Semantic Factors in the Production of Number Agreement

Semantic Factors in the Production of NumberAgreement

Jason Barker1,3, Janet Nicol,1,2 and Merrill Garrett 1

This paper examines the role of semantic factors in the production of subject–verb numberagreement. As an ostensibly grammatical process, number agreement provides an interestingcase for examining the flow and interaction of semantic and syntactic information through thelanguage-production system. Using a sentence-completion task, agreement errors can be elicitedfrom subjects by presenting them with sentence fragments containing a complex noun-phrase, inwhich the nonhead noun is plural (e.g., The key to the cabinets. . . WERE missing.).Previousresearch has demonstrated that the probability of making an error can be affected by varying theproperties of the nouns in the complex noun phrase. By investigating which variables do and donot affect error rates, constraints on the flow of information through the production system canbe inferred. In three experiments, we investigated the possible effects of three different semanticmanipulations of the nouns in the complex NP: animacy, semantic overlap, and plausibility ofmodification by the sentence predicate. We found that both animacy and semantic relatednesshad reliable effects on error rates, indicating that the mechanism involved in implementingagreement cannot be blind to semantic information. However, the plausibility with which eachnoun could serve as the subject of the sentence predicate had no effect on error rates. Takentogether, these results suggest that while semantic information is visible to the agreement mech-anism, there are still constraints on when this information can affect the process. Specifically, itmay be the case that only information contained within the complex NP is considered for the pur-poses of implementing agreement.

KEY WORDS: language production; agreement; speech errors.

91

0090-6905/01/0100-0091$19.50/0 © 2001 Plenum Publishing Corporation

Journal of Psycholinguistic Research, Vol. 30, No. 1, 2001

This research was supported in part by National Multipurpose Research and Training CenterGrant #DC-01409 from the National Institute on Deafness and Other CommunicationDisorders and by the Cognitive Science Program and Department of Psychology at theUniversity of Arizona.1 Department of Psychology, University of Arizona, Tacson, Arizona 85721-0068.2 Department of Linguistics, University of Arizona, Tuscon, Arizona 85721-0068.3 To whom all correspondence should be addressed. e-mail: [email protected].

INTRODUCTION

In the tradition of looking at the distributions of speech errors as a windowinto the architecture of the language-production system (e.g. Garrett, 1975;Cutler, 1988; Levelt, 1983), experimental investigations of subject–verb agree-ment errors have provided a useful paradigm for gathering production data(e.g. Bock & Miller, 1991; Bock & Eberhard, 1993; see Bock, 1995 for areview). Agreement is a particularly worthwhile aspect of speech production toinvestigate since it is an excellent example of one of the main computationalproblems the production system must tackle, namely, specifying and main-taining the structural relationships between lexical items, which may be sub-stantially temporally separated from one another.

In a typical experiment, subjects are provided with sentence fragmentscontaining a complex noun phrase (e.g. The key to the cabinets. . .), whichthey are then to repeat, and provide completions for (e.g., The key to thecabinets was lost). Manipulations of the properties of the complex NP pro-vide the independent variables, while the number of agreement errors pro-duced provides the dependent variable. These errors are often termed“errors of attraction,” based on the view that the nonhead noun “attracts”agreement away from the true head of the phrase.

The most basic finding in this literature is an asymmetry in the preva-lence of errors: they are much more common when the head noun is singularand the nonhead (hereafter called the “distractor” noun) is plural (e.g., Thekey to the cabinetsWERE missing) than in the reverse case (e.g., The keys tothe cabinetWAS missing) (Bock & Miller, 1991; Bock & Eberhard, 1993;Eberhard 1997). The proposed explanation for this asymmetry is that nounsare singular by default and must be specifically marked as plural by the addi-tion of a plural feature (Bock & Miller, 1991; Bock & Eberhard, 1993;Eberhard, 1997). Preambles in which the head noun is singular, and henceunmarked, are vulnerable to interference from any plural feature, which mayexist on other nouns within the complex NP. Sentences in which the headnoun is plural are less prone to interference for two reasons: the head noun isalready marked, and a singular distractor noun, being in its default unmarkedstate, would not possess a disagreeing “singular” feature.

