phrasal ordering constraints in sentence production...

Phrasal Ordering Constraints in Sentence Production: PhraseLength and Verb Disposition in Heavy-NP Shift

Lynne M. Stallings and Maryellen C. MacDonald

University of Southern California

and

Padraig G. O’Seaghdha

Lehigh University

Heavy-NP shift is the tendency for speakers to place long or “heavy” noun phrase direct objectsat the end of a sentence rather than in the canonical postverbal position. Three experiments usingseveral task variations confirmed that length of the noun phrase influenced the ordering of the nounphrase and prepositional phrase during production. We also found that heavy-NP shift was stronglyconstrained by the “shifting disposition” of individual verbs. Verbs that do not require theircomplements (e.g., sentential complements) to appear in an adjacent position yielded more shiftingduring production than did verbs that more frequently appear adjacent to their complements. Analysesof decision/preparation times suggested that shifted and unshifted structures competed for selection.These findings point to the simultaneous activation of lexically derived syntactic representations andordering options in sentence planning. A multiple constraints framework provides a means ofreconciling the existence of competition among ordering options with incremental sentenceconstruction. © 1998 Academic Press

In the novelAnimal Farm, George Orwellgave this description of the activities of the pigSnowball during the animal takeover of thefarm: “Snowball had found in the harness-rooman old green tablecloth of Mrs. Jones’s” (Or-well, 1946, p. 38). This sentence is an exampleof the phenomenon that Kimball (1973) termed“heavy-NP shift,” in which a long or “heavy”direct object noun phrase (NP), such asan old

green tablecloth of Mrs. Jones’s,appears inclause-final position, separated from the verb bysome intervening material, in this case the prep-ositional phrase (PP)in the harness-room.“Shifted” sentences such as Orwell’s contrastwith “unshifted” or “basic order” sentences, inwhich the verb and NP direct object are adja-cent, as inSnowball had found an old greentablecloth of Mrs. Jones’s in the harness-room.Most speakers of English find both the shiftedand basic order structures to be acceptable whenthe direct object NP is long, as it is in Orwell’ssentence. When the direct object NP is short,however, speakers have a strong preference forthe basic order (e.g.,found a tablecloth in theharness-room), and shifted structures (e.g.,found in the harness-room a tablecloth) aretypically judged to be very awkward or ungram-matical. Whereas Ross (1967) pointed to NPcomplexity and Kimball (1973) pointed to NPlength as the key factor in speakers’ preferencesfor the alternative structures, Hawkins’ (1994)analysis of a small text corpus suggested that

This research was supported by NSF Grants DBS-9120415and SBR95-11270 to the second author. We thank thereviewers for helpful comments on previous versions of thismanuscript, and Devina Acharya, Jerry Cortrite, Jared Lay-port, Miranda Lim, Sally Paine, Mariana Sapera, and DanielStallings for assistance in data collection and transcriptcoding. We are also grateful to Jack Hawkins and TomWasow for very helpful discussions of corpus research andheavy-NP shift, and to Hiroyuki Oshita for discussions ofheavy-NP shift in Japanese.

Address correspondence and reprint requests to MaryellenMacDonald, Hedco Neuroscience Building, University ofSouthern California, Los Angeles, CA 90089-2520. [email protected]. The stimulus materials for the exper-iments reported here can also be downloaded from http://siva.usc.edu/coglab/papers.html.

3920749-596X/98 $25.00Copyright © 1998 by Academic PressAll rights of reproduction in any form reserved.

JOURNAL OF MEMORY AND LANGUAGE 39, 392–417 (1998)ARTICLE NO. ML982586

the relative length of the NP and the othermaterial in the verb phrase (such as a PP) was abetter predictor of structure choice than proper-ties of the NP alone. Hawkins found that, inde-pendent of NP length and complexity, NPs didnot shift much until they exceeded the length ofthe other material by at least four words.

The heavy-NP shift phenomenon poses inter-esting problems for researchers in language pro-duction. If speakers choose shifted vs basic or-der structures based on the relative length of thephrases in the verb phrase (Hawkins, 1994),then they would appear to have rather detailedquantitative information concerning the phrasalcontents of to-be-uttered sentences during thestages of production in which syntactic struc-tures are chosen. Investigations of heavy-NPshift during on-line speech production couldtherefore be very revealing about the nature andtime course of production operations. This pa-per presents three experiments exploring sev-eral factors that may constrain the use ofheavy-NP shift during production.

Why Shift?

A wide variety of hypotheses in the linguisticand psycholinguistic literature have been of-fered concerning the reasons why heavy-NPshift should exist. A number of researchers haveassumed that shifting is triggered by a syntacticproperty of the long NP, such as its length orrelative length in number of words (Kimball,1973; Hawkins, 1994), or its syntactic complex-ity (Ross, 1967). Alternatively, shifting may bemotivated by syntactic–prosodic interactionrather than by strictly syntactic processes. Zecand Inkelas (1990) suggest that heavy NP shiftis an instance of a class of phenomena in whichreordering of constituents yields better prosodiccontours. Other theorists have argued that shift-ing is not a syntactic phenomenon at all. Forexample, Firbas (1966) suggested that long NPsare shifted to the end of the sentence becausethey tend to contain more new information thando shorter NPs, and thus shifting satisfies prag-matic constraints favoring the given–new order-ing of information in English.

A related controversy concerns whethermovement phenomena such as heavy-NP shift

are motivated by accommodation to the needsof the hearer or by constraints on speaker per-formance (Wasow, 1997). Miller and Chomsky(1963), Bever (1970), Kimball (1973), and Fra-zier (1985) have formulated different versionsof the idea that it is difficult for comprehendersto process long and/or complex constituents inthe middle of a sentence. Shifting is thus hy-pothesized to accommodate the needs of thecomprehender—it moves the difficult materialto the end of the sentence, where it is easier toprocess. Hawkins (1990, 1994) has proposed amore elaborate processing efficiency account inwhich speakers order constituents in verbphrases to maximize the proximity of the headsof all constituents to the verb.

Other evidence favors the view that shifting is aproduct of production constraints. Arnold, Wa-sow, Losongco, and Ginstrom (1997) found thatconstituent ordering was influenced by productiondifficulty, as measured by speech disfluencies,suggesting that shifting is, at least in part, a strat-egy invoked when the production process is par-ticularly difficult or the shifted item is less acces-sible. Of course, it is possible to develop a hearer-centered interpretation for these data (one shifts inorder to delay the error-prone segments for thehearer), and in general it is difficult to distinguishclearly between inherent performance constraintsand strategies induced by concern for the hearer.One solution is to propose that listener accommo-dation is really a by-product of a speaker’s com-prehension of her own speech (Levelt, 1989). An-other is to argue that speakers do not take listenerneeds into account during the initial planning of anutterance (Brown & Dell, 1987; Horton & Keysar,1996; V. Ferreira & Dell, 1996). To the extent thatheavy-NP shift is motivated by altruistic concernfor the hearer per se, it will be less revealing offundamental sentence construction processes (seeBock, 1990). Therefore, our bias is to assign NPshift to these fundamental processes, and to con-sider listener accommodation only if this accountfails.

Sentence Production

Much research on production has been con-ducted within the framework of models thatidentify three major stages in sentence produc-

393PHRASAL ORDERING CONSTRAINTS IN PRODUCTION

tion: conception, formulation, and articulation.In addition, there are two distinct substages insentence formulation (Bock & Levelt, 1994;Fromkin, 1971; Garrett, 1975, 1976, 1984,1988; Levelt, 1989). The first substage is com-prised of functional processes,which includethe mapping of conceptual representations ontogrammatical roles such as subject, predicate,and object. These processes entail the retrievalof major content words such as nouns and verbs(Bock, 1987b, 1990; Bock, Loebell, & Morey,1992; Levelt, 1989). The second substage com-prises positional processeswhich associategrammatical roles with particular syntacticstructures preparatory to phonological encodingand articulation. Heavy-NP shift is not directlyaddressed in these models but, as we shall see,it makes sense to conceive of it as a late func-tional or early positional process. The distinc-tion between functional and positional pro-cesses is largely based on analysis of speecherrors. In general, errors involving whole wordand phrase exchanges are ascribed to the func-tional substage: They arise as a result of misas-signing entities such as nouns and noun phrasesto equivalent syntactic roles. For example, twonouns are exchanged in the sentenceI left thebriefcase in my cigar(Garrett, 1980), and twoNPs are exchanged in the sentenceI got into thisguy with a discussion(Garrett, 1980).

In the models of language production, theorder of major syntactic constituents may oftenbe determined in the mapping from conceptualto functional representation in an incrementalfashion (Bock, 1982, 1987b; Bock & Levelt,1994; De Smedt, 1990, 1994; Kempen & Hoen-kamp, 1987; Levelt 1989), such that words thatare more accessible during the utterance-plan-ning stage by virtue of their semantics, fre-quency, or some other factor, will be the first tobe incorporated into the speech plan and willtherefore tend to appear earlier in the sentence.De Smedt (1994) has applied this approach toordering phenomena in general, hypothesizingthat shorter phrases as well as more conceptu-ally accessible ones tend to become availablefor sequencing before longer phrases during theproduction process. Applied to heavy-NP shift,longer NPs should tend to appear after more

readily accessible PPs. De Smedt’s approachsuggests that accessibility is affected by thenumber of words in the phrases, but an incre-mental accessibility account could also be de-veloped for some of the other approaches dis-cussed above. For example, more complexprosodic or syntactic structures may be lessaccessible than simpler structures, and new in-formation may be less accessible than old infor-mation. However, Wasow’s (1997) English cor-pus analyses revealed numerous examples ofheavy-NP shift that are not predicted by some orall of the accounts sketched here, in that someshifted NPs were not longer than the PP, werenot syntactically complex, were not phonologi-cally complex, and did not appear to containnew information. Wasow concluded that shift-ing is not triggered by any single factor, sug-gesting to us that a multifactor theory may berequired.

