Journal of Experimental Psychology:Learning, Memory, and Cognition2001, Vol. 27, No. 5, 1223-1237

Copyright 2001 by the American Psychological Association, Inc.0278-7393/01/S5.00 DOI: 1O.1O37//O278-7393.27.5.1223

Dissociation Between Syntactic and Semantic Processing DuringIdiom Comprehension

Robert R. PetersonIllinois State University

Curt BurgessUniversity of California, Riverside

Gary S. DellUniversity of Illinois at Urbana-Champaign

Kathleen M. EberhardUniversity of Notre Dame

Syntactic and semantic processing of literal and idiomatic phrases were investigated with a primingprocedure. In 3 experiments, participants named targets that were syntactically appropriate or inappro-priate completions for semantically unrelated sentence contexts. Sentences ended with incomplete idioms(kick the . . . ) and were biased for either a literal (ball) or an idiomatic (bucket) completion. Syntacticallyappropriate targets were named more quickly than inappropriate ones for both contextual biases,suggesting that syntactic analysis occurs for idioms. In a final experiment, targets were either concrete(expected) or abstract (unexpected) nouns. For literal sentences, the abstract targets were named moreslowly than the concrete targets. In contrast, there was no concreteness effect for idiomatic sentences,suggesting that the literal meaning of the idiom is not processed. Overall, the results provide evidence fordissociation between syntactic and semantic processing.

It is generally agreed that language comprehension involves theconstruction of increasingly abstract representations of a messageover time. In reading, for example, visual input might be repre-sented successively in terms of its elementary visual features,letters and letter clusters, lexical items, syntactic structures, prop-ositions and, ultimately, an integrated representation combiningconceptual information from the text with real-world knowledge.The idea that processing flows across a series of representationallevels has had a profound influence on the direction that psycho-linguistic theorizing has taken over the past 20 years. Specifically,the issues that have been seen as central to the field have beenprecisely those that focus on the nature of and interdependenciesamong different processing levels.

In this article, we focus on the interdependencies that may existbetween syntactic and semantic levels of processing. In the psy-cholinguistic literature, this general issue has often manifesteditself in terms of the more specific debate over the modular natureof syntactic analysis. Specifically, some researchers have arguedthat the parser operates in a strictly bottom-up fashion, in the sense

Robert R. Peterson, Department of Psychology, Illinois State Univer-sity; Curt Burgess, Department of Psychology, University of California,Riverside; Gary S. Dell, Beckman Institute, University of Illinois at Ur-bana-Champaign; Kathleen M. Eberhard, Department of Psychology, Uni-versity of Notre Dame.

Preparation of this article was supported by National Institutes of HealthGrant DC-00191 and National Science Foundation Grant SBR 98-73450.We thank Keith Rayner and Susan Garnsey for comments and ludyAllen-Davis for work on the article.

Correspondence concerning this article should be addressed to Gary S.Dell, Beckman Institute, University of Illinois at Urbana-Champaign, 405North Mathews Avenue, Urbana, Illinois 61801. Electronic mail may besent to gdell@s.psych.uiuc.edu.

that initial syntactic decisions cannot be influenced by higher-levelsemantic considerations (Ferreira & Clifton, 1986; Frazier, 1985;Frazier, Clifton, & Randall, 1983; Frazier & Clifton, 1996; Rayner,Carlson, & Frazier, 1983). Other researchers, however, have en-visioned a quite different system, in which syntactic and semanticprocessors are highly interdependent. According to this view, theoutput from one processor can guide directly the other's ongoingcomputations (Altmann & Steedman, 1988; Crain & Steedman,1985; Garnsey, Pearlmutter, Myers, & Lotocky, 1997; Mac-Donald, Pearlmutter, & Seidenberg, 1994; Taraban & McClelland,1988; Trueswell, Tanenhaus, & Garnsey, 1994).

The most common approach for exploring the modularity ques-tion has been to investigate the processing of sentences containinglocal structural ambiguities. At issue in this research is whethersemantic context can bias the way that ambiguities are resolvedduring first-pass reading. Although this approach has proven to bea fruitful one, the modularity issue has, nevertheless, failed to beresolved, as proponents of both the modularity and the interactivecamps have accumulated empirical support for their positions.

In this study, we took a somewhat different approach for study-ing the interdependencies between syntactic and semantic analy-ses. Specifically, we investigated the syntactic and semantic ex-pectations that accompany two kinds of highly constrainingcontexts, involving the computation of an idiomatic meaning anda literal meaning. Consider, for example, the following two sen-tences:

1. The soccer player slipped when he tried to kick the ball.2. The man was very old and feeble and it was believed that he wouldsoon kick the bucket.

In both sentences, the final word is very predictable. However, thenature of this predictability is intuitively quite different. In theliteral sentence, at the syntactic level, ball is the direct object of the

1223

1224 PETERSON, BURGESS, DELL, AND EBERHARD

transitive verb kick. The semantics of kick indicate that this directobject should be a physical object and that it plays the thematicrole of patient—that which is kicked. Finally, the meaning of theearlier part of the sentence suggests that this physical object is asoccer ball. The idiomatic example is different. It has the transitiveverb kick, and it has a direct object, bucket. The meaning of thephrase, however, bears little relation to its syntactic structure. Thepredictability of bucket in this example is due to the system'srecognition that kick the bucket, meaning die, is consistent with theearlier part of the sentence.

The lack of correspondence between the syntax and the seman-tics of idiomatic phrases creates something of a dilemma for thecomprehension system. Although the syntax could create a struc-tural description of the idiom in a straightforward manner—kickthe bucket and, indeed, most idioms are syntactically wellformed—the semantic processor, if it used that description in thestandard way, would incorrectly generate the literal interpretation.To prevent the computation of the literal meaning, the semanticprocessor either would have to ignore structural information orwould have to signal the syntactic processor to cease its structuralanalysis (or both). Whether the comprehension system can actuallyuse either of these options obviously depends on the generalrelationship between syntax and semantics. Understanding theprocessing differences between phrases such as kick the bucket andkick the ball may therefore provide insight into the general char-acter of the comprehension system.

Our experiments aimed to examine processing in sentences suchas those shown above (sentences 1 and 2) just before the highlypredictable word (e.g., ball or bucket). We did this by using thesekinds of sentences, stripped of their final words, as primes fornaming visually presented targets. However, rather than the ex-pected continuations as targets, our targets were continuations withnear-zero probability that either preserved or violated the syntacticand semantic categories set up by the context. For example, ourfirst three experiments examined syntactic expectations. Afterreading (Experiment 1) or hearing (Experiments 2 and 3) " . . . soc-cer .. . kick the," or " . . . old man ... kick the" participants saw anoun (e.g., town) or a verb (e.g., grow) that they read aloud asquickly as possible. Any advantage for nouns therefore indexed theextent to which the noun category was expected in those contexts.We refer to this difference as the syntactic priming effect. Relatedforms of syntactic priming have been demonstrated with a reading-aloud or naming task (e.g., Boland, 1993; O'Seaghdha, 1997; West& Stanovich, 1986) and lexical decision (e.g., O'Seaghdha, 1989;Wright & Garrett, 1984). The syntactic priming effects resultingfrom sentence or phrasal primes are quite robust in naming, pos-sibly because sentence production processes are especially suscep-tible to syntactic relations (O'Seaghdha, 1997). When the prime isjust a single word, however, syntactic priming effects in namingare more variable (e.g., Carello, Lukatela, & Turvey, 1988; Sei-denberg, Waters, Sanders, & Langer, 1984).

In addition to assessing syntactic expectations in highly con-strained idiomatic and literal sentences, we also examined seman-tic effects (Experiment 4). Participants heard the same kinds ofincomplete sentence primes as in the earlier experiments, butinstead of comparing noun and verb targets, we sought a concep-tual priming effect. Naming times to noun targets that designatedconcrete objects (e.g., shelf) were compared to noun targets refer-ring to abstract concepts (e.g., truth). All sentences included verbs

such as kick or prepositions such as under, whose literal selectionalrestrictions required concrete objects. For literal sentence primes,slower naming times for the abstract target (truth) than for theconcrete one (shelf) would indicate the presence of literal semanticexpectations. For idiomatic primes, finding a concrete-abstractdifference would indicate that these semantic expectations are stillin play even though there is no literal kicking. The presence ofconceptual priming would therefore reveal computations that arerelated to the literal meaning of the idiom. Unlike syntactic prim-ing, which has been frequently studied (although not in idioms),this kind of conceptual priming has, to our knowledge, never beensought experimentally.

These priming manipulations should enable us to answer twoquestions. First, are syntactic expectations present during the pro-cessing of predictable literal and idiomatic phrases? The case ofthe idiom is particularly illuminating given its odd syntactic-semantic mapping. Second, is the semantic processor compelled toderive the literal interpretation of an idiom on the basis of itssyntactic structure? Below, we outline three possible outcomes ofthese experiments and relate these to theory. Each outcome isassociated with a particular view or model of the relationshipbetween syntax and semantics. The predictions made by thesemodels are summarized in Table 1.

Syntactic Dominance Model

According to this model, the syntactic system obligatorily ana-lyzes grammatical strings and makes the outcome of this analysisavailable to a semantic processor. The semantic processor is drivenby the incoming structural analysis and, as a result, tends to derivea literal interpretation. In the syntactic dominance model, thecomputation of syntactic structure has a great deal of momentum.Therefore, despite contextual information that indicates that thecomputation may not be necessary, it nonetheless continues. Also,the syntactic analysis has the power to impel corresponding se-mantic interpretations. Just as a rolling ball's momentum is trans-ferred to objects that it contacts, syntactic momentum is transferredto the development of literal semantic expectations.

Several specific models of parsing emphasize the primacy andinertia of syntax and thus are compatible with the syntactic dom-inance model. For example, in the garden-path model (e.g., Fra-zier, 1987; Frazier & Rayner, 1982; Rayner et al., 1983) and itsrecent descendant, construal theory (Frazier & Clifton, 1996),input is initially structured by a constituent building module that

Table 1Predictions of the Three Models Regarding Syntactic andConceptual Priming in Literal and Idiomatic Contexts

Processing model

Syntactic dominanceSemantic dominanceAutonomy

Syntactic

Literalcontext

++0+

priming

Idiomaticcontext

+0+

Conceptual

Literalcontext

+++

priming

Idiomaticcontext

+00

Note. + = prediction of a priming effect; 0 = prediction of no priming;+0 = prediction of either result.

IDIOM COMPREHENSION 1225

pays attention only to the grammatical category of words. Theparse can be later revised when the semantic interpretation of theparse is found to be inappropriate, but this process is associatedwith effort and, in reading, with regressive eye movements.

With respect to our experiments, the syntactic dominance modelpredicts that syntactic priming will be obtained in both literal andidiomatic conditions. It also predicts that conceptual priming willoccur in both conditions. Even though the context makes it clear inthe idiom condition that, for example, kick the refers to dying, thesyntactic analysis proceeds normally and, more importantly, ini-tiates the literal semantics associated with this analysis.