The exact mechanism by which normal agreement, and by extension,errors, occurs has not yet been fully fleshed out. Vigliocco and colleagues(Vigliocco, Butterworth, & Semenza, 1995; Vigliocco & Nicol, 1998) haveput forward a “feature-percolation” model, which proposes that the number ofthe verb is specified by a check of the features that exist at the highest nodeof the complex NP. Normally, a plural subject transmits (or “percolates”) itsplural feature up to the highest node, where it is then copied to the verb. In

92 Barker, Nicol, and Garrett

Semantics and Agreement 93

the case of errors, the plural feature on the distractor noun is mistakenly trans-mitted up to highest NP node, which then leads the system to mistakenly gen-erate a plural verb. Eberhard (1997) extended this proposal by casting theprocess within an activation framework, helping to determine an explanationof why the plural feature on the distractor noun can be mistakenly associatedwith the head N-P node. Two aspects of Eberhard’s proposal are importanthere. First, features associated with the heads of noun phrases are assumed tobe more highly activated than those associated with nonheads. Second, anyrelevant features associated with nonheads are assumed to create noise in thesystem, which increases the probability that the agreement mechanism willmake an error when trying to assess the existence of activated features on thehead NP node. This leaves us with a probablistic view of the occurrence ofagreement errors, where various lexical, structural, or conceptual factors mayact to increase or decrease the amount of noise created by plural featuresassociated with nonheads, and thus increase or decrease error probabilities.The exact architecture of and flow of information through the production sys-tem will dictate which kinds of information can and cannot create interferenceand thus by investigating the factors which affect error rates. inferences canbe made about how the system must be organized.

Several such factors have been investigated, although not always withinthe activation framework outlined above. Early work in this area (Bock &Miller, 1991; Bock & Cutting, 1992; Beck & Eberhard, 1993) suggestedthat agreement in English was almost exclusively governed by syntactic fac-tors, with semantic and phonological factors being invisible to the process.This was taken as evidence for quite constrained interactions between dif-ferent types of information within the production system (Bock & Miller,1991). For instance, the use of “pseudoplural” words (words ending withan/s/ or /z/ sound) in the distractor noun position failed to elicit any agree-ment errors at all (e.g., The player on the course. . ., indicating that low-level phonological cues were not driving the agreement process (Bock &Miller, 1991). In another experiment, Bock and Miller (1991) investigatedthe possible effects of conceptual factors in agreement by comparing singlevs. distributed referent noun phrases, as in (1) and (2):

(1) Single referentThe key to the cabinets. . .

(2) Distributed referentThe label on the bottles. . .

In both (1) and (2), the head NP is syntactically singular, since the headnoun is unmarked. However, (2) is conceptuallyplural, in that it most plau-sibly refers to the same type of label on many bottles. This is in contrast to

(1), which clearly refers to one key that opens numerous cabinets. No dif-ference in error rates was found between the two sentence types and thiswas taken as evidence that semantic information was not accessible to theagreement mechanism.

In contrast, syntactic manipulations exert a marked effect on the agree-ment process. For instance, Bock and Cutting (1992) found that if the dis-tractor noun is separated from the head noun by a clause boundary (e.g., Themessage that they expelled the students. . .) error rates decline substantially.Further, Nicol and Barker (1996) demonstrated that even without crossing aclause boundary, error rates are reduced if the distractor noun does not havea close syntactic link with the head noun (e.g., The pilot at the airbase nearthe hangars. . .). Taken together, these findings suggest that agreement inEnglish is primarily a syntactic process, which can be affected by syntacticvariables but is largely blind to nonsyntactic ones (e.g., conceptual number,phonology).

However, as often happens, further research complicated the pictureand in this domain has made the case for agreement as a “syntax only”process difficult to maintain. Bock and Cutting (1992) found that concep-tual factors could indeed affect the agreement process under certain condi-tions. They found that error rates were higher for sentences such as (3),which contained singular collective head nouns, than for those such as (4),which contained singular count nouns:

(3) The jury for the trials. . .(4) The judge for the trials. . .