Studies of Word Order

To our knowledge, there have been no exper-imental studies of heavy-NP shift, so that themost relevant production research is providedby studies that have examined the ordering ofwords rather than whole phrases (e.g., Bock,1986, 1987a; Bock & Warren, 1985; Kelly,Bock, & Keil, 1986). For example, McDonald,Bock, and Kelly (1993) investigated both se-mantic and phonological influences on wordorder in several constructions. They manipu-lated the animacy, length, and stress pattern ofthe nouns in active and passive sentences suchasA farmer purchased a refrigerator/A refrig-erator was purchased by a farmer,in conjunc-tive phrases within sentences,The key and themanager were nowhere to be found/The man-ager and the key were nowhere to be found,aswell as in isolated phrases such asmanager andkey/key and manager.They examined the orderof the two nouns in recall of the sentences orphrases and found robust effects of animacy,especially in the active/passive sentences wherethe choice of phrase order affected the assign-ment of grammatical roles. That is, participantstended to place animate before inanimate nounsin recall. However, animacy had little effect onthe ordering of nouns within a conjunctive

394 STALLINGS, MACDONALD, AND O’SEAGHDHA

phrase when the phrase was part of a sentence.McDonald et al. concluded that animacy hadlittle or no effect on word order choices that didnot entail grammatical role assignment.

In contrast to animacy, the length and stresspatterns of the nouns had very little effect onthe order of recall in any of the conditions, butMcDonald et al. did find some sensitivity tolength effects in an acceptability rating task.For example, participants judged conjoined NPsto be more acceptable when a shorter nounpreceded a longer noun (e.g.,book and refrig-erator) than when this order was reversed, con-sistent with other length effects found in accept-ability judgments (Cooper & Ross, 1975; Pinker& Birdsong, 1979). In production tasks, how-ever, semantic effects such as animacy appearto play a role in the choice of word order,especially if order entails assignment of gram-matical roles, but there seems to be only a veryrestricted role for phonological effects such asword length.

Though word and phrase ordering are notnecessarily guided by the same productionmechanisms, these studies of word orderingyield several implications for heavy-NP shift.First, there are clear effects of noun accessibil-ity, including accessibility modulated by ani-macy, on word and phrase ordering choices thatentail grammatical role assignment, but not onword ordering choices that do not affect roleassignment. As heavy-NP shift is the rare caseof phrase ordering that does not affect grammat-ical role assignment (i.e., the direct object NPretains this role whether shifted or not), wepredicted that shifting would not be sensitive toanimacy effects. Second, word length effectsare weak or nonexistent in studies of word or-dering. The evidence that word length has littleeffect on word order in sentences and phrasesseems, on the face of it, to be inconsistent withheavy-NP shift, which by definition involves asubstantial role for phrase length or relativephrase length in the choice of shifted vs basicsyntactic structure (Hawkins, 1994; Kimball,1973; Wasow, 1997). However, the processesinvolved in weighing the “heaviness” of phrasesneed not include an assessment of the length ofindividual words. In their proposed prosodic

account of NP shift, Zec and Inkelas (1990)stipulate that full phonological information isnot available to the decision-making process.Likewise, Sternberg, Monsell, Knoll, andWright (1978; see also Wright, 1990; F. Fer-reira, 1991) found that the number of words ina speech plan, rather than the length of thewords, determined production latency, andWheeldon and Lahiri (1997) showed that initi-ation time for short prepared sentences of fixedlength is determined by number of phonologicalwords (see Nespor & Vogel, 1986). Thus, theordering of phrases within a sentence may besensitive to phrase length defined in word orphonological word units, even though the orderof words within phrases is not sensitive to thelength of those units. Clearly, there is a need forproduction data on phrase ordering processes inheavy-NP shift to illuminate the relation of thiscase to previous phrase length and word orderstudies.

Verb Disposition

Our goal is not merely to capture heavy-NPshift in the laboratory, but to manipulate factorsthat modulate it and so provide some constraintson interpretations of the effect. Given our ex-pectation that animacy should not strongly in-fluence the effect, we turned to the comprehen-sion literature for guidance. In recent sentencecomprehension research, argument structureproperties of verbs have been accorded increas-ing importance in ambiguity resolution and sen-tence interpretation (F. Ferreira & McClure,1997; Garnsey, Pearlmutter, Myers, & Lotocky,1997; MacDonald, 1994; Trueswell, Tanen-haus, & Kello, 1993). On this basis, we devel-oped the hypothesis that argument structureproperties of verbs similarly affect productionprocesses and, specifically, heavy-NP shift.

Our approach relates most directly to lexicalconstraint-based theories of ambiguity resolu-tion (see MacDonald, Pearlmutter, & Seiden-berg, 1994; Tanenhaus & Trueswell, 1995, forreviews), which have emphasized that it is notstrictly the number of alternative argumentstructures but rather the relative frequency ofuse of the alternatives in the language that in-fluences the difficulty of sentence processing


(e.g., Garnsey et al., 1997; MacDonald, 1994;Trueswell et al., 1993). The constraint-basedapproach argues for detailed lexical representa-tions that encode a wide variety of frequency-sensitive information. There is some uncertaintyin the current literature concerning whether thesame lexical representations are used in produc-tion and comprehension (e.g., Bock et al., 1992;Bock & Levelt, 1994; Butterworth, 1989; Dell,Schwartz, Martin, Saffran, & Gagnon, 1997;Dell & O’Seaghdha, 1994; Levelt, Roelofs, &Meyer, in press), but even if these representa-tions are separable, the frequency of their useand their experience should be strongly corre-lated. Therefore, production phenomena such asheavy-NP shift are likely to be influenced bysome of the same lexical properties that havebeen identified in studies of comprehension. Forexample, both comprehenders and speakersshould make use of lexical information con-cerning the frequency with which verbs tend toparticipate in particular syntactic constructions.Thus, when a verb is selected during the pro-duction process, frequency-weighted lexical in-formation will become activated and could con-strain the choice of a syntactic structure.

Following this reasoning, we developed averb disposition hypothesis,which states thatindividual verbs carry with them information onthe history of their participation in shifted struc-tures and that this history influences the likeli-hood of their allowing heavy-NP shift. Thus, forheavy-NP shift sentences such as Orwell’sSnowball had found in the harness-room an oldgreen tablecloth of Mrs. Jones’s,a speaker’sactivation of the verbfound should partiallyactivate both the shifted and the basic syntacticordering options for the direct object and loca-tion arguments of the verb. Furthermore, thedegree of activation of each option should be afunction of the frequency with whichfoundhaspreviously participated in the alternative struc-tures. Thus one component of our hypothesis isthat verbs that have previously been deployed inshifted verb phrases (VPs) will be more likely toappear in shifted VPs in the future. Second, averb’s shifting disposition is hypothesized to bea function not only of previous shifting but alsoof previous experiences in certain other struc-

tures that share a crucial feature with heavy-NPshift. We identify these other structures as onesin which the verb and its complement are non-adjacent.

One nonadjacent structure is the verb–par-ticle construction in English, in which a particlesuch asin, up, out,or onmay intervene betweena verb and its direct object NP, as inthrow outthe trash,and cut up the vegetables.Linguistshave long noted the similarity between the verb-–particle construction and heavy-NP shift inthat the alternative phrase orders Verb–Parti-cle–NP and Verb–NP–Particle are strongly de-termined by the length of the NP (Hawkins,1994; Wasow, 1997). Thus a verb’s frequentparticipation in Verb–Particle–NP sequences, inwhich the particle intervenes between the verband the NP, may leave the verb more disposedto participate in heavy-NP shift structures.

A second nonadjacent construction that mayaffect shifting disposition involves sententialcomplements (S-complements), typically intro-duced bythat,such asMary said that Bill wouldsing,or Mary learned that she would be allowedto go hiking. Whereas each S-complement isadjacent to its verb in these two examples, it istypically nonadjacent when the verb is modifiedby a PP or an adverb, as inMary said in a loudvoice that Bill would sing,or Mary learnedyesterday that she would be allowed to go hik-ing. These sentences avoid ambiguities andsound much more natural than in an alternativeorder in which the verb and S-complement areadjacent, as inMary said that Bill would sing ina loud voiceandMary learned that she wouldbe allowed to go hiking yesterday.By virtue ofthese frequent constituent orderings [V PP S-complement] and [V adverb S-complement],verbs that take S-complements frequently occurin nonadjacent structures. We propose that par-ticipation in nonadjacent S-complement struc-tures will increase a verb’s disposition to par-ticipate in heavy-NP shift structures.

