenberg & McClelland, 1989). It is a well-attested fact in the word 135 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. recognition literature that while low frequency exception w ords such as caste take longer to nam e than low-frequency regular w ords such as waste, this difference is practically elim inated for high frequency w ords, so that it does not take subjects any longer to nam e have than gave. Thus, the results from Experiments 2 and 3 pattern closely w ith well-known findings in w ord recognition. Notice also that this interaction was also present in some of the sentence comprehension studies reviewed in Chapter HI. For instance, W right (1969) found that subjects had more trouble answering comprehension questions that w ere form ulated in a different grammatical voice from target sentences, and that this difficulty was m uch more pronounced in the passive question condition. Similar asymmetries were identified by Prentice (1966) and Taimenbaum and Williams (1968b). Thus, it seems plausible that the high structural frequency of the active voice could prevent certain effects from appearing in tasks involving unam biguous actives. The findings reported here have three major implications for the issues reviewed in the introduction. One is that the investigation of unambiguous sentences has revealed a pattern that is consistent w ith constraint-based models but not w ith serial models of parsing such as the garden-path model. Even in the m ost liberal interpretation of Frazier’ s theory, one finds that frequency is only allowed to play a role after the parser has already selected one interpretation of an ambiguous sentence. W ith unambiguous sentences, there w ould be even less reason for frequency to intervene, and effects such as those of Experiments 2 and 3 would be left unexplained. The second im plication is that the interaction obtained between structural and contingent frequency, which is similar to results from w ord 136 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. recognition w ork, provides support for the view that syntactic and lexical processing should be integrated (MacDonald, Pearlm utter & Seidenberg, 1994). Finally, these data dem onstrate that there is no distinction betw een ambiguous and unam biguous sentences in terms of processing. Both of these types of structures are processed using multiple sources of probabilistic inform ation. 2, The processing of them atically am biguous sentences The lack of verb bias effects in the active sentences in Experiments 2 and 3 was attributed to the high frequency of the active frame. If this claim is correct, then given a situation in which the frequency difference between the two structures is not as large as that between actives and passives, verb bias effects should appear. Accordingly, Experiment 4 investigated the processing of argum ent vs. adjunct ambiguities in thematically am biguous passive sentences such as the car was parked by the parking attendant/the building entrance. S e r i a l and parallel models m ade different predictions about the way in which these sentences should be processed. Thus, the garden-path model handles argum ent vs. adjunct ambiguities in separate m odules. The argument reading w ould fall under the principle of late closure since arguments are prim ary phrases. Here, no effect of lexical bias is predicted and the argum ent sim ply attaches to the m ost local verb, which is also the main verb. The adjunct reading, however, falls under the principle of construal because adjuncts are optional sentence fragments. In this case, the adjunct should be associated according to both structural and non-structural information. N on-structural inform ation includes semantic inform ation, but ^ Notice that this ambiguity is not structural. Both the adjunct and argument PPs follow the same syntactic structure. The difference lies in the interpretation of the PP. 137 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. not frequency of occurrence. Therefore no effect of contingent frequency is predicted for the adjunct interpretation either. As shown in Chapter IV, the effect of contingent frequency was observed in the low-frequency structure, in this case the adjunct interpretation. Adjunct-bias verbs had a reliable advantage over argum ent- bias verbs in passive sentences ending with an adjunct. Furtherm ore, the effect of contingent frequency w as also observed in the high-frequency structure, in this case, the argum ent interpretation. Argument-bias verbs had a significant advantage over adjunct-bias verbs in passive sentences ending w ith an argument. Furthermore, the effect of structural frequency was significant, with argum ent interpretations taking significantly less time to process than adjunct interpretations, regardless of verb bias. The results of Experiment 4 thus indicate that both contingent and structural frequency can impact sentence comprehension. This result, while predicted by the constraint-based m odel, presents a serious challenge for serial parsers such as the garden-path model. Furtherm ore, the constraint-based m odel provides a unified account for ambiguous, unam biguous and thematically am biguous sentences, while the garden-path m odel divides structures into types that fall under different modules. 3. Representational Issues The experimental research undertaken in this dissertation is consistent w ith the view that com prehenders are sensitive to statistical properties of language, and that these properties are exploited during on-line sentence comprehension. The question that arises given this view of sentence comprehension is w hether both structural and contingent frequencies are 138 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. kept track of simultaneously. This is an im portant representational issue. The constraint-based model has been criticized by Frazier (1995, Frazier & Clifton, 1996) for not being specific enough about how frequencies are computed. Fortunately, the work reported here can help shed light on this issue. Recall that while actives are more fi-equent than passives, there exist both active- and passive-bias verbs. Notice also that the vast majority of verbs are active- bias. Knowledge of this contingent frequency amounts to knowledge of structural frequency. In other w ords, knowing that actives are m ore frequent than passives falls out of know ing that m ost verbs are active-bias. Therefore, keeping track of local co-occurrence frequencies can result in knowledge of global, structural frequencies. 