This was taken as evidence that conceptual factors can indeed sometimesgovern the agreement process. Further, using the items from the Bock andMiller’s (1991) singular vs. distributed referent manipulation, Vigliocco andcolleagues were successful in demonstrating a conceptual effect in Italian(Vigliocco, Butterworth & Semenza, 1995), Spanish (Vigliocco, Butterworth,& Garrett, 1996), and Dutch (Vigliocco, Hartsuiker, Jamera, & Kolk, 1996b).Eberhard (1999) was able to demonstrate the effect in English and explainsthe prior difficulty in demonstrating this effect in English as being relatedto the concreteness and imagibility of the distributed referent NPs (cf.Nicol, Teller, & Greth, 2000).

What do these findings indicate about the production system? Eberhard(1997) argues that the fact that both grammatical and conceptual numbercan affect agreement places the agreement process at the functional level ofprocessing (Garrett, 1988; Bock & Levelt, 1994). At the functional level,the utterance-to-be is represented as lemmas—abstract word representationsthat contain semantic information and basic syntactic category information,and markers delineating roles such as subject, object, etc. Importantly, ele-



ments at the functional level are not yet ordered in any linear sense; thistakes place at the next level of processing, the positional level.

On balance, research to date supports the proposal that agreement can begoverned both by syntactically and conceptually specified number informa-tion and this argues against claims that the interaction between syntactic andsemantic information within the production system is highly limited. Note,however, that the type of semantic effect, which has been demonstrated, isstill critically tied to information about the plurality of the head noun phrase.That is, the question has been whether a phrase, which is grammaticallymarked as singular, can still be construed for the purposes of agreement asplural if the phrase as a whole implies plurality (e.g., as in “The picture onthe postcards. . .” or “The jury at the trial. . .”).

A separate question, and one that is of specific interest if one assumes anactivation based framework for agreement (such as that put forward byEberhard, 1997), is whether semantic factors not involving number per secanattract agreement errors away from the head noun in the same way that syn-tactic factors can. As was discussed above, simply having a plural feature onthe distractor noun will cause a certain percentage of errors (e.g. Bock &Miller, 1991). This percentage can be reduced if the distractor noun is placedin a position that makes it syntactically more distant from the head (e.g. Bock& Miller, 1991, Nicol, 1995; Nicol & Barker, 1996. However, can the seman-tic properties of the sentence affect error rates in this way? This is quite a sep-arate issue from the question of “conceptually mediated agreement” as it hastypically been addressed in the relevant research, because it is not concernedwith the contributions of other sources of information about number (e.g. con-ceptual), but rather with what kinds of factors can cause the system to giveundue weight to number features present on the distractor noun. In otherwords: what types of factors can increase the likelihood that the agreementmechanism will select its number properties from the wrong noun in thecomplex noun phrase? This is the specific question that will be addressed inthis paper.

One previous study that did address this question was Bock and Miller’s(1991) second experiment, in which they examined the role of animacy in theagreement process. Their reasoning was as follows: animacy is highly corre-lated with subjecthood (Clark, 1965; MacWhinney, 1977); therefore, animacycan be considered a good “cue” for subjecthood. If the agreement processidentifies sentence subjects, at least in part, by reference to the kinds of“primitive” semantic features they possess, rather than solely by reference tothe role they have been assigned at the functional level of processing, then itshould be the case that distractor nouns that possess semantic features (suchas animacy) that are highly correlated with subjecthood should be more likelyto lead to agreement errors than those that do not. Bock and Miller (1991)


investigated this possibility by comparing sentences with animate head or dis-tractor nouns, as in (5) and (6):

(5) AI: (animate head–inanimate distractor)The author of the speeches. . .

(6) IA: (inanimate head–animate distractor)The speech by the authors. . .

They found no difference in error rates between these two/types of pream-ble and concluded that the agreement process does not make reference tosemantics, but rather only operates on the basis of the syntactic role (e.g.,subject, object) that has been specified at the functional level of processing.