In sum, we posit that prior “syntactic ex-periences” with any of three nonadjacent con-structions can increase a verb’s disposition toshift: (1) prior heavy-NP shifting, (2) prior par-ticipation in verb–particle constructions, and(3) prior participation in S-complement con-


structions. In the experiments, we chose to in-vestigate the effect of the S-complement con-struction. Undoubtedly, the S-complement con-struction and the other syntactic patterns thatmight affect shifting disposition are associatedwith certain prosodic, semantic, and/or dis-course representations that could affect shifting,but we will not initially explore those here. Ourgoal in Experiment 1 is simply to determinewhether there is any variation in verb shiftingdisposition that can be traced to participation inthe S-complement construction. Many verbsthat participate in S-complement structures canalso take NP direct objects. For such verbs,often termed “NP/S” verbs, participation in S-complements provides a number of opportuni-ties to appear nonadjacent to their complements.Therefore, and all else being equal, NP/S verbsshould more readily accommodate heavy-NPshift than do verbs that do not take sententialcomplements. In a corpus analysis of the PennTreebank data, we found support for the hy-pothesis that NP/S verbs have a higher shiftingfrequency than do verbs that do not take sen-tential complements (MacDonald, Stallings, &O’Seaghdha, 1998). To test the verb dispositionhypothesis in production, we constructed stim-uli manipulating NP length, PP noun animacy,and the type of verb so that it was either anNP/S verb likefound or revealed,or it was averb that does not take sentential complements,such astransferredor delayed.We will term thelatter type of verb an “NP-only” verb, indicatingthat it can take NP direct object complements,but not S-complements.

The idea that verb disposition is a frequencysensitive parameter has several important impli-cations. First, the verb disposition hypothesisnaturally entails the possibility that syntacticproduction processes, in addition to having anincremental component, have acompetitivecomponent whereby the alternative structurescompete for selection (Bates & Devescovi,1989; MacWhinney & Bates, 1989). For thecase of heavy-NP shift, we go further to arguethat competition is intrinsic to the choice ofordering options. That is, given that NP-first isclearly dominant in short NP sentences, but thatthe probability of NP shift increases as a func-

tion of NP length and a variety of more subtleconsiderations (Wasow, 1997), there should besituations where the default and shifted con-structions are roughly equipotent and thereforecompete with one another. This argument doesnot deny that many syntactic structure decisionsmay be effected incrementally (De Smedt,1990, 1994; V. Ferreira, 1996) or that many NPplacements are decided incrementally. Weclaim only that where syntactic ordering optionsare in question, competition should occur in atleast some proportion of “tough decision” cases.

A second issue concerns how and why therecord of nonadjacency should be generalizedover such different syntactic structures asheavy-NP shifted sentences, verb–particle con-structions, and nonadjacent S-complement sen-tences. Recent comprehension work shows thatcomprehenders have acquired distributional in-formation over a number of different “grains” ofdetail, from simple lexical and structural fre-quencies to far more complicated combinatorialconstraints, such as the frequency of a word ina particular syntactic structure in a particulardiscourse context (MacDonald, 1997; Mac-Donald et al., 1994; Tabor, Juliano, & Tanen-haus, 1997). A system that can represent con-tingent frequency information across thesevaried grains could encode nonadjacency infor-mation while also encoding specific informationabout each syntactic structure. It might do sobecause the structures that we posit as contrib-uting to shifting disposition share importantproperties. In addition to the property of non-adjacency itself, these structures share the prop-erty of being active rather than passive, and theyall have postverbal complements. Most impor-tant, from the point of view of production, thenonadjacency property is relevant to the alwayspressing question “What comes next?”

Experiment Overview

Because long and complex sentences arenecessary to elicit heavy-NP shift, we were re-luctant to use the recall paradigm common tomany of the word order studies reviewed above.We therefore developed a variant of a sentenceconstruction procedure used by Dell andO’Seaghdha (1992) to begin our exploration of


shifting. In pilot work, participants read sen-tence fragments on a computer screen andformed a plan for producing the complete sen-tence. More specifically, they indicated bypressing one of two keys whether they intendedto produce a [Subject–Verb–NP–PP] or [Sub-ject–Verb–PP–NP] sequence. We manipulatedNP length and animacy of the noun in the PPand found that the former but not the latterinfluenced the probability of shifting. Havingestablished that heavy-NP shift occurs in thetask, we manipulated the same factors in Exper-iment 1, plus the key factor of the type of verbheading the verb phrases. Experiment 2 simpli-fied the task by removing the requirement tomake an explicit choice prior to speaking. Ex-periment 3 examined whether shifting obtainsin a recall paradigm similar to that used in theword order literature, and showed, in conjunc-tion with the findings of Experiments 1 and 2,that our initial reluctance to use a recall task wasunfounded.

EXPERIMENT 1: VERB DISPOSITION, NPLENGTH, AND PP NOUN ANIMACY

This experiment was designed to investigatethe effects of verb disposition, noun phraselength, and animacy of the noun within a prep-ositional phrase on choice of basic [NP–PP] orshifted [PP–NP] order of these phrases. Partic-ipants constructed sentences from phrases ap-pearing on a computer screen, choosing eitherthe basic order in which a verb and direct objectNP were adjacent and a PP followed the NP[S–V–NP–PP] or the shifted order in which thePP intervened between the verb and the directobject NP [S–V–PP–NP]. A second group ofparticipants participated in a ratings task inwhich they reported the relative acceptability ofshifted and basic order structures. The use of therating task for comparison with the productiontask was prompted by McDonald et al.’s (1993)findings of different effects of word length andanimacy in rating and production tasks.

Our first goal in the production experimentwas to assess the effect of a manipulation of NPlength (2 vs 10 words) in this task. We expectedparticipants to choose the shifted structure more

often in the long NP condition than in the shortNP condition.

The PPs contained animate or inanimatenouns (e.g.,to Johnvs at lunch). We will termthis variablePP animacy,recognizing that it isnot the PP itself that is animate but rather theconcept represented by the noun within the PP.According to production theory (e.g., Bock,1987b; Garrett, 1984; Levelt, 1989), animacyaffects accessibility of words and influencesprocesses such as grammatical role assignmentthat occur relatively early in sentence formula-tion. Heavy-NP shift does not affect grammat-ical role assignment, but speakers may have atendency to put PPs with animate nouns earlierin a verb phrase and so produce more shiftedstructures. Of course, because we did not expectshort NPs to shift, any effect of animacy shouldbe stronger in the long NP conditions.

As discussed in the introduction, the key ma-nipulation was verb disposition, operationalizedas the effect of verb type (NP/S or NP-only).The verb disposition hypothesis states that ex-perience with nonadjacent structures affects theextent to which shifting will be observed and, asreviewed above, NP/S verbs have had morenonadjacent experiences than NP-only verbs.We therefore predicted that NP/S verbs wouldappear in more shifted structures than NP-onlyverbs. Shifting should be largely confined tolong NPs, yielding an interaction of NP lengthand verb type in the structure choice data.

In addition to choice of structure, we mea-sured both decision time to make the choice andsentence initiation time following a cue to beginspeaking. The competition view we proposestates that decision times should be longer whentwo alternatives are activated to roughly thesame degree (see e.g., Kawamoto, 1993). Incontrast, strictly incremental structure buildingcan only benefit from the flexibility of havingmore options (e.g., V. Ferreira, 1996) and soshould not care about relative activation levels.Because more shifting should occur in the NP/Sconditions, we predicted that decision timesshould be longer for the NP/S conditions thanfor the NP-only conditions. And because mostshifting should occur in the long NP conditions,there should be an interaction between verb type


and length in the decision latencies. In addition,because it may take longer to activate a shiftedalternative in the face of competition from themore common basic order, longer choice timesfor shifted than for basic order choices could bediagnostic of competition between alternativephrase orders.

In keeping with previous evidence that thecomplexity of prepared utterances affects initi-ation time (F. Ferreira, 1991; Sternberg et al.,1978), we predicted that voice initiation timeswould reflect overall sentence complexity. Thatis, we expected initiation latencies to be longerin the long NP than in the short NP conditions.Because they reflect plan implementation ratherthan plan selection, initiation latencies shouldnot be sensitive to verb type. Beyond that, pre-diction is difficult. If initiation latencies areprimarily sensitive to the structural complexityof some portion of the beginning of a complexsentence (F. Ferreira, 1991), latencies to beginshifted S–V–PP–NP sentences might be rela-tively fast. Note that this would also be com-patible with the view that shifting is motivatedby a strategy of deferring the more complexlong NP constituent to ease the processing bur-den (see Wasow, 1997). However, because wedo not know how the sentences in this experi-ment may be partitioned for production andbecause the working memory commitment oftracking a displaced NP in our paradigm mayincrease initiation latencies, we cannot testmore specific hypotheses.

Method

Participants.Ninety-six University of South-ern California undergraduates were either paid

or received extra credit in psychology coursesfor their participation. Half of the participantsserved in the production experiment and half inthe ratings task. All were native speakers ofEnglish.

Materials. We constructed 40 experimentalitems, each consisting of three components: asubject and verb, a noun phrase, and a preposi-tional phrase. An example is shown in Table 1,and all items are listed in the Appendix.

Each of the 40 subject–verb phrases was 2–3words in length (e.g.,Julie displayed, The con-tractor presented). We selected 20 verb pairsfor the subject–verb phrases. Each pair con-tained one NP-only verb and one NP/S verb(e.g., transferred and revealed, respectively).We excluded alternating datives such asgive,because these verbs can participate in phraseorders other than those that were under investi-gation here. The two groups of verbs werematched for length in syllables. The NP-onlyverbs had a frequency of 36 occurrences permillion in the Francis and Kucera (1982) cor-pus, and the NP/S verbs had a frequency of 47,a nonsignificant difference,t , 1. A subject NP,two direct object NPs manipulating length, andtwo PPs manipulating animacy, were written foreach pair. Short direct object NPs were twowords long and contained 2–4 syllables. Thelong direct object NPs (10 words and 13 sylla-bles) in Experiment 1 contained only prenomi-nal adjectives and prepositional phrases.

