Journal of Experimental Psychology: 1993, Vol. 19, No ...€¦ · Lexical Preference From...

Journal of Experimental Psychology:Learning, Memory, and Cognition1993, Vol. 19, No. 3,528-553

Cop> right 1993 b> the American l\>chulogicdl Association. Itv1)278-7393/93/5.3.00

Verb-Specific Constraints in Sentence Processing: Separating Effects ofLexical Preference From Garden-Paths

John C. Trueswell, Michael K. Tanenhaus, and Christopher Kello

Immediate effects of verb-specific syntactic (subcategorization) information were found in across-modal naming experiment, a self-paced reading experiment, and an experiment in which eyemovements were monitored. In the reading studies, syntactic misanalysis effects in sentencecomplements (e.g., "The student forgot the solution was. . .") occurred at the verb in the comple-ment (e.g., was) for matrix verbs typically used with noun phrase complements but not for verbstypically used with sentence complements. In addition, a complementizer effect for sentence-complement-biased verbs was not due to syntactic misanalysis but was correlated with howstrongly a particular verb prefers to be followed by the complementizer that. The results supportmodels that make immediate use of lexically specific constraints, especially constraint-basedmodels, but are problematic for lexical filtering models.

Many aspects of language comprehension take place rap-idly, with readers and listeners making commitments to atleast partial interpretations soon after receiving linguistic in-put (e.g., Altmann & Steedman, 1988; Crain & Steedman,1985; Frazier, 1989; Frazier & Fodor, 1978; Marslen-Wilson& Tyler, 1987; Tyler, 1989). The on-line nature of compre-hension has important consequences for syntactic processing(parsing). Immediate interpretation requires some local syn-tactic commitments even though sentences often containtemporary ambiguities. As a result, readers and listeners willoccasionally make incorrect commitments that will requirerevision when, an ambiguity is resolved at a later point inprocessing. The frequency of syntactic misanalysis orgarden-pathing will depend on the types of commitmentsmade by the system and the information used to determinethese commitments. A system that makes complete syntacticcommitments using only a restricted domain of syntactically

John C. Trueswell, Michael K. Tanenhaus, and ChristopherKello (now at Department of Psychology, University of California,Santa Cruz), Department of Psychology, University of Rochester.

This work was partially supported by National Institutes ofHealth (NIH) Grant HD 27206 and by NIH Grant EY 01319 to theCenter for Visual Science.

We are grateful to Susan Garnsey for providing us with prelim-inary versions of her completion norms as well as for helpfulsuggestions and comments on the article. The completion normsreported here were collected by Christopher Kello as part of hissenior honors thesis in cognitive science. Experiment 1 is an ex-tension of work that Kello conducted for his thesis. This articlebenefited from suggestions by the reviewers, Chuck Clifton,Fernanda Ferreira, Maryellen MacDonald, and Chuck Perfetti.Some of this work was reported at the November 1991 meeting ofthe Psychonomics Society, San Francisco.

Correspondence concerning this article should be addressedto John C. Trueswell, Department of Psychology, University ofRochester, Rochester, New York 14627. Electronic mail may besent to [email protected].

relevant information will necessarily make more mistakesthan a system making only partial commitments using awider range of constraining information.

Individual words are one of the richest sources of syntac-tically relevant information. Perhaps the most importantclass of lexically specific constraints are those associatedwith verbs. Verbs provide both semantic and syntactic con-straints on the kinds of complements or arguments withwhich they can occur. If immediately accessed and used, thisinformation would be useful in ambiguity resolution. It couldalso be used to coordinate many of the different sources oflinguistic and nonlinguistic knowledge that need to be in-tegrated during on-line comprehension (Marslen-Wilson,Brown, & Tyler, 1988; Tanenhaus, Garnsey, &Boland, 1991;Tanenhaus & Carlson, 1989; Tanenhaus, Carlson, &Trueswell, 1989).

This article focuses on verb-specific syntactic constraints.Following Chomsky (1965), we refer to lexically specificinformation about potential syntactic complements as sub-categorization information. Verb subcategorization informa-tion is particularly informative in languages like English inwhich a verb typically precedes most of its complements. Insome cases, the syntactic analysis of an otherwise ambigu-ous word or phrase can be disambiguated by the subcate-gorization properties of the preceding verb, as Example Iillustrates.

a. The doctor visited the child .b. The doctor insisted the child

(Example 1)

In Example la the noun phrase "the child" is a noun phrasecomplement (i.e., the object of the verb), whereas in lb thesame phrase is the subject of a sentence complement (e.g.," . . . the child needed braces"). Visited permits a noun phrasecomplement but not a sentence complement, whereas just theopposite is true for insisted.

Verbs often permit several different types of complements.For example, both the verbs remember and claim permit ei-ther a noun phrase complement or a sentence complement.

528

VERB INFORMATION AND SENTENCE PROCESSING 529

a. The doctor remembered the idea . . . . (Example 2)b. The doctor suspected the idea . . . .

As a result, in 2a and 2b the relationship between the nounphrase "the idea" and the preceding verb is ambiguous.Nonetheless, the verb still provides potentially useful infor-mation. Remembered is more frequently used with a nounphrase complement, whereas suspected is more frequentlyused with a sentence complement. As Ford, Bresnan, andKaplan (1982) have suggested, frequency information likethis would be extremely useful in ambiguity resolution. In-deed, in 2a there seems to be a strong preference to take "theidea" as the object of the verb, whereas in 2b the preferenceis for "the idea" to be the subject of a sentence complement.

When verb subcategorization information is accessed andused has important implications for parsing. A parser thattakes subcategorization information into account in makinginitial syntactic commitments will necessarily make fewermistakes than a parser that uses subcategorization informa-tion as a filter to evaluate commitments made without the useof verb-specific information (Mitchell, 1987). How sub-categorization information is used also has implications fordifferent classes of parsing models, which can be highlightedby contrasting two currently influential approaches to am-biguity resolution: constraint-based models and two-stagemodels.

Constraint-based models treat ambiguity resolution as acontinuous constraint-satisfaction process (McClelland, St.John, & Taraban, 1988; Taraban & McClelland, 1990). Incurrent structurally driven variants, syntactically relevant in-formation provides evidence about partially activated alter-natives (MacDonald, 1992; Pearlmutter & MacDonald,1992; Spivey-Knowlton, Trueswell, & Tanenhaus, 1992;Trues well, Tanenhaus, & Garnsey, 1992). Because verbs area source of strong local constraints, one would expect to seeimmediate effects of subcategorization as well as effects ofcategory-based configurational constraints. The relative con-tributions of these sources of information will depend ontheir reliability. In addition to subcategorization effects, thereshould also be effects of lexical co-occurrence patterns, suchas the frequency with which a verb occurs with a particularword in a given structure.

Two-stage models assume that parsing proceeds serially,with only one structure under active consideration. Duringthe first stage, a restricted domain of syntactically relevantinformation is used to posit an initial structure. This structureis then evaluated and, if necessary, revised. The evaluationand revision stage can make use of information that is notused in initial structure building. The two-stage model thatis currently most influential in the literature is the garden-path model, originally proposed by Frazier and Rayner(1982). In this model, initial structure building is guided bya small set of maximally general attachment principles. Forexample, one such principle, minimal attachment, states thatat points of ambiguity, the parser prefers to build the structurewith the fewest number of nodes consistent with the grammarof the language. Garden-paths occur whenever the structurethat conforms to the relevant attachment principles turns outto be incorrect.

Within the architecture of the garden-path model, lexicallyspecific information is most naturally used in the revisionstage. As Frazier (1987, 1989) has argued, using subcate-gorization information in initial structure building wouldlimit the scope and generality of the attachment principlesand require many attachment decisions to be delayed. Thus,most proponents of the garden-path model have argued forlexical-filtering approaches (Clifton, Speer, & Abney, 1991;Ferreira & Henderson, 1990, 1991; Frazier, 1987; Frazier &Clifton, 1989). Early use of subcategorization information isnot, however, incompatible with a two-stage serial approachto parsing. For example, Ford et al. (1982) combined theimmediate use of lexically specific information with a dif-ferent set of attachment principles than those argued for byFrazier and colleagues.

There is a growing body of psycholinguistic literature fo-cusing on the use of verb-based information in parsing. Manyof these studies find evidence for the early use of both thesemantic and the syntactic aspects of verb argument struc-ture. (For a recent review see Boland & Tanenhaus, 1991.)However, two recent studies report results that would seemto provide striking support for lexical filtering (Ferreira &Henderson, 1990; Mitchell, 1987). In these studies, readersseemed to be garden-pathed because they initially parsed anoun phrase as the object of a verb that does not permit a nounphrase complement.

Mitchell (1987) used sentences such as Examples 3a and3b in a self-paced reading study. The sentences were pre-sented in two segments; the slashes mark the segmentation.

a. After the child visited the doctor/prescribeda course of injections. (Example 3)

b. After the child sneezed the doctor/prescribeda course of injections.

In Mitchell's study, the first segment in 3a was read morerapidly than the first segment in 3b, whereas the oppositepattern obtained for the second segment. Mitchell argued thatthe results from the first segment demonstrate that readersinitially treated the noun phrase "the doctor" as a noun phrasecomplement for both transitive verbs (visited) and intransi-tive verbs (sneezed). This misanalysis was then revised bysubcategorization information for the intransitive verbs, re-sulting in longer reading times to the first half of 3b. Fortransitive verbs, reanalysis did not begin until the secondsegment.

While these results are consistent with delayed use of lex-ical information, they are far from decisive (Gorrell, 1991).Subordinate clauses such as "after the patient sneezed . . . "are typically followed by commas. The absence of a commaafter the verb may have indicated to subjects that the clausewas continuing (e.g., "after the child sneezed several times. . . "). This would have been reinforced by the segmentationthat was used. The absence of the comma and the presen-tation segmentation, therefore, would have conflicted withthe subcategory information of sneezed and the content of thenoun phrase "the doctor," which, taken together, provide ev-idence for a clause boundary after the verb. This conflictcould have resulted in elevated reading times. Indeed, in asubsequent study, Mitchell (1987) reported that these effects

530 J. TRUESWELL, M. TANENHAUS, AND C. KELLO

are eliminated when a comma is inserted after the verb. Onecould, of course, argue that commas are needed preciselybecause subcategory information is not sufficient to preventmisanalysis. However, it is unclear how one can resolve thequestion of whether it is the absence of the comma that causesthe difficulty or the absence of verb information that makesthe comma important. It seems unlikely, then, that experi-ments with sentences that violate standard punctuation prac-tice will provide definitive evidence regarding when sub-categorization information is accessed and used.

Ferreira and Henderson (1990) reached the same conclu-sions as Mitchell, using the noun phrase/sentence comple-ment ambiguity, which does not hinge on punctuation. (SeeRayner & Frazier, 1987, and Kennedy, Murray, Jennings, &Reid, 1989, for other relevant work examining this ambi-guity.) They used materials like those in Example 4. Eachsentence contained a verb that prefers to be used with a nounphrase complement (e.g., suspect) or a verb that prefers to beused with a sentence complement (e.g., pretend). The verbwas followed by a sentence complement that did or did notcontain the complementizer that (e.g., "[that] Jack ownscredit cards").

a. She suspected Jack owns credit cards. (Example 4)b. She suspected that Jack owns credit cards.c. She pretended Jack owns credit cards.d. She pretended that Jack owns credit cards.

Ferreira and Henderson used these materials in an eye-tracking study and two self-paced reading studies. The ques-tion that they addressed was whether readers would initiallytreat "Jack" as a noun phrase (NP) complement, regardlessof verb type. If so, reading times at the disambiguating verb"owns" should be longer in the sentences without comple-ments. This was, in fact, the result that Ferreira and Hend-erson observed in the eye-tracking study and in a self-pacedreading study with a one-word moving window. A secondself-paced reading study with an accumulating display didnot produce reliable results.

Ferreira and Henderson (1990) concluded that readers ini-tially attached the subject of the sentence complement("Jack") as the object of the preceding verb, regardless ofverb bias. This attachment would be predicted by minimalattachment because establishing an NP complement requirespostulating fewer nodes than a sentence (S) complement.When the reader encountered the verb of the complement(e.g., "owns"), reanalysis was required because tensed verbsin English require a subject.

Ferreira and Henderson's (1990) results differed, however,from results obtained by Holmes, Stowe, and Cupples (1989)with similar materials. Holmes et al. conducted three exper-iments using several different self-paced reading methods.They found increased reading times for sentences withoutcomplementizers for noun phrase complement biased verbs,but not for sentence complement biased verbs. The resultswere clearest when subjects made a grammatical judgmentor when an accumulating display was used. The results wereless clear-cut with moving window presentation. These re-sults were interpreted as evidence for the early use of verbsubcategorization information.

Ferreira and Henderson (1991) attributed the differencesbetween their results and those of Holmes et al. (1989) toprocedural differences among the studies. In particular, theycriticized the self-paced reading methods used by Holmeset al. However, there are several aspects of the Ferreira andHenderson (1990) materials that might have biased their ex-periment against finding immediate verb effects. First, itturns out that the verb-bias manipulation used by Ferreira andHenderson was relatively weak. Second, many of the nounphrases were implausible objects for the noun phrase com-plement biased verbs. Third, about half of the sentences inthe experiments contained sentence complement construc-tions, which could have introduced a set effect in favor ofthis structure. We will return to these issues in more detaillater. We focus here on whether the effects reported inFerreira and Henderson (1990) actually constitute evidencefor a garden-path.

The main result reported by Ferreira and Henderson (1990)was that reading times were longer for sentences withoutcomplementizers, even with sentence complement biasedverbs. However, the source of the complementizer effect forthese verbs might not be a local garden-path. There are atleast two plausible explanations for why such an effect mighthave occurred even if the sentence complement was correctlyparsed.

The first concerns local syntactic complexity. The struc-ture associated with parsing a noun phrase as the subject ofa sentence complement is more complex, that is, requirespostulating more nodes, than the structure associated withparsing it as a noun phrase complement (object). Indeed, thiscomplexity difference is the reason why minimal attachmentpredicts a noun phrase complement preference. Assume thatlocal processing load is partially determined by local nodedensity, as Frazier (1985) has suggested. When a comple-mentizer is present, the sentence complement structure canbe built when the complementizer that is encountered. How-ever, in the absence of a complementizer, the sentence com-plement structure would not be built until the parser encoun-tered the noun phrase, resulting in greater processing loadthan with the complementizer-present condition.' One mightwonder why a parser that has immediate access to subcat-egory information would not posit a sentence complement assoon as it encountered an appropriate verb. In fact, verbs thathave a preference for a sentence complement over a nounphrase complement often have a stronger preference to beused intransitively or with another complement. Thus, it is

1 Note that the node density hypothesis does not predict thatprocessing load will be greater for a complementizer than for adeterminer of comparable length and frequency (e.g., "the maninsisted that" compared with "the man visited some") because thesame number of nodes would be added in either case. A comple-mentizer would fill the comp node for an embedded (S)entence,whereas a determiner would be the (Spec)ifier of a noun phrase.Thus, two nonterminal nodes would be added to a parse tree ineach case. In contrast, at least three nonterminal nodes would berequired when a determiner that follows a verb is part of a nounphrase that is the subject of a sentence complement (Spec, NP,and S).


the noun phrase that provides the first reliable evidence fora sentence complement. Ferreira and Henderson (1990) typ-ically used short one-word noun phrases (e.g., "Jack"), andcomplexity effects in both self-paced reading and in eye-tracking often spill over onto the next word. Therefore, onewould expect a complexity difference between the comple-mentizer-present and -absent conditions to have affectedprocessing at both the noun and the next word, which wastypically the disambiguating verb. Ferreira and Hendersonactually found a suggestion of a complementizer effect atthe noun, although it did not reach significance until theverb.