However, it is possible that there are, in fact, effects of animacy thatthe Bock and Miller (1991) study may have missed. This is because that studyonly looked at a subset of the possible combinations of animacy (animate–inanimate; inanimate–animate). Since animacy is so intricately linked withsubjecthood, it is possible that the baseline conditions for each of these twoconditions are different; that is, sentences in which both nouns are animatemay differ from those in which both are inanimate. This data would be neces-sary in order to fully interpret the animacy-mismatch conditions presented inBock and Miller (1991). Further, since animacy is so highly correlated withsubjecthood, it could be the case that most of its effects will be centered noton the animacy of the distractor noun per se, but rather of the subject nounitself. Bock and Cutting (1992) found a result somewhat similar to this whenthey demonstrated that singular collective nouns did not serve well as error-inducing distractor nouns (e.g., The trial by the jury. . .), but when the samesingular collective noun was used as the subject, in conjunction with a syntac-tically plural distractor noun, error rates do increase relative to sentences usinga simple count noun as the subject, as in (7) and (8), respectively

(7) Collective noun subjectThe jury at the trials. . .

(8) Count noun subjectThe judge at the trials. . .

Our main goal in this study was to test a set of items containing a full coun-terbalancing of the animacy variable across the subject and distractor nounsin the complex NP.

EXPERIMENT 1

The purpose of the first experiment was to investigate the role of ani-macy in agreement, by using a set of items, which consisted of all possible

combinations of animacy. There are three core predictions for these results.First, following the logic in Bock and Miller (1991), it is possible that itemswith animate distractor nouns will have higher error rates than those withinanimate distractors. On the other hand, it could be that it is the animacy ofthe subjectnoun that is of most importance, predicting that preambles withanimate subject nouns would be less prone to error than those with inanimatesubject nouns because, due to the correlation between animacy and subject-hood, animates make better subjects. In other words, it is not that animatenouns prefer to be subjects, but rather that subjects prefer to be animate.Finally, it could be that there exists an interaction between these two factors,whereby the animacy of both nouns is relevant to error rates. For instance,while Bock and Miller (1991) found that IA and AI items were equivalent, itcould still be the case that IA items have higher error rates than II items, andthat AI items have fewer errors than AA items.

Method

Participants

Sixty-four University of Arizona undergraduates participated in thestudy for course credit.

Materials

Twenty-four sentence preamble octuplets were constructed in whichthe subject noun was singular, the distractor noun was either singular orplural, the subject noun was either animate or inanimate, and the distractornoun was either animate or inanimate. Each octuplet was assigned two tar-get predicate words, one for the animate-headed items and one for the inan-imate-headed items. This was necessary due to the difficulty of findingpredicates that work equally well with both animate and inanimate nouns.An example octuplet can be seen in (9). The sentence codes are as follows:(AA) = animate subject; animate distractor; (AI) = animate subject, inani-mate distractor; (II) = inanimate head, inanimate distractor, (IA) = inani-mate head, animate distractor, (ss) = singular subject, singular distractor;(sp) = singular subject, plural distractor).

(9a) AA-ss The girl behind the teacher. . . smart(9b) AA-sp The girl behind the teachers. . . smart(9c) AI-ss The girl behind the desk. . . smart(9d) AI-sp The girl behind the desks. . . smart(9e) II-ss The blackboard behind the desk. . . erased(9f) II-sp The blackboard behind the desks. . . erased


(9g) IA-ss The blackboard behind the teacher. . . erased(9h) IA-sp The blackboard behind the teachers. . . erased

These items were counterbalanced across eight presentation lists, such thateach variant appeared on a different list. Due to the large number of itemsand conditions in this study, the counterbalancing was done in such a waythat head noun animacy was a between subjects variable and this allowedus to only conduct items analyses on the data. For reasons we will discussbelow, we do not believe that this factor influenced our results. In additionto the experimental items, there were also 64 other items that comprised aseparate experiment (not reported on herein) and a number of filler items toequalize certain counterbalancing factors.