In the animate prepositional phrases, thepreposition was alwaysto, expressing a Goalthematic role, as into Mary.The prepositions inthe inanimate condition, includingat, on, in,with, and by, expressed Time, Location and

TABLE 1

Example Materials for Experiment 1

Condition NP & NP/S verb pair PP NP

Animate short transferred/revealed to Leigh the graphsAnimate long transferred/revealed to Leigh some more specific plans for a brand new defense plantInanimate short transferred/revealed at dawn the graphsInanimate long transferred/revealed at dawn some more specific plans for a brand new defense plant

Note.The subject NP for this item wasJanet.


Manner roles, as inat ten, in person,etc. A fewof the Manner items, such asvery quickly,wereadverbial phrases.

The animate and inanimate prepositionalphrases were matched as closely as possible innumber of words and syllables with the shortNP for each item. In the animate condition,there was no difference between the length ofthe NP and the PP in the short condition. In thelong condition, the difference was 8 words and9–11 syllables. In the inanimate condition,there was a 0 word and 1–3 syllable differencein the short condition and no less than an 8 wordand 8 syllable difference in the long condition.

Design and procedure.The production taskinvolved four two-level factors: NP length (2 vs10 words), verb type (NP/S or NP-only), screenposition (NP top vs NP bottom), and animacy ofthe PP noun. PP animacy was manipulated be-tween participants and the others were manipu-lated within.

The task employed a variation on a con-strained production paradigm in which partici-pants prepared and then uttered sentences usingphrases presented on a computer screen (Dell &O’Seaghdha, 1992; F. Ferreira, 1994; V. Fer-reira, 1996). On each trial, three left-justifiedphrases appeared at top, center, and bottomscreen locations. We instructed participants toput the phrases together to form a sensible sen-tence, informing them that some sentenceswould make more sense with the phrase orderMIDDLE–TOP–BOTTOM and others wouldwork better with the order MIDDLE–BOT-TOM–TOP. With the exception of possessiveapostrophes, there was no punctuation. The sub-ject–verb phrase was always in the center posi-tion. It was underlined to indicate that it consti-tuted the beginning of the sentence. Theposition of the other two phrases was counter-balanced across participants and items.

Ten practice and 37 filler items were con-structed so that only one ordering of the top andbottom phrases yielded a grammatical sentence.For 60% of these items, the phrase order MID-DLE–TOP–BOTTOM was the grammaticalone, and for the remaining 40% the order MID-DLE–BOTTOM–TOP was grammatical. Thisratio was adopted to counteract a tendency ob-

served in pilot work to favor the order MID-DLE–BOTTOM–TOP.

Participants pressed a key to initiate eachtrial. The three phrases appeared on the screen.Participants read the phrases and then pressedone of two keys to indicate in which order theyintended to say them, the MIDDLE–TOP–BOTTOM order or the MIDDLE–BOTTOM–TOP order. In this way, we obliged participantsto choose explicitly between the alternative or-ders at a definable moment. We did this becausewe were interested both in the choice of orderand in the difficulty of the choice, as indexed bydecision latency. In addition, we wanted toavoid the possibility that participants would ini-tiate speech before fully reading the displays.

Following the keypress, the screen wasblanked for a 1 sinterval. The word GO thenappeared in large font in the center of thescreen. At the GO prompt, the participant ut-tered the sentence aloud into a microphonelinked to a voice key that recorded the time toinitiate the sentence. The voice initiation timewas measured from the appearance of the GOprompt to the start of the participant’s utterance.Because participants had already decided on theorder of production, latency to respond shouldindex plan complexity and not the planningprocess. Triggering the voice key also causedthe screen display of the sentence fragments toreappear. Participants thus did not have to mem-orize the entire sentence verbatim but only hadto remember the first several words and thechosen order of phrases. Once a participant fin-ished uttering the entire sentence, the experi-menter recorded the participant’s choice ofphrase order with a key press; equipment errorswere recorded by hand. Following the 10 prac-tice items, the 40 experimental and 37 filleritems were presented in random order. Partici-pants completed this portion of the experimentwithout a break in a 15–20 minute session.

Rating task.For the rating task, a secondgroup of 48 participants received the same 40experimental items as in the production task,fully counterbalanced across subjects for NPlength, verb type, animacy, and left/right pageposition of shifted and unshifted items. Shiftedand unshifted structures appeared side by side


on the page, as shown below, and participantscircled a number between 1 and 7, indicatingtheir judgment of which order seemed the more“natural” way to express the thoughts in thesentences. Ten items appeared on each page ofthe questionnaire, and page order was randomlyassigned for each participant. Conditions ap-peared in random order on each page, exceptthat two pages contained only items in whichthe shifted structure was on the left, and theother two pages contained items in which theshifted structure was on the right.

Don introduced atbrunch a coupleof old collegefriends and amath teacher.

1 2 3 4 5 6 7 Donintroduced acouple of oldcollege friendsand a mathteacher atbrunch.

Results and Discussion

Three dependent variables were analyzed in theproduction experiment: (1) the percentage ofshifted [PP NP] orders that participants chose, (2)decision times in choosing a phrase order, as mea-sured from the initial presentation of the threephrases to the participant’s keypress indicating achoice of order, and (3) time to initiate the utter-ance, as measured from the GO prompt. Thesedata were analyzed as a function of length, verbtype, animacy, and screen position (NP top vs NPbottom). Except where noted, screen position hadno effect. In the following analyses, 1.2% of thetrials were excluded due to short phrase orderdecision times (,500 ms), and 1.8% were ex-cluded due to inconsistency between the initialchoice of phrase order and the order subsequentlyproduced.

Phrase order choice.As shown in Fig. 1, therewere clear effects of both length and verb type.Participants shifted about four times as often in thelong NP condition as in the short NP condition,F1(1,46) 5 50.81,p , .001; F2(1,39) 5 94.60,p, .001. Participants also shifted more than twiceas often with the NP/S verbs as with NP-onlyverbs,F1(1,46) 5 50.77,p , .001; F2(1,39) 538.84,p , .001. The verb disposition by lengthinteraction shown in Fig. 1 was also reliable,F1(1,46) 5 16.28,p , .001; F2(1,39) 5 20.36,p , .001, reflecting the concentration of the verb

type effect in the long NP conditions. These dataclearly support our main hypothesis that NP/Sverbs should be associated with more heavy-NPshifts than NP-only verbs. There were no effectsof animacy,Fs, 1, and this factor did not interactwith any others.

Although we did not expect a strong effect ofanimacy, we were concerned that the argument/adjunct status of PPs was correlated with animacy,as some syntactic ordering choices are influencedby thematic structure (F. Ferreira, 1994). We con-ducted a post hoc analysis to determine whetherthe effect of verb type and the absence of an effectof animacy were related to variations in the argu-ment/adjunct status of PPs across the conditions.For each item with each verb and PP combination,we coded the PP as an argument or an adjunctrespective to its corresponding NP-only and NP/Sverbs.1 All of the inanimate items were adjuncts.

1 We adopted semantic selection criteria (Dowty, 1982)such that argument status was assigned to those PPs thatwere entailed by their verbs. Otherwise, adjunct status wasassigned. Thus, “to Leigh” is an argument of the verb“transfer,” since transferring a document requires a recipi-ent. We examined two separate argument codings: a strictargument interpretation in which questionable argumentswere coded as adjuncts and a looser interpretation in whichquestionable arguments were coded as arguments. We pre-sented the data from the looser coding in the text becausethat coding yields the larger number of arguments, but thepattern of results did not change with the stricter coding.

FIG. 1. Production of heavy-NP shifted structures as a func-tion of animacy, length, and verb disposition in Experiment 1.


Of the animate items, those with NP-only verbscontained 16 arguments and 4 adjuncts, whileitems with NP/S verbs had a very similar pattern,15 arguments and 5 adjuncts. To determinewhether argument status affected shifting rates,we reanalyzed the shifting data for items in theanimate condition with argument status as a factorand found that it had no effect on shifting. Argu-ment and adjunct PPs showed very similar shiftingrates—15.4 and 17.5% respectively,F2(1,38) 51.47,p . .20. The argument status factor did notinteract with any others in this analysis. In sum,phrase order choice was affected by NP length andby verb disposition, but not by the animacy ofnouns in prepositional phrases.

Decision times.Table 2 shows the phrase orderdecision times presented, for informational pur-poses, according to whether participants chose thebasic or shifted orders. Due to the infrequency ofshifting, the mean shifted choice times are calcu-lated over a large number of empty cells and mustbe viewed with caution. Therefore, we collapsedover structure choice in the main analysis of de-cision times. As expected, decision times werelonger in the long NP conditions (5362 ms) than inthe short NP conditions (4187 ms),F1(1,46) 553.52,p , .001;F2(1,39)5 74.61,p , .001. Thiseffect is at least partly attributable to the fact thatparticipants took longer to process the material inthe long NP condition. More interestingly, therewas also a main effect of verb disposition, suchthat decision times were longer for sentences con-

taining NP/S verbs (5038 ms) than for those con-taining NP-only verbs (4511 ms),F1(1,46) 522.57,p , .001; F2(1,39) 5 10.71,p , .005.However, there was no interaction of verb typeand NP length,Fs, 1. Consistent with the choiceevidence, there was no effect of animacy on de-cision time, and animacy did not interact with theother factors,Fs , 1.