139 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. REFERENCES Allen, Jv MacDonald, M. C., Seidenberg, M. S., & Christiansen, M. H. (1996). Probabilistic constraints in acquisition. Paper presented at the ninth annual CUNY sentence processing conference. Allen, J., & Christiansen, M. H. (1996). Integrating m ultiple cues in w ord segmentation: A connectionist model using hints. Proceedings of the Eighteenth A nnual Conference of the Cognitive Science Society. Lawrence Erlbaum. Altmann, G., & Steedm an, M. (1988). Interaction w ith context during hum an sentence processing. Cognition, 30, 191-238. Anisfeld, M., & Klenbort, I. (1973). On the functions of structural paraphrase: The view from the passive voice. Psychological Bulletin, 2, 117-126. Attneave, P. (1953). Psychological probability as a function of experienced frequency. Journal of Experimental Psychology, 46, 81-86. Baron, J. (1988). Thinking and deciding. N ew York: Cambridge U niversity Press. Bates, E., & Devescovi, A. (1989). Crosslinguistic studies of sentence production. In E., Bates, and B., MacWhinney (Eds.), The crosslinguistic study of sentence processing (pp 224-256). Cambridge: Cam bridge University Press. Bates, E., & M acW hinney, B. (1987). Competition, variation and language learning. In B. M acW hinney (Ed.), Mechanisms of language acquisition. Hillsdale, NJ: Lawrence Erlbaum. Bates, E., & M acW hiimey, B. (1989). The crosslinguistic study of sentence processing. Cambridge: Cambridge University Press. Bates, E., & Wulfeck, B. (1989). Crosslinguistic studies of aphasia. In B. MacWhiimey and E. Bates (Eds.), The Crosslinguistic Study of Sentence Processing (pp 328-371). Cambridge: Cambridge University Press. 140 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Bates, E., Friederid, A., & Wulfeck, B. (1987). Grammatical m orphology in aphasia: evidence firom three languages. Cortex. 23, 545-574. Bates, E., Harris, C., Marchman, V., & Wulfeck, B. (1995). Production of complex syntax in norm al aging and Alzheimer's Disease. Language and Cognitive Processes, 10, (5), 487-539. Berlin, B., Boster, J. S., & O’ Neill, J. P. (1981). The perceptual bases of ethnobiological dassification: Evidence from A guaruna Jivaro ornithology. Journal of Ethnobiology, 1, 95-108. Berman, R (1985). The acquisition of Hebrew. In D. I. Slobin, (Ed.), The crosslinguistic study of language acquisition. Volume 1: The data, (pp 255- 371). Hillsdale, NJ: Erlbaum. Bever, T. G. (1970). The cognitive basis for linguistic structure. N ew York: W iley. Bever, T. G. (1988). The psychological reality of grammar: a student’ s-eye view of cognitive sdence. In W. Hirst, (Ed.), The making of cognitive sciences. Essays in honor of George A. Miller (pp 112-142). Cambridge: Cam bridge University Press. Bever, T. G., & McElree, B. (1988). Empty categories access their antecedents during comprehension. Linguistic Inquiry, 19 (1), 35-43. Bever, T. G., Straub, K., Shenkman, K., Kim, J. J., & Carrithers, C. (1989). The psychological reality of NP-trace. Manuscript. Amherst, Mass: University of M assachusetts. Boland, J. E., & Tanenhaus, M. K. (1991). The role of lexical representations in sentence processing. In G. B. Simpson (Ed.), Understanding word and sentence (pp. 331-366). Amsterdam: North-Holland. Branigan, H. P., Pickering, M. J., Liversedge, S. P., Stewart, A. P., & Urbach, T. P. (1995). Syntactic priming: Investigating the mental representation of language. Journal of Psycholinguistic Research, 24 (6), 489-506. 141 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Britt, M. A., Perfetti, C. A., Garrod, S., & Rayner, K. (1992). Parsing in discourse: Context effects and their limits. Journal of Memory and Language, 31 (3), 293-314. Broadbent, D. E., & Gregory, M. (1968). Visual perception of w ords differing in letter bigram frequency. Journal of Verbal Learning and Verbal Behavior, 7, 569-571. Caplan, D., Baker, C., & Dehaut, F. (1985). Syntactic determ inants of sentence comprehension in aphasia. Cognition, 21 (2) 117-175. Caramazza, A., & Miceli, G. (1991). Selective im pairm ent of thematic role assignment in sentence processing. Brain and Language, 41 (3), 402-436. Carreiras, M., & Clifton, C. (1993). Relative clause interpretation preferences in Spanish and English. Language and Speech, 36 (4), 353-372. Carrithers, C. (1986). The special status of non canonical sequences. Doctoral Dissertation. Colum bia University. Carrithers, C. (1989). Syntactic complexity does not necessarily make sentences harder to understand. Journal of Psycholinguistic Research, 18 (1), 75-87. Carrithers, C., & Bever, T. G. (1984). Eye fixation patterns during reading confirm theories of language comprehension. Cognitive Science, 8 (2), 157- 172. Chang, P. R. (1980). Active m emory processes in visual sentence comprehension: Clause effects and pronominal reference. Memory and Cognition, 8, 58-64. Chomsky, (1957). Syntactic structures. The Hague: M outon. Chomsky, N, (1981). Lectures on government and binding. Dordrecht: Foris. Clark, H. H. (1965). Some structural properties of sim ple active and passive sentences. Journal of Verbal Learning and Verbal Behavior, 4, 365-370. 142 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Clifton, C. (1993). Thematic roles in sentence parsing. Canadian Journal of Experimental Psychology, 45 (2), 222-246. Clifton, C., & Frazier, L. (1986). The use of syntactic inform ation in filling gaps. Journal of Psycholinguistic Research, 15 (3), 209-224. Clifton, C., Frazier, L., & Connine, (1984). Lexical expectations in sentence com prehension. Journal of Verbal Learning and Verbal Behavior, 23 (6) 696- 708. Clifton, C., Kurcz, I., & Jenkins, J. J. (1965). Grammatical relations as determ inants of sentence sim ilarity. Journal of Verbal Learning and Verbal Behavior, 4, 112-117. Clifton, C., Speer, S., & Abney, S. P. (1991). Parsing arguments: Phrase structure and argum ent structure as determinants of initial parsing decisions. Journal of Memory and Language, 30 (2), 251-271. Connine, C., Ferreira, F., Jones, C., Clifton, C., & Frazier, L. (1984). Verb frame preferences: Descriptive norms. Journal of Psycholinguistic Research, 13 (4), 307-319. Coren, S., & Porac, C. (1977) Fifty centuries of right-handedness: The historical record. Science, 198, 631-632. Corley, M. M. B., & Corley, S. (submitted). Cross-linguistic and intra-Iinguistic evidence for the use of statistics in hum an sentence processing. Crain, S., & Steedman, M. (1985). O n not being led up the garden-path: the use of context by the psychological syntax processor. In D. R. Dowty, L. Kartunnen, and A. M. Zwicky, (Eds.)Natural language processing (pp320-358). Cambridge: Cambrige University Press. Cuetos, F., & Mitchell, D. C. (1988). Cross-Linguistic differences in parsing: Restrictions on the use of the Late Closure strategy in Spanish. Cognition, 30, 73-105. 