The second explanation hinges on the fact that many verbs,especially in written English, are typically followed by thecomplementizer that when they are used with a sentencecomplement. Constraint-based systems would predict thatlexically specific co-occurrence information is coded by thelanguage-processing system (Hindle & Rooth, 1990), per-haps in terms of activation levels for expected words or struc-tures (Elman, 1991; MacDonald, 1992; Trueswell et al.,1992). If this is the case, the processing system would oftennot expect a sentence complement unless it encountered athat. If" unexpected words or structures take longer to pro-cess than more expected structures, an overall complemen-tizer effect would emerge. Of course, the complexity andrftaj-expectation explanations are not incompatible with oneanother.

These alternative explanations for the complementizer ef-fect present a methodological challenge, which extends be-yond the noun phrase/sentence complement ambiguity.Greater processing complexity for ambiguous as comparedwith unambiguous structures is standardly taken as evidencefor syntactic misanalysis. However, this effect might arise forother reasons, an idea that was initially explored in the earlysentence-processing literature (e.g., Hakes, 1972). Thus, itbecomes crucial to separate processing load effects that aredue to syntactic misanalysis from those that are not.

The research reported here addressed three questions:First, how rapidly does subcategorization information be-come available to the processing system? Second, is there acomplementizer effect for sentence complement biasedverbs that is independent of syntactic misanalysis, and if so,can it be separated from syntactic misanalysis effects?Third, when is subcategorization information used in resolv-ing the attachment of a noun phrase that would otherwise beambiguous?

To answer these questions, we made use of three differentmethodologies: cross-modal naming, self-paced reading, andmonitoring eye movements during reading. In Experiment 1,we used a cross-modal naming task to investigate if, withunambiguous sentences, we could find both evidence for im-mediate effects of verb subcategorization information andevidence for a complementizer effect. The results demon-strated a complementizer effect that was correlated with howstrongly individual verbs prefer to be used with a that whenthey are followed by a sentence complement. In Experiment2, we used self-paced reading to investigate whether sub-categorization information can be used to correctly parse anotherwise ambiguous noun phrase and to investigate whether

a complementizer effect would also be found in reading. Theresults produced both of these effects and again indicated thatthe complementizer effect was due to expectations for a that.In Experiment 3, the same materials were used in an eye-movement study in order to replicate the self-paced readingresults under more natural reading conditions. Monitoringeye movements also allowed us to use measures that arearguably more sensitive to initial parsing.

Experiment 1

This experiment had two goals. The first was to determinewhether subcategorization information is accessed rapidlyenough to affect the processing of the word that immediatelyfollows a verb. The second goal was to determine whetherdeleting the complementizer that makes processing a sen-tence complement more difficult, even when no potentialambiguity is involved.

We took advantage of the fact that pronouns in Englishhave different morphological forms depending on their case.For example, the form for the singular masculine pronoun ishim when it is used as an object pronoun (accusative case)and he when it is used as a subject pronoun (nominativecase). Thus, him is an NP complement when it follows a verb,whereas he is the subject of a sentence complement. He andhim were used as targets in a cross-modal integration studyin which subjects heard an auditory fragment and then nameda visually presented target word. The auditory fragmentsended with either an NP-bias verb or an S-bias verb (e.g., "theold man accepted/insisted . . .") or with a complementizer(e.g., "the old man accepted/insisted that..."). Prior researchusing this paradigm has demonstrated that naming latenciesto target words that are ungrammatical continuations of thecontext are longer than naming latencies to grammatical con-tinuations (Cowart, 1987; Tyler & Marslen-Wilson, 1977;West & Stanovich, 1986).

For fragments that end with that, latencies to name himshould be longer than latencies to name he, regardless of verbtype. This is because the nominative pronoun {he) is gram-matical in this environment, whereas the accusative pronoun{him) is not. This data pattern would demonstrate the sen-sitivity of the task and also show that subjects are able tomake rapid use of case information.

Naming latencies to he and him following fragments with-out a complementizer (e.g., "the old man accepted/in-sisted . . .") can be used to diagnose whether subcategori-zation information is used. If this information is used, sub-jects should be sensitive to whether the case of the pronounis congruous with the subcategorization properties of theverb. Naming times to him provide the clearest evidence.If subcategorization information is available, naming laten-cies should be longer for S-bias verbs than for NP-bias verbs(e.g., "the old man insisted" . . . him > "the old man ac-cepted" . . . him). Moreover, for S-bias verbs, naming laten-cies to him should be similar regardless of whether the frag-ment ends with a complementizer (e.g., "the old maninsisted" . . . him = "the old man insisted that" . . . him).However, if subcategorization information is not available,


naming latencies to him should be similar for S-bias andNP-bias verbs.

Naming latencies to he can be used to diagnose whetheror not there is a complementizer effect. If there is a com-plementizer effect, naming times to he should be longer whenthe fragment ends with a verb than when it ends with a com-plementizer. Crucially, this effect should be seen for theS-bias verbs as well as for the NP-bias verbs (e.g., "the oldman insisted" . . . he > "the old man insisted that" . . . he).

In sum, this experiment should reveal whether there is acomplementizer effect even under conditions where there isnot a potential attachment ambiguity. Evidence for a com-plementizer effect would come from a main effect of com-plementizer presence on naming times to the subject pronounhe, including an effect of complementizer presence for S-biasverbs. In addition, use of subcategorization informationwould be reflected in an interaction between verb type andcase for fragments without complementizers.

Table 1Example of Target Stimuli for Experiment 1

Visually presented pronoun

Method

Subjects

Thirty-six undergraduates from the University of Rochester par-ticipated in the experiment for course credit. All subjects werenative English speakers.

Preliminary Normative Study

A sentence completion study was conducted on a separate groupof 14 University of Rochester undergraduates. Target items con-sisted of 50 sentence fragments of the type shown in Example 5.

John insisted[subject completion was written here]. (Example 5)

Each target item began with a person's name and ended with a verb.Fifty verbs were tested. The 50 experimental items were intermixedwith a set of 30 fillers that were composed of a variety of gram-matical constructions. Subjects were asked to write down, asquickly as possible, the first completion that came to mind. Eachsurvey was later scored by hand. Completions were grouped intoone of three categories: (a) sentence complement completions, (b)noun phrase complement completions, and (c) completions usingsome other construction (typically a prepositional phrase or infin-itival complement).

Appendix A presents the results. None of the verbs we tested werealways completed with sentence complements. Typically, the verbshad some completions that were neither a sentence complement nora noun phrase complement (e.g., about phrases, infinitive comple-ments, etc.). We used verbs that preferred sentence complements (Sbias) and verbs that preferred noun phrase complements (NP bias).Appendix 1 indicates which verbs were used in each study.2

Materials

Of 16 verbs that were selected, 8 were NP-bias verbs and 8 wereS-bias verbs. An additional constraint was placed on the selectionof S-bias verbs. If the verb could ever be used with a noun phrasecomplement, the complement could not be animate. Three ofthe S-bias verbs had a few NP complement completions: realize(e.g., "Bill realized his goal"), decided (e.g., "Bill decided the dis-

Auditory sentence fragment

Sentence-bias verbComplementizer absent:"The old man insisted"Complementizer present:"The old man insisted that"

Noun-phrase-bias verbComplementizer absent:"The young boy observed"Complementizer present:"The young boy observed that"

Accusativecase

Him

Him

Him

Him

Nominativecase

He

He

He

He

pute between . . . "), and implied (e.g., "Bill implied the answerwhen . . . "). The other 5 S-bias verbs were never completed withan NP complement. Each of the 16 verbs was incorporated into asentence fragment that began with a noun phrase consisting of adeterminer, an adjective, and a common noun (e.g., "the old man")and ended with either a verb ("the old man insisted") or the com-plementizer that ("the old man insisted that . . . "). Each fragmentwas then paired with two targets: the nominative pronoun he andthe accusative pronoun him. Table 1 presents an example stimulusset. Altogether then, there were two within-item factors, comple-mentizer and case, and one between-item factor, verb type. The fourfragment-target combinations for each item were rotated throughfour presentation lists. The 16 experimental trials on each list wereconstructed so that no subject would encounter the same sentencefragment twice. The 16 experimental materials in each list werecombined with 42 distractor trials of the same construction for atotal of 58 trials. Each target trial was followed by at least 1 dis-tractor trial. The distractors consisted of 30 sentence fragments ofa variety of syntactic types and 12 other fragments, 6 of which hadan accusative pronoun target and 6 of which had a nominative pro-noun target. Distractor trials and experimental trials appeared inpseudorandom order in each list.

The sentences were produced by a rehearsed male native Englishspeaker and digitally recorded onto a Macintosh II computerequipped with a Farallon MacRecorder digitizer. The speech wassampled at 22 kHz and edited using Farallon SoundEdit software.Digital stereo channels were used. One channel contained the sen-tence fragment, and the other channel contained a brief 1 -kHz tonethat coincided with the offset of the final syllable of the sentence

2 Readers may be concerned that the materials (and the correla-tional analyses we present later) are based on results from a sampleof only 14 subjects. Susan Garnsey at the University of Illinois hasrecently conducted an extensive completion study using 107 sub-jects. All of the verbs that we used were included in her study.Garnsey's results for these verbs were strikingly similar to ours.For the verbs used in our experiments, the percentage of NPcomplement completions and the percentage of S complementcompletions were highly correlated between the two norming stud-ies; for percentage of NP completions, Pearson R = 0.935, p < .01;for percentage of S completions, Pearson R = 0.916, p < .01. Inaddition, the percentage of sentence complement completions thatcontained a that was also highly correlated, Pearson R = 0.767,p < .01. When we removed the one verb that had discrepant results(decided), the Pearson R was 0.971.


Table 2Naming Latencies (in ms) and Percentages of Trials Judged Not to Be Good Continuations for Experiment 1

Verb type

Noun phrase biasSentence bias

M

He

Latency

499486493

Complementizer present

% notgood

132

Him

Latency

533539536

M

% notgood Latency

97 51699 51398

% notgood

4951

He

Latency

532519526

Complementizer absent

Him

% notgood Latency

60 49218 53239 512

% notgood

79451

M

Latency

512526

% notgood

3456

fragment. Both channels were recorded onto a stereo cassette tapeusing a cassette tape recorder.

Procedure

Subjects were tested individually. Each subject was seated infront of an IBM PC or PC compatible computer equipped with aDigitry CTS timing system. The subject wore headphones con-nected to the cassette tape recorder. On each trial, the auditorysentence fragment was presented binaurally. The inaudible tone atthe end of the sentence fragment triggered presentation of the targetword, which appeared in the center of screen. The subject thennamed the word aloud as fast as possible into a microphone, andnaming latency was recorded using a voice key connected to theDigitry response box.

After naming the word aloud, the visual target was replaced withthe words "Good Continuation?" The subject responded by pressingeither a YES button or a NO button on the Digitry response box. Theexperimenter was present in the room and recorded any trials onwhich (a) the subject named the word incorrectly, (b) an extraneousnoise such as a cough caused the voice key to trigger too soon, and(c) the voice key failed to respond when the word was named.

Results

As mentioned earlier, we were able to identify only eightS-bias verbs that met our criteria. With the design we used,there was a maximum of only two data points per conditionper subject. In order to reduce the variability in the data, wedid not analyze data from any subject who had missing dataon any of the experimental trials. The experiment was rununtil 36 subjects met this criterion.3 Missing data occurredon about 5% of the trials and was nearly always due to trialsin which the subject did not speak loudly enough to triggerthe voice relay.

Naming Latencies

The naming latency data were analyzed in analyses of vari-ance (ANOVAs) using subjects and items as random factors.Each ANOVA contained four factors: verb type, complemen-tizer (present or absent), case, and a grouping factor (list inthe subject analysis and item group in the item analysis).Table 2 contains the mean naming latencies for eachcondition.

An omnibus ANOVA revealed that case of pronoun was theonly main effect that approached significance, F,(l, 32) =3.18, p< A,MSe = 5,104; F2(l, 8) = 4.60, p< A,MSe =

873. There were significant interactions between verb typeandcaseofpronoun,F,(l,32) = 7.03,p< .05,MSe = 3,473;F2(l , 8) = 5.67, p < .05, MSe = 4,953, and between com-plementizer presence and case of pronoun, F](l, 32) = 20.20, p < .01, MSe = 2,889; F2(l, 8) = 22.49, p < .01, MSe

= 5,270. The interaction between complementizer presenceand verb type was marginal in the subject analysis but notsignificant in the item analysis, F]( 1,32) = 2.7'5,p = A,MSe

= 1,991; F2(l, 8) = 2.75, p > .1, MSe = 608. The tripleinteraction among case of pronoun, complementizer pres-ence, and verb type was marginal in the subject analysis butnot significant in the item analysis, F^ l , 32) = 3.34,p< A,MSe = 1,677; F2(l, 8) = 1.67, p> . l ,MSe = 878.

We will first consider separately the results from fragmentswith and without the complementizer. When the fragmentended with a complementizer (left half of Table 2), him tooklonger to name than he for both verb types. This was re-flected in a significant effect of case, F ^ l , 32) = 13.02,p < .01, MSe= 5,229; F2(l, 8) = 21.38, p < .01; MSe =694; no effect of verb type (Fs < 1); and no interaction withverb type, F,(l, 32) = 2.26, F2(l, 12) = 1.20. This patternwas expected because accusative marked pronouns cannotbe in the subject position of a sentence complement. Thus,this result demonstrates the sensitivity of the experimentand also demonstrates that subjects make immediate use ofcase information.

The pattern for the fragments without the complementizerwas clearly different (right half of Table 2). The fastest nam-ing times were to him following the NP-bias verbs. The slow-est naming times were to him following an S-bias verb andto he following an NP-bias verb. Naming times to he fol-lowing an S-bias verb were intermediate. This resulted in areliable interaction between case and verb type, F](l, 32) =7.25, p < .05, MSe = 3,684; F2(l, 8) = 7.08, p < .05, MSe

= 706. Simple effects tests revealed that the effect of verbtype for the naming of him was reliable in the subject analysisand marginally significant in the item analysis, F,(l, 32) =9.89, p < .01, MSe = 3,013; F2(l, 8) = 4.12, p < A,MSe

= 1,038. There were no reliable effects for the naming of he(Fs < 1). The effects for him demonstrate that verb subcat-egorization information is available when the pronoun is be-ing processed. The results for he suggest that there is also

3 Sixty subjects were run to meet this criterion. An analysis thatincluded subjects with missing data revealed the same pattern asreported here, although it was somewhat noisier.

534 J. TRUESWELL, M TANENHAUS, AND C. KELLO

a complementizer effect associated with establishing a sen-tence complement. In order to explore this in more detail, itwill be useful to consider first the judgment results and thenthe combined results of judgments and latencies for the pro-noun he.

"Good Continuation " Judgments

Table 2 also presents the mean percent no responses to the"Good Continuation?" prompt in the eight conditions of theexperiment. As shown in the table, the type of pronoun hada large effect on judgments when the complementizer thatwas present: When the target was the nominative pronoun he,almost all of the trials were judged as good continuations ofthe sentence; conversely, when the target was him, almostnone of the trials were judged as good continuations of thesentence. When the complementizer was absent, verb typehad a large effect on the judgments to the pronoun him: Whenthe sentence fragment contained an NP-bias verb, almost allof the hims were judged as good continuations, whereas al-most none of the hims were judged as good continuationswhen the fragment had an S-bias verb. The judgments to thepronoun he showed the opposite effect. Only 18% of the heswere judged as not being good continuations when the frag-ment contained an S-bias verb, whereas 60% of the hes werejudged as not being good continuations when the fragmentcontained an NP-bias verb. The average percent no responsesfor each subject and each item were analyzed in separateANOVAs with four factors: list (four lists) or item group(four groups), verb type (NPbias or S bias), complementizerpresence (that or no that), and target pronoun type (accu-sative [him] or nominative [he]).