Procedure

All items were displayed visually in the following manner: a fixationcross was first presented in the center of the screen for 300 msec. This wasthen replaced by the target predicate word, which remained on the screenfor 600 msec. Finally, this was replaced by the sentence fragment, whichremained on the screen for 1200 msec. A blank screen was displayed forroughly 100 msec in between each change in the visual display. Subjects wereinstructed to read the target predicate word and sentence fragment silently tothemselves and to then speak aloud a complete sentence, beginning with thesentence fragment and including the predicate word somewhere within thecompletion. Subjects moved on to the next item by pressing the space bar ofan ordinary keyboard and no constraints were placed upon the length of theirresponses or the speed at which they proceeded to the next item. The itemswere presented on a computer-controlled display using the DMDX system(developed by J. I. Forster & K. I. Forster at the University of Arizona).Subject responses were recorded on audio cassette tape.

Scoring

Subject responses were transcribed by research assistants, who wereunaware of the specific manipulations involved in the study. Responseswere then scored by placement into four categories: Correct (C) the subjectrepeated the sentence fragment accurately, included the target predicate intheir completion, and used a correctly inflected verb; agreement error (AG),as above, except that the subject used a verb, which did not agree in num-ber with the subject; uninflected (U), As above, except that the subject useda verb, which was uninflected for number; other (Oth), this category encom-passed fragment repetition errors and other miscellaneous types of errors,which made the item uncodable within the three other categories.



Results

The results can be seen in Table I. A two (head noun animacy, animatevs. inanimate) by two (distractor noun animacy, animate vs. inanimate) by two(distractor noun number, singular vs. plural) ANOVA was conducted on eachof the four response categories.

For (AG), there was a main effect of distractor noun number [F(1,46) =59.22, p < .001], an interaction of distractor noun number and subject nounanimacy [F(1,46) = 5.33, p < .03], and an interaction of subject and distrac-tor noun animacy [F(1,46) = 3.84, p < .05]. There was also a near significantmain effect of subject noun animacy [F(1,46) = 3.69, p < .06] and a near sig-nificant three-way interaction of all variables [F(1,46) = 3.51, p < .06].4 Forthe (C) responses, there was only a main effect of distractor noun number[F(1,46) = 37.43, p < .001]. No significant differences existed for the (U)responses, and only a main effect of distractor noun number was found for(Oth) responses [F(1,46) = 12.83, p < .01]. This latter effect is due mostly tothe fact that speakers are much more likely to omit a plural ending on a noun,than they are to add one.

Discussion

The main effects in our data seem to indicate clearly that the agreementprocess is sensitive to animacy variables. However, the interaction of subject

Table I. Subject Responses for Experiment 1.a

AG C U Oth

AA-ss The girl behind the teacher . . . 1 156 7 28AA-sp The girl behind the teachers. . . 14 133 9 36AI-ss The girl behind the desks. . . 0 153 8 31AI-sp The girl behind the desk. . . 8 138 7 39II-ss The blackboard the desk. . . 0 151 8 33II-sp The blackboard the desks. . . 25 121 6 40IA-ss The blackboard behind the teacher. . . 0 154 11 27IA-sp The blackboard behind the teachers . . . 14 124 6 48

a Scoring codes at are as follows: (C) correct; (AG) agreement error; (U) uninflected; (Oth)other; (ss) singular subject, singular distractor; (sp) singular subject, plural distractor.

4 Due to the fact that there was only one error in the SS (singular subject noun, singular dis-tractor noun) across the entire experiment, it is simpler to deal with the two-way interactionthat omits distractor noun number from the equation. Since the baseline in all four animacyconditions is essentially zero, it seems unnecessary to consider the SS conditions in the inter-actions and therefore, these will not be dealt with further.

and distractor noun animacy is not easily explained within the predictionswe laid out for the experiment. We will take each of these findings in turn.First, it is clear that the animacy of the distractor noun, in and of itself, doesnot affect error rates. In line with Bock and Miller’s (1991) original find-ings, animate distractor nouns were no more likely to cause errors thaninanimate ones, as evidenced by the complete lack of a main effect fordistractor noun animacy (if we collapse across conditions, we see that itemswith animate distractor nouns led to 28 errors, while those with inanimatesled to 33). On the other hand, the animacy of the subject noun does appearto play a role, as there was a marginal main effect of head noun animacy,and a significant interaction with distractor noun animacy (collapsing acrossconditions, we see that items with animate subjects led to 22 errors whilethose with inanimate subjects led to 39 errors). Having an animate noun insubject position appears to provide some insulation against the possibility ofmaking an error. Regarding the counterbalancing of head noun animacy, wedo not believe this affected our results for two reasons. First, there are noeffects of animacy for either the (U) or (OTH) response categories. If thetwo groups of subjects were behaving differently, it is unclear why this dif-ference would manifest itself only in the number of agreement errors made,and not in other response categories. Second, the filler items were identicalon all lists and there were no differences in performance on these itemsacross the two groups of subjects.