Semantic Dominance Model

In this model, the syntactic analysis has very little momentum.The semantic processor, in computing an interpretation, can by-pass or terminate the syntactic analysis, depending on the contex-tual constraints present in the input. Thus, this model is heir toapproaches to language comprehension that minimize the influ-ence of syntax (e.g., Schank, 1984). In such models, syntacticinformation is computed only when there is some uncertaintyabout meaning. Thus, in the kick the example, the system maydetermine the syntactic characteristics of the words following kickonly when there is a conceptual role associated with kicking thatneeds to be filled. In the idiom condition, there is no such role.

With respect to our experiments, the semantic dominance modelpredicts that syntactic priming will not be in evidence in the idiomcondition, since retrieval of the figurative meaning of the idiomwill allow the semantic processor to terminate further syntacticanalysis. In fact, there is some reason to expect that syntacticpriming will be curtailed in the literal condition as well, given thestrong semantic context present. In such a context, there may be noneed to consult syntactic information any further. With respect toconceptual priming, the semantic dominance model predicts aneffect only in the literal condition. There should be no computationof literal semantic expectations in the idiom condition.

Syntax-Semantics Autonomy Model

According to this model, the syntactic and semantic processorsare capable of independent operation. Neither dominates the other.At the syntactic level, analysis continues as long as the incomingwords form a grammatical string. Even if the meaning of theutterance has already been determined, the syntax continues. Inthis sense, the autonomy model is like the syntactic dominancemodel. It differs from the syntactic dominance model, though,because it assumes that the syntactic analysis does not compel thecreation of a literal semantic representation. The semantic proces-sor can, by consideration of context and the meanings of lexicalitems, assemble a representation of the meaning on its own.

The autonomy model is most closely related to a model ofsyntactic-semantic interaction developed by Boland (1993, 1997).Boland's approach, called the concurrent model, emphasizes that asemantic representation can be produced before the parser hassettled on a syntactic representation. The semantic processor doesthis by considering thematic role information associated with verbsand context (see Burgess & Lund, 1994; MacDonald et al., 1994;Tanenhaus, Carlson, & Trueswell, 1989). The syntax also worksby itself. It computes syntactic structures, at least initially, without

considering any semantic information. Eventually, the syntacticand semantic representations are matched against one another, andinconsistent pairs are discarded. A consequence of the autonomyof syntax and semantics in this approach is that the two processorscan develop independent expectations. This situation is seen mostclearly in the idiom condition in our experiments, in which a nounis expected in a direct object slot (because kick is a transitive verb),but there is no corresponding semantic expectation for a concreteobject (because the semantics of kick the bucket have no suchthematic role).

The autonomy model is also at least partly consistent withinteractive parsing models that allow semantic information to havean early impact on the resolution of syntactic ambiguities (e.g.,MacDonald et al., 1994; Trueswell & Tanenhaus, 1994). The keyhere is that there is no syntactic ambiguity in phrases such as kickthe bucket. Hence, any semantic influence from either a literal ora figurative interpretation would not change the syntactic process-ing. If, however, an interactive model did allow semantics evenmore influence, for example, to eliminate syntactic processes oncea figurative meaning has been identified, then such a model wouldcount as a semantic dominance model rather than as an autonomymodel with our nomenclature.

The autonomy model predicts that syntactic priming will befound in both the idiom and the literal conditions in our experi-ments. The autonomy of syntax ensures that syntactic expectationswill continue to be generated, even for the idiom condition. Con-ceptual priming, however, should differ in the literal and idiomconditions. For the literal condition, the semantic processor ex-pects a concrete object (e.g., what is being kicked); hence, abstracttarget nouns mismatch expectations and should be associated withslower times than concrete ones. With idioms in strong contexts,the semantic processor can settle quite early on the idiomaticmeaning (e.g., kick the means dying); hence, there is no particularexpectation for a concrete noun to fill a semantic role. As a result,there will be no conceptual priming.

Because we used idioms to examine syntactic and semanticexpectations, it is worthwhile to review the several studies in theliterature that have investigated the role of literal processing duringthe comprehension of idioms. To a large extent, the idiom-processing literature has not addressed the relationship betweensyntactic structure and literal processing; hence, it is difficult torelate its findings to our concerns. Nonetheless, the models thathave evolved in idiom-processing studies do show some corre-spondence with our three models.

Idiom-Processing Studies

Swinney and Cutler (1979) proposed that idioms are stored assingle entries in the mental lexicon and are retrieved in the samemanner as any other lexical item. Thus, according to these re-searchers, the comprehension of an idiom does not require that itsfigurative meaning be actively computed, since the meaning can beaccessed directly from the lexicon. Importantly, Swinney andCutler also assumed that the retrieval of an idiom's figurativemeaning occurs in parallel with the computation of its literalmeaning. In their model, however, figurative processing concludesmore quickly than literal processing, since the direct access of thefigurative meaning is simpler than the compositional analysisrequired to derive a literal interpretation. In their experiment,

1226 PETERSON, BURGESS, DELL, AND EBERHARD

Swinney and Cutler instructed participants to judge whether astring of words formed a meaningful phrase. They found thatresponses were faster for idiomatic phrases (e.g., see the light) thanfor phrases that had only a literal interpretation (e.g., get the light),thus supporting their parallel-access model.

Several researchers have argued that parallel access occurs notonly for idioms presented in isolation but also for idioms presentedin contexts that clearly bias the idiom's interpretation (e.g., Estill& Kemper, 1982; Ortony, Shallert, Reynolds, & Antos, 1978).This claim has generally been based on demonstrations, such asSwinney and Cutler's (1979), that idioms are no more difficult andperhaps are easier to process than are literal phrases. The claim thatthe literal meaning of an idiom is retrieved even in contexts inwhich it is inappropriate suggests that literal processing is oblig-atory. Thus, the parallel-processing view appears to be most com-patible with the syntactic dominance model outlined above, inwhich the derivation of a syntactic structure of an idiom compelsthe semantic processor to derive a literal description of the idiom.

There is, however, an alternative interpretation of Swinney andCutler's (1979) data. Gibbs (1986) suggested that faster responsesto idiomatic phrases could indicate that retrieval of the figurativemeaning of an idiom always precedes literal processing, with theliteral meaning being derived only if the figurative meaning isfound to be anomalous given the immediate context. Thus, Gibbssuggests a serial-processing model, rather than a model in whichmeanings are retrieved in parallel.

Gibbs' (1986) model is most compatible with the semanticdominance and autonomy models, since both of these modelspropose that the semantic processor can terminate the literal pro-cessing of an idiomatic string. There are, however, differencesbetween Gibbs' view of idiom processing and the account pro-vided by these two models, and it is important to highlight thesedifferences. First, Gibbs claims that the figurative meanings ofidioms can be retrieved without any analysis of their literal mean-ings. This view, however, does not specify how the comprehensionsystem decides to terminate ongoing literal analysis. In contrast,the semantic dominance and autonomy models are explicit withrespect to this issue. They both propose that it is the retrieval andintegration of a figurative meaning itself that signal the semanticprocessor to cease literal processing. Thus, these two modelspredict that literal processing will occur during the comprehensionof an idiomatic string at least until the figurative nature of thestring is appreciated.

Another difference between Gibbs' (1986) model and the twomodels that we have proposed is that Gibbs' model does not dealexplicitly with syntactic processing; thus, it is difficult to deter-mine whether his model is most consistent with the semanticdominance model (in which an idiom is not fully analyzed syn-tactically) or the autonomy model (in which an idiom does un-dergo a full syntactic analysis). In other work, however, Gibbs andGonzales (1985) investigated how the processing of an idiomvaries as a function of the idiom's syntactic frozenness (see alsoReagan, 1987; Swinney & Cutler, 1979). Syntactic frozennessrefers to the extent to which an idiom can undergo syntactictransformations and still retain its figurative meaning. It appearsthat there is a great deal of variability across idioms in terms oftheir level of frozenness. Some idioms can undergo few transfor-mations (e.g., John kicked the bucket can mean that John died; thebucket was kicked by John cannot). Other idioms, however, are

much more flexible (e.g., both her father laid down the law and thelaw was laid down by her father seem to mean that her father gavestrict orders). In addition to these extremes, idioms can also beintermediate in their level of syntactic flexibility (retaining theirmeaning under some, but not all, transformations). Thus, frozen-ness appears to be a continuous rather than a strictly dichotomousdimension (Fraser, 1970; Gibbs & Gonzales, 1985; Reagan, 1987).

This idea that idioms vary in their degree of syntactic flexibilitywould seem to suggest that the semantic dominance and autonomymodels might be too extreme in their views of syntactic process-ing. Rather than the all-or-none view taken by these two models,it might be better to think of syntactic processing in more contin-uous terms, with idioms undergoing different degrees of structuralanalysis depending on their level of frozenness (and/or degree ofsemantic decomposability). It is important to note, however, thatthese different perspectives regarding the continuity of syntacticprocessing are not necessarily incompatible. The apparent differ-ences arise because the various models focus on slightly differentquestions. In our models, we are interested in whether the syntacticprocessor automatically computes the structure of an idiomaticphrase, but we do not directly address the question of whether afull syntactic description is required to access an idiom's figurativemeaning. However, it is this latter question that was of specificinterest to Gibbs and Gonzales (1985). These questions are logi-cally independent, however. For example, it is possible that thesyntactic processor obligatorily performs a complete structuralanalysis of an idiom (consistent with the autonomy model), eventhough such an analysis is required only for extremely unfrozenidioms in order for their figurative meanings to be retrieved. Wewill return to a discussion of syntactic frozenness followingExperiment 3.

Overall, then, it is difficult to use the results from existingstudies to evaluate the feasibility of the three models of processingthat we present here. Although results from previous studies havebeen used to test models that share some similarities with thosethat we outline here, the data from those experiments do notprovide direct evidence regarding on-line syntactic and semanticprocessing. Our goal in this set of studies was to collect compre-hension data that can provide exactly this kind of evidence.

Our first three experiments examined syntactic priming in idi-omatic and literal contexts, and the fourth experiment examinedconceptual priming in the same contexts. The first experiment useda visual presentation of the primes, which were either literal (thesoccer player slipped when he tried to kick the) or idiomatic (theman was very old and feeble and it was believed that he wouldsoon kick the). The targets, also presented visually, were seman-tically unrelated nouns (e.g., town) or verbs (e.g., grow). Thepresence of syntactic priming with the literal prime will demon-strate the sensitivity of the naming task to syntactic categoryexpectations. If such priming is also obtained in the idiom condi-tion, the syntactic dominance and autonomy models will receivesupport.