All of the main effects and interactions were significant orapproached significance, including the effect of verb type,F,(l, 32) = 40.50,/? < .01,MSe = 263; F2(l, 8) = 77.72,p < .01, MSe = 30; the effect of complementizer presence,F,(l, 32) = 6.47, p < .05, MSe = 302; F2{\, 8) = 4.24, p< A,MSe = 99; the effect of case of pronoun, F,(l, 32) =612.10,/? < .01,MSe = 341;F2(1,8) = 1,581.84,/? < .01,MSe = 30; the interaction between verb type and case ofpronoun, F,( 1,32) = 223.20, p< .01, MSe = 336;F2(1,8)= 512.92, p < .01, MSe = 30; the interaction between verbtype and complementizer presence, F\(l, 32) = 35.27, p <.01, MSe = 237; F2(l, 8) = 20.84,/? < .01; MSe = 99; theinteraction between case of pronoun and complementizerpresence, F,(l, 32) = 297.28,/? < .01, MSe = 293; F2(l, 8)= 338.05, p < .01, MSe = 84; and the triple interactionbetween verb type, complementizer presence, and case ofpronoun, F,(l, 32) = 256.27,/? < .01, MSe = 293; F2(l, 8)= 191.58,/? < .0l,MSe = 84.

The Pronoun He

As can be seen in Table 2, naming latencies to the pronounhe showed no hint of an interaction between verb type andcomplementizer presence: He was easier to name when acomplementizer was present. An analysis using the namingtimes to he revealed an effect of complementizer presence,

F , ( l ,32)= 18.93,/?<.01,MSe = 1.998; F2(l, 8) = 18.72,p < .01, MSe = 389; but no effect of verb type, F,(l, 32)= 1.53, MSe = 3,283; F2(l, 8) = 0.23, MSC = 608. Thefactors did not interact (Fs < 1).

Although the percentage of no responses to he increasesfor both verb types when the complementizer is absent, theNP bias show a much larger increase. An analysis on theseresponses revealed a significant interaction between com-plementizer presence and verb type, Fx (1, 32) = 45.22, p <.0\,MSe = 369; F2(1, 8) = 21.68,/? < .01, MSe = 152. Thesmall effect of complementizer presence for the S-bias verbswas significant in the subject and marginal in the item anal-yses, F,(l , 32) = 9.68,/? < .01, MSe = 434; F2( 1,4) = 7.20.p < . 1, MSe = 154, and the effect of complementizer pres-ence for the NP-bias verbs was significant in both analyses,F,(l,32) = 123.79,/? < .Ol.A/5e = 495; F,( 1.4) = 87.12,p < .01, MSe = 151.

The combined results for he were quite revealing. For theS-bias verbs, subjects clearly had more difficulty naming thepronoun when it was not preceded by a complementizer,despite the fact that the case information was unambiguousand despite the fact that verb subcategory information wasavailable, as revealed by the results with him. Therefore, thelonger naming times to he in the complementizer-absent con-dition in S-bias sentences are not likely to be due to a mis-analysis effect. This is supported by the fact that subjectsusually judged he to be a good continuation after the S-biasverbs. While similar latency effects obtained for the stimuliwith NP-bias verbs, these elevated naming times to he in theabsence of a complementizer may be due to a mixture of acomplementizer effect and a grammaticality effect. Supportfor this comes from the fact that 60% of the he trials werejudged to be bad continuations in the complementizer-absentcondition. However, the grammaticality effect and the com-plementizer effect were clearly not additive.

Discussion

The results indicate that verb subcategorization informa-tion was accessed rapidly. In addition, there was a comple-mentizer effect that was unrelated to syntactic misanalysis.For the fragments without a complementizer, there was aclear interaction between verb type and case of the targetpronoun, indicating that subcategorization information wasused. Moreover, naming times to object pronouns {him) wereequally slow following S-bias verbs with and without com-plementizers. Evidence for an increase in processing diffi-culty in the absence of a complementizer came from a maineffect of complementizer on naming times to he. Crucially,naming times to he were longer following an S-bias verbwithout a complementizer than an S-bias verb with a com-plementizer. This effect cannot be attributed to a temporarysyntactic misanalysis for three reasons: (a) subcategorizationwas clearly available, (b) subjects made immediate use ofcase information, and (c) the pronoun he is unambiguouslycase-marked as a subject pronoun.

The conclusion that subcategorization information be-comes available shortly after a verb is recognized is sup-


ported by several recent studies. Shapiro and colleagues havefound that lexical decision times to a visually presented targetare affected by the argument structure of the preceding verbin an auditory fragment (Shapiro, Nagel, & Levin, in press;Shapiro, Zurif, & Grimshaw, 1987, 1989; but cf. Schmauder,Kennison, & Clifton, 1991). Gorrell (1991) reported fasterlexical decision times to noun targets that followed visuallypresented sentence fragments when the fragment ended in anobligatorily transitive verb (e.g., permit) as compared withan obligatorily intransitive verb (e.g., remain). Trueswell(1993) used the verbs from the present experiment as targetsin a cross-modal naming study in which the congruence ofthe verb as a continuation of an auditory fragment dependedon the verb's subcategorization properties. Naming latenciesto NP-bias and S-bias verbs did not differ in contexts inwhich they were both syntactically permissible (e.g., "the oldman . . . accepted/insisted"). However, latencies to S-biasverbs were longer than NP-bias verbs in contexts in whichonly a transitive verb was grammatical (e.g., "the old manwas . . . accepted/*insisted").

The locus of the effects obtained in the present study meritscomment. It seems likely that some of the effects are postlex-ical and that they are due in part to the language productionsystem (Tanenhaus & Lucas, 1987;West'&Stanovich, 1986).In addition, the "Good Continuation?" judgment may havecontributed to the effects, perhaps by making subjects moresensitive to subcategorization information than they mighthave otherwise been. However, Boland (1991, in press) alsofound that subcategorization affected cross-modal naming ina task that did not include a continuation judgment. InBoland's study, nominative and accusative plural pronouns(e.g., they and them) were presented as targets followingauditory fragments that ended in either an S-bias verb or anNP-bias verb that did not permit a sentence complement(e.g., visited). Verb type affected naming times to accusativebut not nominative pronouns, the same pattern reported here.Boland also showed that cross-modal naming was not af-fected by either thematic congruity or plausibility, in contrastto lexical decision, which showed effects of both of thesevariables. This suggests that the cross-modal task is primarilysensitive to syntactic processes. Boland did not, however,include a complementizer condition. It also remains to beseen whether cross-modal naming would be selectively sen-sitive to syntactic congruity when continuation judgmentsare used.

While these are interesting issues, what is crucial for thepresent investigation is that a complementizer effect oc-curred even when subcategorization information was beingused. Thus, a complementizer effect could easily be mistakenfor a garden-path under conditions where the resolution of apotential attachment ambiguity depends on subcategoriza-tion constraints.

Earlier, we hypothesized two different types of explana-tions for why a complementizer effect could occur for S-biasverbs in the absence of any syntactic misanalysis. One ex-planation was that building the structure for a noun phrasethat is the subject of a sentence complement requires morelocal structure building at the noun phrase when there is nota complementizer. The second explanation was that there is

100

n

0 20 40 60 80 100 120Percentage of S complement completions with a "that"

Figure 1. Scattergram plotting the he complementizer effect (inms) against the percentage of sentence complement completionsthat began with a that for each sentence (S) bias verb.

a general preference for a sentence complement to be intro-duced by a that, although the preference varies across verbs.If this information is coded by the processing system in termsof activation levels for potential structures associated withthe verb, then, after some verbs, the parser would not beexpecting a sentence complement unless it encountered athat, whereas for other verbs a that-\ess sentence comple-ment would be a likely structure.

For the S-bias verbs used in this experiment, 58% of thesentence complement completions contained a that. How-ever, the preference for a that varied across verbs and it wasunrelated to the strength of the verb bias; that is, it was un-related to either the percentage of sentence complement com-pletions or noun phrase complement completions. As a result,it is possible to perform post hoc correlations to test betweena local complexity and a r/zar-preference explanation for thecomplementizer effect for S-bias verbs.4 If the complemen-tizer effect for naming he with S-bias verbs is actually dueto a lexical preference for a that, the size of the effect forindividual verbs should vary with the strength of their thatpreference. S-bias verbs that tend not to be used with a thatin a sentence complement should show little or no comple-mentizer effect, whereas S-bias verbs that tend to be usedwith a that should show a large complementizer effect. Fig-ure 1 presents a scattergram in which the complementizereffect is plotted against the percentage of sentence comple-ment completions that contained a that. A regression analysisbetween these two measures revealed a significant positivecorrelation, Pearson r = 0.734, F(l, 6) = 7.38, p < .05, MSe

= 6,035, with a slope of 0.81 and an intercept of-17.7 ms.An additional regression analysis was conducted on a trans-formation of the percentages (\og(p/l-p)) because the per-

4 The same correlation with that preference cannot be performedon NP-bias verbs because almost all sentence complement com-pletions with these verbs contained a that, resulting in primarily100% that preferences (see Appendix A). Moreover, the data in thisgroup are too sparse, because, by definition, an NP-bias verb washardly ever completed with a sentence complement.


centage range is bounded. The correlation was also signif-icant, Pearson r = 0.829, F(l, 6) = 13.13, p < .05, MSe =7,511. The that preference for S-bias verbs did not correlatewith differences in verb bias within the S-bias group (i.e., thepercentage of sentence complement completions, nounphrase complement completions, or their differences), Fs <1 for transformed and untransformed percentages.

The results of these correlations strongly support the that-preference explanation. The size of the complementizer ef-fect for individual S-bias verbs was strongly correlated withthe degree of that preference, and there was no suggestionof a residual complementizer effect. These correlations, inconjunction with the latency results, demonstrate that theprocessing system is sensitive to both subcategory informa-tion and subtle patterns of lexical co-occurrence.

Experiment 2

This experiment used self-paced reading times to inves-tigate when subcategorization information is used in resolv-ing the attachment of a noun phrase that would otherwise beambiguous. It was important to determine whether that-preference effects also occur in reading and, if so, are re-sponsible for effects that have been previously interpreted asmisanalysis effects due to delayed use of verb information.

This study was similar in design to the reading studiesconducted by Holmes et al. (1989) and Ferreira and Hen-derson (1991). We compared reading times to sentential com-plements following verbs that prefer NP complements andverbs that prefer S complements. The sentence complementwas preceded by the complementizer that or directly fol-lowed the matrix verb. All of the sentence complements be-gan with a two-word noun phrase that was a plausible objectfor the sentences with NP-bias verbs. Sample sentences arepresented in Example 6:

a. The student forgot (that) the solution wasin the back of the book. (NP-bias) (Example 6)

b. The student hoped (that) the solution wasin the back of the book. (S-bias)

For the NP-bias verbs without a complementizer, readersshould initially interpret the noun phrase after the verb as anNP complement. Consequently, reanalysis would be requiredwhen the following verb provides unambiguous syntacticevidence that the noun phrase was actually the subject of asentence complement. The results for the S-bias verbs de-pend on when subcategorization information is used. If useof subcategorization information is delayed, the results forthe S-bias verbs should be similar to those for the NP-biasverbs. If subcategorization information is used quickly, as theprevious experiment suggests, the noun phrase will be cor-rectly analyzed as the subject of a sentence complement.Thus, reanalysis would not be required at the verb. In ad-dition, given the results of Experiment 1, there should be a//jar-preference effect at the noun phrase for sentences withS-bias verbs without a complementizer.

Ferreira and Henderson (1990) used self-paced readingwith similar sentences but found only delayed effects of verb

bias. However, there are a number of problems with the Fer-reira and Henderson materials that could have contributed tothis result. First, their verb bias manipulation was relativelyweak. Our sentence completion norms revealed that of the 20S-bias verbs used by Ferreira and Henderson, 2 verbs actu-ally had a strong noun phrase bias (NP bias), and 5 verbsshowed little or no preference toward either complementizertype. Overall, Ferreira and Henderson's S-bias verbs had15% noun phrase complement completions and 43% sen-tential complement completions. Of the 20 NP-bias verbsused by Ferreira and Henderson, 4 verbs showed a strongsentential complement bias (S-bias), and 3 verbs showed lit-tle or no preference. Overall, Ferreira and Henderson's NP-bias verbs had 58% NP completions and 23% S completions.

The importance of using strongly biased verbs is high-lighted by a recent study by Shapiro et al. (in press). Shapiroet al. showed that there are clear individual differences inpreferred completions for verbs with different argumentstructures. In their experiments, which used a cross-modallexical decision paradigm, no verb effects emerged whenverbs were divided according to group preferences. How-ever, clear effects emerged when verb preferences were de-termined individually for each subject. The likelihood ofverb effects being masked by individual differences is great-est when verb bias is relatively weak.

Second, inspection of the materials published in Ferreiraand Henderson (1990) revealed that about half of the nounphrases in the NP-bias sentences were implausible direct ob-jects of the verb (e.g., "Sue wrote Iowa elected better peo-ple," "She admitted fish like the water too"). Because plau-sibility can have immediate effects on ambiguity resolution(Trueswell et al., 1992), this could have reduced the size ofthe misanalysis effect for NP-bias verbs.

Finally, of the 152 sentences used in the Ferreira and Hend-erson experiment, 80 were target sentences with sentencecomplements. Repeated use of sentence complements couldhave introduced set effects in favor of this construction(Mehler & Carey, 1967).

The cumulative effect of weakly biased verbs, nounphrases that were implausible objects, and the repeated useof sentence complements could have been to reduce thegarden-path for Ferreira and Henderson's NP-bias sentences,while leaving intact a that preference effect for S-bias sen-tences. Because Ferreira and Henderson (1990) typicallyused only a single word noun phrase, a that preference forS-bias verbs is likely to appear on the disambiguating verbrather than immediately on the noun. These factors takentogether could have resulted in longer reading times at thedisambiguating region for both S-bias and NP-bias verbs inthe complementizer-absent conditions, with little evidence ofstrong reanalysis effects for either verb type.

In contrast, the present study used strongly biased verbs,noun phrases that were plausible objects for NP-bias verbs,and a relatively small proportion of sentences with sentencecomplements. We also used two-word noun phrases to in-crease the likelihood that a r/raf-preference effect would beresolved prior to the disambiguating region (i.e., the verbphrase in the sentence complement).


Method

Subjects

Forty subjects from the University of Rochester participated inthe experiment for course credit. All subjects were native Englishspeakers.

Materials

The experimental materials were generated from 20 verbs usedin the sentence completion norms (see Appendix B). An exampleset appears in Example 6, given earlier. Each sentence began witha simple noun phrase, containing a determiner and a noun. The firstfactor in the experiment was whether the matrix verb in the sentencehad bias toward taking a noun phrase object (NP bias) or a sentencecomplement (S bias). Two additional verbs that also showed nocompletions using a noun phrase object were added to the S-biasgroup. Also, 2 additional verbs that tend to be completed with directobjects were added to the NP-bias verb group. On the basis of theearlier sentence completion study, the S-bias verbs had 4% nounphrase completions and 64% sentence complement completions,whereas the NP-bias verbs had 78% noun phrase completions and9% sentence complement completions. Following the verb was asentence complement that began with a definite NP. The secondfactor in the experiment was whether this sentence complement waspreceded with the complementizer that or not. The noun phrases thatbegan the sentence complements were chosen to be highly plausible(potential) objects for the sentences with NP-bias verbs. In a sep-arate norming study, sentence completions for fragments endingwith the noun phrase (e.g., "The student hoped/forgot the solution. . .") resulted in 99% noun phrase complement completions for thefragments with NP-bias verbs and 88% sentence complement com-pletions for the S-bias verbs. These percentages were based on com-pletions from 10 subjects.