An analysis of the basic order choice timestells a similar story. Decision times were longerin the long NP conditions (5304 ms) than in theshort NP conditions (4149 ms),F1(1,46) 558.10,p , .001; F2(1,39) 5 31.20,p , .001and for sentences containing NP/S verbs (5006ms) than for those containing NP-only verbs(4447 ms),F1(1,46) 5 33.86,p , .001; F2(1,39) 5 7.27,p , .01. Furthermore, there was noeffect of animacy on decision time and therewere no interactions (Fs , 1).

One interpretation of the effect of verb type isthat the longer decision times for NP/S verbssimply reflect longer reading times for thesesentences. This pattern could be due to theadded complexity of having to resolve the ar-gument structure ambiguity inherent in NP/Sverbs, in contrast to the unambiguous NP-onlyverbs. Alternatively, the effect may lie in thedecision process, such that decision times arelonger when there is greater competition be-tween alternative phrase orders. The frequentparticipation of NP/S verbs in constructions thatinvolve a nonadjacent complement would yield

TABLE 2

Decision (Experiment 1) and Ready Response (Experiments 2–3) Times in ms (with Standard Deviations),as a Function of Order, NP Length, and Verb Type

Experiment 1 Experiment 2 Experiment 3

NP-Only NP/S NP-Only NP/S NP-Only NP/S

Short Long Short Long Short Long Short Long Short Long Short LongBasic 3883 4969 4416 5458 3618 5109 3851 6083 3643 12923 4070 13124

(1754) (2576) (2019) (2907) (1570) (2390) (1606) (3908) (1053) (7391) (1592) (9012)Shifted 4881 6678 5374 6107 4082 5333 3928 6701 4499 18051 4408 18860

(3750) (4026) (4572) (3654) (2034) (2822) (1678) (4494) (1561) (13662) (1315) (13199)Overall 3925 5097 4448 5628 3625 5281 3829 6263 3643 13130 4062 13618

(1767) (2788) (2111) (2806) (1584) (2437) (1500) (4038) (1040) (7647) (1586) (8395)

* The overall means are reported from analyses that collapse over order.


partial activation of both alternative structuresand competition between them, slowing re-sponse times, whereas the NP-only verbs shouldhave much less activation of the nonadjacentstructure and thus little competition.

However, the competition hypothesis makesspecific predictions concerning decision timesconditionalized on choice of the basic vs theshifted structure. When the basic order is cho-sen, decision times should be shortest whenthere is little activation of the shifted structure.Thus, decision times for the basic order shouldbe shorter for sentences with rarely shiftedverbs than for frequently shifted verbs. Simi-larly, the competition account predicts that adecision in favor of the shifted structure shouldbe made more quickly in at least some caseswhen there is substantial activation of theshifted structure, that is, with frequently shiftedverbs, than when there is little activation of thisstructure, as with rarely shifted verbs. With sofew instances of shifting for NP-only verbs, it isimpossible to analyze decision times in an anal-ysis of variance with Decision (basic vs shifted)and Verb type as factors. Instead, we used cor-relational analyses to examine the relationshipbetween shifting frequency and the latency todecide on basic and shifted structures. For eachexperimental item with each verb (recall thateach item appeared with two different verbs),we counted the number of times the item wasuttered in shifted form. These frequencies areshown in the Appendix. The range of shiftingfrequencies was 0–11, (mean5 3.9) out of atotal of 24 presentations (12 of which were inthe short condition, for which shifting was veryrare). Most, but not all, frequently shifted itemscontained NP/S verbs, and the items that shiftedwith moderate frequency (i.e., 4–5 times) con-tained about equal numbers of NP-only andNP/S verbs.

We correlated the shifting frequencies withdecision times in the long NP condition for thebasic order and for the shifted order. Forchoices in favor of the basic order, we found arobust positive correlation between shifting fre-quency and decision time,r (79) 5 .33, p ,.005. This pattern is the one predicted by thecompetition hypothesis: The more frequently

shifted an item, the longer the decision time infavor of the unshifted, basic order. The compe-tition hypothesis predicts a correlation in theopposite direction for choices for the shiftedorder: The more frequently shifted items shouldtend to have shorter decision times than rarelyshifted items. The correlation turned out to be asmall negative one,r (72)5 2.19,p 5 .11. Therelative weakness of this correlation may be dueto the fact that, because the basic order is thedefault, the decision to shift may involve com-petition on individual trials even for many of themost frequently shifted verbs. The importantpoint is the difference between the shifted andbasic order correlations rather than the exactvalues of the coefficients. The contrast betweenthe negative correlation between shifting dispo-sition and shift latency and the positive corre-lation between shifting disposition and nonshiftlatency is consistent with the existence of acompetitive process.

Following this reasoning, we tested whetherthe two correlations were significantly differentfrom one another. The Z1* statistic (Steiger,1980) is appropriate for this case, in which thetwo correlations share one factor (shifting fre-quency), and independence between the corre-lations cannot be assumed. The test showed thatthe two correlations were indeed significantlydifferent, Z1* 5 2.94, p , .005. Overall, thecorrelational analyses support the competitionhypothesis, suggesting that both the shifted andunshifted orders are partially activated and com-pete with one another in our production task. Tothe extent that such results generalize to spon-taneous production, they argue against strictlyincrementalist accounts of phrase ordering andpotentially against strongly incrementalist ac-counts of sentence construction in general.

Voice initiation times.Table 3 shows voiceinitiation times corresponding to the decisiontimes in Table 2. Trials on which the voice keyfailed to register the participant’s voice orwhere the RT was greater than 3000 ms wereexcluded (an additional 2.9% of the trials).Analysis of variance showed a main effect ofNP length, such that participants took longer tobegin speaking in the long condition (744 ms)than in the short condition (699 ms),F1(1,46)5


8.25,p , .01; F2(1,39) 5 5.1, p , .05. Therewere no effects of animacy or verb dispositionon initiation time, except for a small uninter-pretable interaction between verb dispositionand screen position in the participants analysisonly.

The NP length effect is consistent with otherevidence that initiation time is sensitive to ut-terance complexity under conditions where theutterance is maximally prepared (e.g., F. Fer-reira, 1991; Sternberg et al., 1978). In thepresent context, it provides useful confirmationthat, although participants knew that the sen-tence ingredients would be redisplayed follow-ing utterance initiation, they genuinely plannedthe sentences before beginning to speak. Itseems likely that participants were committedto producing the SV phrase in all conditions;beyond that, the overall complexity of the ut-terance (defined by whether or not it containeda long NP) rather than the order of phrasesdetermined initiation time. The evidence doesnot allow us to conclude whether this is a gen-eral characteristic of planning for this kind ofsentence or whether it is due to the fact thatparticipants knew that the NP and PP would beredisplayed following sentence initiation.

Rating data.Figure 2 shows the rating data,in which higher ratings indicate greater prefer-ence for the shifted order. Consistent with theproduction data, there were main effects of bothverb disposition and length. Participants pre-

ferred the shifted order more when the sentencecontained an NP/S verb (2.84) than when itcontained an NP-only verb (2.55),F1(1,46) 521.22,p , .001; F2(1,39) 5 10.98,p , .005.Likewise, they preferred the shifted order morein the Long NP condition (3.59) than in theShort NP condition (1.80),F1(1,46) 5 106.25,p , .001;F2(1,39)5 390.58,p , .001.

As in the production study, there was no maineffect of animacy. There was a small interactionbetween animacy and length, reliable only inthe items analysis,F1(1,46) 5 2.90, p . .05;

TABLE 3

Voice Initiation Times in ms (with Standard Deviations), as a Function of Order, NP Length,and Verb Type for Experiments 1–3

Experiment 1 Experiment 2 Experiment 3

NP-Only NP/S NP-Only NP/S NP-Only NP/S

Short Long Short Long Short Long Short Long Short Long Short LongBasic 685 732 708 758 711 708 687 730 530 655 539 672

(220) (236) (259) (277) (239) (240) (235) (240) (141) (196) (126) (210)Shifted 768 752 798 725 808 729 845 759 889 666 521 617

(382) (390) (527) (268) (666) (259) (657) (282) (158) (480) (137) (175)Overall 687 739 711 750 714 707 706 707 532 657 539 660

(222) (226) (261) (246) (243) (219) (244) (238) (143) (204) (123) (180)

* The overall means are reported from analyses that collapse over order.

FIG. 2. Mean acceptability ratings for heavy-NP shiftedstructures as a function of animacy, length, and verb dispo-sition in Experiment 1.


F2(1,39) 5 10.12,p , .005. However, the na-ture of the interaction was not consistent withour theoretical expectations: In the short NPcondition, shifted animates werelesspreferred(1.63 mean rating) than shifted inanimate items(1.98 mean rating), and there was no effect ofanimacy in the long NP condition. Thus, thisweak interaction does not provide evidence insupport of the idea that PP animacy increasespreference for shifted NP sentences. In contrastwith McDonald et al.’s (1993) finding that peo-ple preferred animate words early even thoughthey did not produce them earlier, it seems thatPP animacy has little or no effect on preferencefor the shifted vs basic syntactic structure.