Experiment 1

MethodParticipants. Thirty-two undergraduates at the University of Rochester

served as participants and received class credit for participation. All par-

IDIOM COMPREHENSION 1227

ticipants were native English speakers and had normal or corrected-to-normal vision. These conditions applied to participants in all fourexperiments.

Stimuli. Thirty-two idiomatic phrases ending with a noun were se-lected from Boatner, Gates, and Makkai (1975). For each idiom, a corre-sponding literal phrase was created by changing the final word of the idiomto a different noun. For example, we derived the literal phrase kick the ballfrom the idiomatic phrase kick the bucket (see the Appendix for a fulllisting of the stimuli). The final nouns for the idiomatic and literal phraseswere balanced for length (the means were 4.7 and 4.8 characters for theidiomatic and literal phrases, respectively) and frequency (the mean logfrequencies were 1.57 and 1.58, respectively). For each of the 32 phrasepairs, two sentences were written: one ending with the idiomatic phrase andone ending with the literal phrase. The idiomatic and literal sentences werebalanced for length, with means of 17.9 and 18.2 words, respectively. Thesentences were strongly biased so that the final word of the phrase washighly predictable. Sentences 1 and 2 are examples of idiomatic and literalsentences, respectively.

1. The man was old and feeble and it was believed that he would soonkick the bucket.2. The soccer player slipped when he tried to kick the ball.

A norming study was conducted to ensure that there was no differencebetween the idiomatic and literal sentences in the predictability of the finalword. We removed the last word from each of the sentences and instructedparticipants to complete the resulting sentence fragments. Each participantwas presented with either the literal or the idiomatic version of each ofthe 32 sentence pairs, with the restriction that 16 of the sentences wereliteral and 16 were idiomatic. Two different presentation lists were createdso that, across lists, both the literal and the idiomatic forms of each itemwere presented. Each list was given to 45 participants in booklet form, andparticipants wrote their completions in the booklet. The results of the studyare straightforward. For each of our sentences, the word that was mostfrequently given as a completion by participants was the same word that wehad used to complete the sentence (e.g., bucket or ball). Further, theproportion of times that this word was given did not differ between theidiomatic and the literal sentences (the mean proportion was .87 for bothsentence types), indicating that the two sentence versions were equallyconstraining.

Thirty-two monosyllabic nouns and 32 monosyllabic verbs were se-lected to serve as target items. The nouns and verbs were matched in termsof printed frequency (Kucera & Francis, 1967) and length (the nouns andverbs each had an overall mean log frequency of 1.96 and a mean lengthof 4.3 characters). In an additional norming study, we collected partici-pants' naming times for each of the 64 target words presented in isolation.Twenty-two participants were tested, and each participant responded toall 64 targets. The verb targets were named slightly more quickly than thenoun targets (417 versus 425 ms, respectively), and this difference wasmarginally significant, F(\, 21) = 4.01, MSE = 178, p = .058. It isimportant to note that this effect is in a direction opposite that of thesyntactic priming effect that is sought in Experiment 1. Targets in Exper-iment 1 are presented at the end of a sentence context in which a noun isexpected. Hence, the noun targets should be named more quickly than theverb targets if there is any syntactic priming. Given that the verbs werenamed more quickly in our pilot study, any syntactic effect that may arisein Experiment 1 cannot be attributed to baseline differences between nounand verb targets.

Design and procedure. The design of the experiment was a 4 x 2 X 2factorial with list (four stimulus lists) as a between-subjects factor andgrammatical category (noun and verb targets) and sentence type (idiomaticand literal) as within-subject factors.

The final noun from the idiomatic and literal sentence contexts wasremoved, and the resulting sentence fragments served as primes in theexperiment. Each idiomatic-literal sentence pair was randomly assigned a

noun-verb target pair, with the restriction that neither the noun nor the verbbe semantically related to either of the sentences. An initial stimulus listwas created by randomly assigning eight of the items to each of the fourSentence Type X Grammatical Category conditions. Three additional listswere created by rotating items across the four conditions such that, acrossthe four lists, all of the items appeared in each of the experimentalconditions.

Sixty-four filler trials were randomly interspersed within these lists,making a total of 96 trials per list. The fillers were included so that thetargets would sometimes be sensible completions to the sentence contexts.Sixteen of the fillers had sentence primes that ended with an idiomaticphrase, whereas the other 48 fillers had literal primes. We included moreliteral than idiomatic fillers in order to discourage participants from ex-pecting idiomatic phrases on any given trial in the experiment. Targets forthe idiomatic fillers were always the expected completions for the idioms.For the literal fillers, 32 had targets that were sensible completions for thesentence contexts, whereas 16 had semantically unrelated targets (althoughall of these were syntactically appropriate). Thus, an entire stimulus listconsisted of 32 idiomatic primes and 64 literal primes, with half of each ofthese types being followed by a target that was a meaningful completionfor the sentence prime.

Participants were tested individually and were seated in front of an IBMPC at a distance of approximately 60 cm. On each trial, participants presseda response button with their left index finger to display the sentence contextand then read the context aloud at a normal reading pace. As they weresaying the last word of the context, the participants pressed a differentresponse button with their right index finger. Four hundred millisecondsafter the button was pressed, the sentence context disappeared and a targetword appeared one space to the right of where the last word of the contexthad been (i.e., the target appeared where it would have been, had it beenpart of the context). The appearance of the target initiated a millisecondtimer (Digitry CTS system). Participants were instructed to name the targetword as quickly as possible. Their naming response stopped the timer andcleared the computer screen. After 3 s, the next trial began.

To ensure that participants were attending to the meaning of the sen-tences, a comprehension task was given to them immediately following thecompletion of a trial for 20 randomly selected filler trials with meaningfulcompletions. For this task, a statement was presented on the computerscreen, and the participants were asked to decide whether it was true orfalse with respect to the sentence that they had read on the preceding trial.

Participants were given 20 practice trials. None of the practice trialscontained any idiomatic phrases. Half of the trials had targets that weresensible completions for the sentence primes, whereas the other half hadtargets that were semantically unrelated to their contexts. All the targetswere syntactically correct completions.

Results and Discussion

For each participant and for each noun-verb target pair, themean response latency and error proportion were computed foreach Sentence Type X Grammatical Category condition. Reactiontimes were excluded from the analysis if (a) the response waslonger than 2 s, (b) a mechanical failure occurred on the trial, or (c)the participant made an error. Reaction times greater than or lessthan 2 standard deviations from the participant's mean were re-placed by the 2-standard-deviation cutoff score. A 4 (list) X 2(sentence type) X 2 (grammatical category) mixed analysis ofvariance was conducted on the participants' mean latencies anderror rates (reported here as Fl), with list serving as a between-subjects factor (e.g. Pollatsek & Well, 1995). A parallel analysiswas conducted on the item means (reported here as F2), differingfrom the subject analysis only in that item group rather than list

1228 PETERSON, BURGESS, DELL, AND EBERHARD

served as the between-items factor. For all analyses, an alpha of.05 was used.

Participants' mean naming latencies are shown in Figure 1. Theverb targets were named more slowly than the noun targets (599and 575 ms, respectively). This main effect of grammatical cate-gory was significant, Fl(l, 28) = 18.15, MSE = 1,044, F2(l,28) = 4.43, MSE = 2,817. Thus, a syntactic priming effect wasobtained. The interaction between grammatical category and sen-tence type was not significant, both Fl and F2 < 1, indicating thatthe magnitude of the syntactic effect did not differ between theidiomatic and the literal sentences. Thus, there is no evidence thatnormal syntactic processing is in any way usurped during thecomprehension of an idiom.

For figuratively biased contexts, the mean error proportionswere .06 for verb targets and .02 for noun targets. For literallybiased contexts, the proportions were .07 and .02, respectively.Only the main effect of grammatical category was significant,Fl{\, 28) = 13.97, MSE = .01, F2(l, 31) = 8.59, MSE = .01.Consistent with the latency data, the verb targets resulted in ahigher proportion of errors than did the noun targets (.07 and .02,respectively). Also consistent with the latency data is the fact thatthe interaction between sentence type and grammatical categorywas not significant, both Fs < 1. Thus, the error data provideadditional support for the claim that idioms undergo normal syn-tactic processing.

The existence of syntactic priming with idiomatic sentences isconsistent with the predictions made by the syntactic dominanceand syntactic-semantic autonomy models. Both of these modelsassume that the syntactic processor obligatorily parses an idio-matic phrase. In contrast, the results of Experiment 1 fail tosupport the semantic dominance model. In this latter model, pars-ing of an idiomatic string is discontinued once the phrase has beenfiguratively interpreted; therefore, the observed syntactic primingeffect in Experiment 1 makes this model less tenable. However, itis premature to reject the model based solely on the results ofExperiment 1, since the semantic dominance model assumes thatparsing is terminated only if the literal meaning of the idiom is nolonger being computed. Thus, disproving the model requires test-ing for the derivation of literal meaning. We address this issue inExperiment 4.

In the next two experiments, we attempted to replicate thefindings from Experiment 1 using somewhat different conditionsin order to establish the reliability and generality of the syntacticpriming effect.

Experiment 2

Experiment 1 involved a visual presentation of the sentencecontexts. However, it seems likely that idioms are more typicallyfound in speech than in written text. If so, participants in Exper-

Experiment 1Target Word

• VerbDNoun

Idiom LiteralSentence Type

Experiment 3

710-,Target Nonword

• "Verb"• "Noun"

IdiomSentence Type

Experiment 2Target Word

• Verb• Noun

Idiom LiteralSentence Type

Experiment 4Target Word• AbstractD Concrete

Idiom LiteralSentence Type

Figure 1. Naming latencies for all conditions in Experiments 1 through 4.

IDIOM COMPREHENSION 1229

iment 1 might have parsed the idiomatic strings literally because ofthe novelty of the idioms in written form. Thus, the observedsyntactic priming effect for idiomatic phrases might have been anartifact of the mode of presentation of the sentences. To rule outthis possibility, Experiment 2 involved a cross-modal primingtechnique in which the sentence contexts were presented auditorilyand the targets were presented visually.

An additional advantage of the cross-modal technique is that itallows greater control of the delay between the presentation of thesentence context and the onset of the target. In Experiment 1,participants pressed a button as they began to say the last word ofthe context, and the target appeared 400 ms later. This methodentails some variability with regard to when the button pressactually occurs. With the cross-modal technique, we were able totime precisely the onset of the target relative to the offset of theauditory signal corresponding to the sentence context. In Experi-ment 2, targets were presented immediately following the offset ofthe final word of the context.

MethodParticipants. Thirty-two undergraduates at the University of Rochester

served as participants.Stimuli. The sentence primes and target words from Experiment 1 were

used. The sentences were recorded on one channel of a Sony TC-270 stereotape recorder. A separate tape was made for each of the four stimulus listsused in Experiment 1. For each of the tapes, brief 1000-Hz tones wereplaced on a channel separate from the sentence contexts such that the tonescoincided with the offset of the final word of each context. The point ofoffset was determined by running the tapes slowly over the head of areel-to-reel tape recorder.1 In addition to the critical and filler sentences,the comprehension questions used in Experiment 1 were also recorded oneach of the tapes.