Each target sentence was followed by a sentence that naturallycontinued the description or story begun by the first sentence. Fourpresentation lists were constructed by pseudorandomly combiningthe 20 experimental materials with 40 distractor texts for a total of60 trials. Each target text was followed by at least 1 distractor text.Each distractor text consisted of a sentence pair that sounded naturaltogether. Each text contained a variety of syntactic structures andverbs with various subcategorizations. Ten of the distractors hadinitial sentences that (a) contained a matrix verb that subcategorizesfor a sentence complement (and was different from the experimentalverbs) and (b) had a noun phrase complement in object position,following the verb. This was done to avoid introducing a bias towardthe sentence complement construction. Thus, out of all the sen-tences in the study, 17% (20 of 120) had a sentence complementimmediately following the main verb.

Procedure

The stimuli were presented on an IBM or IBM-compatible PCequipped with a color monitor and Digitry CTS timing system andresponse box. The monitor was set to a text mode of 80 charactersper line. The subject was seated approximately 72 cm from themonitor (although this distance was not controlled), and the visualangle for each character was approximately 22 min arc. At thebeginning of each trial, the entire text was displayed on the screen,with each character (except spaces) covered by a single dash (-).The subject would then press a button marked SCROLL to uncover andread the first word. With each press of the SCROLL button, the nextword in the sentence would be uncovered and the previous word

would be covered up with dashes again. This was repeated until theend of the text. Reaction times were collected for the reading of eachword in each target sentence. After about a third of the texts, a yes/nocomprehension question was displayed on the screen, and the sub-ject responded by pressing YES or NO on the button box. Subjectswere given feedback concerning whether their answer was correct.

Results

Figure 2 presents the self-paced reading times to the matrixverb and the first four words of the sentence complement foreach of the four conditions. (Reading times to the word thatwere not collected.) The mean reading times for each subjectand item were analyzed in separate ANOVAs with four fac-tors: list (four lists) or item group (four groups); verb type(NP bias, S bias); complementizer presence (that or no that);and word position (matrix verb, the, noun, be verb, and thenext word). We will first discuss overall effects and inter-actions and then look at separate analyses for each wordposition.

Overall Analysis Across Word Position

An effect of word position was significant in the subjectand item analysis, F,(4, 33) = 4.52, p < .01, MSe = 6,124;F2(4, 13) = 5.73, p < .01, MSe = 2,425. We report Hyunh-Feldt adjusted probability levels for all effects involvingword position, because it has more than two levels that arenot independent of one another. However, we report non-adjusted degrees of freedom. There was also an overall effectof complementizer presence, Fi(l, 36) = 13.85, p < .01,MSe = 5,942; F2(l, 16), = 6.34, p < .05, MSe = 6,467; noeffect of verb type (Fs < 1); and no interaction between thesefactors (Fs < 1). However, these factors did interact withword position. There were significant interactions betweenverb type and word position, F](4, 33) = 3.48, p < .05, MSe

= 3,045; F2(4, 13) = 2.86, p < .05, MSe = 1,860; andcomplementizer presence and word position, F,(4, 33) =2.91, p < .05, MSe = 3,821; F2(4, 13) = 3.76, p < .05, MSe

= 1,480. In addition, the triple interaction between verb type,

450

440

430

I 420

~ 410

•s 400

4 390

& 380

370-

360

350

— • — NPbias - absent "The student forgot the..."— E J - - NPbias - present "The student forgot that the..."

— A-- - Sbias -- absent "The student hoped the..."— & • - • Sbias - present "The student hoped that the..."

forgothoped

the solution was

Word position

Figure 2. Word-by-word self-paced reading times (in ms). (NP= noun phrase; S = sentence.)


complementizer presence, and word position was significant,F,(4, 33) = 53\,p< .01, MSe = 3,241; F,(4, 13) = 6.24,p < .0\,MSe = 1,377.

Reading Times at the Matrix Verb

There were no effects or interactions between factors at themain verb of the sentence (Fs < 1).

Reading Times at the Determiner

At the determiner {the), there was an immediate increasein reading times for sentences without the complementizerthat as compared with those sentences containing the that,with the largest difference occurring for the S-bias verbs.There was a main effect of complementizer presence at thisposition, F,(l, 36) = 15.12,p < .01,MSe = 3,821; F-,{\, 16)= 10.97,/? < .01, MSe = 2,630; no effect of verb type, F,(l,36) = 1.51, MSe = 4,102; F2(l, 16) = 0.50, MSe = 6,154;and no interaction between these factors, F,(l, 36) = 1.36,MSe = 4,110; F2(l, 16) = 1.13, MSe = 2,417. An exami-nation of the simple effects revealed no reliable differencebetween the verb types for the sentences without a comple-mentizer, F,(l, 36) = 2.00, MSe = 5,883; F2(l, 16) = 1.05,MSe = 5,525.

Reading Times at the Noun

At the noun, the S-bias sentences showed a smaller dif-ference between sentences without the complementizer thatand sentences with the complementizer that, and no differ-ence in the reading times for NP-bias sentences. There wasa significant effect of verb type for the subject analysis butnot for the item analysis, F,(l, 36) = 4.46, p < .05, MSe =3,163; F2(l, 16) = 0.75, MSe = 9,327; and no effect ofcomplementizer presence, Fj(l, 36) = 1.31, MSC = 4,666;F2(l, 16) = 1.06, MSe = 2,896. The interaction between verbtype and complementizer did not approach significance,F,(l, 36) = 1.88, MSe = 3,868; F2(l, 16) = 1.50, MSe =2,422. An examination of the simple effects revealed that thedifference between the verb types for the sentences withouta complementizer was significant in the subject analysis, butnot significant in the item analysis, F,(l, 36) = 5.06, p <.05, MSe = 4,118; F2(l, 16) = 1.67, MSe = 6,216.

Reading Times at the Complement Verb

At the complement verb, there were no effects or in-teractions that reached significance, all Fs < 1, exceptcomplementizer presence, F,(l, 36) = 2.63, MSe = 2,010;F2(l, 16) = 1.73, MSe = 1,545.

Reading Times at the Final Position

At the word after the be verb, NP-bias sentences showeda large increase in reading times for sentences that did notcontain a that as compared with sentences that did. S-biassentences did not show an increase. This difference was

realized in the analysis as a significant interaction betweenverb type and complementizer presence, F,(l, 36) = 10.34.p< .0\,MSe = 5,317; F2(l, 16) = 10.96, p< ,01,A«c =2,518. There was also an effect of complementizer presence,F,(l, 36) = 8.34, p < .01, MSe = 6,900; F2(l, 16) = 9.43,p < .01, MSe = 3,038, whereas the effect of verb type onlyapproached significance in the subject analysis, F, (1, 36) =3.54, p < .10, MSe = 6,098; F+ 2( l , 16) = 1.63, MSe =6,652. Simple effects revealed that the interaction betweenthese factors was due to a significant effect of complemen-tizer presence in the NP-bias sentences, FjO, 36) = 11.10,p < .01, MSe = 10,137; F2(l, 16) = 13.28, p < .01, MSe

= 4,236; and no effect of complementizer presence in theS-bias sentences (Fs < 1).

Discussion

The results clearly demonstrated that subcategorization in-formation is used to determine the correct attachment of anoun phrase that would be otherwise ambiguous between anNP complement and the subject of a sentence complement.When the preceding verb had a strong bias in favor of an NPcomplement, subjects experienced difficulty when the NPwas disambiguated as the subject of a sentence complement:NP-bias sentences without a complementizer showed a largeelevation at the word after the verb in the sentence comple-ment, whereas no such difficulty obtained for S-bias verbs.(Moving-window reading time differences often appear oneor two words downstream, especially when the reading rateis reasonably fast.) This pattern of results suggests that thenoun phrase was taken to be a noun phrase complement whenit followed an NP-bias verb and the subject of a sentencecomplement when it followed an S-bias verb.

In addition, readers had more difficulty at the determinerthe immediately after the verb when the complementizer wasabsent, regardless of verb type. Although reading times to thedeterminer for S-bias sentences without a complementizerwere numerically longer than reading times at the determinerfor NP-bias sentences, the reading times were statisticallyindistinguishable, and both were reliably greater than thecomplementizer-present sentences. This effect is difficult toexplain with either a lexically based parsing proposal or alexical filtering proposal. The small effect for NP-bias sen-tences cannot be attributed to either a misanalysis effect ora complexity effect because the preferred NP complementconstruction is syntactically less complex. The elevations atthe determiner, however, may be explained by the differencein lexical content in the word prior to the determiner. Thelonger reading times at the determiner may simply be aspillover effect from the verb. In the complementizer-absentstimuli, the was preceded by a complex content word, a verb,whereas in the complementizer-present stimuli, the was pre-ceded by a high-frequency function word that. These dif-ferences in content are likely to be realized in elevations inreading times, which will spill over onto the following word.

Finally, readers had more difficulty at the head of a nounphrase following an S-bias verb when there was not a com-plementizer: Reading times at the noun for S-bias sentences


without a complementizer were numerically and statisticallydifferent from the corresponding NP-bias sentences.

One might argue that the effect at the noun is due to rapidlexical filtering. On this account, the parser attached the nounphrase as the object of the verb and then immediately revisedthe analysis. Recall, however, that a similar complementizereffect was found in Experiment 1 for pronouns that are un-ambiguously the subject of a sentence complement. For alexical filtering hypothesis to be compelling, then, it wouldfirst be necessary to demonstrate an effect above and beyondwhat would be expected for unambiguous structures.

Fortunately, it is possible to provide a more direct test ofthe locus of the complementizer effect by examining the cor-relation between the size of the effect and the strength of thethat preference for the S-bias verbs. If the effect is a that-preference effect, as in Experiment 1, there should be a cor-relation between strength of preference and the size of thecomplementizer effect. Indeed, regression analyses usingthese factors revealed a positive correlation between the pref-erence for a that and the effect size at the noun. Table 3presents these correlations for both the complementizer ef-fect within S-bias sentences (complementizer-absent sen-tences minus complementizer-present) and the verb-type ef-fect for complementizer-absent stimuli (S bias minus NP biasfor sentences without a complementizer). Neither of theseeffects correlated with verb biases (all Fs < 1). We alsoexamined correlations at the determiner. As expected, read-ing times did not correlate with either verb preferences or thatpreferences (all Fs < 1), further suggesting that the effectsat this position were independent of factors related to verbtype.

The regression analyses that we conducted to examine cor-relations with that preference at the noun allowed a criticalprediction made by the lexical filtering hypothesis to be test-ed. The lexical filtering hypothesis predicts that a nounphrase following an S-bias verb will be initially treated as anNP complement, resulting in a need for subsequent reanal-ysis. This reanalysis should result in greater processing dif-ficulty for noun phrases following S-bias verbs without com-plementizers as compared to with complementizers. Arevision effect should occur even for S-bias verbs that typ-ically occur without a complementizer. Therefore, the zerointercept for the regression equation relating that preference

Table 3Correlations of Self-Paced Reading Time Differences atthe Noun With That Preference for Experiment 2

Independentmeasure Slope Intercept Pearson r F(\, 8) p <

Complementizer effect for sentence-bias sentences% that 1.29 -44.7 .597 4.44 .07Trans(%) 10.6 +18.8 .507 2.77

Verb-type effect for sentences without a complementizer% that 1.84 -74.0 .607 4.67 .06Trans(%) 18.4 +20.0 .594 4.37 .07

Note. Trans(%) = transformed percentages. Intercepts for trans-formed data do not correspond to 0% that preference.

to the difference between complementizer-present and-absent conditions should be positive for the untransformeddata. In fact, there was no hint of a residual complementizereffect. The intercept was actually negative. Note that the zerointercept of transformed data does not correspond to a 0%that preference. In fact, this zero intercept corresponds toabout a 50% that preference and should therefore be a pos-itive value. The predicted value corresponding to an actual0% that preference in these regressions is also slightly neg-ative. This was also true for all other regressions with trans-formed percentages reported in this article.

Experiment 3

This experiment used the same materials as those used inExperiment 2. Instead of collecting self-paced reading times,we monitored eye movements as subjects read each sentence.There are several reasons why it was important to replicatethe self-paced reading results with eye-movement data. Self-paced reading with single-word presentation forces readersto fixate on every word and does not allow parafoveal pre-view of upcoming material. As Ferreira and Henderson(1990) suggested, it is possible that readers may rely on lex-ical information more in single-word, self-paced readingthan they would with whole-sentence presentation. In gen-eral, studies that have directly compared self-paced readingwith a single-word moving window and eye-tracking withwhole-sentence presentation have found similar results (e.g.,Britt, Perfetti, Garrod, & Rayner, 1992; Ferreira & Clifton,1986; Ferreira & Henderson, 1990). However, there are someexceptions. For example, the effects of some thematic anddiscourse constraints on syntactic ambiguity resolution withrelative clauses (e.g., "the student spotted by . . . ") are re-duced or eliminated when single-word presentation preventsthe reader from parafoveally viewing a disambiguating prep-osition (Burgess & Tanenhaus, 1992; Spivey-Knowlton etal., 1992; Trueswell et al., 1992).

Eye-movement data can also provide several types of mea-sures that can be used to determine the time course withwhich information is used in reading (Rayner, Sereno,Morris, Schmauder, & Clifton, 1989). Some of these mea-sures might be sensitive to small effects that other taskswould be less likely to detect. Thus, it could be argued thatthere is a small misanalysis effect for S-bias verbs that mightbe detected with eye movements, but not with self-pacedreading times.

Method

Subjects

Twenty-four University of Rochester students participated in theexperiment. Subjects were paid $7 for each hour of their time.

Equipment

The eye movements of each subject were recorded using a Stan-ford Research Institute Dual Purkinje Eyetracker (Fifth Genera-tion). The eyetracker transmitted information concerning horizontal


and vertical eye position angle to a Macintosh II computer equippedwith an analog to digital conversion board. Eye position was de-termined by sampling every millisecond both the horizontal andvertical eye angles and blink signals from the eyetracker. At the endof each trial, fixation positions and durations were computed andstored to disk. Each fixation was represented by an x and v screencoordinate, a starting time, and an ending time. Although eye move-ments were recorded from the right eye, viewing was binocular.Stimuli were displayed on a 13-inch AppleCoIor High ResolutionRGB monitor.

Materials

The materials were identical to those used in the previousexperiment.

Procedure

Because small head movements decrease the accuracy of theeyetracker, a bite bar was made for each subject at the beginningof the testing session. Subjects were given instructions and seatedin front of the computer screen, with the subject's eyes approxi-mately 64 cm from the screen. All sentences appeared in mixed casein Courier 14-point font. The visual angle of each character wasslightly greater than 12 min arc, allowing for one character reso-lution from the eyetracker position signals. At the beginning of theexperiment, the brightness of the monitor was adjusted to the sub-ject's comfort. The eyetracker was aligned and the signal from theeyetracker was calibrated with the screen coordinates. During thecalibration procedure, the subject fixated on a series of screen po-sitions, with the computer sampling the eyetracker at each position.These samples were then used by the computer to derive a set oflinear equations that converted the horizontal eye position signalinto horizontal screen coordinates and the vertical signal into ver-tical screen coordinates.

Each trial consisted of the presentation of the sentence pair. Eachline contained no more than 65 characters, and all the critical scor-ing regions appeared on the first line of text. Before the sentenceswere presented, a fixation cross was displayed at the starting po-sition of the first sentence. The subject fixated on the cross andpressed the computer mouse button to display the sentences. Thesubject read the sentences silently and then pressed the button againto signal that he or she was finished. After each sentence pair, thecalibration was checked by displaying a line trace controlled by thesubject's eye movements and the fixation cross of the next trial. Theline trace was a line that was continually drawn out on the screenindicating the subject's latest eye position. If the experimenterjudged that the eye position did not adequately line up, the computerwas recalibrated. (Recalibrations were typically not necessary.) Onabout a third of the trials, a yes/no question appeared on the screenprior to the line trace test. Subjects answered the question by mov-ing the mouse into either a YES box or a NO box and clicking themouse button. Subjects were given feedback as to whether theiranswer was correct. Each reading session lasted approximately 25min. Subjects were allowed to release from the bite bar betweensentences at any time during the experiment. Subjects usually tookone or two breaks.