143 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Cuetos, F., Mitchell, D. C., & Corley, M. M. B. (in press). Parsing in different languages. In M. Carreiras, J. Garda-Albea, and N . Sabastian-Galles, (Eds.), Language processing in Spanish. M ahwah, N.J.: Lawrence Erlbaum. DeM uth, K. (1989). M aturation and the acquisition of the Sesotho passive. Language, 65 (1), 56-80. DeVicend, M., & Job, R. (1993). Some observations on the universality of the late-dosure strategy. Journal of Psycholinguistic Research, 22 (2), 189-206. DeVicend, M., & Job, R. (1995). An investigation of late dosure: The role of syntax, thematic structure and pragm atics in initial interpretation. Journal of Experimental Psychology: Learning, Memory and Cognition, 21 (5), 1303-1321. Eberhard, K. M., Spivey-Knowlton, M. J., Sedivy, J. C., & Tanenhaus, M. K. (1995). Eye-movements as a window into real-time spoken language com prehension in N atural Contexts. Journal of Psycholinguistic Research, 24 (6), 409-435. Elman, J. L. (1993). Learning and developm ent in neural networks: the im portance of starting small. Cognition, 48, 71-99. Ferreira, F. (1994). Choice of passive is affected by verb type and animacy. Journal of Memory and Language, 33 (6), 715-736. Ferreira, F., &Clifton, C. (1986). The independence of syntactic processing. Journal of Memory and Language, 25 (3), 348-368. Ferreira, F., & Henderson, J. M. (1991). Recovery from misanalyses of garden- path sentences. Journal of Memory and Language, 30 (6), 725-745. Ferreira, F., & Henderson, J. M. (1993). Reading processes during syntactic analysis and reanalysis. Canadian Journal of Experimental Psychology, 47 (2), 247-275. Finegan, E. (1982). Form and function in testam ent language. In R. J. Di Pietro, (Ed.), Linguistics and the professions. N orw ood, N. J.: Ablex. 144 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Fodor, J. Dv & Frazier, L. (1980). Is the hum an sentence parsing m echanism an ATN? Cognition, 8, 417-459. Ford, M., Bresnan, J., & Kaplan, R. M. (1982). A competence-based theory of syntactic closure. In J. Bresnan, (Ed ), The mental representation of grammatical relations (pp 727-796). Cambridge: MTT Press. Forster, K. I., & Chambers, S. (1973). Lexical access and naming time. Journal of Verbal Learning and Verbal Behavior. 12, 627-635. Forster, K. I., & Olbrei, I. (1973). Semantic heuristics and syntactic analysis. Cognition, 2, 319-348. Fraser, C., Bellugi, U., & Brown, R. (1963). Control of gram m ar in imitation, com prehension, and production. Journal of Verbal Learning and Verbal Behavior, 2, 121-135. Frazier, L. (1978). On comprehending sentences: syntactic parsing strategies. Ph. D. dissertation. University of Connecticut. Bloomington, Indiana University Linguistics Club, 1979. Frazier, L. (1987). Sentence processing: A tutorial review. Coltheart, M. (Ed.), Attention and Performance XII: The Psychology of Reading (pp 559-586). Hillsdale, NJ. Erlbaum. Frazier, L. (1989). Against lexical generation of syntax. Marslen-Wilson, W. (Ed.) Lexical Representation and Process (pp 505-528). Cambridge. MIT Press. Frazier, L. (1994). Sentence (re)-analysis. Unpublished manuscript. Frazier, L. (1995). Constraint satisfaction as a theory of sentence processing. Journal of Psycholinguistic Research, 24 (6), 437-468. Frazier, L., & Clifton, C. (1989). Successive cyclicity in the grammar and the parser. Language and Cognitive Processes, 4, 93-126. Frazier, L., & Clifton, C. (1996). Construal. Cambridge: MIT Press. 145 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Frazier, L., & Flores d ’Areals, G. B. (1989). Filler-driven parsing: A study of gap filling in Dutch. Journal of Memory and Language, 28 (3), 331-344. Frazier, L., & Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally am biguous sentences. Cognitive Psychology, 14, 178-210. Frazier, L., & Rayner, K. (1987). Resolution of syntactic category ambiguities: Eye m ovem ents in parsing lexically am biguous sentences. Journal of Memory arui Language, 26 (5), 505-526. Frazier, L. Clifton, C., & Randall, J. (1983). Decision principles and structure in sentence comprehension. Cognition, 13 (2), 187-222. Frazier, L., Taft, L., Roeper, T., Clifton, C., & Ehrlich, K. (1984). Parallel structure: A source of facilitation in sentence comprehension. Memory and Cognition, 12 (5), 421-430. Frederiksen, J. R., & KroU, J. F. (1976). Spelling and sound: Approaches to the internal lexicon. Journal of Experimental Psychology: Human Perception and Performance, 2, 361-379. Gam ham , A. (1985). Psycholinguistics: Central Topics. London: M ethuen. Gamsey, S. M. Pearlmutter, N. J., Myers, E., & Lotocky, M. E. (1996). The relative contributions of verb bias and plausibility to the com prehension of tem porarily am biguous sentences. In press. Journal of Memory and Language. Gemsbacher, M. A. (1984). Resolving 20 years of inconsistent interactions between lexical familiarity and orthography, concreteness and polysemy. Journal of Experimental Psychology, 113 (2), 256-281. Gibson, E. (1992). On the adequacy of the competition model. Review Article. Language, 68 (4), 821-830. 146 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Gibson, E. A. F. (1991). A computational theory of human linguistic processing: Memory limitations and processing breakdown. Doctoral dissertation. Carnegie Mellon University. Gibson, E., & Hickok, G. (1993). Sentence processing w ith empty categories. Language and Cognitive Processes, 8 (2), 147-161. Gibson, E., Hickok, G., & Schiitze, C. T. (1994). Processing empty categories: A parallel approach. Journal of Psycholinguistic Research, 23 (5), 381-405. Gibson, E., Schiitze, C. T., & Salomon, A. (1996). The relationship betw een the frequency and the processing complexity of linguistic structure. Journal of Psycholinguistic Research, 25 (1), 59-92. Gilboy, E., Sopena, J-M., Clifton, C., & Frazier, L. A rgum ent structure and association preferences in Spanish and English complex NPs. Cognition, 54, 131-167. Givôn, T. (1979). On Understanding Grammar. N ew York: Academic Press. Givôn, T. (Ed.) (1994). Voice and inversion. A m sterdam : John Benjamins. (Gordon, B. (1985). Subjective frequency and the lexical decision latency function: Implications for m echanism s of lexical access. Journal of Memory and Language, 24 (6), 631-645. Gough, P. B. (1965). Grammatical transformations and speed of understanding. Journal of Verbal Learning and Verbal Behavior, 4, 107-111. Gough, P. B. (1966). The verification of sentences: The effects of delay of evidence and sentence length. Journal of Verbal Learning and Verbal Behavior, 5, 492-496. Greenberg, J. H. (1963). The languages of Africa. Bloomington: Indiana University Press. Greenberg, J. H. (1963). Universals of language. Conference report. New York: Dobbs Ferry. 