Design and procedure. The experimental design was the same as inExperiment 1. The only difference in procedure from the previous exper-iment was that the sentences were presented auditorily. On each trial in theexperiment, participants heard, over a pair of headphones, an announce-ment of the trial number, followed by the sentence context. At the end ofthe last word of the context, the tone from the other channel of the recorder(which was not heard by participants) was fed into a voice-activated relay,which was interfaced with the computer. When the tone activated the relay,the computer immediately presented the target word on the computerscreen (thus, there was a 0-ms context—target interstimulus interval). As inExperiment 1, participants were instructed to name the target word asquickly as possible. Following 20 trials, the participants responded to thecomprehension question in the same manner as in Experiment 1.

Results and Discussion

Analyses identical to those performed in Experiment 1 wereconducted. Mean naming latencies are given in Figure 1. The maineffect of grammatical category was significant both with subjects,Fl(l, 28) = 13.07, MSE = 1,193, and with items, F2(l,28) = 4.84, MSE = 2,223. Consistent with the results of Experi-ment 1, the verb targets were named more slowly than the nountargets (530 and 508 ms, respectively). The magnitude of thiseffect was numerically larger for the idiomatic sentences than forthe literal sentences (29 and 16 ms, respectively), although thisinteraction was not significant in either the subject or the itemanalysis, Fi ( l , 28) = 1.22, MSE = 1,008, F2(l, 28) = 2.38,MSE = 606. Thus, as in Experiment 1, the latency data clearly

demonstrate that there is no diminution of the syntactic primingeffect when idiomatic contexts are used.

For figuratively biased contexts, the mean error proportionswere .05 for verb targets and .03 for noun targets. For literallybiased contexts, the proportions were .04 and .02, respectively.The effect of grammatical category was significant with subjects,Fl(\, 28) = 4.26, MSE = .003, but not with items, F2(l,28) = 1.20, MSE = .01. The interaction between sentence type andgrammatical category was nonsignificant in both analyses, bothFs < 1.

Overall, the results of Experiment 2 were consistent with thoseof Experiment 1. In both studies, the syntactic priming effect foridiomatic sentences was as large as (or larger than) the effect forliteral sentences. Thus, regardless of the modality of presentationof the prime, robust syntactic expectations are present in bothkinds of sentences.

Experiment 3

Despite the consistency of the results of Experiments 1 and 2,there is a possible criticism of the studies that could qualify ourconclusions. Perhaps the syntactic priming effect that was obtainedin these experiments was not a reflection of the syntactic incom-patibility of the verb targets. Rather, the priming effect might havereflected facilitation in naming the nouns because of a bias in ourmaterials that resulted in the nouns being more semantically re-lated than the verbs to the general meaning of the sentences.Although this possibility seems unlikely, we nevertheless decidedto replicate the findings of Experiment 2 by using nonword targetsthat have suffixes primarily associated with a verb or a noun. Forthe "verb" nonwords, we used the suffix -ed, which is a salientinflectional marker for past tense. For the "noun" nonwords, weused the suffix -er, which is the derivational morpheme for agen-tive.2 Thus, the -ed and -er nonwords functionally paralleled theverb and noun targets used in Experiments 1 and 2. A syntacticpriming effect would be reflected, therefore, by longer naminglatencies for the -ed nonwords than for the -er nonwords when theyfollowed the sentence contexts used in the first two experiments.Because the targets are nonwords differing only in their suffixes,an obtained syntactic priming effect could not be attributed todifferences in the targets' semantic relationships to the sentencecontexts.

MethodParticipants. Forty-eight University of Rochester undergraduates

served as participants.Stimuli. Thirty-two pairs of nonwords were created. One of the non-

words within each pair ended with -ed. The other item within the pair wasidentical to the first except that it ended with -er rather than -ed. Anexample of such a pair are the nonwords glatted and glatter, respectively.

1 One should note that any error in tone placement is orthogonal to theexperimental contrasts of interest, given that both syntactically appropriateand inappropriate targets are tested with the same auditory sentence primes.

2 Of course, the suffix -er is also used with comparative adjectives.However, the ambiguity in the interpretation of the -er suffix is not aproblem in this study, since both nouns and adjectives are syntacticallyappropriate continuations of our sentence contexts.

1230 PETERSON, BURGESS, DELL, AND EBERHARD

As we did for the word targets used in Experiments 1 and 2, we collectedparticipants' naming latencies for the nonword targets presented in isola-tion. Pilot participants were presented with either the -ed or the -ernonword from each of the 32 nonword pairs. In addition, they werepresented with 64 filler targets corresponding to the filler trials to be usedin the primary experiment. These filler targets consisted of 16 nonwords(none of which ended with -ed or -er) and 48 words. Twenty-five partic-ipants were tested, although 1 participant was excluded because of anexcessive number of errors on the critical nonword trials. The mean naminglatencies for the -ed and -er nonwords were 595 and 599 ms, respectively.This difference was not significant, both Fl and F2 < 1. The errorproportions for the -ed and -er nonwords were .09 and .07, respectively. Aresponse was considered an error if there was any lack of fluency in aparticipant's response or if the participant used a pronunciation for thenonword that differed from the pronunciation that we had intended. Thedifference in error proportions did not approach significance, Fl(l,20) = 1.66, MSE = .00, F2(l, 20) = 1.65, MSE = .00. These data suggestthat there are no baseline differences between the two types of nonwords;hence, any differences obtained in Experiment 3 can legitimately beattributed to the influence of the preceding sentence contexts.

Design and procedure. The procedures for this experiment were iden-tical to those used in Experiment 2. However, the filler targets werechanged somewhat so that, across an entire stimulus list, there would be anequal number of word and nonword targets (of the 64 fillers, 48 were wordsand 16 were nonwords). Of the filler targets that were words, half weresensible completions for the sentence contexts and half were not (althoughall of these latter targets were syntactically correct completions).

Results and Discussion

Analyses identical to those performed in Experiments 1 and 2were conducted. Mean naming latencies are given in Figure 1. Themain effect of grammatical category was significant both withsubjects, F i ( l , 44) = 19.60, MSE = 2,667, and with items, F2(l,28) = 7.28, MSE = 2,857. Consistent with the effects found inprevious experiments, the -ed nonwords were named more slowlythan the -er nonwords (692 and 659 ms, respectively). The inter-action between grammatical category and sentence type was sig-nificant in both analyses, Fl{\, 44) = 5.80, MSE = 2,181, F2(l,28) = 5.93, MSE = 1,598, indicating that the syntactic primingeffect was larger for idiomatic than for literal sentences. Simpleeffect analyses revealed that the 49 ms priming effect for idiomaticsentences was highly significant, both with subjects and withitems, Fl{\, 44) = 17.50, MSE = 3,324, F2(\, 28) = 14.96,MSE = 1,950. The 17-ms effect for literal sentences was signifi-cant with subjects only, Fi ( l , 44) = 4.43, MSE = 1,523, F2(l,28) < 1.

For figuratively biased contexts, the mean error proportionswere .16 for verb targets and .10 for noun targets. For literallybiased contexts, the proportions were .17 and .11, respectively.The effect of grammatical category was significant both withsubjects and with items, F i ( l , 44) = 9.17, MSE = .03, F2(l,28) = 8.21, MSE = .05. No other effects approached significance.

We have once again demonstrated a robust syntactic primingeffect for idiomatic phrases. Interestingly, this priming was ob-tained with nonword stimuli, showing that syntactic category ex-pectations are not strictly lexical (e.g., the word the is associatedwith the word town). Moreover, in each of the experiments that wehave conducted, the syntactic effect for idiomatic primes waslarger than the effect for literal primes, although this differencewas statistically significant only in Experiment 3. Given the con-

sistency of this pattern of data across the three experiments, itseems reasonable to speculate about whether the processing of anidiomatic phrase involves a heightened sensitivity to the phrase'ssyntactic structure. Why might this effect occur? We assume thatthe representation of an idiom in memory includes informationregarding both its meaning (i.e., that kick the bucket means die)and its syntactic structure. When one encounters the idiom in abiasing context, it is possible that the mental representation of theidiom is accessed relatively quickly (likely before one has beenexposed to the entire idiom, given the fact that idioms are severalwords long). The retrieval of the idiom makes available theidiom's syntactic structure and thus allows the comprehensionsystem to project the upcoming structure of the remainder of theidiomatic phrase in a relatively determinate fashion. A literalphrase, since it does not have a preexisting mental representation,may be less efficient at generating specific syntactic expectations.Idiomatic phrases, therefore, may make available syntacticallyrelevant information more robustly than literal phrases; as a result,they may actually be more, not less, susceptible to interferencewhen syntactic violations occur. However, as our data are some-what equivocal on this point, we emphasize the more conservativeconclusion, that idiomatic phrases are at least as susceptible tointerference from syntactic violations as are literal phrases.

A related but distinct issue concerning syntactic structure isfrozenness. As noted in the introduction, Gibbs and Gonzales(1985) have presented evidence suggesting that syntactically fro-zen idioms are easier to process than are unfrozen idioms. Onepossible explanation for this processing difference is that frozenidioms do not undergo a full syntactic analysis during comprehen-sion. If this view is correct, then one might expect that frozenidioms would be the least susceptible to syntactic priming effects.However, the fact that we obtained syntactic effects for idiomaticphrases at least as large as if not larger than those for literal phrasessuggests that this possibility might not be correct. Even fairlyunfrozen idioms are less flexible syntactically than are their literalcounterparts; hence, one would have expected that the syntacticeffects for idiomatic phrases would have been smaller than theeffects for literal phrases, regardless of the overall level of frozen-ness of the idioms. Alternatively, perhaps frozenness aids in thegeneration of syntactic expectations. The more frozen an idiom,the more its syntactic structure is directly stored with it. In thisview, perhaps frozenness leads to more syntactic priming. In anycase, it would be valuable to examine this issue with our data.

To test directly whether frozenness affects the syntactic process-ing of idiomatic strings, we obtained frozenness ratings for each ofour idioms. Sixty participants were given either the literal or theidiomatic sentence from each of our 32 sentence pairs, along witha syntactically transformed version of the sentence. The transfor-mation was either passive, when possible, or another noun-phrasemovement, when passive was not grammatical. The participantsrated how similar in meaning the original and transformed sen-tences were. For example, the passive transformation applied tothe kick the bucket example would lead to a comparison betweenthat he would soon kick the bucket and that the bucket would soonbe kicked by him. A five-point rating scale was used, with 1 beingvery different in meaning and 5 being very similar in meaning.Overall, the mean rating for the idiomatic sentences was lowerthan that for the literal sentences (2.67 and 3.81, respectively), andthis difference was highly significant, F(l, 31) = 62.20,

IDIOM COMPREHENSION 1231

MSE = 0.34. Thus, it is clear that the idiomatic phrases are lessflexible syntactically than are the corresponding literal phrases.This is an important demonstration, since it might have beenargued that we obtained large priming effects for our idioms in theprevious experiments because our idioms were extremely unfro-zen. Given the results of these ratings, this notion does not seem tobe a viable interpretation of our data.