Results and Discussion

The results from this study are divided into three sections:tota] processing effects, initial processing effects, and re-processing effects.

Total Processing

The test sentences were divided into four regions: (a) thesubject noun phrase of the sentence, (b) the matrix verb, (c)the subject noun phrase of the sentence complement, and (d)the three words following the noun phrase. Table 4 presentsmean total reading times for each region. Total reading timescorresponded to the total amount of time spent within a re-gion, including rereads of a region. As shown in the table,subcategory biases associated with the matrix verbs clearlyinfluenced reading times. For NP-bias verbs, sentences with-out the complementizer took longer to read than sentenceswith the complementizer. For S-bias verbs, sentences withoutthe complementizer took only slightly longer to read thansentences with the complementizer. In addition, sentenceswith a that also appear to show a small effect of verb type,with NP-bias sentences taking slightly longer to read thanS-bias sentences, especially in the final region.

Subject and item means were entered into separateANOVAs with four factors: list (four lists) or item group(four groups); region (matrix verb, sentence complementsubject NP, and sentence complement verb [three words]);verb type (S bias and NP bias); and complementizer (presentand absent). When collapsing across all of the scoring re-gions, there was a reliable interaction between verb type andcomplementizer presence, F,(l, 20) = 11.76, p < .01, MSe

= 27,681; F2(l, 16) = 7.32, p < .05, MSe = 43,429. Inaddition, there were main effects of verb type, F,(l, 20) =23.10,/? < ,01,MSe = 40,706; F2(l, 16) = 10.42, p< .01,MSe = 71,994; and complementizer presence, F)(l, 20) =30.63, p < .0l,MSe = 25,422; F2(l, 16) = 28.99, p < .01,MSe = 22,083. Simple effects revealed a clear effect of com-plementizer presence for the NP-bias sentences, F,( 1,20) =25.88,/? < .01,MSe = 40,787; F2(\, 16) = 24.09,/? < .01,MSe = 38,616. The effect of complementizer presence for theS-bias sentences, however, was marginal in the subject anal-ysis and not significant in the item analysis, F,( 1,20) = 3.95,

Table 4Mean Total Reading Times (in ms) for Each ScoringRegion for Experiment 3

Scoring region

"The student"

"forgot/hoped"

"the solution"

"was in the"

Verb type

S biasNP bias

M

S biasNP bias

M

S biasNP bias

M

S biasNP bias

M

Complementizer

Absent

459541500

440554497

557677617

7611,073

917

Present

446471459

433433433

523555539

688793741

A

+ 13+70

+7+ 121

+34a

+ 122

+73a

+280

M

453506

437494

540616

725933

Note. S = sentence; NP = noun phrase.a Difference correlated with lexically specific thai preferences.


p < .10, MSe = 12,315; F2(l, 16) = 1.04, MSC = 26,896.Finally, there was an overall effect of scoring region, F,(l,20) = 116.28, p < .01, MSe = 23,421; F2(l, 16) = 70.00,p < .01, MSe = 32.727.

Analysis by region. Similar ANOVAs were conducted ateach scoring region. To conserve space, only the analysis ofthe subject NP and verb phrase regions of the sentence com-plement will be reported. At the noun phrase region, therewas a main effect of verb type, F,(l, 20) = 7.03, p < .05,MSe = 20,034; F2(l, 16) = 4.60, p < .05, MSe = 26,510;and complementizer presence, F,(l, 20) = 6.71, p < .01,MSe = 21,748; F2(l, 16) = 6.41, p < .05, MSe = 16,344;whereas the interaction between these factors did not reachsignificance, F,(l, 20) = 2.91, p = .10, MSe = 16,790;F2(l, 16) = 1.94, MSe = 21,090.

At the final verb phrase region, although there was a clearinteraction between verb type and complementizer presence,F,( l ,20)= 12.84, p < . 0 1 , AfSe = 19,999; F2(l, 16) = 6.29,p < .01, MS,, = 32,413, simple effects revealed an effect ofcomplementizer presence for both the NP-bias verbs,F,(l, 20) = 40.42, p < .01, MSe = 23,344; F2(l, 16) =16.55, p < .01, MSe = 45,850, and the S-bias verbs,F!(l,20) = 5.92, p<. 05, MSe = 10,948; F2(l, 16) = 4.60,p < .05, MSe = 11,740.

Correlations with that preference. Reading times forS-bias verbs were somewhat longer at both the noun phraseand the verb phrase in the sentences without a complemen-tizer in comparison with the sentences with a complemen-tizer. We again examined the correlations between the sizeof the complementizer effect and the strength of that pref-erence for the S-bias verbs (see Table 5). We did not computecorrelations with the NP-bias sentences without a comple-mentizer because both the noun phrase and the verb phraseregions contain long second-pass reading times that are dueto the garden-path found in this condition. As in Experiments1 and 2, there was a correlation with degree of that prefer-ence, no correlation with strength of verb bias, and no ev-idence of a residual complexity effect in the regression equa-tion. Thus, it appears that all minor elevations found in S-biassentences can be accounted for as a f/iaf-preference effect.

Table 5Correlations of Total Reading Time DifferencesWith That Preference for Experiment 3Independent

measure Slope Intercept Pearson r F(l, 8) p <

Complementizer effect for sentence-bias sentencesat noun phrase region

% that 1.83 -80.0 .604 4.58 .07Trans(%) 20.4 +12.6 .654 5.97 .05

Complementizerat

% that 1.31Trans(%) 14.2

effect for sentence-bias sentencesverb phrase

-7.9+58.3

region.667.703

6.407.81

.05

.05Note. Trans(%) = transformed percentages. Intercepts for trans-formed data do not correspond to 0% that preference.

Initial Processing

The total reading times provide clear evidence that readersused verb subcategory information to mediate processing dif-ficulties associated with sentence complements without acomplementizer. However, total reading times do not dif-ferentiate between initial processing and secondary process-ing (rereads). Thus, total times do not provide a clear measureof when verb information influences parsing commitments.For this reason, we report several analyses of the eye-movement data that attempt to tap more immediate com-prehension processes, including first-pass reading times foreach region, the landing site positions for fixations that fol-lowed the initial reading of the main verb, and first fixationdurations and probabilities for each word in the noun phraseand the disambiguating region.

First-Pass Reading Times

When the reader entered a scoring region, fixation dura-tions were considered to be part of a first-pass reading if (a)the subject had not read the region before and (b) the subjecthad not already read any of the words beyond that region. Afirst-pass reading time was obtained by summing the dura-tions of all left-to-right fixations in a region plus any re-gressions made to other points within that region. When thereader made an eye movement out of a region (either a re-gressive eye movement to a prior region or a forward move-ment to a following region), first-pass reading was consid-ered complete for that region (Rayner et al., 1989). Table 6presents first-pass reading times for the matrix verb, the sen-tence complement subject NP, and the following three words.As shown in the table, there were little or no differences ateither the matrix verb or at the subject of the sentence com-plement. However, in the final disambiguating region, theNP-bias verb condition showed a considerable increase inreading times for sentences without a complementizer ascompared with sentences with a complementizer. The S-biassentences showed only a small increase for sentences withoutcomplementizers. Subject and item means were entered intoseparate ANOVAs with the same design as the total readingtime analyses.

Overall effects. Collapsing across all four scoring re-gions, the ANOVA revealed no effect of verb type (F, andF2 < 1) and a marginal effect of complementizer presence,F,(l, 20) = 3.94,p < .10, MSe = 10,192; F2(l, 16) = 3.59,p < .10, MSe = 5,808. The interaction between these twofactors was not significant, F,(l, 20) = 2.11, MSe = 7,493;F2(l, 16) = 1.26, MSe = 8,372. There was an overall effectacross scoring regions, F,(2, 36) = 120.98, p < .01, MSe =17,881; F2(2,25) = 66.67,p < .01, MSe = 27,335. The effectof verb type interacted with region, Fi(2, 40) = 7.05, p <.01, MSe = 5,569; F2(2, 32) = 3.53, p < .05, MSe = 8,966,as did the effect of complementizer presence, F,(2, 40) =4.33, p < .05, MSe = 8,563; F2(2, 32) = 5.72, p < .01, MSe

= 5,437. There was no significant interaction among verbtype, complementizer presence, and region, F](2, 32) =1.03, MSS = 7,495; F2(2, 32) = 1.70, MSe = 4,358.


Table 6Mean First-Pass Reading Times (in ms) for EachScoring Region for Experiment 3

ComplementizerScoring region

"forgot/hoped"

"the solution"

"was in the"

Verb type

S biasNP bias

AM

S biasNP bias

AM

S biasNP bias

AM

Absent

347349_2

348

446424+22"435

630719-89675

Present

355342+ 13349

445420-25433

596615-19606

A

-8+7

——

+ la

+4——

+34a

+ 104——

M

351346

—

446422

—

613667

Note. S = sentence; NP = noun phrase.a Difference correlated with lexically specific that preferences.

Analysis by region. There were no significant effects orinteractions at the matrix verb region (all Fs < 1). There werealso no significant effects or interactions at the followingnoun phrase: For effect of verb type, Fi(l, 20) = 2.68, MSe

= 4,916, F2(l, 16) = 1.49, MSe = 4,910; for effect of com-plementizer, Fs < 1; for interaction, Fs < 1. In the finalscoring region, there was a significant effect of verb type,F, (1,20) = 6.89, p< . 05, MSe = 10,086; F2(l, 16) = 4.63,p < .05, MSe = 12,824; and a significant effect of comple-mentizer presence, F,(l, 20) = 6.76,p < .05, MSe = 16,871;F2(l, 16) = 8.51, p < .01, MSe = 9,549. The interactionbetween verb type and complementizer presence did notreach significance, F,(l, 20) = 2.06, MSe = 14,494;F2(l, 16) = 2.61, MSe = 8,164. However, simple effectsrevealed that for sentences with NP-bias verbs, there was asignificant effect of complementizer, F ^ 1, 20) = 6.84, p <.05, M5e= 19,040; F2(l, 16) = 8.15, p<.05,MSe = 11,385,whereas sentences with S-bias verbs showed no effect ofcomplementizer (Fs < 1). Also, first-pass reading timesshowed no reliable effects of verb type in thecomplementizer-present stimuli (Fs < 1).

Correlations with that preference. First-pass readingtimes for S-bias verbs did not show significant elevations atthe noun phrase or the verb phrase in the sentences withouta complementizer as compared with the sentences with acomplementizer. Nonetheless, we again examined the cor-relations between size of the complementizer effect and thestrength of that preference for the S-bias verbs (see Table 7).The difference between the S-bias and NP-bias sentenceswithout a complementizer (i.e., the verb-type effect) was nottested for the verb phrase region because NP-bias sentencesare elevated due to a syntactic misanalysis. As in Experi-ments 1 and 2, there was a correlation with the degree of thatpreference, no correlation with strength of verb bias, and noevidence of a residual complementizer effect in the regres-sion equation. Thus, the small differences for S-bias verbswere again due to a complementizer preference effect.

Summary offirst-pass reading times. The first-pass read-ing times indicated that early processing was influenced byverb information. At the disambiguating verb phrase regionof the sentence complement, readers showed a reliable in-crease in reading times in NP-bias sentences when the com-plementizer was not present, suggesting that they initiallyinterpreted the earlier noun phrase as a direct object. ForS-bias sentences, however, subjects had little difficulty read-ing the final region when the complementizer was notpresent. Although the interaction between verb type andcomplementizer presence did not reach significance at theverb phrase, this lack of a reliable interaction was most likelydue to the small complementizer effect for S-bias verbs.However, this effect correlated with the degree of that pref-erence, and the relation revealed no residual complementizereffect at the zero intercept.

The first-pass reading times at the noun phrase, however,were missing a crucial result found in the self-paced readingstudy. The self-paced reading revealed an early complemen-tizer effect at the noun phrase for the S-bias sentences, whichwas not detected by the first-pass analysis. However, thefirst-pass reading time difference between S-bias sentenceswith and without a complementizer again correlated with thatpreference. The presence of this correlation strongly suggeststhat the no-difference finding in the average reading times isnot simply the result of variability from an insensitive mea-sure. Rather, it appears either that the first-pass reading timesfor the S-bias sentences without a complementizer were sys-tematically decreased or that the first-pass reading times forthe baseline sentences (sentences with a complementizer)were systematically elevated. In fact, it appears that the latterproposal is correct. Additional analyses, described later,clearly reveal that the lack of effects at the noun phrase aredue to preview effects common to reading short functionwords (e.g., that, the, was). In particular, we demonstrate thatreading times to the postverbal noun phrase were artificiallyelevated for all sentences without a complementizer because

Table 7Correlations of First-Pass Reading Time DifferencesWith That Preference for Experiment 3

Independentmeasure Slope Intercept Pearson r F(\, 8) p <

Complementizer effect for sentence-bias sentencesat noun phrase region

% that 1.07 -67.4 .402 1.55 —Trans(%) 17.1 -17.07 .628 5.21 .05

Verb-type effect for sentences without a complementizerat noun phrase region

% that 1.16 -43.1 .529 3.12 —Trans(%) 14.3 +14.4 .638 5.49 .05

Complementizer effect for sentence-bias sentencesat verb phrase region

% that 1.83 -79.9 .604 4.59 .07Trans(%) 20.4 +12.6 .654 5.97 .05

Note. Trans(%) = transformed percentages. Intercepts for trans-formed data do not correspond to 0% that preference.


of basic changes in first-pass eye-movement patterns result-ing from skipping the complementizer that.

Landing Site Positions of Initial FixationsAfter the Verb

There are several differences between natural reading (ineyetracking) and self-paced reading (Burgess, 1991; Ferreira& Clifton, 1986; Ferreira & Henderson, 1990; Just, Carpen-ter, & Woolley, 1982). Of primary interest is that readers ineyetracking studies often skip short function words (e.g., the,that by, was; Just & Carpenter, 1980; Rayner & McConkie,1976; Rayner & Pollatsek, 1987). Presumably, the reader hasrecognized or can predict the function word while reading theprevious word and does not need to centrally foveate it.When two function words appear in a row, it is likely that thesecond of the two will have a higher than normal probabilityof being foveated (i.e., the first function word is skipped,whereas the second is fixated).

In our baseline conditions, the insertion of a that resultedin the presence of two short function words in a row (". . .that the . . ."). Indeed, an analysis of the fixation data afterthe matrix verb revealed that the inclusion of a complemen-tizer had effects on the probability with which the was fix-ated. Without the complementizer (e.g., "the student hopedthe solution . . . " ) , the determiner was typically skipped.When the complementizer was present (e.g., "the studenthoped that the solution . . . " ) , that was often skipped and thethe was frequently fixated. This is of critical importance be-cause it means that we were comparing first-pass readingtimes to noun phrase regions (e.g., "the solution") that haddifferent fixation patterns.

To illustrate this, we collected the landing site position ofthe two forward fixations after any first-pass reading of thematrix verb. These fixations were typically the only fixationscontributing to first-pass reading times at the NP (if the fix-ation was in the region). Figure 3 shows histograms of thelanding site positions of these forward fixations after thematrix verb relative to the right-hand boundary of the verb(i.e., the last letter of the verb). The top panel of Figure 3shows these data when the complementizer was absent, andthe bottom panel shows these data when the complementizerwas present. (An analysis that subdivided the data on thebasis of verb type showed no differences between S-bias andNP-bias verbs.) As one can see, there was little difference inthe landing site positions for either condition. However, thevertical lines define the average word boundaries (a noun wason average 7.8 characters long). The insertion of a that in 3boffsets the character positions of each scoring region by ex-actly 5 characters (space + that). Thus, subjects typicallyfixated on the head noun of the noun phrase when the com-plementizer was absent but fixated on the determiner whenthe complementizer was present.