147 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Greenberg, J. H. (1978). Universals of human language. Stanford: Stanford U niversity Press. Greene, J. M. (1970). The semantic function of negatives and passives. British Journal of Psychology, 61, 17-22. Grice, H. P. (1975). Logic and conversation. In P. Cole, and J. L. M organ (Eds.), Syntax and semantics, 3. N ew York: Academic Press. Griffiths, D. (1854). A grammar of the Malagasy language. W oodbrige: Edw ard Pite. Grimshaw, J. (1990). Argument structure. Cambridge: MIT Press. Hanna, J. E., Barker, C., &Tanenhaus, M. K. (1995). Integrating local and discourse constraints in resolving lexical thematic ambiguities. Poster presented at the eight annual CUNY sentence processing conference. Hanna, J. E., Spivey-Knowlton, M. J., & Tanenhaus, M. K. (1996). Integrating discourse and local constraints in resolving lexical thematic ambiguities. Proceedings of the Eighteenth A nnual Conference of the Cognitive Science Society. Lawrence Erlbaum. Hartlep, K. (1983). Simultaneous presentation of m oving objects in an infant tracking task. Infant Behavior and Development, 6 (1), 79-84. Hasher, L., & Zacks, R. T. (1984). Autom atic processing of fundam ental information: The case of frequency of occurrence. American Psychologist, 39 (12), 1372-1388. Hatano, G., Siegler, R. S., Richards, D. D., & Inagaki, K. (1993). The developm ent of biological knowledge - a multinational study. Cognitive Development, 8 (1), 47-62. Hickok, G. (1993). Parallel processing: Evidence from reactivation in garden- path sentences. Journal of Psycholinguistic Research, 22 (2), 239-249. 148 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Holmes, V. M, Stowe, L., Sc Cupples, L. (1989). Lexical expectations in parsing complement-verb sentences. Journal of Memory and Language, 28 (6), 668- 689. Hopper, P. J., & Thompson, S. A. (1980). Transitivity in gram m ar and discourse. Language, 56 (2), 251-299. Johnson, M. G. (1967). Syntactic position and rated meaning. Journal of Verbal Learning and Verbal Behavior, 6, 240-246. Johnson, M. H., Dziurawiec, S., Ellis, H., & Morton, J. (1991). Newborns preferential tracking of face-like stimuli and its subsequent decline. Cognition, 40 (1-2), 1-19. Johnson-Laird, P. N. (1968). The choice of passive voice in a communicative task. British Journal of Psychology, 59 (1), 7-15. Johnston, J. R. (1985). Cognitive prerequisites: The evidence from children learning English. In D. I. Slobin (Ed.), The cross-linguistic studies of language acquisition (pp. 961-1004). Hillsdale: Lawrence Erlbaum. Juliano, C., & Tanenhaus, M. K. (1993). Contingent frequency effects in syntactic am biguity resolution. In Fifteenth annual conference of the Cognitive Science Society. Hillsdale, NJ: Erlbaum. Juliano, C., & Tanenhaus, M. K. (1994). A constraint-based lexicalist account of the subject/object attachm ent preference. Special issue: Sentence processing: m . Journal of Psycholinguistic Research, 23 (6), 459-471. Just, M. A., & Carpenter, P. A. (1992). A capacity theory of comprehension: Individual differences in w orking memorv. Psychological Review, 99 (1), 122- 149. Just, M. A., & Carpenter, P. A. (1980). A theory of reading: From eye fixations to comprehension. Psychological Review, 87 (4), 329-354. Kawamoto, A. H. (1988). Distributed representations of ambiguous w ords and their resolution in a connectionist network. In S. Small, G. Cottrell and M. K. 149 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. T anenhaus (Eds.)/ Lexical ambiguity resolution: Computational linguistic and psychological perspectives ppl95-228. N ew York: M organ Kauffman. Kawam oto, A. H. (1993). Nonlinear dynam ics in the resolution of lexical am biguity: A parallel distributed processing account. Journal of Memory and Language, 32 (4), 474-516. Keenan, E. L. (1987). Universal grammar: 15 essays. London: Croom Helm. Kelly, M. H., & M artin, S. (1994). Domain-general abilities applied to domain- spedfîc tasks: Sensitivity to probabilities in perception, cognition and language. Lingua 92 (1-4), 105-140. Kilbom , K. (1994). Learning a language late: Second language acquisition in adults. In M. A. Gernsbacher. (Ed.) Handbook of Psycholinguistics (pp 917- 944). San Diego: Academic Press. Kimball, J. (1973). Seven principles of surface structure parsing in natural language. Cognition, 2, 15-47. Kucera, H. & Francis, W. N. (1967). Com putational analysis of present day Am erican English. Providence, R. I.: Brown University Press. K urtzm an, H. (1985). Studies in syntactic ambiguity resolution. U npublished doctoral dissertation. Massachusetts Institute of Technology. Lalami, L., & MacDonald, M. C. (1996). Structural and contingent frequency effects in sentence comprehension. Poster presented at the N inth annual CUNY sentence processing conference. Landauer, T., & Streeter, L. (1973). Structural differences between common and rare words: Failure of equivalence assum ptions for theories of w ord recognition. Journal of Verbal Learning and Verbal Behavior, 12, 119-131. Levin, B. (1993). English verb classes and alternations. Chicago: University of Chicago Press. 150 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Levin, B., & Pinker, S. (1991). Introduction to special issue of cognition on lexical and conceptual semantics. Cognition, 41 (1-3), 1-7. Lichenstein, S., Slovic, P., Fischoff, B., Layman, M., & c Combs, B. (1978). Judged frequency of lethal events. Journal of Experimental Psychology: Human Learning and Memory 4, 551-578. M acDonald, M. C. (1989). Prim ing effects from gaps to antecedents. Language and Cognitive Processes, 4 (1), 35-56. MacDonald, M. C (1993). The interaction of lexical and syntactic ambiguity. Journal of Memory and Language, 32 (5), 692-715. MacDonald, M. C. (1994). Probabilistic constraints and syntactic am biguity resolution. Language and Cognitive Processes, 9 (2), 157-201. MacDonald, M. C., Pearlmutter, N., & Seidenberg, M. S. (1994). The lexical nature of syntactic ambiguity resolution. Psychological Review, 101 (4), 673- 703. MacDonald, M. C., Pearlmutter, N ., & Seidenberg, M. S., (1994). Syntactic am biguity resolution as lexical am biguity resolution. In C. Clifton, L. Frazier and K. Rayner (Eds.), Perspectives on sentence processing (pp. 123-154). Hillsdale, NJ: Erlbaum. M archman, V., Harris, C., Juarez, J., & Bates, E. (1992). Functional constraints on the production of complex syntax. Manuscript. La Jolla: U niversity of California, San Diego, Center for Research in Language. M arcus, M. P. (1980). A theory of syntactic recognition for natural language. Cambridge: MIT Press. M arcus, M., Kim, G., M ardnkiewicz, M. A., MacIntyre, R., Bies, A., Ferguson, M., Katz, K. & Schasberger, B. (1994) The Penn Treebank: Annotating Predicate A rgum ent Structure. Proceedings of the Human Language Technology Workshop, San Francisco, CA: Morgan Kaufmann Publishers Inc. 151 