On the basis of the above ratings, we obtained frozenness scoresfor each of our idioms by subtracting the mean rating for eachidiomatic sentence from the mean rating for its literal control.Thus, the higher the frozenness score, the greater the frozenness ofthe idiom (i.e., a high score indicates that the rating was relativelylow for the idiom relative to the rating for the idiom's literalcontrol). For example, a relatively frozen idiom such as kick thebucket had a score of 1.68 in our sentences. In comparison, thescore for the unfrozen idiom break the ice was —0.69. The fro-zenness scores for the idioms ranged from —0.69 to 3.00. Wecomputed the correlation between the frozenness scores and thesyntactic priming effects obtained for the idioms in Experiments 1,2, and 3. The correlations were -0.09, -0 .01 , and -0.08, respec-tively. None of these correlations approached significance (allhad 28 degrees of freedom and p values of greater than .50, witha two-tailed test). Thus, frozenness did not affect the magnitude ofthe syntactic priming effect. Apparently, on-line syntactic expec-tations are not influenced by the degree of structural flexibility ofa given idiomatic phrase. It is important to keep two things in mindwith respect to this suggestion, however. First, this conclusionspecifically applies to expectations of the final word of an idiompresented in a strongly biasing sentence context. Perhaps frozen-ness plays a role at earlier points in the idiom and in contexts thatare less constraining. Second, this conclusion does not imply thata full syntactic analysis is necessary in order to retrieve thefigurative meaning of any idiomatic phrase. Perhaps idioms dodiffer in terms of the amount of syntactic analysis that is requiredto retrieve their meanings. Our data suggest, however, that regard-less of possible retrieval differences among idioms, the syntacticprocessor will nevertheless fully parse a given idiomatic phrase.

In summary, we have found a syntactic priming effect foridioms with both visual and auditory presentations of the idiomaticprimes and with word and nonword targets. It appears, therefore,that the processing of idiomatic strings and the processing of literalstrings are similar in terms of the syntactic representations that arederived.

Experiment 4

In the final experiment, we tested for conceptual priming. Thisenables us to examine whether the literal meaning of an idiomaticphrase is computed even in a context biasing the idiom's figurativeinterpretation. As sentence primes, we used idiomatic and literalsentences from the previous experiments that had a concrete nounas the expected completion. The meanings of the literal sentencesrequired that a concrete object serve as the completion. For exam-ple, in the sentence context the soccer player slipped when he triedto kick the, the completion must be literally kickable and so mustbe concrete. This semantic bias might make it more difficult toname an abstract target word (e.g., truth) than an unrelated con-crete target word (e.g., shelf). Of primary interest in this experi-ment was whether figuratively biased sentences would show a

conceptual priming effect. Consider, for example, the sentencecontext the man was old and feeble and it was believed that hewould soon kick the. Although the expected completion (bucket) isa concrete noun, the concreteness of the completion is irrelevant tothe idiom's figurative meaning. Thus, one would not expect to finda conceptual priming effect following sentences such as this oneunless the literal meaning of the idiom were derived along with theidiom's figurative interpretation. The presence of conceptual prim-ing is therefore used in this experiment as an indicator of ongoingliteral computation.

The syntactic dominance model claims that the semantic pro-cessor must obligatorily analyze the literal meaning of an idiom onthe basis of structural input from the syntactic processor. Thismodel predicts, therefore, that conceptual priming should occurwith idiomatic sentences. In contrast, the autonomy model as-sumes that a full literal interpretation of an idiom is not derivedeven though its syntactic structure is computed. Thus, this lattermodel predicts that conceptual priming should not arise withidiomatic sentences.

MethodParticipants. Thirty-six University of Rochester undergraduates

served as participants.Stimuli. The sentence pairs from the previous experiments were used,

with the exception of seven pairs in which either the idiomatic or the literalsentence (or both) ended with an abstract noun. A randomly selected eighthpair was excluded so that there were a total of 24 pairs, thus allowing equalnumbers of pairs to be assigned to each of the four Concreteness XSentence Type conditions (see the Appendix for a listing of the eightexcluded items). We reanalyzed the data from the first three experimentsby using just these 24 pairs in order to ensure that these items would showthe same effects as were found with the entire set of 32 pairs. There wasno change in the pattern of the data when the subset of 24 items was used.The magnitudes of the syntactic priming effects for idiomatic and literalsentences, respectively, were 28 and 21 ms in Experiment 1, 23 and 23 msin Experiment 2, and 39 and 22 ms in Experiment 3. A statistical analysisdemonstrated that all of the effects that were significant with the 32-itemset were also significant with the 24-item subset, with one exception: Theinteraction between sentence type and grammatical category in Experi-ment 3 was not significant with the 24-item subset. Overall, then, syntacticpriming is robust with these 24 items for both idiomatic and literal phrases.However, the obtained greater magnitude of priming for the idiomaticstimuli in Experiment 3 is no longer reliable.

Twenty-four concrete words and 24 abstract words were selected fromToglia and Battig's (1978) norms to serve as targets. The concrete wordshad a mean concreteness rating of 6.10, with no item having a rating of lessthan 5.79. The abstract words had a mean rating of 3.16, with no itemhaving a rating of greater than 3.59. The concrete and abstract target wordswere matched in terms of familiarity (the mean familiarity ratings fromToglia and Battig's norms were 6.20 and 6.18 for the concrete and abstracttargets, respectively), number of syllables (both target types had a meansyllable length of 1.75), and number of letters (mean lengths of 5.50characters for the concrete target words and 5.46 characters for the abstracttarget words). The abstract words had a somewhat higher mean logfrequency (1.73) than the concrete words (1.39). As in previous experi-ments, we collected participants' naming times for each of the 48 totaltarget words presented in isolation. Twenty-four participants were tested,and each participant responded to all 48 targets. The abstract targets werenamed slightly more quickly than the concrete targets (432 and 437 ms,respectively), although this difference was not significant, either withsubjects or with items, Fl(l, 21) = 1.70, MSE = 212, F2(l, 46) < 1 (theproportion of naming errors was .01 for both the abstract and the concrete

1232 PETERSON, BURGESS, DELL, AND EBERHARD

words). This lack of a real difference between naming concrete and abstractwords in isolation is in line with other research showing no concretenesseffect for regularly spelled words (Strain, Patterson, & Seidenberg, 1995).Most of our targets were regular words (see the Appendix). Importantly,the slight difference in naming times that we obtained between the abstractand the concrete words was in a direction opposite the effect examined inExperiment 4. In Experiment 4, the expected completions for the sentencecontexts are concrete nouns; hence, if there is a concreteness effect, theabstract targets should be named more slowly than the concrete targets.

Design and procedure. The procedures for this experiment were iden-tical to those used in Experiments 2 and 3. The stimulus tapes used in theprevious two experiments were used again here, although for this experi-ment there were only 24 critical trials. The additional eight critical trialsfrom the previous experiments served as filler trials.

Results and Discussion

Mean naming latencies for abstract and concrete targets follow-ing figuratively and literally biasing contexts are given in Figure 1.Abstract targets were named more slowly than concrete targets(535 and 525 ms, respectively), although this effect was significantwith subjects only, Fl{\, 32) = 4.17, MSE = 799, F2(l,20) = 1.16, MSE = 1,375. Importantly, the interaction betweenconcreteness and sentence type was significant in both the subjectand the item analyses, Fl(\, 32) = 4.77, MSE = 948, F2(l,20) = 5.64, MSE = 934. For literal sentences, abstract words werenamed 21 ms more slowly than concrete words, F/( l , 32) = 8.45,MSE = 925, F2(l, 20) = 6.66, MSE = 952. However, there wasno effect of concreteness for idiomatic sentences, both Fl andF2 < 1. In fact, for idiomatic sentences, the abstract targets werenamed slightly more quickly than the concrete targets (525 and527 ms, respectively), just as we found for these targets named inisolation.

For figuratively biased contexts, the mean error proportionswere .01 for abstract targets and .02 for concrete targets. Forliterally biased contexts, the proportions were .03 and .02, respec-tively. The pattern of errors is consistent with the reaction timedata, although neither of the main effects nor the interactionapproached statistical significance.

The pattern of data from Experiment 4 suggests, therefore, thatthe literal meaning of an idiomatic phrase is computed only inliterally biased contexts. When the context is biased toward anidiomatic interpretation of the phrase, literal analysis appears to beterminated prior to the final word of the idiom. The results also aresignificant in showing that general semantic expectations can beassessed in a primed naming task.

The results of this experiment stand in sharp contrast to thepattern of results found in the first three experiments. In the priorexperiments, idiomatic sentences showed syntactic priming effectsthat were as large as (if not larger than) those for literal sentences.In this experiment, however, a conceptual priming effect wasfound for literal sentences only. Taken together, these resultssuggest that, by the final word of an idiomatic phrase (presented ina strong figuratively biased context), participants continue to mon-itor the syntactic structure of the idiom but do not compute acorresponding literal interpretation of the phrase. This dissociationbetween syntactic and semantic analyses serves as clear evidencefor the syntactic-semantic autonomy model. According to thismodel, the semantic processor terminates its analysis of the literalmeaning of an idiomatic phrase once the figurative meaning of the

idiom has been retrieved. However, the termination of semanticanalysis does not result in subsequent elimination of syntacticprocessing. Therefore, a full syntactic description of an idiom willbe constructed regardless of contextual constraints and regardlessof the status of ongoing literal processing.

General Discussion

Our primary concern was to explore the relationship betweensyntactic and semantic levels of processing. To do so, we focusedspecifically on the on-line interpretation of highly constrainingphrases. Of particular interest was whether idioms presented infiguratively biased contexts undergo normal syntactic analysis andwhether their literal meanings are obligatorily computed. Theresults from our experiments can be summarized easily. In Exper-iments 1, 2, and 3, the magnitude of syntactic priming found withidiomatic phrases was at least as large as that found with literalphrases. Moreover, the priming found with idiomatic phrases wasnot correlated with our measures of the phrases' syntactic frozen-ness. These data suggest that the comprehension system evaluatesthe structural relationships that hold among the component wordsof a string, even though that structure may not always correspondin a direct way to the string's interpretation. In Experiment 4,conceptual priming was found for literal phrases but not foridiomatic phrases, suggesting that the literal meaning of an idiomis not obligatorily computed.