The interaction between subject and distractor noun animacy is moredifficult to account for, since the pattern of results does not fit the kind ofinteraction that we would have predicted based on our initial assumptionsabout the role animacy could be playing. For instance, an interaction inwhich animate distractor nouns caused more errors only for sentences withinanimate subjects would be easily explainable, since this follows from ourpredictions about the role animacy might be playing in each noun position.The pattern of results we obtained paints a very different picture though, inthat it is the (II) condition that leads to, by far, the most errors. None of ourinitial assumptions would predict that this condition should be the mostproblematic for subjects and this caused us to take a second look at theseassumptions.

By using a fully counterbalanced set of animacy items, we introduced apossibly confounding factor into the manipulation. Animacy is but one se-mantic feature and it is likely that pairs of animate or inanimate nouns mayhave considerable overlap of other semantic features as well. Of specificimportance here is that, on average, pairs of animate or inanimate nouns willhave substantially more overlap than pairs containing one animate and oneinanimate noun. It is possible that the degree of semantic overlap between thesubject and distractor nouns may be an independent source of variance in


error rates. The argument for animacy as a potential variable in agreement(put forward by Bock & Miller, 1991) suggests that the production systemmay identify subjects, at least partly, by reference to a statistical record ofwhat kinds of semantic features are usually associated with subjecthood. Anadditional question is whether, in any specific sentence,subject identificationcan be made more difficult when the subject noun shares many semantic fea-tures with the distractor noun. This question gets at the issue of the type anddegree of interaction between lexical representations at the functional level,rather than the issue of statistical biases in the way the system functions as awhole. If the degree of semantic overlap between nouns in the complex nounphrase, indeed, plays a role in the agreement process, then this factor willlikely have contributed variance to the present experiment and may help toexplain the present pattern of results. Therefore, in Experiment 2, we specif-ically investigate the role of semantic overlap in the agreement process.

EXPERIMENT 2

The purpose of this experiment was to investigate whether the degreeof semantic overlap between subject and distractor nouns in a complex NPwould have an effect on the likelihood of an agreement error during sen-tence production. If the agreement process is sensitive to this factor, thiswill be relevant to our interpretation of the results of Experiment 1.

Method

Participants

Forty-eight University of Arizona undergraduates participated in thisstudy for course credit.

Materials

Thirty-two sentence fragment quadruplets were constructed in whichthe subject noun was singular, the distractor noun was either singular orplural, and the subject and distractor nouns had either high- or low-semanticoverlap. High-overlap nouns were those which had substantial overlap ofsemantic features (e.g., sailboat/canoe). These ranged from synonyms tosimply superordinate or subordinate category relations. Each quadruplet wasalso assigned an adjective or past-tense verb, which would serve as the tar-get predicate for the sentence completion the subjects would be asked toprovide. An example quadruplet is shown in (10). The new sentence codes



are as follows: high = high degree of semantic overlap; low = low degreeof semantic overlap).

(10a) ss - high The canoe by the sailboat. . . . damaged(10b) sp - high The canoe by the sailboats. . . damaged(10c) ss - low The canoe by the cabin. . . damaged(10d) sp - low The canoe by the cabins. . . damaged

The plausibility with which each noun could be related to the target predicatewas kept roughly equal across all the items and nearly all of the items con-tained nouns that were either both animate or both inanimate. These itemswere counterbalanced across four presentation lists such that each variantappeared on a different list and equal tokens of each sentence type appearedon each list. In addition, there were 48 other items, which were part of a sep-arate experiment (not reported herein), and 10 filler items, which were usedto equalize our counterbalancing across the entire experiment.