Conclusions.Experiment 1 demonstrated aclear effect of NP length on shifting and, moreimportantly, showed that the length effect ismodulated by verb type. Participants utteredshifted structures more, and rated shifted struc-tures as being more acceptable, when the sen-tence contained an NP/S verb than when itcontained an NP-only verb, especially in thelong NP conditions. Thus our task demonstratedsensitivity to the heavy-NP shift phenomenon,and the experiment provides preliminary sup-port for the verb disposition hypothesis. Deci-sion times were also sensitive to verb disposi-tion but not to its interaction with NP length.However, this might be because of a mixture ofeasy and difficult decisions in the long NP con-ditions. A correlational analysis confirmed thisconjecture by uncovering evidence of competi-tion in the decision process: decisions in favorof the basic order were slower and decisions toshift were relatively fast for more frequentlyshifted verbs. No effect of PP noun animacyobtained, replicating findings in McDonald,Bock, and Kelly (1993) that animacy does notinfluence ordering in structures where grammat-ical role assignment is unaffected. Therefore,we dropped the manipulation of animacy in thefollowing experiments.

Before discussing more general implicationsof our findings, it is important to consolidate ourexperimental grasp of heavy-NP shift. In Exper-iment 1, we endeavored to capture the heavy-NPshift phenomenon and its competitive implica-tions by obligating participants to choose the

order of phrases before beginning to speak. Adrawback of this procedure is that it is some-what artificial, perhaps forcing a degree of com-petition and shifting that does not occur in nat-ural speech. Specifically, the explicit decisionabout order can be viewed as exaggerating therole of a metalinguistic order preference andpossibly obscuring the incremental nature ofunrestricted production. The task therefore mayreflect comprehension and judgment processesmore than the production processes we wishedto capture. To address these concerns, we variedthe procedure in Experiments 2 and 3.

EXPERIMENT 2: READINESS TASK

This experiment was very similar to Experi-ment 1 except for the following modifications.First, because animacy of the prepositional phrasenoun had no effect in Experiment 1, we used onlythe inanimate items in Experiment 2. Second, be-cause the rating data mirrored the choices in Ex-periment 1, we did not collect additional ratings.Most importantly, we removed the explicit orderchoice decision from the production task. In Ex-periment 2, participants were required only toindicate readiness to speak, not to commit to aparticular order of phrases before beginning toproduce the sentences. In general, because thenumber of inconsistencies between order deci-sions and actual productions in Experiment 1 wasvery low (1.8 % of all trials), we expected largelysimilar outcomes in this experiment. We investi-gated three specific issues: (1) whether the effectsof NP length and verb disposition observed inExperiment 1 were artifacts of the explicit choicerequired in that experiment; (2) whether prepara-tion time before indicating readiness, which pro-vides a more indirect measure of competitive pro-cesses underlying sentence production, would stillbe sensitive to verb disposition; and (3) whetherrelaxing the requirement to commit to a sentenceproduction plan would reduce the effect of sen-tence complexity on initiation times.

Method

Participants. Twenty-four University ofSouthern California undergraduates were eitherpaid or received extra credit in psychology


courses for their participation. All were nativespeakers of English.

Materials and procedure.The materials werethe same as the inanimate items from Experi-ment 1. The procedure was similar to that ofExperiment 1 except that participants were notrequired to indicate their choice of phrase orderwith a keypress. Instead, they were instructed toread the phrases, prepare to produce them in asensible order, and press a key when ready tobegin speaking. Following the keypress, thescreen was blanked for 1 s until the GO promptappeared, cueing the participant to begin to utterthe sentence. The participant’s voice triggeredthe timer connected to the microphone andcaused the sentence fragments to reappear onthe screen. The session lasted about 20 minutes.


As in the previous experiment, there werethree dependent variables. The first, percentageof shifted [PP–NP] productions, and the third,voice initiation time, were the same as in Ex-periment 1. The second dependent variable,time to indicate readiness to speak, was a lessexplicit measure of decision time than thechoice required in Experiment 1. In the follow-ing analyses, 2.3% of the trials were excludedbecause time to read the phrases and press thekey was less than 500 ms.

Phrase order choice.As shown in Fig. 3,there were clear effects of both length and verbdisposition on participants’ choice of phraseorders. As in Experiment 1, participants utteredsentences with a heavy-NP shift structure aboutfour times as often in the long NP condition asin the short NP condition,F1(1,23) 5 21.77,p , .001; F2(1,39) 5 54.84, p , .001. Theeffects of verb disposition were also replicated;participants shifted almost twice as often withthe NP/S verbs as with NP-only verbs,F1(1,23)5 9.48,p , .005;F2(1,39)5 15.84,p , .001.Finally, though the NP/S–NP-only differencewas larger in the long NP condition, the verbtype by length interaction was marginal,F1(1,23) 5 3.86,p , .10; F2(1,39)5 3.15,p , .10.Overall, the data replicate the theoretically im-portant results of Experiment 1, showing that

those findings were not due to the explicitchoice requirement.

Ready responses.Although the removal ofthe explicit choice requirement could makepreparation time less sensitive to the processesinvolved in sentence formulation, the same fac-tors should influence preparation time in thisexperiment as choice time in Experiment 1. Ingeneral, the pattern for mean latencies of readyresponses was similar to that of Experiment 1(see Table 2). As before, we analyzed the over-all data, judging the shifted data to be unstable.Preparation time was, of course, longer in thelong NP (5772 ms) than the short NP conditions(3727 ms),F1(1,23) 5 32.38 p , .001; F2(1,39) 5 79.32,p , .001. More importantly, verbtype also had a significant effect. As in Exper-iment 1, participants took longer to prepare sen-tences containing NP/S verbs (5046 ms) than toprepare sentences containing NP-only verbs(4453 ms),F1(1,23)5 5.01p , .05;F2(1,39)55.82 p , .05. There was also a nonsignificanttendency to interaction between length and verbtype,F1(1,23)5 3.12,p , .10,F2 5 2.33,p 5.14. Assuming that the verb type effect reflectsthe same configuration of decision-making dif-ficulty uncovered in our correlational analysisof Experiment 1, this suggests that competitionof the kind we discussed in relation to Experi-ment 1 was also at work here.

FIG. 3. Production of heavy-NP shifted structures with-out explicit decision component as a function of animacy,length, and verb disposition in Experiment 2.


Voice initiation time.Although the structurechoice and preparation time findings with thechoice and ready procedures were very similar,we expected that relaxation of the explicitchoice requirement might reduce the effect ofsentence complexity (length) on voice initiationtimes. Long voice RTs (.3000 ms) and micro-phone errors were excluded from the voice ini-tiation time data (1.5% of trials). Inspection ofTable 2 suggests a rather dramatic difference inbehavior in the basic and shifted conditions.However, because the shifted data are based ontoo few observations, we concentrate here againon the overall data.2

In contrast to Experiment 1, there was nomain effect of length in the analysis of varianceon the overall data. It appears that participantsdid engage in sentence formulation, as indicatedby the sensitivity of “ready” latencies to verbtype, but that they were not as committed to anexplicit sentence production plan as were theparticipants in Experiment 1. It may be thatparticipants were committed only to productionof the SV phrase at the moment of voice initi-ation. If so, this gives some support to the ideathat the choice procedure of Experiment 1forced participants to be more committed andless incremental in their planning than theywould freely elect to be. Despite this, however,the outcomes in terms of choice of phrase orderwere remarkably consonant in Experiments 1and 2. This suggests that the determination ofphrase order may occur at a point in sentenceformulation at which speakers are only partlycommitted to the implementation of the result-ing structure, a conclusion that is compatiblewith most sentence production theories (seee.g., Bock & Levelt, 1994).

Though the task in Experiment 2 was lessrestrictive than that in Experiment 1, it differsfrom the still less restrictive cued recall proce-dure of most word-order studies (e.g., Mc-Donald et al., 1993). We were initially disin-clined to use a cued or immediate (e.g., F.

Ferreira, 1991) recall task to investigateheavy-NP shift, because the difficulty of re-membering very long NPs could bias speakersagainst shifting. However, there is good evi-dence that even immediate recall captures manyof the processes of spontaneous sentence gen-eration (e.g., Potter & Lombardi, 1990). There-fore, to provide a better comparison to otherwork in the literature, we decided to try a recalltask in Experiment 3.

EXPERIMENT 3: PROMPTED RECALL

This experiment was designed to investi-gate the effects of length and verb dispositionin a cued recall task. Participants again readthe phrases on a screen and pressed a buttonwhen they were ready to speak. They thenreceived the subject and verb (SV) phrase(e.g.,Janet revealedfrom Table 1) as a signalto begin speaking, but the NP and PP had tobe recalled from memory. This is quite adifficult task, especially in the long NP con-ditions, and we did not expect error-free re-call. Rather, we were primarily interested inwhether the NP or PP immediately followedthe SV prompt. Given the robust effects in theprevious experiments, we again predicted ef-fects of NP length and verb type on shifting.Note that in this procedure, however, partic-ipants could well decide on the fly whether toproduce the NP or the PP first, even thoughthey were instructed to plan the order ofphrases. The heavy memory demands of therecall task could promote heavy-NP shift bymaking long NPs more difficult to retrieve, orwork against it by shrinking the length of theNPs held in memory. We also expected prep-aration time to be influenced by verb type andespecially by NP length. The prediction forvoice initiation time was less clear. On onehand, this experiment required participants toprepare the sentence to be produced morethan either of the others. On the other hand,the recall cue is the initial SV part of thesentence. Knowing that this would be avail-able as a prompt could desensitize voice ini-tiation latencies to the degree or nature ofpreparation.

2 There was again a small interaction between verb typeand screen position in this experiment. However, it was inthe opposite direction to that observed in Experiment 1.Thus, we judge both effects to be spurious. No interactionwith position occurred in Experiment 3.


Method

Participants. Twenty-four University ofSouthern California undergraduates were eitherpaid or received extra credit in psychologycourses for their participation. All were nativespeakers of English.