These results suggest that syntactic processing operates, at leastin part, in isolation from semantic processing. Specifically, wehave shown that, at a point in time when participants have aban-doned their analysis of an idiom's literal meaning, they neverthe-less continue to monitor the syntactic structure of the phrase. So,for example, given the context it was believed that the old manwould soon kick the, it appears that participants do not expect thesentence completion to be a kickable object but do expect thecompletion to be a noun. Thus, expectations appear to be based onthe grammatical class of an upcoming word, rather than on theword's semantic compatibility with the literal context. This patternsupports what we have referred to as the syntactic-semantic au-tonomy model of processing. In this model, literal processing isterminated when a contextually appropriate figurative meaning ofan idiom is retrieved. Such termination, however, does not elim-inate ongoing syntactic analysis. It is this continuation of syntacticprocessing, despite abandonment of semantic processing, thatserves as a testimony to the isolability of syntactic processing (e.g.,Boland, 1997).

Syntactic isolability has been a crucial assumption underlying amodular view of language processing (e.g., Fodor, 1983; Forster,1979). Our data support this specific aspect of the modularityclaim and argue against models in which syntax is inextricablyintertwined with semantic and/or message-level details (Bates &MacWhinney, 1982). As such, our results serve to strengthen theclaim that the language comprehension system is composed of aseries of distinct processing levels, at least one of which is respon-sible for computing the grammatical relationships that hold amongthe words in a sentence (e.g., Forster, 1979). Although this claimhas been the standard one in the psycholinguistics literature, it hasbeen called into question by a number of researchers. Marslen-Wilson and Tyler (1988), for example, argue that there is littledirect evidence for the claim that an independent level of gram-

IDIOM COMPREHENSION 1233

matical structure intervenes between lexical and discourse repre-sentations. The problem with this claim is that it cannot account forthe fact that participants in our studies were sensitive to thegrammatical compatibility of the final word of an idiomatic phrase,even though the semantics of the word were largely irrelevant (weacknowledge that this claim of semantic irrelevance is crucial;hence, we explore this issue below). The simplest way to accountfor our pattern of data is to postulate a level of processing inde-pendent of semantic interpretation.

By supporting claims of syntactic isolability, our study comple-ments others that have sought to dissociate syntactic and semanticexpectations. For example, McElree and Griffith (1995) used aspeed-accuracy trade-off technique to identify different timecourses for syntactic (subcategorization frames) and semantic (the-matic role) characteristics of verbs. Event-related brain potentialsalso have been applied to the question of syntactic isolability andsome studies have found dissociations in brain potentials consis-tent with the syntactic-semantic distinction (e.g., Ainsworth-Darnell, Shulman, & Boland, 1998; Friederici & Frisch, 2000).

It is important to note that although our experiments support oneaspect of the modularity position (i.e., the isolability of syntacticprocessing), they do not test other critical assumptions of modu-larity and thus should not be seen as a strong test of the positiongenerally. For example, a central tenet of modularity (and one thathas probably resulted in the greatest theoretical debate) is whetherlevels of processing are informationally encapsulated (Fodor,1983), such that the operation of a given level (such as syntax) isnot influenced by higher-level processors. Our experiments sup-port only a specific aspect of this position: We show that the parserhas considerable internal momentum; it cannot be disabled whensemantic and/or message-level processors adopt analyses that areat least partially independent of the grammatical structure of theidiomatic phrase. In particular, our results do not rule out interac-tive models that directly represent both syntactic and semanticlevels but allow semantics to impact very quickly on the resolutionof syntactic ambiguity (e.g., Gamsey et al., 1997; MacDonald etal., 1994; Trueswell et al., 1994). The different meanings associ-ated with a phrase such as kick the bucket are not associated withdifferent syntactic structures. Hence, there is no syntactic ambigu-ity to be resolved. Consequently, our studies are best seen asproviding evidence not for modularity per se but for the claim thatthere exists a level of grammatical analysis that can functionindependently of higher-level interpretive processes. In fact, this isexactly the conclusion that has been reached from a series ofstudies on language production based on syntactic repetition ef-fects (e.g., Bock, 1986; Bock & Loebell, 1990; Bock, Loebell, &Morey, 1992) and idiom blends (Cutting & Bock, 1997).

Our evidence for syntactic isolability is based on the assumptionthat the syntactic processing that we observed with idiomaticphrases occurred in the absence of literal semantic processing. Thisassumption would be violated, however, if in fact the syntacticstructure of an idiom played an integral role in the on-line con-struction of the figurative interpretation of the idiom. Recently,Gibbs and colleagues suggested that, for at least some idioms,figurative interpretation may involve such a structural analysis(Gibbs & Nayak, 1989; Gibbs, Nayak, & Cutting, 1989). Becauseof the importance of this claim to our conclusions, we wish toreview it in some detail in order to evaluate its possible implica-tions for our conclusions.

Gibbs and Nayak (1989) showed that idioms vary among them-selves in their degree of compositionality. Some idioms, such askick the bucket, are in fact noncompositional. The words in theidiom do not correspond in any obvious way to parts of the idiom'smeaning. Other idioms, however, appear to have a structure thatbears a closer relation to the figurative interpretation. Nunberg(1978) referred to these types of idioms as normally decompos-able. For example, the idiom pop the question is closely related instructure to its meaning, uttering a marriage proposal. That is, popcorresponds to the making of an utterance, while the question isassociated with the proposal. Further, some idioms, although se-mantically decomposable, have constituents that are related to theidiomatic meaning in a less direct, metaphorical way. For example,in the idiom spill the beans, there is an obvious parallel betweenthe constituent structure of the phrase and the idiom's interpreta-tion, reveal the secret. However, the semantic relation betweenbeans and secret is obviously not a direct one. Nunberg (1978)referred to these types of idioms as abnormally decomposable.Gibbs et al. (1989) argued that the compositionality of an idiominfluences its processing. Specifically, they demonstrated that par-ticipants can judge decomposable idioms (of both normal andabnormal varieties) as being meaningful strings faster than theycan make the same judgment for either literal control phrases ornondecomposable idioms. On the basis of these findings, Gibbsand colleagues suggested that, for decomposable idioms, and pos-sibly some nondecomposable idioms as well (Hamblin & Gibbs,1999), a compositional analysis of the idiom is used to derive anidiomatic interpretation (however, see Titone & Connine, 1999).The implication of this claim for our study is that, at least fordecomposable idioms, syntactic analysis may be tightly bound tosemantic (figurative) interpretation, thus not permitting us to arguefor the isolability of syntax.

Although we do not deny that for some idioms (e.g., decom-posable ones), extensive compositional analysis might be requiredin order to derive the idiom's meaning, we nevertheless suggestthat this claim is tangential to the syntactic isolability issue. Thisissue, as we have addressed it, is based on whether the strength ofsyntactic expectations that arise during the processing of a phraseis independent of the phrase's semantic analysis (and, by exten-sion, of the phrase's semantic compositionality). Thus, we wereconcerned not with the question of whether some idioms requirecompositional analysis but rather with the issue of whether syn-tactic analysis occurs regardless of a given phrase's level ofcompositionality. To explore this specific issue, we comparedliteral phrases that, by definition, are entirely decomposable withidiomatic phrases. As noted previously, idiomatic phrases tend tofall along a continuum of compositionality (Gibbs & Nayak,1989), ranging from those that are largely nondecompositional tothose that approach the compositionality of literal phrases. If,contrary to the isolability claim, the magnitude of syntactic prim-ing is directly a function of the phrase's semantic compositionality,one would expect diminished priming effects for our idioms com-pared to our literal controls (since, on average, our idioms are lessdecomposable than are literal phrases). In contrast to this predic-tion, we found that our idioms showed syntactic priming effects aslarge as or even larger than those for our literal controls. This resultsuggests that syntactic analysis occurs independently of the se-mantic decomposability of a given phrase and thus provides strongevidence for syntactic isolability.

1234 PETERSON, BURGESS, DELL, AND EBERHARD

To provide further support for this claim, we categorized ouridioms as being nondecomposable, abnormally decomposable, ornormally decomposable on the basis of the norming study of Gibbsand Nayak (1989). Ten of our 32 idioms were contained in their setof idioms (2 were categorized as nondecomposable, 5 were cate-gorized as abnormally decomposable, and 3 were categorized asnormally decomposable). Two additional idioms from our exper-iments were classified as being nondecomposable on the basis ofthe norms provided by Titone and Connine (1994b). The remain-der of our idioms did not appear in either the Gibbs-Nayak orTitone-Connine norms. Using these 12 items, we reanalyzed thedata from Experiments 1, 2, and 3 with decomposability as afactor. Averaging across the three experiments, we found syntacticpriming effects of 42, 29, and 36 ms for nondecomposable, ab-normally decomposable, and normally decomposable idioms, re-spectively (priming effects for the literal controls were 22, 26,and 19 ms, respectively). The interaction between priming anddecomposability did not approach significance, both Fl andF2 < 1. Thus, with at least this subset of our items, there is no hintof an effect of decomposability on syntactic priming—priming fornondecomposable idioms appears to be at least as large as that fordecomposable ones. Note that the magnitude of priming seen withthis 12-item subset is comparable to the priming effects observedwith all of our materials in Experiments 1 through 3, in which theaverage syntactic priming effects were 34 and 19 ms for idiomaticand literal phrases, respectively. These data serve as additionalsupport for the claim that syntactic expectations can be derivedindependently of semantic concerns and thus add further evidencein favor of the syntax-semantics autonomy model.

It is important to note that the autonomy claim does not implythat the semantic compositionality of idioms is irrelevant to theway in which they are processed. It might well be the case that astructural analysis contributes to the meaning of a decomposableidiom, whereas the meaning of a nondecomposable idiom might bederived independently of such an analysis. Our experiments werenot designed to address this specific issue, and so our data cannotbe applied directly to it. Gibbs et al. (1989), however, haveprovided convincing evidence for such a processing difference.Interestingly, one effect that was observed in Gibbs et al.'s (1989)study is particularly compatible with the claim of syntactic iso-lability. Specifically, they found that participants took longer tojudge nondecomposable idioms as being meaningful than they didto judge literal control phrases, or decomposable idioms. To ac-count for this finding, Gibbs et al. (1989) assumed that all idioms(even nondecomposable ones) undergo compositional analysis.For a nondecomposable idiom, then, such an analysis interfereswith the evaluation of the meaningfulness of the idiom, since theanalysis is not compatible with the phrase's figurative meaning.Our data provide some support for this claim and, further, providea framework within which to interpret the interference effect.Specifically, we provide direct evidence that idiomatic phrases(including frozen and/or nondecompositional ones) do in factundergo full syntactic analysis. The interference observed fornondecomposable idioms, then, might arise when the system at-tempts to match the syntactic structure of an idiom with a repre-sentation of its (figurative) meaning (see, e.g., Boland, 1997, for anaccount of syntactic-semantic matching for nonidiomatic phrases).The mismatch between these levels of representation might delaythe decision that the phrase is in fact an acceptable one. Further,

our results make it clear that syntactic analysis of idiomaticphrases is not necessarily tied to construction of a literal interpre-tation of an idiom. Specifically, with our figuratively biased sen-tences, we observed syntactic priming effects in the absence of(literal) conceptual priming.