Procedure and Scoring

Both were equivalent to Experiment 1.

Results

The results can be seen in Table II. Two analyses of variance were con-ducted, one with subjects (F1) and one with items (F2) as the random variable.A 2 (distractor noun number, singular vs. plural) × 2 (degree of semantic over-lap, high vs. low) ANOVA conducted on the (AG) class of responses revealeda significant main effect of distractor noun number [F1(1,45) = 47.57, p <.001;F2(1,31) = 64.88, p < .001), semantic overlap [F1(1,45) = 14.03, p <.001;F2(1,31) = 10.19, p < .003], and an interaction of the two [F1(1,45) = 9.24,p < .004; F2(1,31) = 12.02, p < .002). A similar analysis on (C) responsesrevealed essentially identical results; a significant main effect of distractor

Table II. Subject Responses for Experiment 2a

AG C U Oth

ss-high The canoe by the sailboat. . . 11 326 1 46sp-high The canoe by the sailboats. . . 84 223 0 77ss-low The canoe by the cabin. . . 10 328 1 45sp-low The canoe by the cabins . . . 48 264 0 72

a Scoring codes are equivalent to Experiment 1 with the addition of: (high) = high semanticoverlap; (low) = low semantic overlap.


noun number [F1(1,45) = 63.74, p < .001; F2(1,31) = 58.9, p < .001), seman-tic overlap [F1(1,45) = 10.53 p < .002; F2(1,31) = 9.02, p < .005], and an inter-action of the two [F1(1,45) = 5.27, p < .03; F2(1,31) = 4.89, p < .05]. Therewere too few (U) responses to warrant an analysis. Analysis of the (Oth)responses revealed only a main effect of distractor noun number [F1(1,45) =25.31, p < 001; [F2(1,31) = 27.02, p < .001], with all other F’s < 1.3. Thiseffect is again due simply to the fact that subjects are much more likely to omitan existing plural feature on the distractor noun than they are to add a pluralfeature, which was not present in the preamble.

Discussion

These results clearly support the notion that the degree of semantic over-lap between nouns within the complex noun phrase can affect error rates.Subjects were nearly twice as likely to make an agreement error when thenouns had a high degree of semantic overlap than when they had a lowdegree of overlap. One possible alternative explanation for this finding is thatthe difference in error rates had nothing to do with the effects of semantics onthe agreement process per se, but rather was due simply to a general increasein difficulty or memory load caused by the ease of confusability of the twonouns. However, if this were true, one would expect that all types of errors(e.g., repetition errors, nonsensical sentence completions) should be morecommon for the high semantic overlap items, rather than being restricted toonly agreement errors. However, we found no effect of semantic overlap for(OTH) responses. Increased difficulty was demonstrated only with regard to(AG). The strong effect of semantic overlap demonstrated here necessitates areconsideration of the results of Experiment 1 and we believe that this find-ing, in fact, makes the results of Experiment 1 much clearer.

REEVALUATION OF EXPERIMENT 1

Recall that Experiment 1 left us with an interaction of subject and dis-tractor noun animacy that was not easily explainable within our existingassumptions about the possible role that animacy might play in agreement.Most striking was the unexpected finding that the (II) condition led to themost agreement errors (see Table I for the overall results). While the interac-tion was difficult to explain, we were left with two clear main effects: (1) theanimacy of the distractor noun was not, in and of itself, predictive of errorrates and (2) the animacy of the subject noun was predictive, with sentencepreambles with inanimate subjects leading to significantly more errors thanthose with animate subjects. Together with the results of Experiment 2, this


suggests that our characterization of the independent variables was incorrect.Rather than simply noting whether the subject and distractor nouns are ani-mate or not, what is relevant is (1) whether the subject noun is animate, and(2) whether the distractor noun matchesor mismatchesthe subject in animacy.Two nouns which match in animacy are on average going to have moresemantic overlap than two nouns which do not, and therefore, following fromthe results of Experiment 2, we should expect preambles containing nouns,which match in animacy (II, AA), to lead to more errors than those with nounsthat mismatch in animacy (IA, AI).