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Marcus, M., Santorini, B., & M ardnkiewicz, M. A., (1993). Building a large annotated corpus of English: the Penn Treebank. Computational Linguistics, 19. Marslen-Wilson, W. D., & Tyler, L. K. (1975). Processing structure of sentence perception. Nature, 257, 784-786. Manner, G., Tanenhaus, M K , & Carlson, G. N. (1995). Im plidt argum ents in sentence processing. Journal of Memory and Language, 34 (3), 357-382. McClelland, J. L., St. John, M., & Taraban, R. (1989). Sentence comprehension: A parallel distributed processing approach. Language and Cognitive Processes, 4, SI287-226 McDonald, J., & MacW hinney, B. (1989). Maximum likelihood models for sentence processing. In B. MacW hinney and E. Bates (Eds.) The Crosslinguistic Study of Sentence Processing (pp.397-421). Cambridge: Cambridge University Press. McDonald, J., Bock, K., & Kelly, M. H. (1993). W ord and w orld order: Semantic, phonological and metrical determinants of serial position. Cognitive Psychology, 25, 188-230. McMahon, L. E. (1963). Grammatical analysis as part of understanding a sentence. Unpublished doctoral dissertation. H arvard University. Miller, (1962). Some psychological studies of grammar. American Psychologist, 17, 748-762. Miller, G. A., & Isard, S. D. (1963). Some perceptual consequences of linguistic rules. Journal of Verbal Learning and Verbal Behavior 2, 217-28. Mills, A. E. (1985). The acquisition of German. In D. I. Slobin, {'Ed.),The crosslinguistic study of language acquisition. Volume 1: The data (pp. 141- 254). Hillsdale, NJ: Erlbaum. Mitchell, D. C. (1987). Lexical guidance in human parsing: Locus and processing characteristics. In M. Coltheart, (Ed.), Attention and Performance X: The Psychology of Reading (pp 601-618). London: Lawrence Erlbaum. 152 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Mitchell, D. C. (1994) Sentence Parsing. In M. A. Gemsbacher (Ed.), Handbook of Psycholinguistics. San Diego: Academic Press. Mitchell, D. C., & Cuetos, F. (1991a). Restrictions on Late Closure: The computational underpinnings of parsing strategies in Spanish and English. U npublished M anuscript. Mitchell, D. C., & Cuetos, F. (1991b). The Origins of Parsing Strategies. In C. Sm ith (Ed.), Current Issues in Natural Language Processing (pp. 1-12). Center for Cognitive Science: University of Austin, Texas. Mitchell, D. C., & Green, D. W. (1978). The effects of context and content on im m ediate processing in reading. Quarterly Journal of Experimental Psychology, 30, 609-636. Mitchell, D. C., Cuetos, F., & Zagar, D. (1990). Reading in different languages: Is there a universal mechanism for parsing sentences? In D. Balota, G. B. Flores d'Arcais & K. Rayner, (Eds.), Comprehension processes in reading (pp. 285-302). Hillsdale, N. J. Erlbaum. Mitchell, D. C., Cuetos, F., Corley, M. M. B., & Brysbaert, M. (1995). Exposure- based models of hum an parsing: Evidence for the use of coarse-grained (non- lexical) statistical records. M anuscript subm itted for publication. Miyake, A., Just, M. A., & Carpenter, P. A. (1994). A capacity approach to syntactic comprehension: Maldng normal adults perform like aphasie patients. Cognitive Neuropsychology, 11, 671-717. M organ, J. L., & Saffran, J. R. (1995). Emerging integration of sequential and suprasegm ental inform ation in preverbal speech segmentation. Child Development, 66 (4), 911-936. Paivio, A., Yuille, J., & Madigan, S. (1968). Concreteness, imagery, and m eaningfulness values for 925 nouns. Journal of Experimental Psychology Monograph, 76 (1, Pt. 2). 153 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Pearlm utter, N. J., & MacDonald, M. C. (1995). Individual differences and probabilistic constraints in syntactic ambiguity resolution. Journal of Memory and Language, 34 (4), 521-542. Pickering, M. J., Liversedge, S. P., Branigan, H. P., & van Com pel, R. P. G. (19%). Context effects on the processing of adjuncts and argum ents. Poster presented at the 19% CUNY sentence processing conference. N ew York. Plaut, D. C., McClelland, J. L., Seidenberg, M. S., & Patterson, K. (19%). U nderstanding normal and im paired word reading: Com putational principles in quasi-regular dom ains. Psychological Review, 103 (1), 56-115. Pleszeswka, K. (1985). Der einfluss von grammatikalischen satztransformationen auf das behalten eines vorlesungsinhaltes. [The influence of grammatical transform ation on the retention of lecture material]Sc/iw;cizerisc)ie Zeitschrift fur Psychologie und hre Anwendungen, [Swiss Journal of Psychology and its Applications]44 (1-2), 31-41. Prentice, J. L. (1966). Response strength of single words as an influence in sentence behavior. Journal of Verbal Learning and Verbal Behavior, 5, 429- 433. Ratcliff, R., & Me Koon, G. (1989). Similarity information vs. relational information: Differences in the time-course of retrieval. Cognitive Psychology, 21 (2) 139-155. Rayner, K., & Frazier, L. (1989). Selection mechanisms in reading lexically am biguous words. Journal of Experimental Psychology: Learning, Memory and Cognition, 15 (5), 779-790. Rayner, K. Garrod, S., & Perfetti, C. A. (1992). Discourse influences during parsing are delayed. Cognition, 45 (2), 109-139. Rayner, K., Carlson, M., & Frazier, L. (1983). The interaction of syntax and semantics during sentence processing: Eye-movements in the analysis of sem antically biased sentences. Journal of Verbal Learning and Verbal Behavior, 22, 358-374. 154 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Rice, G. A., & Robinson, D. O. (1975). The role of bigram frequency in perception of w ords and nonwords. Memory and Cognition, 3, 513-518. Roman, G. Pavard, B., & Asselah, B. (1985). Traitem ent perceptif des phrases ambiguës en Arabe. [Perception treatm ent of am biguous phrases in Arabic]Cahiers de Psychologie Cognitive, 5 (1), 5-22. Rubenstein, H., Garfîeld, L., & Millikan, J. (1970). Hom ographie entries in the internal lexicon. Journal of Verbal Learning and Verbal Behavior, 9, 487-494. Rubin, D. C. (1974). The subjective estimation of syllable frequency. Perception and Psychophysics, 16, 193-196. Saegert, S., Swap, W., & Zajonc, R. B. (1973). Exposure, context and interpersonal attraction. Journal of Personality and Social Psychology, 25, 234- 242. Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical cues in language acquisition: W ord segmentation by infants. Proceedings of the Eighteenth Annual Conference of the Cognitive Science Society. Lawrence Erlbaum. Saffran, J. R., N ew port, E. L., & Aslin, R. N. (1996). W ord segmentation: The role of distributional cues. Journal of Memory and Language, 35 (4), 606-621. Schedler, J. K., Jonides, J., & Manis, M. (1985). Availability: Plausible but questionable. Paper presented at the 26th annual m eeting of the Psychonomic ^ d e ty . Boston, MA. Schwartz, M. P., Saffran, E. H., & Marin, O. S. H. (1980). Fractioning the reading process in dementia: Evidence for w ord-spedfic print-to-sound assodations. In M. Coltheart, K. E. Patterson and J. C. Marshall (Eds.). Deep dyslexia (pp. 259-269). London: Routledge and Kegan Paul. . Seidenberg, M. S., & McClelland, J. L. (1989). A distributed, developmental model of w ord recognition and naming. Psychological Review, 96 (4), 523-568. Seidenberg, M. S., Plaut, D. C., Petersen, A. S., & McClelland, J. L. (1994). Nonword pronundation and models of w ord recognition. Journal of 155 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Experimental Psychology: Human Perception and Performance, 20 (6), 1177- 1196. Shapiro, B. J. (1969) The subjective estimation of relative w ord frequency. Journal of Verbal Learning and Verbal Behavior, 8, 248-251. Siewierska, A. (1984). The passive: a comparative linguistic analysis. London, Croom Helm. Slobin, D. I. (1966a). Grammatical transformations and sentence comprehension in childhood and adulthood. Journal of Verbal Learning and Verbal Behavior, 5, 219-227. Smith, F. (1965). Reversal of m eaning as a variable in the transform ation of gram m atical sentences. Journal of Verbal Learning and Verbal Behavior, 4, 39-43. Spivey-Knowlton, M. J, Hanna, J. E., & Tanenhaus, M. K. (1996). M odeling discourse context effects: A m ultiple constraints approach. Poster presented at the ninth annual CUNY sentence processing conference. Spivey-Knowlton, M. J., & Sedivy, J. (1995). Resolving attachm ent ambiguities w ith m ultiple constraints. Cognition, 55 (2), 227-267. Spivey-Knowlton, M. J., Trueswell, J. C., & Tanenhaus, M. K. (1993). Context effects in syntactic ambiguity resolution: Discourse and semantic influences in parsing reduced relative clauses. Canadian Journal of Experimental Psychology, 47 (2), 276-309. St. John, M., & Gemsbacher, M. A. (1995). Syntactic comprehension— Practice makes perfect and frequency makes fleet. Center for Research in Language Technical Report #9501. La Jolla, California: University of California, San Diego. Steedman, M., & Altmann, G. T. (1989). Ambiguity in context: A reply. Special Issue: Parsing and interpretation. Language and Cognitive Processes, 4 (3-4) SI105-SI122. 156 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Stevens, K. L., & MacDonald, M. C. (1994). Aging and the use of context during ambiguity resolution. Poster presented at the 1994 Cognitive Aging Conference, Atlanta, GA. Stowe, L. A., Tanenhaus, M. K., & Carlson, G. N. (1991). Filling gaps on-line: Use of lexical and semantic inform ation in sentence processing. Language and Speech, 34 (4), 319-340. Streim, N. (1988). The effects of syntactic priming on the production of active and passive sentences. U npublished doctoral dissertation. U niversity of Wisconsin, Madison. Svartvik, J. (1966). On voice in the English verb. The Hague, M outon. Tanenhaus, M. K , & Carlson, G. N. (1989). Lexical structure and language comprehension. In W. Marslen-W ilson, (Ed.), Lexical representation and process (pp529-561). Cambridge: MIT Press. Tanenhaus, M. K., & Trueswell, J. C. (1995). Sentence comprehension. In J. L. Miller, and P. D. Eimas (Eds ), Speech, Language and Communication (pp. 217-262). San Diego, Academic Press. Tanenhaus, M. K., Carlson, G., & Trueswell, J. C. (1989). The role of thematic structures in interpretation and parsing. Special Issue: Parsing and interpretation. Language and Cognitive Processes, 4 (3-4), SI211-SI234. Tannenbaum, P. H., & Williams, P. (1968). Generation of active and passive sentences as a function of subject and object focus. Journal of Verbal Learning and Verbal Behavior, 7, 246-250. Tannenbaum, P. H., & Williams, F. (1968). Prom pted word replacem ent in active and passive sentences. Language and Speech, 11, 220-229. Taraban, R , & McClelland, J. L. (1987). Conspiracy effects in word recognition. Journal of Memory and Language, 26 (6), 608-631. Taraban, R , & McClelland, J. L. (1988). Constituent attachment and them atic role assignment in sentence processing: Influences of content-based expectations. Journal of Memory and Language, 27 (6), 597-632. 157 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Thibadeau, R., Just, M. A., & Carpenter, P. A. (1982), A m odel of the time course and content of reading. Cognitive Science, 6, 157-203. Thompson, S. A. (1987). The passive in English: A discourse perspective. In R. Channon and L. Shockey (Eds.), In honor of lise Lehiste (pp. 497-511). Dordrecht: Foris. Thornton, R. B., Gil, M., & MacDonald, M. C. (1996). Constraint-based models and modification ambiguities. Foster presented at the N inth annual CUNY sentence processing conference. Traxler, M. J., & Pickering, M. J. (1995). Evidence against statistical parsing. Poster presented at the 8th annual CUNY conference on sentence processing, Tucson, AZ. Trueswell, J. C., & Tanenhaus, M. K. (1991). Tense, tem poral context and syntactic am biguity resolution. Language and Cognitive Processes, 6 (4), 303- 338. Trueswell, J. C., Tanenhaus, M. K., & Kello, C. (1993). Verb specific constraints in sentence processing: Separating effect of lexical preference from garden- paths. Journal of Experimental Psychology: Learning, Memory and Cognition, 19 (3), 528-553. Trueswell, J. C., Tanenhaus, M. K., & Gamsey S. M. (1994). Semantic influences on parsing: Use of thematic role information in syntactic am biguity resolution. Journal of Memory and Language, 33 (3), 285-318. Tyler, L. K., & Marslen-Wilson, W. D. (1977). The on-line effects of semantic context on syntactic processing. Journal of Verbal Learning and Verbal Behavior, 16, 683-692. Valentine, T. and Moore, V. (1995). Nam ing faces: The effects of facial distinctiveness and surnam e frequency. Quarterly Journal of Experimental Psychology: Human Experimental Psychology, 48A (4), 849-878. 158 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Vemberg, E. M., Ewell, K. K , Beery, S. H., & Freeman, C. M. (1995). Aversive exchanges with peers and adjustm ent during early adolescence: Is disclosure helpful? Child Psychiatry and Human Development, 26 (1), 43- 59. Wall Street Journal Materials, Copyright 1987,1988,1989. Dow Jones Inc. W arner, J., & Glass, A. (1987). Context and distance-to-disambiguation effects in ambiguity resolution: Evidence &om grammaticality judgments of garden- path sentences. Journal of Memory and Language, 26 (6), 714-738. Warren, P., Grabe, E., & N olan, F. (1995). Prosody, phonology and parsing in closure ambiguities. Language and Cognitive Processes, 10 (5), 457-486. Wiggs, C. L. (1993). Aging and m em ory for frequency of occurrence of novel, visual stimuli: Direct and indirect measures. Psychology and Aging, 8 (3), 400- 410. Witkowski, S. R., & Brown, C. H. (1983). Marking-reversals and cultural importance. Language, 59 (3), 569-582 W right, P. (1969). Transformations and the understanding of sentences. Language and Speech, 12, 156-166. Yu, K , & Blake, R. (1992). Do recognizable figures enjoy an advantage in binocular rivalry? Journal of Experimental Psychology: Human Perception and Performance, 18 (4), 1158-1173. Zagar, D., & Pynte, J. (1992). The role of semantic information and of attention in processing syntactic ambiguity: Eye-movement study. Paper presented at the 5th conference of the European Society for Cognitive Psychology. Paris. Zechmeister, E. B., King, J. F., Gude, C., & Opera-Nadi, B. (1975). Ratings of frequency, familiarity, orthographic distinctiveness and pronundability for 192 surnames. Instrumentation and Psychophysics, 7, 531-533. Zipf, G. K. (1935). The psycho-biology of language. Boston: H oughton Mifflin. 