Two ANOVAs were conducted on these data. One analysiswas done on the character position landing site data as shownin the figure. As expected, there was no effect of comple-mentizer presence or verb type for either fixation (Fs < 1).An identical analysis was done using character position withrespect to the leftmost boundary of the NP region. These data

40

in

O 30

20-1

Xi

I 10z

0'

40

"the" head nounwan length » 7.8char

verb phrase

1st fixation2nd fixation

llllll..!

§30••2s

20

.210-

"that"

ll

"the" head nounnean length = 7.8char

verb phrase

1st fixation2nd fixation

5 10 15 20# of characters to the right ofmain verb (S-bias or NP-bias)

25

Figure 3. Number of first two forward eye-movement landingsites after first-pass reading of matrix verb as a function of char-acter position. (The top panel shows sentences without the com-plementizer that; the bottom panel shows sentences with the com-plementizer that; char = character; S = sentence; NP = nounphrase.)

were the same, except that when the complementizer waspresent, five character positionc were deducted from thelanding site. As one might expect, ihcie was a significanteffect of complementizer presence for both fixations: firstfixation, F,(l , 20) = 494.56, p < .01, MSe = 1.203;F2(l, 16) = 399.43,/? < .01, MSe = 1.272; second fixation,F,(l , 20) = 96.54, p < .01, MSe = 8.316; F2(l, 16) =174.41, p < .01, MSe = 3.977; there was no effect of verbtype (Fs < 1). By the third forward fixation (not shown here),average landing sites began to readjust so that there was littledifference in the position of the third fixation for both sen-tence types.

This analysis demonstrates that although fixation characterpositions did not differ between the test items and their base-lines, they did differ with respect to the critical scoring regionboundaries. Thus, the probability of initially landing on oneword or another within the two-word scoring region of thenoun phrase was influenced by whether the complementizerwas present.This pattern of forward eye movements makesit extremely likely that first-pass reading times in this regionwere artificially elevated for sentences with a complemen-tizer. The initial fixation in these conditions tended to be on


the less informative determiner, forcing either a longer thannormal reading time to recognize the head noun parafoveallyor a second fixation in the region to fully foveate the noun.Thus, these changes in fixation probabilities elevated first-pass reading times on the entire noun phrase when the com-plementizer was present, effectively masking any elevationsin processing load due to the absence of the complementizer.Despite these elevations in the baseline, correlations on thisnonsignificant complementizer effect at the noun phrase stillrevealed a that preference effect for S-bias verbs.

First Fixation Durations and ComparisonsWith Ferreira and Henderson (1990)

Ferreira and Henderson (1990) did not report first-passanalyses. Instead, they reported the duration of the first fix-ation for each word. First fixation duration is the duration ofthe initial fixation of a first-pass reading of a region. Thus,any first-pass refixations in a region do not contribute to thismeasure. In order to permit a direct comparison to Ferreiraand Henderson's results, we will also report first fixationduration data.

For convenience, Ferreira and Henderson's (1990) meanfirst fixation durations for the three words following the ma-trix verb are presented in Table 8. Recall that these are theambiguous noun (1), the disambiguating verb (2), and thenext word (3) in Example 7.

Bill hoped/wrote (that) Jill arrived safely today.1 2 3 (Example 7)

There was a main effect of complementizer presence at thedisambiguating verb (Position 2). Fixation durations at thisposition were longer for sentences without a complementizerthan for sentences with a complementizer, regardless of verbtype. No reliable effects or interactions were found in the firstor third positions.

Table 9 reports mean first fixation durations for the fivewords following the matrix verb in our study. These words

Table 8Mean First Fixation Duration (in ms)From Ferreira and Henderson (1990) for the Noun, theDisambiguating Sentence Complement Verb,and the Next Word

Table 9Mean First Fixation Duration (in ms)for the Determiner, the Head Noun,the Disambiguating Sentence Complement Verb,and the Next Two Words for Experiment 3

Word position

Noun

DisambiguatingSC verb

Next word

Verb type

S biasNP bias

M

S biasNP bias

M

S biasNP bias

M

Complementizer

Absent

214217216

230215223

321308314

Present

208200204

200201201

288302295

A

+6+ 17

+30+ 14

+33+6

M

211209

215208

305305

Word position

Determiner

Head noun


Next word

Next word

Verb type

S biasNP bias

M

S biasNP bias

M

S biasNP bias

M

S biasNP bias

M

S biasNP bias

M

Complementizer

Absent

240240240

279286283

263281272

272287278

264287276

Present

237244241

251259255

253246250

264267266

260272266

A

+3-4

+28+27

+ 10+35

+6+20

+4+ 15

M

239242

265273

258264

268277

262280

Note. S = sentence; NP = noun phrase; SC = sentence com-plement.


were the determiner (1), the head noun (2), the disambigu-ating word (3), and the next two words (4 and 5).

The student hoped/forgot (that) the solution was in theback of the book. 1 2 3 4 5

(Example 8)The means showed an effect of complementizer presence atthe head noun (Word Position 2). Sentences without a com-plementizer were slower than sentences with a complemen-tizer. Recall that this effect was not found in the first-passanalyses of the entire noun phrase region. However, the fix-ations durations at the verb phrase region (Word Positions 3,4, and 5) revealed a pattern very consistent with the first-passanalysis. Fixation durations for NP-bias sentences without acomplementizer were longer than those for the other threesentence types. ANO VAs at each word position revealed onlya reliable effect of complementizer presence at the headnoun, F,(l, 20) = 12.77,/? < .01, MSe = 1,382; F2(l, 16)= 24.39, p < .01, MSe = 553. No other reliable effectsor interactions were found. However, at the third position(the disambiguating word), there was a marginal effect ofcomplementizer, F,(l, 20) = 4.27, p < .10, MSe = 2,949;F2(l, 16) = 3.92, p < .10, MSe = 2,387. This effect wascarried by an effect in the NP-bias verbs, F,(l, 20) = 3.92,p < .10, M5e = 3,907; F2(l, 16) = 4.12, p < .10, MSe =2,182; but not the S-bias verbs (Fs < 1).

Statistically, then, the first fixation pattern in our study wasidentical to that of Ferreira and Henderson (1990). In bothstudies, the second word after the matrix verb showed a maineffect of complementizer. Fixation durations at this positionwere longer for sentences without a complementizer than forsentences with a complementizer, regardless of verb type.


This similarity suggests that the Ferreira and Henderson(1990) results and the present results are not as dramaticallydifferent as one might expect. However, because our maineffect of complementizer occurred prior to the disambigu-ating word (on the head noun) rather than on the disambig-uating verb, the results suggest that the elevations in firstfixation durations for sentences without a complementizer inboth studies were not due to syntactic misanalysis.

Three questions are raised by the first fixation results in ourexperiment. First, why were first fixation durations in theverb phrase region too noisy to find any reliable differences?Second, why did first fixation durations differ for the headnoun in the complementizer-present and -absent conditions?Third, why did first-pass reading times at the noun phrase notshow a similar elevation? An analysis of the probability offixating on each word provided a plausible answer to eachof these questions.

As mentioned earlier, it is well established in the literaturethat although the majority of words in a sentence are fixated,many words are skipped (Just & Carpenter, 1980; Rayner &McConkie, 1976; see also Rayner & Pollatsek, 1987).Shorter words are less likely to be fixated (Rayner & Mc-Conkie, 1976). First fixation durations for a word are basedon only those trials in which the subject fixated on the wordwithout having already fixated on any other words furtheralong in the sentence. Table 10 presents the probability of afirst fixation for each of our critical word's positions.

Clearly, certain words tended to be skipped, especially thedeterminer and the disambiguating word, which was typi-cally a be verb or modal. The probability of fixating on thesewords is consistent with fixation probability data reported by

Table 10The Probability of a First-Pass Readingfor the Determiner, the Head Noun,the Disambiguating Sentence Complement Verb,and the Next Two Words for Experiment 3

Word position

Determiner

Head noun


Next word

Next word

Verb type

S biasNP bias

M

S biasNP bias

M

S biasNP bias

M

S biasNP bias

M

S biasNP bias

M

Complementizer

Absent

0.380.450.42

0.970.910.94

0.460.540.50

0.770.780.78

0.690.700.70

Present

0.600.640.62

0.960.850.91

0.370.530.45

0.750.790.77

0.740.690.72

A

-0.22-0.19

+0.01+0.06

+0.09+0.01

+0.02-0.01

-0.05+0.01

M

0.490.55

0.970.88

0.420.54

0.760.79

0.720.70

Rayner and McConkie (1976) and Just and Carpenter (1980).Thus, it should be clear why the only reliable effect for firstfixation durations occurred at the head noun of the nounphrase. It was the only position that subjects consistentlyfixated. The probabilities at all other word positions were lowenough that subject and item cell means represented rela-tively few observations, and often represented different setsof subjects or different sets of items. Thus, the means will berelatively noisy at these positions. In fact, the probability ofa first-pass fixation on a word was sometimes low enoughthat subject and item cell means had no data. In other words,either a subject consistently skipped a word in a given subjectcondition, or all subjects consistently skipped a word in agiven item condition. Thus, in order to perform the prior firstfixation duration ANOVAs, certain item cell means and cer-tain subject cell means had to be replaced by an estimatedvalue.5 For subject means, the following percentage of cellshad to be estimated: 10% at the determiner, 1% for the headnoun, 8.3% at the disambiguating modal. For the item meansthe percentages were 6% at the disambiguating modal and5% at the final word. Missing data are less of a problem withlarger scoring regions. For example, the probability of a first-pass reading on each trial for the scoring regions in the anal-yses we reported earlier was close to 1.0.

Now consider the second question, namely, why was therean effect of the complementizer on the initial fixations to thehead noun? As would be expected by the landing site anal-yses reported earlier, the probability of a reader fixating onthe determiner (shown in Table 10) was considerably higherwhen the complementizer was present compared with whenit was absent, F,( 1,20) = 6.47,p< .05, MSe = 0.1568; F2(\,16) = 21.98,/? < .01, MSe = 0.0389. When that was present,subjects tended to skip that and land on the and then moveon to the head noun. When that was absent, subjects oftenskipped the and landed directly on the head noun. As a con-sequence, subjects were much more likely to fixate on theword preceding the noun when there was a complementizerpresent, permitting them to have a clear preview advantagewhen they moved to the noun. Inhoff and Rayner (1986),Lima (1987), and Pollatsek, Rayner, and Bolota (1986) haveall demonstrated that the first fixation on a word is shorterwhen visual information about the word is availableparafoveally on the prior fixation than when the visual in-formation is not available. (See Rayner & Pollatsek, 1987,for a complete review.) Thus, the complementizer effect atthe head noun is likely to be due to a preview effect.

If the complementizer effect at the noun was in fact dueto a preview effect (or some other low-level source), then wewould not expect it to correlate with any of the variablesmeasured in the completion norms. In fact, this was the case.The complementizer effect at the noun in both the S-bias and


5 Missing data were estimated using the stepwise regressionestimation routine found in the BMDP statistical package (subrou-tine pam). Estimations of missing subject and item cell means werebased on correlations with all other means, including each sub-ject's or item's overall mean, and the overall mean within a givenposition. It should be noted that subject and item means did notchange much when missing data were replaced.


NP-bias sentences did not correlate with the percentage ofNP, S, or that completions or any differences or transfor-mations. Thus, it appears that these effects do not arise fromparsing considerations.

Although Ferreira and Henderson (1990) did not reportfirst fixation probabilities, it is likely that the complementizereffect at the verb in their experiment was also due to landingsite differences.6 When the complementizer was absent, sub-jects may have tended to skip the noun and land directly onthe disambiguating verb, whereas when the complementizerwas present, subjects may have been more prone to skip thethat, land on the noun and then the verb, providing a previewadvantage for the verb. Ferreira and Henderson's nouns wererelatively short (averaging 4.7 characters). Although we areunaware of a study reporting fixation probabilities for dif-ferent lengths of content words alone, Rayner and McConkie(1976) found that, in general, words under 6 characters inlength have less than a 0.6 probability of fixation. So, if theprobability of fixating on the noun differed in the comple-mentizer-absent and -present conditions in the Ferreira andHenderson study, they are likely to have found the samepattern of differences at the verb that we found at the noun.

Finally, the third question concerning our data was why nocomplementizer effect was seen for the first-pass analyses ofthe entire noun phrase region. Although first fixations werefaster on the head noun when the complementizer was absentcompared with when it was present, the fixation probabilitiesrevealed that these faster fixations were also accompanied bya greater probability of having a fixation on the prior de-terminer. Thus, overall, these shorter fixations were accom-panied by more fixations on the determiner, making first-passreading times in these conditions slightly longer, and effec-tively masking the first fixation difference. To demonstratethis informally, we multiplied the probabilities of a fixationat the determiner and the noun with their respective firstfixation durations, and then added these products together foreach condition. The results revealed no differences betweenthe complementizer-present and complementizer-absentconditions (NP bias without complementizer: 376 ms; NPbias with complementizer: 376 ms; S bias without comple-mentizer: 362 ms; S bias with complementizer: 383 ms).

For the most part, the analysis of the first fixation datarevealed patterns that were consistent with both the first-passdata and the landing site data. The only discrepancy, a com-plementizer effect at the head noun, was most likely due toeye-movement landing site differences. A similar effect islikely to have occurred in the Ferreira and Henderson (1990)study, resulting in first fixation elevations that appeared to bedue to a syntactic misanalysis but were actually due tochanges in the probability of the fixating on different words.

Reprocessing (Rereads)

Regressive eye-movement pattern. Eye-movement pat-terns from our study also indicated that sentences withouta complementizer had large reanalysis effects (rereads)when the verb was NP-bias versus when it was S-bias. ForNP-bias sentences without a complementizer, 38% of first-pass readings of the final word of the disambiguating verb

Table 11Mean Second-Pass Reading Times (in ms)for Each Scoring Region for Experiment 3

Scoring region

"The student"

"forgot/hoped"

"the solution"

"was in the"

Verb type

S biasNP bias

M

S biasNP bias

M

S biasNP bias

M

S biasNP bias

M

Complementizer

Absent

78149114

93205149

11!254183

131354243

Present

877079

789185

78136107

91178135

A

-9+79

+ 15+ 114

+33+ 118

+40+ 176

M

83110

83148

95195

111266

Note. S = sentence; NP = noun phrase.

phrase region ended with a regressive eye movement ascompared with 19% when the sentence contained a com-plementizer. For S-bias sentences without a complemen-tizer, 16% of the first-pass readings ended with a regres-sive eye movement as compared with 15% when thesentence contained a complementizer. Such a pattern sug-gests that a garden-path occurred only in the NP-bias sen-tences without a complementizer.

Second-pass reading times. Table 11 presents second-pass reading times in milliseconds for all four regions.Second-pass reading times reflect any rereads of these re-gions. As seen in the table, NP-bias sentences without a com-plementizer showed large elevations in second-pass readingtimes with respect to the other three conditions. Subject anditem means were entered into separate ANOVAs with fourfactors: list (four lists) or item group (four groups); region(subject NP, matrix verb, sentence complement subject NP,and sentence complement verb [three words]); verb type (Sbias and NPbias); and complementizer (present and absent).We will report overall effects and interactions and then ef-fects at each scoring region.

When collapsing across all four scoring regions, theANOVA revealed a significant interaction between verb typeand complementizer presence, F}(1, 20) = 7.96, p < .05,MSC = 31,255; F2(l, 16) = 7.58, p < .05, MSe = 33,469.

6 It should be noted that for unrelated reasons Ferreira andHenderson (1990) embedded a contingent display manipulationinto their eye-movement study (see Henderson & Ferreira, 1990).On two thirds of the experimental trials, a nonword was presentedparafoveally (the next word over) when subjects fixated on thedisambiguating verb. This nonword rapidly changed into the ap-propriate word while subjects moved their eyes to foveate thenonword position. Henderson and Ferreira (1990) noted that sub-jects occasionally volunteered that they had detected a displaychange. It is conceivable that the detection of display changesmight have influenced overall fixation patterns.