Materials and procedure.The same inani-mate items as in Experiment 2 were used in thisexperiment. The stimulus displays and sequenceof events in a trial were similar to those of theprevious experiments up to the point whereparticipants indicated readiness to speak. At thestart of each trial, three phrases appeared on thescreen, arranged as in the previous experiments.Participants were instructed to read the phrases,arrange them in a sensible order, and prepare toproduce the resulting sentence from memorywhen cued to speak. Participants pressed a keyto indicate when they were ready to beginspeaking. As in previous experiments, this key-press was followed by a 1 s blank interval,which was in turn followed by a prompt tobegin speaking. However, unlike in the previ-ous experiments, the speaking prompt was notthe word GO but rather the Subject–Verb phraseof the sentence, which reappeared in its mid-screen location. When participants saw this cue,they were to say the entire sentence from mem-ory as accurately as possible. As soon as par-ticipants began speaking, the Subject–Verbphrase disappeared and they completed the sen-tence as best they could. None of the phrasesreappeared on the screen once the prompt dis-appeared, and the screen remained blank untilthe experimenter pressed a key recording theorder (basic or shifted) in which the sentencehad been uttered. The sessions were tape-re-corded for later transcription and lasted between30 and 45 minutes.


In addition to the usual dependent variables,we also analyzed the participants’ utterances fornumber of words produced across conditions,based on transcripts of the recorded experimen-tal sessions. In the following analyses, 1.5% ofthe trials were omitted because time to read thephrases and press the key was less than 500 ms.

Due to the nature of the task, participants some-times changed, deleted, or added words to theoriginal stimulus items. In 2.9% of the trials,participants failed to recall any words in one ofthe constituents (either the PP or the NP). Thesetrials, as well as an additional two trials inwhich participants produced sentential comple-ment constructions, were excluded from theanalyses. Moreover, due to failure of the taperecorder, some trials were not recorded duringthe testing of some participants, resulting in theloss of an additional 5.6% of the trials in theanalyses that were based on the transcripts.

Phrase order choice.As shown in Fig. 4, theoverall rate of shifting was lower than in previ-ous experiments, but the pattern of resultsclosely replicated the pattern in the previousstudies. Participants shifted about 15% of thetime in the long NP condition as opposed to justover 1.5% of the time in the short NP condition,F1(1,23) 5 25.3,p , .001; F2(1,39) 5 29.56,p , .001. The effect of verb type also repli-cated; participants again shifted about twice asoften with the NP/S verbs as with NP-onlyverbs,F1(1,23)5 15.06,p , .005;F2(1,39)58.9,p , .01. Most importantly, the verb dispo-sition by length interaction shown in Fig. 4 wasalso reliable,F1(1,23) 5 11.13, p , .005;F2(1,39)5 5.03,p , .05.

Number of words produced.The number of

FIG. 4. Production of heavy-NP shifted structures in aprompted recall task as a function of animacy, length, andverb disposition in Experiment 3.


words produced in each condition was calcu-lated from the transcripts and is shown in Table4. Corrections, hesitations, and repetitions werenot included in word counts. As shifting isstrongly influenced by length, we were less con-cerned with whether the phrases were recalledperfectly accurately than with the number ofwords that were uttered. For example, for theNP thirty paintings, accurate reproduction ofthis NP would be counted as a two-word NPutterance, as would the slightly inaccuratetwenty paintings,andabout thirty paintingswasscored as a three-word NP. Participants’ pro-ductions of the PPs were extremely accurate.Production of short NPs was also very accurate,but long NP production was less so. The tableshows that speakers produced a mean of 8.6words in the long (10 word) NP conditions. Thisresult suggests that the lower rate of shifting inthis study may have been influenced by thesomewhat shorter NPs actually produced in thelong conditions. Importantly, however, thesedata provide no evidence that the effect of verbtype on structure choice is due to the number ofwords produced in each sentence, as there wasno difference in the number of words producedfor NP-only and NP/S verbs,F’s , 1. As al-ways, the means for the shifted conditions re-flect very few observations and should beviewed with caution.

Ready responses.The striking feature ofthe readiness response data is the very longpreparation times in the long NP conditions(see Table 2). These long RTs are evidentlydue to the additional memory requirements ofthe procedure, and this pattern yielded astrong effect of NP length,F1(1,23)5 48.41,p , .001; F2(1,39) 5 365.47,p , .001. In

addition, there is some evidence of a verb-type effect in the participants analysis, suchthat ready responses were longer with NP/Sthan NP-only verbs,F1(1,23)5 4.52,p , .05;but F2 , 1. There was no interaction of NPlength and verb type,Fs , 1.

Voice initiation time.In this analysis, five trialswere excluded due to long voice RTs (.3000 ms)or microphone errors. Initiation times (see Table3) were generally shorter in this study comparedto the previous ones, probably because the firstwords of the sentence were presented as the recallcue. The short RTs also suggest that speakerswere strongly committed to producing the sen-tence they had formulated and did not take advan-tage of the possibility of using the SV prompt as adelaying tactic. In addition, participants tooklonger to initiate speaking in the long conditions(659 ms) than in the short conditions (535 ms),F1(1,23)5 32.75,p , .001;F2(1,39)5 71.92,p ,.001. There was no effect of verb disposition oninitiation time, and there were no interactions,Fs , 1. These results thus replicate the effect oflength on voice initiation times in Experiment 1and are consistent with our suggestion that thefailure to find this effect in Experiment 2 was dueto a lesser degree of explicit planning in thatexperiment. Experiment 1 ensured extensive plan-ning by having participants make an explicit struc-ture choice prior to speaking, and the presentexperiment forced planning by requiring that theparticipants report the sentences from memory,but Experiment 2 had neither of these features.However, as we argued above, this does not meanthat participants in Experiment 2 did not engage inthe processes leading to determination of phraseorder prior to speaking. Rather, it means only thatthey did not incur the memory commitments con-

TABLE 4

Experiment 3. Mean Number of Words Produced in the NP (with Standard Deviations),as a Function of the Order Produced, NP Length, and Verb Type

Short NPs (2 words) Long NPs (10 words)

NP-only verbs NP/S verbs NP-only verbs NP/S verbs

Basic 2.00 (.04) 2.01 (.06) 8.54 (.99) 8.46 (1.21)Shifted 2.00 (0) 2.00 (0) 8.80 (1.99) 8.77 (1.60)


sequent on the phrase order decision, relying in-stead on the reappearance of the phrases to helpimplement it.

Conclusions.Our reservations about the fea-sibility of using a sentence recall task to studyheavy-NP shift appear to have been unfounded.We again observed separate and interactive verbdisposition and length effects on choice ofphrase order. The effect of verb disposition onpreparation time was not strong, at least byitems, in this experiment. However, there was aclear effect of NP length on voice initiationtime, indicating that participants were stronglycommitted to what they would produce on re-ceiving the SV prompt.

GENERAL DISCUSSION

Our findings can be summarized ratherstraightforwardly. First, the experiments con-firmed that the choice of the heavy-NP shiftedvs basic order (NP–PP) structure during speechproduction is constrained by the length of theNP in the verb phrase. Because we manipulatedthe length of only the NP, our findings do notspeak to the issue of whether it is NP length orrelative NP–PP length that is important. Wemerely note that the corpus analyses of Hawkins(1994) and Wasow (1997) favor the latter view.Second, we found that the probability ofheavy-NP shift was conditioned by the “shiftingdisposition” of verbs. Shifting disposition, inturn, appears to be determined by the frequencywith which a particular verb has previouslybeen used in structures in which the verb and itscomplement are not adjacent. Third, decisionlatencies were sensitive to the shifting frequen-cies of verbs: Frequently shifted items in Ex-periment 1 yielded longer decision times infavor of the basic order and relatively shortdecision times in favor of the shifted order,compared to rarely shifted items. This resultsuggests that the alternative structures com-peted for activation during sentence formula-tion. Fourth, voice initiation times reflect thecomplexity of an explicit plan for sentence pro-duction and the degree of commitment to thisplan. Initiation times were longer when the sen-tences contained long NPs (Experiment 1) andespecially when participants knew that they

would have to reproduce these long NPs frommemory (Experiment 3), but not when an ex-plicit commitment to produce the NP in a par-ticular sentence position was not required (Ex-periment 2). Fifth, the animacy of the noun inthe prepositional phrase had little or no effect onthe choice of phrase order or on preference forshifted vs unshifted structures in the rating task(Experiment 1). This is consistent with Mc-Donald et al.’s (1993) conclusion that animacyinfluences grammatical function assignment butnot word order per se.

A remarkable feature of the data is their con-sistency across several different task variations.The determinants of shifting were the same,independent of whether speakers committed inadvance to producing the NP and PP in a par-ticular order (Experiment 1), merely indicatedreadiness to speak the sentence (Experiment 2),or prepared to reproduce the NP and PP frommemory (Experiment 3). In addition, rated pref-erences for shifted and unshifted sentences (Ex-periment 1) patterned like the production data.This consistency can be interpreted in twoways. It can be argued that all of the tasksreflect a late “postsyntactic” editing process of akind that is likely to be exercised in deliberatewriting more than in spontaneous speech. Al-ternatively, it can be argued that all of the tasksreflect the processes that give rise to heavy-NPshift in spontaneous production. In this view,heavy-NP shift is not a mere stylistic nicety butinstead reflects fundamental processes of sen-tence formulation. Although we acknowledgethat both interpretations are currently sustain-able, we suggest that to the extent that the verbdisposition hypothesis is supported, the balanceof plausibility favors the fundamental processview. This is because there is no obvious mech-anism in the postsyntactic editing account thatcould account for the kind of detailed sensitivityto the history of prior use that is implied by theverb disposition hypothesis. We now discusssome of the more salient outcomes and impli-cations of our findings in more detail.