159 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. APPENDIX I: HOPPER AND THOMPSON'S PARAMETERS Param eter Explanation and Example Participants High: 2 or more participants Low: 1 participant No transfer at all can take place unless two participants are involved. Contrast: hugged vs. died Kinesis High: action Low: state Actions can be transferred from one participant to another; states cannot. Contrast: hugged vs. liked. Aspect High: action finished Low: action in progress An action viewed from its end-point, i.e. a telic action, is more effectively transferred to a patient than one not provided by such an endpoint Contrast: ate vs. eating. Punctuality High: one stage Low: transitional stages Actions carried o u t w ith no transitional phase between inception and completion have a more marked effect on their patients than actions which are inherently on-going. Contrast: kick vs. carry. V olitionality High: volitional Low: involuntary The effect on the patient is typically m ore apparent when the agent is presented as acting purposefully. Contrast: wrote vs. forgot A ffirm ation High: affirm ative Low: negative This is the affirm ative/negative param eter. Contrast: / did it vs. I didn 't do it. 160 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Mode High: action asserted as occurring in real world Low: action that could occur in other worlds This refers to the distinction between realis' and 'irrealis' encoding of events. An action which either did not occur or which is presented as occurring in a non-real (contingent) w orld, is obviously less effective than one w hose occurrence is actually asserted as corresponding directly w ith a real event. Agency High: good agent Low: poor agent Participants high in agency can affect a transfer of an action in a w ay that those low in agency cannot. Contrast: George startled me vs. the picture startled m e Affectedness of object High: patient is atfected Low: patient is only partly affected The degree to which an action is transferred to a patient is a function of how completely the patient is affected. Contrast: I drank up the milk vs. / drank some of the milk. Individuation of object High: e.g. animate Low: e.g. inanimate This refers to both the distinctness of the patient from the agent and to its distinctness from its own background. Contrast: proper, human, animate, concrete, singular, count, referential, definite vs. common, inanim ate, abstract, plural, mass, non-referential. 161 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. APPENDIX II: EXPERIMENT 1 VERB RATINGS Note: Num bers given in the 'Active Usage’ and Passive Usage’ columns are h-om the Penn Treebank database. Verb R ating Active Usage Passive Usage Percent accompany 4.13 11 14 56 accuse 3.52 46 20 30.303 acquit 4.11 4 4 50 add 3.53 551 28 4.836 adjust 3.58 19 29 60.417 advise 3.23 37 8 17.778 affect 4.46 69 33 32.353 am aze 43 9 2 1 33.333 anger 3.11 5 4 44.444 announce 3.9 241 54 18.305 annoy 3.46 2 0 0 a p p a ll 4.79 0 2 100 appoint 4.44 7 25 78.125 approach 3.09 22 1 1 33.333 approve 3.97 146 61 29.469 arrest 3.27 7 14 66.667 ask 2.86 178 20 10.101 assign 3.74 6 13 68.421 base 3.83 7 290 97.643 begin 2.93 481 7 1.434 benefit 3.86 69 0 0 bother 338 16 1 5.882 boycott 3.39 0 1 100 b ty 2.97 • # # call 2.62 440 22 4.762 caution 3.14 26 0 0 challenge 3.16 26 8 23.529 cheer 3.02 5 3 37.5 cite 3.79 90 16 15.094 clean 3.16 17 3 15 climb 233 103 0 0 close 3.32 517 28 5.138 compliment 3.37 1 0 0 conduct 3.79 27 24 47.059 OMiftont 3.2 5 3 37.5 congratulate 3.15 2 0 0 console 4.08 1 0 0 contact 3.58 6 4 40 convict 3.95 11 18 62.069 crush 3.24 7 4 36.364 dam age 4.05 14 22 61.111 defend 2.91 54 1 1.818 d elig h t 4.3 3 0 0 design 4.13 23 94 80.342 162 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. destroy 358 16 7 30.435 d ev asta te 459 1 2 66.667 d isab le 4.18 1 1 50 d isap p o in t 4.48 7 6 46.154 discourage 4.16 14 4 22.222 d istrib u te 3.86 23 25 52.083 disturb 3.72 5 3 373 earn 252 158 6 3.659 elect 4.43 17 60 77.922 em barrass 435 0 0 # embrace 333 16 1 5382 em ploy 432 28 18 39.13 encourage 336 53 19 26.389 energize 339 0 1 100 en tertain 3.83 3 0 0 entice 436 7 0 0 escort 3.37 1 0 0 estim ate 3.18 160 54 25.234 expel 4.15 2 5 71.429 file 3.56 164 105 39.033 find 3.21 356 51 12331 fine 4.09 4 17 80.952 fin ish 2.37 67 1 1.471 flabbergast 4.75 0 1 100 force 3.05 83 74 47.134 forgive 3.11 4 1 20 frighten 3.72 3 0 0 fru strate 3.94 2 6 75 gain 2.91 211 2 .939 grant 3.24 19 10 34.483 greet 3.7 1 4 80 guide 3.67 7 2 22.222 h e lp 3.32 440 23 4.968 h ire 3.85 47 18 27.692 hug 2 1 0 0 h u rt 3.75 55 61 52.586 id en tify 4.14 49 22 30.986 im ita te 4.04 1 0 0 im press 4.97 6 2 25 im prove 3.83 132 2 1.493 include 3.02 525 33 5.914 incorporate 3.8 8 7 46.667 in fu riate 2.8 2 0 0 injure 3.66 2 13 86.667 instruct 3.38 7 2 22.222 insure 4.13 15 17 53.125 interrupt 3.83 1 4 80 in tim id ate 4.46 1 0 0 investigate 3.47 19 3 13.636 in v ite 3.5 17 10 37.037 issue 3.75 105 70 40 kidnap 3.75 2 3 60 kiss 2.71 0 0 • know 1.88 302 114 27.404 163 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. lik e 2.02 101 0 0 locate 2.86 4 19 82.609 m islead 4.19 3 2 40 name 3.19 42 194 82.203 nominate 4.76 6 3 33.333 note 3.97 234 13 5263 offer 3.38 295 74 20.054 offset 4.72 47 35 42.683 open 2.44 200 1 .498 overw helm 436 4 1 20 p ay 2.69 511 106 17.18 perm it 338 45 18 28.571 phone 1.76 6 0 0 praise 3.41 12 2 14.286 precede 433 4 1 20 predict 3.44 121 2 1.626 preoccupy 435 1 2 66.667 prepare 3.61 39 29 42.647 prevent 3.91 95 1 1.042 price 3.44 15 142 90.446 promote 3.83 34 4 10.526 protect 3.83 62 11 15.068 publish 4.19 44 25 36.232 punish 3.38 6 1 14.286 question 3.75 34 1 2.857 quote 2.77 15 89 85.577 reassure 3.41 4 3 42.857 receive 2.66 341 31 8333 recognize 2.72 29 0 0 recommend 3.66 50 3 5.66 recruit 4.16 2 2 50 refer 4.41 21 4 16 relax 4.12 11 0 0 release 3.57 36 54 60 relieve 3.93 10 7 41.176 remind 