Our results also have important implications for specific modelsof idiom processing. The advantage of using an on-line primingprocedure is that it can provide a relatively detailed analysis ofdifferent levels of processing during comprehension (see alsoCacciari & Tabossi, 1988; Titone & Connine, 1994a). As a result,our data are able to extend the level of theoretical detail incorpo-rated in models of idiom comprehension. Our results are probablymost compatible with a model of on-line processing outlined byCacciari and Tabossi (1988) and extended by Titone and Connine(1994a). These researchers suggested that an idiom is mentallyrepresented as a configuration of lexical items, in which a singleword (or perhaps a subset of words), the key, is more relevant thanthe others for recognizing the idiom. During on-line comprehen-sion, the initial portion of the idiom is interpreted literally, until thekey itself is processed. At that point, the idiom is recognized, andits figurative interpretation is made available. Assuming that thefigurative meaning is compatible with the context and the literalmeaning is not, the remainder of the idiom may not undergofurther literal analysis. The termination of literal processing seemsparticularly compatible with the results that we observed with ourfiguratively biased contexts. Our data, however, allow a refine-ment of this position. Specifically, it appears that the abandonmentof literal processing is specific to semantic analysis: Syntacticprocessing appears to persist in a normal fashion. Despite thisevidence, however, it is clear that further research is needed todetermine the precise role that syntactic analysis plays in idiomcomprehension.

Given the success that we and others (Cacciari & Tabossi, 1988;Titone & Connine, 1994a) have had with the priming procedure, itseems likely that such an on-line approach will be a fruitful one(and perhaps even a necessary one) for exploring increasinglydetailed aspects of syntactic and semantic processing of both literaland idiomatic phrases. It is worth noting that our version of thepriming procedure is somewhat different from that used in otheridiom studies and may in fact be more suitable for exploring therelationship between literal processing and figurative processing.Specifically, other priming studies have tapped literal processingby testing for priming for target words that are semantically relatedto a specific word contained in the idiomatic phrase (e.g., kick thebucket priming PAIL). The problem with such an approach is thatthe presence of this kind of lexical priming does not necessarilyimply that the literal meaning of the phrase as a whole is beingconstructed (see Gibbs et al., 1989, and Titone & Connine, 1994a,for a discussion of this limitation of the lexical priming procedure).In contrast, we explored priming for target words presented ascompletions for strongly biased sentence fragments. The targetwords themselves were not the expected completions for thefragments but rather were unexpected continuations that eitherviolated or were consistent with the syntactic and (literal) semanticstructures set up by the context. Priming therefore presumablyreflects the extent to which the target can (or cannot) be legiti-mately integrated into the phrasal structure built to that point(O'Seaghdha, 1997; West & Stanovich, 1986). Thus, our task tapsphrase-level analyses rather than simple lexical retrieval processes.

IDIOM COMPREHENSION 1235

It is exactly this level of analysis that is at issue in the idiom-processing literature; thus, our version of the priming task is, inthis sense, a particularly appropriate one.

In summary, we have presented evidence that syntactic andsemantic levels of processing can operate in relative independencefrom one another, at least under the conditions of our experiments,which involved highly constraining contexts. To separate theselevels, we focused on the processing of idiomatic phrases, in whichthe syntactic-semantic mapping is often indirect. Our data clearlyshowed that, in figuratively biased contexts, idioms undergo com-plete syntactic analysis, regardless of their syntactic productivity(i.e., frozenness) and even when their literal meaning is not beingconstructed. Our success at isolating syntactic and semantic levelsof representation suggests that studies of idiom processing canprovide important insights into the operation of the languagecomprehension system.

References

Ainsworth-Darnell, K., Shulman, H. G., & Boland, J. E. (1998). Dissoci-ating brain responses to syntactic and semantic anomalies: Evidencefrom event-related potentials. Journal of Memory and Language, 38,112-130.

Altmann, G., & Steedman, M. (1988). Interaction with context duringhuman sentence processing. Cognition, 30, 191-238.

Bates, E., & MacWhinney, B. (1982). Functionalist approaches to gram-mar. In E. Wanner & L. R. Gleitman (Eds.), Language acquisition: Thestate of the art (pp. 173-218). Cambridge, England: Cambridge Univer-sity Press.

Boatner, M., Gates, J., & Makkai, A. (1975). A dictionary of Americanidioms. Hauppauge, NY: Barren's Educational Series.

Bock, J. K. (1986). Syntactic persistence in language production. CognitivePsychology, 18, 355-387.

Bock, J. K., & Loebell, H. (1990). Framing sentences. Cognition, 35, 1-39.Bock, J. K., Loebell, H., & Morey, R. (1992). From conceptual roles to

structural relations: Bridging the syntactic cleft. Psychological Re-view, 99, 150-171.

Boland, J. E. (1993). The role of verb argument structure in sentenceprocessing: Distinguishing between syntactic and semantic effects. Jour-nal of Psycholinguistic Research, 22, 109-132.

Boland, J. E. (1997). The relationship between syntactic and semanticprocesses in sentence comprehension. Language and Cognitive Pro-cesses, 12, 423-484.

Burgess, C , & Lund, K. (1994). Multiple constraints in syntactic ambigu-ity resolution: A connectionist account of psycholinguistic data. Pro-ceedings of the Cognitive Science Society (pp. 90—95). Hillsdale, NJ:Erlbaum.

Cacciari, C , & Tabossi, P. (1988). The comprehension of idioms. Journalof Memory and Language, 27, 668-683.

Carello, C , Lukatela, G., & Turvey, M. T. (1988). Rapid naming isaffected by association but not by syntax. Memory & Cognition, 16,187-195.

Crain, S., & Steedman, M. (1985). On not being led up the garden path:The use of context by the psychological parser. In D. Dowty, L. Kar-tunnen, & A. M. Zwickey (Eds.), Natural language parsing (pp. 320-358). Cambridge, England: Cambridge University Press.

Cutting, J. C , & Bock, J. K. (1997). That's the way the cookie bounces:Syntactic and semantic components of experimentally elicited idiomblends. Memory & Cognition, 25, 57-71.

Estill, R., & Kemper, S. (1982). Interpreting idioms. Journal of Psycho-linguistic Research, 6, 559-568.

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

Fodor, J. A. (1983). The modularity of mind. Cambridge, MA: MTT Press.Forster, K. I. (1979). Levels of processing and the structure of the language

processor. In W. E. Cooper & E. C. T. Walker (Eds.), Sentence pro-cessing: Psycholinguistic studies presented in Merrill Garrett (pp. 2 6 -85). Hillsdale, NJ: Erlbaum.

Fraser, B. (1970). Idioms within a transformational grammar. Foundationsof Language, 6, 22-42.

Frazier, L. (1985). Modularity in the representation of hypotheses. Pro-ceedings of the Northeastern Linguistics Society, 131-144.

Frazier, L. (1987). Theories of syntactic processing. In J. L. Garfield (Ed.),Modularity in knowledge representation and natural language process-ing (pp. 291-307). Cambridge, MA: MIT Press.

Frazier, L., & Clifton, C. E., Jr. (1996). Construal. Cambridge, MA: MITPress.

Frazier, L., Clifton, C , & Randall, J. (1983). Filling gaps: Decisionprinciples and structure in sentence comprehension. Cognition, 13, 187-222.

Frazier, L., & Rayner, K. (1982). Making and correcting errors duringsentence comprehension: Eye movements in the analysis of structurallyambiguous sentences. Cognitive Psychology, 14, 178-210.

Friederici, A. D., & Frisch, S. (2000). Verb argument structure processing:The role of verb-specific and argument-specific information. Journal ofMemory and Language, 43, 476-507.

Garnsey, S. M., Pearlmutter, N. J., Myers, E., & Lotocky, M. A. (1997).The contributions of verb bias and plausibility to the comprehension oftemporarily ambiguous sentences. Journal of Memory & Language, 37,58-93.

Gibbs, R. (1986). Skating on thin ice: Literal meaning and understandingidioms in conversation. Discourse Processes, 9, 17-30.

Gibbs, R., & Gonzales, G. (1985). Syntactic frozenness in processing andremembering idioms. Cognition, 20, 243-259.

Gibbs, R., & Nayak, N. P. (1989). Psycholinguistic studies on the syntacticbehavior of idioms. Cognitive Psychology, 21, 100-138.

Gibbs, R., Nayak, N. P., & Cutting, C. (1989). How to kick the bucket andnot decompose: Analyzability and idiom processing. Journal of Memoryand Language, 28, 576-593.

Hamblin, J. L., Gibbs, R. W., Jr. (1999). Why you can't kick the bucket asyou slowly die: Verbs in idiom comprehension. Journal of Psycholin-guistic Research, 28, 25-39.

Kufera, H., & Francis, W. (1967). Computational analysis of present-dayAmerican English. Providence, RI: Brown University Press.

MacDonakl, M. C , Pearlmutter, N. J., & Seidenberg, M. S. (1994). Thelexical nature of syntactic ambiguity resolution. Psychological Review,101, 676-703.

Marslen-Wilson, W., & Tyler, L. K. (1988). Against modularity. In J.Garfield (Ed.), Modularity in knowledge representation and naturallanguage processing (pp. 37-62). Cambridge, MA: Bradford Books.

McElree, B., & Griffith, T. (1995). Syntactic and thematic processing insentence comprehension. Journal of Experimental Psychology: Learn-ing, Memory, and Cognition, 21, 134—157.

Nunberg, G. (1978). The pragmatics of reference. Bloomington, IN: Indi-ana University Linguistics Club.

Ortony, A., Shallert, D., Reynolds, R., & Antos, S. (1978). Interpretingmetaphors and idioms: Some effects of context on comprehension.Journal of Verbal Learning and Verbal Behavior, 17, 465-477.

O'Seaghdha, P. G. (1989). The dependence of lexical relatedness effects onsyntactic connectedness. Journal of Experimental Psychology: Learn-ing, Memory, and Cognition, 15, 73-78.

O'Seaghdha, P. G. (1997). Conjoint and dissociable effects of syntactic andsemantic context. Journal of Experimental Psychology: Learning, Mem-ory, and Cognition, 23, 807-828.

Pollatsek, A., & Well, A. D. (1995). On the use of counterbalanced designsin cognitive research: A suggestion for a better and more powerful

1236 PETERSON, BURGESS, DELL, AND EBERHARD

analysis. Journal of Experimental Psychology: Learning, Memory, andCognition, 21, 785-794.