Word and Phrase Order

We observed a strong effect of NP length onthe probability of choosing the shifted PP–NP


phrase order. In contrast, McDonald et al. didnot find effects of word length on word order.This seeming discrepancy is not surprising,however, if word and phrase length effects ariseat different stages of production, as indeed theyappear to do. In standard theories of sentenceproduction, phrase ordering should occur at anearly stage, before the phonological propertiesof words become fully available, so that itshould be sensitive to aspects of syntactic–pro-sodic planning that precede retrieval of wordphonology. Earlier we suggested that the mostlikely locus is late in the functional or very earlyin the positional substage of grammatical en-coding (e.g., Bock & Levelt, 1994), and cer-tainly preceding determination of word orderand phonological encoding. This implies thatthe “heaviness” of a long noun phrase must beassessed in terms of conceptual, syntactic, orprosodic complexity rather than in terms of thelength in phonological units of the words thatcomprise it. In support of this conclusion, initi-ation times of prepared utterances are sensitiveto complexity as indexed by number of words,but not to wordlengthper se (see Sternberg etal., 1978). Further, when utterance length isheld constant, number of phonological wordsrather than length in units below the level of thephonological word determines initiation time(Wheeldon & Lahiri, 1997). Thus, to paraphraseRoss (1967) and McDonald et al. (1993), wordand phrase ordering may pertain to differentparts of the syntactic world.

Verb Disposition

The claim that speakers have access to infor-mation concerning how frequently a verb hasappeared in a syntactic structure is broadly com-patible with current approaches to production inwhich sentence construction is lexically driven(e.g., Bock, 1987b; Bock & Levelt, 1994; Lev-elt, 1989), but it has not been explored in detailin the production literature. By contrast, contin-gent frequency information—the frequencywith which a word appears in a particular syn-tactic context—has become increasingly impor-tant in theories of sentence comprehension(MacDonald et al., 1994; Tanenhaus & True-swell, 1995). For example, MacDonald et al.

(1994) suggested that the representations ofverbs include information about how frequentlya verb has appeared in different environments,including alternative tenses, active vs passivevoice, and alternative argument structures suchas taking an NP complement vs an S-comple-ment. They assumed that the frequency infor-mation is represented via weighted links to rep-resentations in the lexicon, (e.g., a verb that istypically used in the active voice would have astrong link to the “active” representation and aweak link to the “passive” representation,whereas a verb that is used more often in bothvoices would have more even links to eachrepresentation of voice). The weights on theselinks determine the extent to which alternativeinterpretations, such as active and passive, willbe activated when a verb is encountered.

Although properties of individual verbs havenot received much attention in the production lit-erature, the work presented here is part of a grow-ing body of research that investigates the effects ofverbs on choice of syntactic structure (e.g., F.Ferreira, 1994; V. Ferreira, 1996). These otherstudies differ from the present one in that theyinterpret effects as a function of whole classes ofverbs, such as alternating datives (V. Ferreira) andtheme-experiencer verbs (F. Ferreira), whereas wehave suggested that an individual verb’s historyinfluences production. At this point, our data arestill open to a verb-class explanation: There couldbe something special about NP/S verbs as a classthat makes them more likely to shift than NP-onlyverbs. For example, the fact that NP/S verbs havemultiple argument structures (taking both NPs andS-complements) might somehow make theseverbs more amenable to shifting, or perhaps thefact that NP/S verbs form a coherent semanticclass, typically expressing cognition or communi-cation of knowledge, might cause these verbs toshift more than NP-only verbs. Hypotheses of thissort would not entail any claim that the frequen-cies of syntactic configurations into which theyenter are represented with individual verbs. Con-versely, previous results that have been interpretedto indicate verb class effects (F. Ferreira, 1994; V.Ferreira, 1996) could be given an individual-verbinterpretation. Additional research is necessary to


determine which alternative is preferable for anyof these findings.

Assuming for now that structural frequencyis represented with individual verbs, it will beimportant to account for why some verbs tend tooccur more often in nonadjacent structures thanothers. There may be several different explana-tions for these frequency differences. These in-clude the increased opportunities that NP/Sverbs have to appear in nonadjacent structures,compared with NP-only verbs. Moreover, someverbs may tend to appear with long NPs, orprosodically prominent NPs, or NPs that typi-cally convey new information, so that they haveparticipated in heavy-NP shift more than otherverbs. Ultimately, we would like to be able toprovide a complete account of the factors thatpromote shifting in the majority of environ-ments. However, our current proposal merelystates that whatever these factors may be, thehistory of nonadjacent structure usage is repre-sented with individual verbs, and this historyaffects the extent to which a verb will engage inheavy-NP shift. The fact that these preferences,acquired in other contexts, were expressed inour experiments even though participants wereprovided with sentence ingredients and there-fore knew they were dealing with an NP struc-ture, appears to be rather strong evidence of theforce of the verb effects. This, in turn, suggeststhat knowledge of what constituents may bedeferred, regardless of the reason, is of vitalinterest to speakers. Heavy-NP shift may pro-vide a particularly sensitive test of the verbdisposition hypothesis because it involves anordering preference independent of syntacticrole assignment, so that verb preferences maybe less constrained than in other contexts.

Our claims for the importance of nonadja-cency bear some similarity to Wasow’s (1997)observations concerning collocations betweenverbs and PPs. Wasow suggested that the fre-quency of shifting rests in part in the frequencyof V–PP phrases such askeep in mindandtakeinto account.In other words, Wasow proposeda frequency-based account focusing on the ad-jacency of the verb and PP, in contrast to ournonadjacency hypothesis. We do not doubt thatthe frequency of co-occurrence of adjacent

phrases plays an important role in production;for example the frequent production of colloca-tions such askeep in mindcould make the PPinmind quite accessible when the verbkeep isreadied for production. There also appears to bea definite role for nonadjacency in shifting,however. A simple adjacency approach predictsthat optionally intransitive verbs, such asrun,move, walk, drive,etc., which frequently occuradjacent to PPs such asto the store,should havea disposition to shift by virtue of the frequentV–PP co-occurrence. The nonadjacency ac-count does not make this prediction, because theV–PP sequences in intransitive sentences suchasMary walked to the storedo not contain therequisite nonadjacent verbal complement. Inother corpus based work (MacDonald et al.,1998), we have found that optionally intransi-tive verbs appear in shifted structures extremelyrarely, suggesting that the mere co-occurrenceof verbs and adjacent PPs does not fully deter-mine a verb’s shifting disposition.

Competition in Production

While we do not deny the important role ofaccessibility of constituents such as NPs in thechoice of syntactic structure, our data point to aclear competitive component in the ordering ofconstituents as well. The competition in thiscase is tied to individual verbs, specifically thefrequency with which each verb participates inthe adjacent-complement and nonadjacent com-plement structures. On this view, verbs thatrarely appear in nonadjacent structures engen-der little competition between alternatives,whereas verbs that more frequently appear innonadjacent structures should engender partialactivation of the alternative structures and thepotential for competition among them. Compe-tition should have two results: (a) effects ondecision times and (b) greater sensitivity to con-textual factors that might promote one or theother structure. We see (a) in the effects of verbdisposition on preparation time, especially inExperiment 1, where verbs with stronger shift-ing dispositions yielded longer decision timesfor the basic order and relatively short decisiontimes for the shifted order compared to verbswith weaker shifting dispositions. One can see


result (b), greater sensitivity to context, in theinteraction between verb disposition and NPlength in our choice data and also in F. Fer-reira’s (1994) data. The theme-experiencer verbstested in Ferreira’s study have a higher fre-quency usage in the passive voice than simpletransitive verbs. The manipulation of anotherfactor, subject NP animacy, had a greater effecton choice of active vs passive syntactic struc-ture for the theme-experiencer verbs than for thesimple transitive verbs. We attribute the theme-experiencer verbs’ sensitivity to animacy to thepartial activation of the active and passive voicefor these verbs. The simple transitive verbs, bycontrast, have little activation of the passivevoice and therefore the animacy manipulationdid little to promote the use of passive structureswith these verbs. This result is strongly remi-niscent of verb–context interactions in languagecomprehension, in which the degree to whichcontextual manipulations, such as noun ani-macy, affect syntactic ambiguity resolution var-ies as a function of the ambiguous verb in thesentence, specifically the frequency with whichthe verb participated in alternative syntacticstructures (Garnsey et al., 1997; MacDonald,1994; Trueswell, Tanenhaus, & Garnsey, 1994).

The evidence of competition also makes pre-dictions for two different kinds of speech error.Analyses of speech errors (e.g., Garrett, 1975,1988) have generally operated within the frame-work of canonical phrase orders and so providelittle guidance in theorizing about the specific caseof heavy-NP shift. However, Arnold et al. (1997)have plausibly suggested that speech disfluenciesmay be diagnostic of sentence formulation diffi-culties that could result in heavy-NP shift. In ad-dition, our verb disposition findings make a spe-cific prediction for another well known class ofspeech error, whole word exchanges. Exchangesbetween same category words (e.g., nouns, adjec-tives, etc.) should be proportionate to the degree ofcompetition between the phrases to which theybelong in syntactic planning. For example, ex-changes between

phrasal ordering constraints in sentence production...

Documents