In addition, there were significant effects of verb type,F,(l, 20) = 19.83,p < .01,MSe = 35,999; F2(l, 16) = 8.79,p < .01, MSe = 62,677; and complementizer presence,F,(l,20) = 20.46,/?< .01,MSe = 23,477;F2(1,16) = 24.18,p < .01, MSe = 18,425. Simple effects revealed a significanteffect of complementizer presence for the NP-bias sentences,F,(l,20) = 18.55,p< .01,MSe = 38,281;F2(1,16) = 20.06,p < .01, MSe = 34,190; but not the S-bias sentences,F,(l, 20) = 1.15, MSe = 16,451; F2(\, 16) = 0.76, MSe =17,704. In addition, there was a significant effect of verb typefor sentences that did not contain a complementizer, F, (1,20)= 15.75, p < .01, MSe = 57,321; F2(l, 16) = 11.01,p < .01, MSe = 70,539. Sentences containing a comple-mentizer showed an effect of verb type in the subject analysesbut not in the item analyses, F](l, 20) = 6.03,p < .01, MSe

= 9,933; F2(l, 16) = 1.11, MSe = 25,608. There was asignificant effect of scoring region, F,(3, 36) = 11.06, p <.01, MSe = 11,811; F2(3, 54) = 14.46, p < .01, MSe =10,660. In addition, the effect of verb type interacted withscoring region, F,(3, 43) = 8.24, p < .01, MSe = 7,748;F2(3, 57) = 11.91, p < .01, MSe = 6,028; and the effect ofstructure interacted with scoring region, F,(3, 59) = 3.47, p< .05, MSe = 6,318; F2(3,48) = 3.25,/? < .01,MSe = 5,809.

Analysis by region. Separate analyses were conductedfor each scoring region. To conserve space, the same basiceffects obtained at each region: an interaction between verbtype and complementizer presence, overall effects of bothverb type and complementizer, an effect of complementizerpresence for sentences with NP-bias verbs but not with S-biasverbs, and an effect of verb type for sentences without acomplementizer. The only effect that varied across positionwas an effect of verb type for sentences containing comple-mentizers. At the first two regions, there was no effect (Fs< 1). However, at the final two regions there was an effectbut only in the subject analyses: At the NP, F t( l , 20) = 8.41,p < .01, MSe = 4,779; F2(l, 16) = 2.25, MSe = 10,330; atthe second verb, F,(l, 20) = 9.81, p < .01, MSe = 9,220;F2(l, 16) = 3.46, p < .10, MSe = 21,479.

Discussion

The results demonstrated clear and immediate effects ofverb subcategorization. Consider first the results for the sen-tences with NP-bias verbs. These sentences showed largereanalysis effects when the complementizer was absent.First-pass reading times at the disambiguating verb phrasewere longer for sentences without a complementizer than forsentences with a complementizer. Near the end of the verbphrase, readers were more likely to make a regressive eyemovement for the sentences without complementizers. Inaddition, second-pass and total reading times for sentenceswithout a complementizer were elevated in comparison withall other sentence types. All of these results indicate thatsubjects initially incorporated the noun phrase after an NP-bias verb as the direct object of the verb. The subsequent verbphrase then forced a reanalysis.

The S-bias sentences, on the other hand, showed no evi-dence of a misanalysis effect. First-pass reading times at thenoun phrase and the following verb phrase regions were not

reliably longer for sentences without a complementizer ascompared with those with a complementizer. Regressive eyemovements at the verb phrase were equally likely for sen-tences with and without a complementizer. Moreover,second-pass reading times were not reliably different. Theonly reliable complementizer effect occurred for total read-ing times at the verb phrase. This effect and the nonsignif-icant effects in first-pass reading times at the noun phrase andthe verb phrase all correlated with the that preference forindividual verbs. As in Experiment 2, the regression equa-tions did not reveal the residual reanalysis effect that wouldbe predicted by lexical filtering models.

Along with the evidence concerning the use of verb in-formation, the results of the present study also have impor-tant methodological implications for eye-movement studies.An analysis of landing site probabilities revealed that theinsertion of a short function word (that) altered the proba-bilities of landing on particular words within the followingscoring region. This result is likely to have direct bearing onother eye-movement studies of parsing. Many parsing stud-ies compare first-pass reading times across scoring regionswith the same linguistic content but with different characterpositions. Thus, as in our study, scoring regions are typicallyoffset because of the insertion of one or two syntacticallydisambiguating function words prior to the critical scoringregions. Under these conditions, comparisons of ambiguousand unambiguous structures can both mask real effects andintroduce spurious differences. In the present study, we haveargued that preview advantages arising from landing site dif-ferences on the determiner decreased first fixation durationsat the head noun when it followed a complementizer. More-over, the same landing site differences resulted in increasedfirst-pass reading times to the entire noun phrase when thecomplementizer was present, because fixations on the headnoun were typically accompanied by a prior fixation on thedeterminer. Both of these fixations would have contributedto the first-pass analysis of the noun phrase.

General Discussion

The research reported here provided clear answers to thethree questions that we outlined earlier. First, subcategori-zation information becomes available almost immediatelyafter a verb is recognized. Experiment 1 demonstrated thatsubcategorization information was available rapidly enoughto influence integration effects in cross-modal naming whenthe visual target immediately followed the verb.

Second, there is processing difficulty associated with that-less sentence complements that is not due to syntactic mis-analysis. Pronouns that unambiguously determine a sentencecomplement reading (e.g., he) were more difficult to namewhen preceded by an S-bias sentence fragment without acomplementizer as compared with an S-bias sentence frag-ment with a complementizer. This effect cannot be attributedto a temporary syntactic misanalysis for two reasons: First,the results with him demonstrated that subcategorization in-formation is available, and second, the pronoun he is un-ambiguously a subject pronoun. Moreover, correlations be-tween the completion norms and the complementizer effect


indicated that the effect depended on the degree to which theverb was preferably followed by sentence complement con-taining a complementizer.

Third, the subcategorization properties of the verb imme-diately affect the syntactic analysis of a noun phrase thatwould otherwise be ambiguous. Experiments 2 and 3 dem-onstrated clear syntactic misanalysis effects when a nounphrase that was a plausible object of an NP-bias verb turnedout instead to be the subject of a sentence complement. How-ever, no misanalysis effects obtained when the same nounphrase followed an S-bias verb. Instead, reading times to thenoun phrase were elevated in comparison with sentences inwhich the noun phrase was preceded by a complementizer.This complementizer effect was again correlated with thedegree to which the verb prefers to be used with a comple-mentizer when it occurs in a sentence complement construc-tion. Finally, the eye-tracking data also showed small ele-vations in reading times at the verb phrase for S-bias verbs.These effects, like those at the noun phrase, were correlatedwith complementizer preference.

This set of results is clearly problematic for lexical filteringmodels. In the model proposed by Ferreira and Henderson(1991), lexically specific information is used to guide syn-tactic misanalysis only after the parser encounters syntacticinformation that is inconsistent with the initial parse. Ourresults are clearly inconsistent with this prediction. Subcat-egorization information is accessed as soon as a verb is en-countered, and it has immediate effects on the processing ofinformation that follows the verb. Moreover, no syntacticmisanalysis occurs when readers encounter information thatis consistent with the subcategorization properties of theverb, but would have been inconsistent with an anal-ysis based on major category information and attachmentstrategies.

Models in which lexical filtering occurs more rapidly alsorun into serious problems (Frazier 1987, 1989; Mitchell,1989). One might attempt to explain our data by proposingthat a noun phrase following a sentence complement verb isinitially parsed as the object of the verb, with reanalysis tak-ing place within the 400- to 600-tns window that it takesreaders to process a two-word noun phrase. This would pre-dict elevated reading times to noun phrases that follow verbsthat are strongly biased against an NP complement but nosubsequent misanalysis effect. Although this story superfi-cially describes the data, it runs into serious empirical andtheoretical problems.

The first empirical problem comes from our finding thatthe processing of the subject of a sentence complement (inthe absence of a complementizer) was more difficult than theprocessing of a noun phrase complement, even when thepossibility of syntactic misanalysis did not exist (i.e., whenthe complement was a case-marked pronoun). The secondproblem comes from the, finding that, for S-bias verbs, themagnitude of the complementizer effect (the difference inprocessing times to noun phrases when there was not a com-plementizer compared with when there was a complemen-tizer) correlated with the degree to which the verb preferredto be followed by a complementizer but not with the strengthof the subcategorization bias. Moreover, the regression equa-

tions with both unambiguous noun phrases (Experiment 1)and potentially ambiguous noun phrases (Experiments 2 and3) showed no hint of the residual complementizer effect thatis required by the lexical filtering hypothesis. Given thesefacts, a rapid revision hypothesis is untenable.

The theoretical difficulties for lexical filtering models areequally serious. There are two reasons why a language-processing system might initially ignore lexically specificsyntactic constraints. The first reason is that lexically specificinformation might be accessed too slowly to be reliably usedby the system when it is making initial commitments. How-ever, this is clearly not the case given the data that we andothers have presented. The second reason is that ignoringlexically specific information might increase the speed andefficiency of the system by allowing the system to builda structure more rapidly, thereby reducing memory load(Frazier, 1989). However, this argument is valid only for aserial parser in which complete commitments either arebased on major category information (and thus can be madeimmediately) or must await definitive evidence about sub-categorization information (which can be delayed for severalwords or even phrases). Recent results suggest that thisstrong serial assumption is unlikely to be correct (Altmann& Steedman, 1988; Gorrell, 1989, 1991; MacDonald, Just,& Carpenter, 1992; Ni &Crain, 1990;Trueswelletal., 1992).Moreover, it is hard to defend the proposal that building andimmediately revising a structure places fewer demands onworking memory than building a structure that is consistentwith the relevant linguistic information, especially when thatinformation has just been accessed.

The results presented here complement recent studies find-ing that other aspects of verb-based information are also im-mediately accessed and used in sentence processing (for arecent review see Boland & Tanenhaus, 1991). These studiesconverge on a view in which recognition of a verb makesavailable combinatory syntactic and semantic informationthat allows the processing system to make partial commit-ments while taking into account relevant semantic, syntactic,and discourse-based information (Carlson & Tanenhaus,1988; MacDonald, 1992; Marslen-Wilson, Brown, & Tyler,1988; Tanenhaus, Boland, Mauner, & Carlson, in press;Tanenhaus & Carlson, 1989; Tanenhaus, Carlson, &Trueswell, 1989; Tanenhaus, Garnsey, & Boland, 1991;Tyler, 1989). Empirical results supporting this view includeevidence that the semantic fit of potential arguments is usedin syntactic ambiguity resolution with relative clauses (Bur-gess, 1991; MacDonald, 1992; Pearlmutter & MacDonald,1992; Trueswell et al., 1992); evidence that verb-based con-trol information is used immediately (Boland, Tanenhaus, &Garnsey, 1990); and evidence that the semantic fit betweena verb and potential arguments is used in filler-gap assign-ment (Tanenhaus, Boland, Garnsey, & Carlson, 1989). Re-sults such as these all depend to some degree on the avail-ability of verb-specific syntactic constraints.

The fact that verb subcategorization information is im-mediately accessed and used does not, however, mean thatall syntactic commitments are determined by lexically spe-cific information as opposed to information based on syn-tactic categories. For example, in constraint-based ap-


preaches to parsing, the relative strength of category effectsand subcategory effects in different environments will de-pend on which type of information is more reliable. Thus, wewould expect to find certain environments in which verb-specific constraints have weak or delayed effects and otherenvironments in which they would have immediate effects.This would be consistent with the statistical tuning hypoth-esis advanced by Mitchell and Cuetos (1991). The that-preference effect that we report is further evidence that thelanguage-processing system is subtly tuned to lexically spe-cific co-occurrences. As we mentioned earlier, this type ofeffect is naturally accommodated within constraint-basedframeworks. The presence of such an effect suggests that itwill be fruitful to explore how lexical co-occurrence patternsaffect the parsing of other structures. It will also be importantto explore explanations for why such patterns might occur inthe language, such as the relative frequency of a verb, thesemantic characteristics of the verb, the propositional contentof the complement, and so on. (See Elsness, 1984, for someinteresting observations.) Finally, it will be important to de-termine whether f/zaf-preference effects also occur with lessstrongly based sentence complement verbs.

The current results have several significant methodolog-ical implications for empirical investigations of sentence pro-cessing. It will be important for future eye-tracking studiesto take into account information about landing sites, espe-cially when making comparisons across structures with dif-ferent sequences of words. The fixation pattern for identicalmaterial can be altered dramatically by the insertion of aword or two. This can lead to real differences being masked,but can also introduce spurious differences. (In fact, both ofthese effects were observed in Experiment 3.) Other readingmethodologies, in particular, self-paced reading with single-word presentation, have different albeit related problems,because subjects are forced to fixate each word and do nothave the advantage of parafoveal information.

Our results also highlight the importance of obtaining nor-mative data for the materials used in sentence-processingexperiments. By their very nature, the dimensions measuredby norms are continuous rather than discrete. In several re-cent articles, Pearlmutter and MacDonald (1992; Mac-Donald, 1992) have argued for treating these variables ascontinuous and using tools suited to continuous variables,such as regression analyses, in conjunction with on-line mea-sures. Our results clearly support this theoretical and meth-odological approach.

In addition, our results underscore a simple point about thelinking assumption between processing load and underlyingcomprehension processes that has perhaps been underappre-ciated in recent work in parsing. Local increases in process-ing difficulty can be due to a variety of factors, only one ofwhich is syntactic misanalysis. Taking a garden-path expla-nation as the default will lead one to underestimate the in-formation that the system makes use of during sentence pro-cessing and overestimate the amount of syntactic misanalysisthat actually occurs.

Finally, the research presented here underscores the valueof using multiple methodologies. While each of the meth-odologies that we used has potential problems of interpre-

tation associated with it, the results from each task helpedshed light on the interpretation of the results obtained withthe other tasks. Moreover, the fact that similar patterns ofresults obtained with three different methods makes it un-likely that the results were due to task-specific strategies. Inaddition, because the tasks are not equally well suited for allproblems (e.g., only cross-modal naming can be used tostudy spoken language processing), it is reassuring to knowthat the results from each task lead to a similar conclusion.

In sum, we have provided evidence that verb subcatego-rization information is rapidly accessed and used in syntacticambiguity resolution. We have also demonstrated a verb-specific co-occurrence effect. Both of these effects are prob-lematic for lexical filtering models, and both are naturallymodeled within a constraint-based framework. However, itis important to note that we intentionally selected verbs withextremely strong biases in order to clearly establish whetherverb information would be used by the processing system.We also chose nouns that were highly plausible objects forthe NP-bias verbs. In addition, for most of these verbs andobjects, there was a clear situational shift between the in-terpretation where the noun phrase was the object (e.g., "ac-cepting an award") and the interpretation where the nounphrase was the subject of a sentence complement ("acceptingthat the award"). Thus, the results obtained here, namelystrong initial commitments consistent with the subcategori-zation properties of the verb, and clear misanalysis effectswhen the preferred structure was disconfirmed, could bemodeled by any system that is immediately sensitive to lex-ically specific information and that makes at least provisionalon-line commitments to consistent interpretations. In futureresearch it will be important to better understand the natureof the commitments that are typically made during sentenceprocessing as well as how different sources of informationare weighted in determining these commitments. Answeringthese questions will require treating dimensions such as sub-categorization preference, semantic fit between an argumentand possible argument positions, and situational differencesamong alternative analyses as continuous variables.

References

Altmann, G. T. M, & Steedman, M. J. (1988). Interaction withcontext during human sentence processing. Cognition, 30, 191-238.

Boland, J. (1991). The use of lexical knowledge in sentence pro-cessing (Doctoral dissertation, University of Rochester, 1991).Dissertation Abstracts International, 52/1 IB, 6111.

Boland, J. (in press). The role of verb argument structure in sentenceprocessing: Distinguishing between syntactic and semantic ef-fects. Journal of Psycholinguistic Research.