2.78 13 0 0 report 3.75 536 72 11.842 reprim and 3.57 0 0 • resist 2.47 19 2 9524 rew ard 3.71 7 1 12.5 rob 3.93 0 6 100 scandalize 4.72 1 0 0 scare 3.34 9 0 0 schedule 4.3 29 100 77.519 sentence 3.84 2 16 88.889 serve 3.28 111 8 6.723 shoot 3.23 22 5 18.519 show 3.21 355 13 3533 s ta rt 3.09 243 9 3571 strengthen 4.39 36 2 5.263 stress 3.66 26 1 3.704 stun 4.09 6 1 14.286 subvert 3.87 2 0 0 sue 2.78 26 6 18.75 164 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. surround 4.27 2 6 75 suspect 2.81 25 2 7.407 su4%nd 3.63 37 27 42.188 sw allow 333 4 2 33.333 teach 3.12 30 2 6.25 th rea ten 351 44 6 12 transform 4.27 21 4 16 transport 3.95 7 2 22.222 trau m atize 4.31 1 0 0 treated 4.47 25 10 28571 trouble 4.09 2 5 71.429 upset 351 4 2 33533 urge 3.18 60 4 6.25 vex 4.09 0 0 • vote 3.9 98 1 1.01 w aive 2.93 12 4 25 w ant 2.24 552 2 .361 warm 3.9 5 0 0 w arn 4 61 1 1.613 welcome 3.42 16 4 20 win 2.78 193 7 3 5 worry 3.33 63 9 12.5 w rite 4.66 132 26 16.456 165 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. APPENDIX III; STIMULI FOR EXPERIMENTS 2 AND 3 Note: Stimuli are given in their active form. The first verb is active-bias and the second verb is passive-bias. The passive versions of sentences included a by-phrase for Experiment 2 but w ere agentless for Experiment 3. 1. After several discussions, the entire committee asked/elected the competent lawyer to represent the firm. 2. Late yesterday m orning, the tall m an phoned/kidnapped the British ambassador during a visit to the U nited Nations. 3. After the meeting, the company chairman advised/appointed the new executive to study the long proposal. 4. According to Sam, the federal judge sued/convicted the controversial Congressman for bribery in a much publicized case. 5. Before/After the deliberations, the city court rem inded/acquitted the plastic surgeon of the malpractice charges. 6. Late last night, the conference participants congratulated/ nom inated the young attorney o n /fo r her exceptional public relations skills. 7. According to Jane, the famous scientists hugged/intim idated the prom ising physicist after she/they w on the prestigious award. 8. Earlier this morning, the Dallas court instructed/sentenced the tabloid journalist to complete several hours of com m unity service. 9. At the agency, the perky secretary approached/greeted the new clients when they walked in. 10. After careful inspection, the agriculture agency urged/fined the new farm er to use better/for using harm ful insecticides. 11. Late yesterday morning, the weird vagrant called/robbed the attractive wom an who was waiting /w hile she was at the railway station. 12. Sometime last week, the handsom e m an phoned / kidnapped the young tourist while she was in the hotel lobby. 166 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 13. On our recom mendation, the team coach punished/expelled the ball player for using drugs. 14. To everyone's delight, the top executive advised/appointed the prom inent banker on the new commission for the business. 15. During the sum m er, the state court sued/convicted the software company for/of discrim ination against women. 16. Earlier this week, the Sacramento judge rem inded/acquitted the nervous defendant of all the charges. 17. During the celebration, the soccer players congratulated/nom inated the goal keeper o n /fo r the much-coveted Popularity trophy. 18. Sometime this afternoon, the old judge instructed/sentenced the angry man to pay a fine of over $2,000. 19. Late that day, the tour guide approached/greeted the bus passengers with a smile. 20. Last Tuesday afternoon, the town baker called/robbed the old lady about all the m oney/of all her money. 21. According to Sam, the strict principal punished/expelled the three students for cheating on the exam. 22. W hen everyone approved, the tired workers asked/elected the new representative about/for handling negotiations w ith the management. 23. After the performance, the dance teacher hugged/ intim idated the talented ballerinas because she was very im pressed/everyone was very impressive. 24. In downtown L.A., the street cop urged/fined the careless driver to stop at red lights/for stopping at green lights. 167 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. APPENDIX IV: STIMULI FOR EXPERIMENT 4 Note: The first verb is adjunct biased while the second verb is argum ent biased. The first continuation is an adjunct while the second continuation is an argum ent of the verb. 1. The Mercedes was parked/driven by the apartm ent building/parking attendant. 2. The car was parked/driven by the building entrance/com pany driver. 3. The painting was located/hung by the big w indow /interior decorator. 4. The picture was located/hung by the large bookcase/m useum employee. 5. The magazines were piled/stacked by the back door/teenage boy. 6. The books were piled/stacked by the broken typew riter/biology student. 7. The boxes w ere stored/opened by the front door/com pany movers. 8. The CDs were stored/opened by the Sony stereo/general manager. 9. The fountain was situated/built by the philosophy library/hired contractors. 10. The well was situated/built by the little barn/construction workers. 11. The pennies were scattered/collected by the white lam p/old lady. 12. The seeds were scattered/collected by the orange tree/good farmer. 13. The chemical was m easured/researched by Robinson's latest m ethod/new scientist. 14. The fluid was m easured/researched by McCarthy's new technique/lab technician. 15. The report was illustrated/m odified by adding nice figures/the gifted student. 16. The argum ent was illustrated/m odified by providing clear exam ples/the speech writer. 168 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 17. The transaction was com pleted/review ed by eleven thirty a.m ./the bank clerks. 18. The article was com pleted/review ed by the m idnight deadline/UCLA professor. 19. The treasure w as hidden/claim ed by the apple tree/young children. 20. The gold was hidden/claim ed by the mine entrance/old explorer. 169 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.

Asset Metadata

Creator Lalami, Laila (author)

Core Title Frequency in sentence comprehension

Contributor Digitized by ProQuest (provenance)

Degree Doctor of Philosophy

Degree Program Linguistics

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

Tag language, linguistics,oai:digitallibrary.usc.edu:usctheses,OAI-PMH Harvest,psychology, experimental

Language English

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c17-266820

Unique identifier UC11352853

Identifier 9733083.pdf (filename),usctheses-c17-266820 (legacy record id)

Legacy Identifier 9733083.pdf

Dmrecord 266820

Document Type Dissertation

Rights Lalami, Laila

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