Rayner, K., Carlson, M., & Frazier, L. (1983). The interaction of syntaxand semantics during sentence processing: Eye movements in the anal-ysis of semantically biased sentences. Journal of Verbal Learning andVerbal Behavior, 22, 358-374.

Reagan, R. T. (1987). The syntax of English idioms: Can the dog be puton? Journal of Psycholinguistic Research, 16, 417-441.

Schank, R. C. (1984). The cognitive computer. Reading, MA: Addison-Wesley.

Seidenberg, M. S., Waters, G. S., Sanders, M, & Langer, P. (1984). Pre-and post-lexical loci of contextual effects on word recognition. Memory& Cognition, 12, 315-328.

Strain, E., Patterson, K., & Seidenberg, M. S. (1995). Semantic effects insingle-word naming. Journal of Experimental Psychology: Learning,Memory, and Cognition, 21, 1140-1154.

Swinney, D., & Cutler, A. (1979). The access and processing of idiomaticexpressions. Journal of Verbal Learning and Verbal Behavior, 18,523-534.

Tanenhaus, M. K., Carlson, G. N., & Trueswell, J. C. (1989). The role ofthematic structures in interpretation and parsing. Language and Cogni-tive Processes, 4, 211-234.

Taraban, R., & McClelland, J. L. (1988). Constituent attachment and

thematic role assignment in sentence processing: Influences of content-based expectations. Journal of Memory and Language, 27, 597-632.

Titone, D. A., & Connine, C. M. (1994a). Comprehension of idiomaticexpressions: Effects of predictability and literality. Journal of Experi-mental Psychology: Learning, Memory, and Cognition, 20, 1126-1138.

Titone, D. A., & Connine, C. M. (1994b). Descriptive norms for 171idiomatic expressions: Familiarity, compositionality, predictability, andliterality. Metaphor and Symbolic Activity, 9, 247-270.

Titone, D. A., & Connine, C. M. (1999). On the compositional andnoncompositional nature of idiomatic expressions. Journal of Pragmat-ics, 31, 1655-1674.

Toglia, M. P., & Battig, W. F. (1978). Handbook of semantic word norms.Hillsdale, NJ: Erlbaum.

Trueswell, J. C , & Tanenhaus, M. K. (1994). Toward a lexical frameworkof constraint-based syntactic ambiguity resolution. In C. E. Clifton, L.Frazier, & K. Rayner (Eds.), Perspectives on sentence processing (pp.155-179). Hillsdale, NJ: Erlbaum.

Trueswell, J. C , Tanenhaus, M. K., & Garnsey, S. M. (1994). Semanticinfluences on parsing: Use of thematic role information in syntacticambiguity resolution. Journal of Memory and Language, 33, 285-318.

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

Wright, B., & Garrett, M. (1984). Lexical decision in sentences: Effects ofsyntactic structure. Memory & Cognition, 12, 31-45.

Appendix

Sentence Pairs

Thirty-two figurative-literal sentence pairs were used in the experi-ments. For each item, the first sentence is the figuratively biased sentence,and the second sentence is literally biased. The third entry for each itemlists (a) the idiom's frozenness scores, (b) the verb and noun targets usedin Experiments 1 and 2, (c) the "verb" and "noun" nonwords used inExperiment 3, and (d) the abstract and concrete noun targets used inExperiment 4 (note that for items not used in Experiment 4, asterisks areshown in place of the target pair).

1. The man was upset when his wife got a job because he believed that intheir family, he should be the one to wear the pants.

At graduation, the boy was in a big hurry and forgot his cap, but at leasthe did not forget to wear the gown.

-0.19; SEND-MILE BLIDDED-BLIDDER HONESTY-BANANA2. The girl continued to shoplift even though her friends had told her that

she was playing with fire.After having several talks with his son about the dangers of fire, the

father was upset when he found him playing with matches.0.38; SIT-HOUR SURDED-SURDER *****

3. They were late for the party so the man told his wife to hurry up andshake a leg.

The enthusiastic new father leaned over the crib and tried to get theinfant to shake a rattle.

1.75; SING-SONG VIRKED-VIRKER MEMORY-CAMERA4. The business was running quite smoothly so the manager warned the

disgruntled secretary not to rock the boat.The only way the mother could get her baby to go to sleep was to lay

him down and gently rock the cradle.-0.06; LEND-SHIRT FLASTED-FLASTER GUILT-PEARL

5. It was only midseason, but the rookie pitcher already had 12 wins underhis belt.

The old man didn't believe in banks so he kept his money under hismattress.

1.69; ASK-CHILD CLEPPED-CLEPPER TREND-SNAKE6. John said he could beat anyone at tennis so Bill was determined to make

him eat his words.It was noon, but the construction worker was so busy that he couldn't

take time out to eat his lunch.1.57; TEND-YARD DARFED-DARFER *****

7. After living together for several years, John and Mary finally decided totie the knot.

The man wrapped the birthday present but when he got to the ribbon, hedidn't know how to tie the bow.

0.69; SPEAK-DEATH DITTED-DITTER FREEDOM-TRACTOR8. The student was failing all of his classes because he never bothered to

crack a book.The hiker tried to be very quiet as he watched a squirrel crack a nut.1.63; WEEP-HORN YATTED-YATTER DUTY-CANDY

9. The pompous orator spoke about the ills of society until he was blue inthe face.

The weather forecaster announced that tomorrow there would be fewerclouds and you would see more blue in the sky.

1.00; HIRE-GATE PREEDED-PREEDER GRIEF-BRICK10. Before John arrived at the party, nobody was socializing, but John's

quick wit soon broke the ice.The boy threw a rock at the house and broke the window.-0.69; HANG-TRUTH DORTED-DORTER WISDOM-PENCIL

IDIOM COMPREHENSION 1237

11. Bill and John had been enemies for several years, but they finallydecided to bury the hatchet.

After devouring the steak, the dog trotted out to the backyard to burythe bone.

1.24; SWEAR-CHEEK TECKED-TECKER BELIEF-CEILING12. The mother knew that a birthday party for twenty toddlers would soon

have her climbing the walls.Fearing that her son might fall, the mother went outside and told the

boy to quit climbing the tree.0.63; STIR-CORN BISHED-BISHER METHOD-COTTON

13. With so many young girls imitating her, Madonna look-alikes are adime a dozen.

The old man told the waitress about the days when coffee was just adime a cup.

1.56; PRAY-BUNK BAPPED-BAPPER *****14. Jane thought the man looked familiar, but when she tried to think of his

name she drew a blank.The masked man boldly approached the bank teller and quickly drew

a gun.0.81; CHOOSE-SOURCE JEDDED-JEDDER *****

15. Reformed alcoholics constantly fear that they might someday fall offthe wagon.

At the playground, the babysitter didn't want to push little Jenny toohigh for fear that she might fall off the swing.

1.12; COPE-AUNT MORPED-MORPER GLORY-EAGLE16. Following her husband's successful surgery, the woman spoke to the

doctor and thanked him from the bottom of her heart.The mother told the girl that she couldn't come into the house until she

scraped the mud from the bottom of her shoes.0.25; CLING-TANK PLICKED-PLICKER FACT-BELL

17. The taxi driver had yet another accident, so the boss had no choice butto give him the ax.

The robber ordered the clerk to open the cash register and give him themoney.

3.00; TEACH-HAIR FLIRKED-FLIRKER RUMOR-BASKET18. The boy was angry at his sister and finally told her to go fly a kite.

The pilot had lost his nerve and didn't know if he would ever again beable to go fly a plane.

0.69; SHUT-JOY ZILLED-ZILLER MANNER-JACKET19. The boy kept talking during class so the teacher finally told him he had

better hold his tongue.The smell of the stockyards was so bad that, when John walked by, he

had to hold his nose.1.32; WIPE-BEER FOUTED-FOUTER YEAR-GRASS

20. At the campaign headquarters, everyone was busy making phone callsexcept the lazy man who wouldn't lift a finger.

The bellboy had to quit his job at the hotel because he hurt his backwhen he went to lift a suitcase.

1.93; BLESS-BLADE JOADED-JOADER MOMENT-CANNON21. In the emergency, the woman panicked and lost her head.

At the door, the woman fumbled in her purse and finally realized thatshe must have lost her keys.

1.87; OWE-MUD VIDDED-VIDDER FATE-TENT22. The politician agreed to talk to the reporter about the scandal, but only

if it was completely off the record.After dinner, the mother told her daughter to clear the dishes off the

table.1.14; GREET-NERVE KESSED-KESSER *****

23. Hank asked Steve if he wanted to go to the fights, but Steve declined,saying that boxing just wasn't his cup of tea.

The cake was much too sweet because the girl had put in an extra cupof sugar.

2.37; SPEND-PEACE PESHED-PESHER FANTASY-PIANO24. The man was very old and feeble and it was believed that he would

soon kick the bucket.The soccer player slipped when he tried to kick the ball.1.68; GROW-TOWN GLATTED-GLATTER TRUTH-SHELF

25. Mr. and Mrs. Smith considered a separation because their marriagewas on the rocks.

Just before the rodeo was to begin, the cowboy put the saddle on thehorse.

1.41; BRING-NIGHT KLAITED-KLArrER AMBITION-TOBACCO26. John was very grumpy this morning and must have gotten up on the

wrong side of the bed.When the Englishman first arrived in the States, he would often get

confused and drive on the wrong side of the road.0.18; DINE-POND NARTED-NARTER LOSS-DOOR

27. Everyone avoided arranging the office Christmas party and when Joewas asked to do it, he also tried to pass the buck.

When Joe was served his dinner, he had the pepper, but he had to askMary to pass the salt.

1.56; LEARN-WIFE TIBBED-TIBBER *****28. Curt was madly in love with Kristine and decided that tonight was the

night he would pop the question.The bottle of champagne slipped out of Cult's hand when he tried to

pop the cork.-0 .31; SEEK-FLOOR ZELDED-ZELDER *****

29. When the 24-year-old man began to date the 16-year-old girl, everyonesaid he was robbing the cradle.

Bonnie and Clyde were sentenced to 10 years in the state penitentiaryfor robbing the bank.

2.06; FAIL-HALL GRASHED-GRASHER JOY-DESK30. When she found out that her husband knew about his surprise party,

the wife demanded to know who had spilled the beans.While the waiter went to get the cream and sugar, the customer

accidentally spilled the coffee.1.37; LOSE-ART SHORKED-SHORKER THEME-TONGUE

31. After the worker got his raise, he thought about asking for a vacation,but he decided not to press his luck.

The man had a job interview, so he went to the drycleaners and askedthem to clean and press his suit.

1.75; WARN-CARD BRIFFED-BRIFFER *****32. His business was nearly bankrupt and Henry knew that his dreams

were quickly going down the drain.The elevator in the old building was out of order so the woman had to

go down the stairs.1.37; PROVE-EARTH CRANDED-CRANDER PASSION-ENGINE

Received November 8, 2000Revision received January 22, 2001

Accepted January 25, 2001