Boland, J., & Tanenhaus, M. K. (1991). The role of lexical repre-sentations in sentence processing. In G. B. Simpson (Ed.), Un-derstanding word and sentence (pp. 331-366). Amsterdam:North-Holland.

Boland, J., Tanenhaus, M. K., & Garnsey, S. M. (1990). Evidencefor the immediate use of verb control information in sentenceprocessing. Journal of Memory and Language, 29, 413—432.

Britt, M. A., Perfetti, C. A., Garrod, S., & Rayner, K. (1992). Parsingand discourse: Context effects and their limits. Journal of Mem-ory and Language, 31, 293-314.


Burgess, C. (1991). Interaction of semantic, syntactic and visualfactors in syntactic ambiguity resolution (Doctoral dissertation,University of Rochester, 1991). Dissertation Abstracts Interna-tional, 52/02B, 1098.

Burgess, C , & Tanenhaus, M. K. (1992). The interaction of se-mantic and parafoveal information in syntactic ambiguity reso-lution. Unpublished manuscript.

Carlson, G. N., & Tanenhaus, M. K. (1988). Thematic roles andlanguage comprehension. In W. Wilkens (Ed.), Syntax and se-mantics (Vol. 21, pp. 263-300). San Diego, CA: Academic Press.

Chomsky, N. (1965). Aspects of a theory of syntax. Cambridge, MA:MIT Press.

Clifton, C, Jr., Speer, S., & Abney, S. P. (1991). Parsing arguments:Phrase structure and argument structure as determinants of initialparsing decisions. Journal of Memory and Language. 30, 251-271.

Cowart, W. (1987). Evidence for an anaphoric mechanism withinsyntactic processing: Some reference relations defy semantic andpragmatic considerations. Memory & Cognition, 15, 318-331.

Crain, S., & Steedman, M. J. (1985). On not being led up the gardenpath: The use of context by the psychological parser. In D. Dowty.L. Karttunen, & A. Zwicky (Eds.), Natural language parsing:Psychological, computational, and theoretical perspectives (pp.320-358). Cambridge, England: Cambridge University Press.

Elman, J. L. (1991). Representation and structure in connectionistmodels. In G. T. M. Altmann (Ed.), Cognitive models of speechprocessing: Psycholinguistic and computational perspectives(pp. 345-382). Cambridge, MA: MIT Press.

Elsness, J. (1984). That or zero? A look at the choice of object clauseconnective in a corpus of American English. English Studies, 65,519-533.

Ferreira, F., & Clifton, C. (1986). The independence of syntacticprocessing. Journal of Memory and Language, 25, 348-368.

Ferreira, F., & Henderson, J. M. (1990). The use of verb informationin syntactic parsing: A comparison of evidence from eye move-ments and word-by-word self-paced reading. Journal of Exper-imental Psychology: Learning, Memory, and Cognition, 16, 555—568.

Ferreira, F., & Henderson, J. M. (1991). How is verb informationused during syntactic processing? In G. B. Simpson (Ed.), Un-derstanding word and sentence (pp. 305-330). Amsterdam:North-Holland.

Ford, M., Bresnan, J., & Kaplan, R. (1982). A competence basedtheory of syntactic closure. In J. Bresnan (Ed.), The mental rep-resentation of grammatical relations (pp. 727-796). Cambridge,MA: MIT Press.

Frazier, L. (1985). Syntactic complexity. In D. Dowty, L. Karttunen,& A. Zwicky (Eds.), Natural language parsing (pp. 129-189).Cambridge, England: Cambridge University Press.

Frazier, L. (1987). Sentence processing: A tutorial review. In M.Coltheart (Ed.), Attention and performance XII: The psychologyof reading (pp. 559-586). Hillsdale, NJ: Erlbaum.

Frazier, L. (1989). Against lexical generation of syntax. In W. D.Marslen-Wilson (Ed.), Lexical representation and process (pp.505-528). Cambridge, MA: MIT Press.

Frazier, L., & Clifton, C. (1989). Successive cyclicity in the gram-mar and the parser. Language and Cognitive Processes, 4, 93-126.

Frazier, L., & Fodor, J. D. (1978). The sausage machine: A newtwo-stage parsing model. Cognition, 6, 291-325.

Frazier, L., & Rayner, K. (1982). Making and correcting errorsduring sentence comprehension: Eye movements in the analysisof structurally ambiguous sentences. Cognitive Psychology, 14,178-2)0.

Gorrell, P. (1989). Establishing the loci of serial and parallel effects

in syntactic processing. Journal of'Psxcholinguistic Research, 18,61-73.

Gorrell, P. (1991). Subcategorization and sentence processing. In R.C. Berwick. S. P. Abney, & C. Tenny (Eds.), Principle basedparsing (pp. 279-300). Dordrecht, The Netherlands: Kluwer Ac-ademic Press.

Hakes, S. D. T. (1972). Effects of reducing complement construc-tion on sentence comprehension. Journal of Verbal Learning andVerbal Behavior, 11, 278-286.

Henderson, J. M., & Ferreira, F. (1990). The effects of foveal pro-cessing difficulty on the perceptual span in reading: Implicationsfor attention and eye movement control. Journal of ExperimentalPsychology: Learning, Memory, and Cognition, 16, 417-429.

Hindle, M., & Rooth, M. (1990). Structural ambiguity and lexicalrelations. In Proceedings of the 28th annual meeting of the As-sociation of Computational Linguistics (pp. 229-336). Cam-bridge, MA: MIT Press.

Holmes, V. M., Stowe, L.. & Cupples, L. (1989). Lexical expec-tations in parsing complement-verb sentences. Journal of Mem-ory and Language, 28, 668-689.

Inhoff. A. W., & Rayner. K. (1986). Parafoveal word processingduring eye fixations in reading: Effects of word frequency. Per-ception & Psychophvsics, 40. 431--4-39.

Just, M. A.. & Carpenter. P. A. (1980). A theory of reading: Fromeye fixations to comprehension. Psychological Review, 87. 329-354.

Just. M. A., Carpenter, P. A., & Woolley, J. (1982). Paradigms andprocesses in reading comprehension. Journal of ExperimentalPsychology: General. Ill, 228-238.

Kennedy, A., Murray, W. S.. Jennings, F.. & Reid, C. (1989). Parsingcomplements: Comments on the generality of the principle ofminimal attachment. Language and Cognitive Processing, Pars-ing and Interpretation. 4. 21—49.

Lima, S. D. (1987). Morphological analysis in sentence reading.Journal of Memory and Language, 26. 84-99.

MacDonald, M. (1992). Probabilistic constraints and syntactic am-biguity resolution. Manuscript submitted for publication.

MacDonald, M., Just, M. A.. & Carpenter, P. A. (1992). Workingmemory constraints on the processing of syntactic ambiguity.Cognitive Psychology, 24, 56-98.

Marslen-Wilson, W. D.. Brown. C, & Tyler, L. K. (1988). Lexicalrepresentations in language comprehension. Language and Cog-nitive Processes. J, 1-16.

Marslen-Wilson, W. D., & Tyler, L. K. (1987). Against modularity.In J. Garfield (Ed.), Modularity in knowledge representations andnatural language understanding (pp. 37-62). Cambridge, MA:MIT Press.

McClelland, J. L., St. John. M.. & Taraban. R. (1989). Sentencecomprehension: A parallel distributed processing approach. Lan-guage and Cognitive Processes, 4, 287-336.

Mehler. J., & Carey, P. (1967). Role of surface and base structurein the perception of sentences. Journal of Verbal Learning andVerbal Behavior, 6, 335-338.

Mitchell, D. C. (1987). Lexical guidance in human parsing: Locusand processing characteristics. In M. Coltheart (Ed.), Attentionand performance XII: The psychology of reading (pp. 601-618).Hillsdale, NJ: Erlbaum.

Mitchell, D. C. (1989). Verb-guidance and other lexical effects inparsing. Language and Cognitive Processes, 4, 123-154.

Mitchell, D. C , & Cuetos, F. (1991). The origins of parsing strat-egies. Unpublished manuscript.

Ni, W., & Crain, S. (1990). How to resolve structural ambiguities.Proceedings to NELS, 20 (2), 414^27.

Pearlmutter, N. J., & MacDonald. M. C. (1992). Plausibility effectsin syntactic ambiguity resolution. Proceedings of the 14th Annual


Conference of the Cognitive Science Society (pp. 498-503). Hills-dale, NJ: Erlbaum.

Pollatsek, A., Rayner, K., & Bolota, D. A. (1986). Inferences abouteye movement control from the perceptual span in reading. Per-ception & Psychophysics, 40, 123-130.

Rayner, K., & Frazier, L. (1987). Parsing temporarily ambiguouscomplements. Quarterly Journal of Experimental Psychology,39A, 657-673.

Rayner, K., & McConkie, G. W. (1976). What guides a reader's eyemovements? Vision Research, 16, 829-837.

Rayner, K., & Pollatsek, A. (1987). Eye movements in reading: Atutorial review. In M. Coltheart (Ed.), Attention and performanceXII (pp. 327-362). Hillsdale, NJ: Erlbaum.

Rayner, K., Sereno, S. C , Morris, R. K., Schmauder, A. R., &Clifton, C. (1989). Eye movements and on-line language com-prehension processes. Language and Cognitive Processing, Pars-ing and Interpretation, 4, 21—49.

Schmauder, R., Kennison, S., & Clifton, C. (1991). On the condi-tions necessary for observing argument structure complexity ef-fects. Journal of Experimental Psychology: Learning, Memory,and Cognition, 17, 1188-1192.

Shapiro, L. P., Nagel, H. N., & Levin, B. A. (in press). Preferencesfor a verb's complements and their use in sentence processing.Journal of Memory and Language.

Shapiro, L. P., Zurif, E., & Grimshaw, J. (1987). Sentence pro-cessing and the mental representation of verbs. Cognition, 27,219-246.

Shapiro, L. P., Zurif, E., & Grimshaw, J. (1989). Verb representationand sentence processing: Contextual impenetrability. Journal ofPsycholinguistic Research, 18, 223-243.

Spivey-Knowlton, M. J., Trueswell, J. C , & Tanenhaus, M. K.(1992). Context effects in syntactic ambiguity resolution: Parsingreduced relative clauses. Manuscript submitted for publication.

Tanenhaus, M. K., Boland, J. E., Garnsey, S. M., & Carlson, G. N.(1989). Lexical structure in parsing long-distance dependencies.Journal of Psycholinguistic Research, 18, 37-50.

Tanenhaus, M. K., Boland, J. E., Mauner, G., & Carlson, G. N. (inpress). More on combinatory lexical information: Thematic struc-ture in parsing and interpretation. In G. T. M. Altmann (Ed.),

Cognitive models of speech and processing. Volume 2: Psycho-linguistic and computational perspectives. Cambridge, MA: MITPress.

Tanenhaus, M. K., & Carlson, G. N. (1989). Lexical structure andlanguage comprehension. In W. D. Marslen-Wilson (Ed.), Lexicalrepresentation and process (pp. 529-561). Cambridge, MA: MITPress.

Tanenhaus, M. K., Carlson, G. N., & Trueswell, J. C. (1989). Therole of thematic structures in interpretation and parsing. Lan-guage and Cognitive Processing, Parsing and Interpretation, 4,211-234.

Tanenhaus, M. K., Garnsey, S. M., & Boland, J. (1991). Combi-natory lexical information and language comprehension. In G. T.M. Altmann (Ed.), Cognitive models of speech processing: Psy-cholinguistic and computational perspectives (pp. 383-408).Cambridge, MA: MIT Press.

Tanenhaus, M. K., & Lucas, M. M. (1987). Context effects in lexicalprocessing. Cognition, 25, 213-234.

Taraban, R., & McClelland, J. (1988). Constituent attachment andthematic role assignment in sentence processing: Influences ofcontent-based expectations. Journal of Memory and Language,27, 1-36.

Trueswell, J. C. (1993). (Untitled.) Unpublished doctoral disserta-tion, Department of Psychology, University of Rochester, Roch-ester, NY.

Trueswell, J. C , Tanenhaus, M. K., & Garnsey, S. M. (1992). Se-mantic influences on parsing: Use of thematic role informationin syntactic ambiguity resolution. Manuscript submitted for pub-lication.

Tyler, L. K. (1989). The role of lexical representations in languagecomprehension. In W. D. Marslen-Wilson (Ed.), Lexical repre-sentation and process (pp. 439^62). Cambridge, MA: MITPress.

Tyler, L. K., & Marslen-Wilson, W. D. (1977). The on-line effectsof semantic context on syntactic processing. Journal of VerbalLearning and Verbal Behavior, 16, 683-692.

West, R. F, & Stanovich, K. E. (1986). Robust effects of syntacticstructure on visual word processing. Memory & Cognition, 14,104-112.

(Appendix A begins on next page)


Appendix A

Verb

AcceptAdviseConfirmForgetLearnMaintainObserveRecallRememberRevealTeachUnderstandWarnWrite

BoastClaimDecide"HintHopeImplyInsistPretendRealizeWish

AdmitAgreeArgueAssertBelieveBragConfessDenyDisputeDoubtDreamFeelFigureGuessInferMentionNoticePrayPredictPromiseProtestRemarkSpeculateSupposeSuspectThink

Experiment

NP-bias1,2,31, 2, 32, 31,2, 32, 32, 312, 31, 2, 32, 31112, 3

% completions

Noun phrase Sentential

verbs from sentence9393

10057647986

1005779

100865057

completion0700

21211400

210

14210

Other

study

700

4315000

43000

2943

Sentence-bias verbs from sentence completion studyI, 2, 32, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2 , 3, 2, 3

2, 3

All other

07700

210070

verbs from sentence

3607

50210

435093360

360

434343500

577

430

210

290

43717

71647971509386

completion57297

4329214343

057212143215057502943

070

21937143

57228629360

29500

14

study7

71867

507914777

794364367

500

710

9350

1005870

57

That preference

100

10066

100

100

100100

100600

100299170576217

63100100675067

100100—5066661766

100100867550

—100—100466083

Note. That preference = percentage of sentence complement completions starting with that.a Because of an initial scoring error, decide was accidentally included as a sentence-bias word.


Appendix B

Target Sentences From Experiments 2 and 3

1. a. The waiter insisted/confirmed (that) the reservation was made yesterday,b. The scientist insisted/confirmed (that) the hypothesis was being studied.

2. a. The chef claimed/remembered (that) the recipe would require using fresh basil,b. Mr. Smith claimed/remembered (that) the directions would need to be changed.

3. a. The general pretended/revealed (that) the weapon was ready to be used,b. The professor pretended/revealed (that) the answer was not correct.

4. a. The accountant hinted/advised (that) the client was cheating on his taxes,b. The attorney hinted/advised (that) the defendant was planning to jump bail.

5. a. The author boasted/wrote (that) the novel was likely to be a best-seller.b. The lawyer boasted/wrote (that) the memo was from the president of the company.

6. a. The defendant wished/accepted (that) the verdict would be decided soon,b. The man wished/accepted (that) the award would go to his brother.

7. a. The gardener decided/maintained (that) the lawn was in good shape.b. The mechanic decided/maintained (that) the engine was working reliably.

8. a. The student hoped/forgot (that) the solution was in the back of the book,b. The woman hoped/forgot (that) the address was in the directory.

9. a. The student realized/learned (that) the language was spoken only in one province,b. The apprentice realized/learned (that) the skill was quite marketable.

10. a. The teacher implied/recalled (that) the answer was very complicated.b. The poet implied/recalled (that) his childhood was very unhappy.

Note. The same verb never appeared twice in the same list.

Received May 1, 1992Revision received August 24, 1992

Accepted October 5, 1992 •

Journal of Experimental Psychology: 1993, Vol. 19, No ...€¦ · Lexical Preference From...

Documents

Transcript of Journal of Experimental Psychology: 1993, Vol. 19, No ...€¦ · Lexical Preference From...