To investigate this possibility, we recategorized the conditions fromExperiment 1 as outlined above. A 2 (head noun animacy, animate vs. inani-mate) by two (animacy match, head and distractor nouns match vs. mismatchin animacy) 1 × 2 (distractor noun number, singular vs. plural) ANOVA onthe data from Experiment 1 revealed a near significant main effect of subjectnoun animacy [F(1,46) = 3.68, p < .06]; a main effect of animacy match[F(1,46) = 3.83, p < .05); and, of course, a main effect of distractor nounnumber [F(1,46) = 59.27, p < 001]. There was no interaction of head nounanimacy and animacy match at all (F < .49).

Interpreted in this way, the results of Experiment 1 are much more clear.Two additive factors appear to be at play in this experiment. First, items withnouns that match in animacy (we call these “+ match”) have higher error ratesthan those that do not. Second, it is also true that items, which have inanimatesubject nouns (“+ inanimate subject”) are, overall, more likely to produceerrors than those that have animate subject nouns. These two factors can addi-tively affect error rates, such that an item with the characteristics +match and+inanimate subject will be the most susceptible to errors, while an item whichis −match and −inanimate subject will be the least. Items which have eithercharacteristic will likely fall somewhere in between. This is in fact the patternof errors here and a recoded example of the data can be seen in Table III. Notethat the II condition has the most errors, the AI the least, while the other twoconditions fall in between. On-way analyses of variance confirmed that the II(+, +) condition had significantly more errors than the AI (−, −) condition[F(1,46) = 6.55, p < .02). As would be predicted, the AA and IA conditionsfell somewhere in between and did not differ significantly from either the IIor AI conditions: the II condition had marginally more errors than either theAA (−+) condition [F(1,46) = 2.43, p < .11] or IA (+−) condition [F(1,46) =2.52, p < .12] and the AI condition did not differ significantly from either theIA or AA conditions (F’s < 1.2)

Taken together, these results suggest that both semantic overlap andhead noun animacy additively affect agreement error rates. This arguesstrongly that semantic features associated with lexical items are in compe-

tition at the stage at which agreement is computed and that the agreementmechanism is sensitive to this competition.

Although it seems clear that the agreement process is sensitive to lexi-cal semantics (or at least semantic similarity of lexical items), an obviousfollow up question is whether “sentence-level” semantic information canalso affect the agreement process. What we mean by “sentence level” is thetype of semantic relationship that is expressed through the relations betweenlexical items, rather than in the lexical entries themselves. Our question iswhether semantic information of this type can have an impact on processesinvolved in grammatical encoding, such as agreement. One specific test ofthis is whether the degree to which both nouns in the complex noun phrasebear plausible relations to the sentence predicate may also be predictive oferrors. If so, this would indicate that the agreement mechanism can receiveinput from high-level semantic representations of the entire utterance, ratherthan being simply prone to interference by various lexical level semanticfactors (e.g., animacy and semantic overlap). In addition to addressing apotentially interesting question about the production system in general, it ispossible that this very factor had an effect on the results of Experiment 1:since we provided target predicates for each of our sentence preambles, sub-jects were constrained to form a sentence on the basis of that predicate. Itis possible that subjects would be more likely to make an error on items


Table III. Subject Responses for Experiment 1 with Alternate Item Codinga

Anim SubjectMatch Inamin AG C U Oth

− − AI-ss The girl behind 0 153 8 31the blackboard. . .

− − AI-sp The girl behind 8 138 7 39the blackboards. . .

− + IA-ss The blackboard behind 0 154 11 27the teacher. . .

− + IA-sp The blackboard behind 14 124 6 48the teachers. . .

+ − AA-ss The girl behind 1 156 7 28the teacher. . .

+ − AA-sp The girl behind 14 133 9 36the teachers. . .

+ + II-ss The blackboard the desk. . . 0 151 8 33+ + II-sp The blackboard the desks. . . 25 121 6 40

a Additional codes are as follows: (Anim Match +) = the two nouns match in animacy; (AnimMatch −) = the two nouns mismatch in animacy; (Subject Inanim +) = sentence subject isinanimate; (Subject Inanim −) = sentence subject is animate.

where both nouns in the complex noun phrase could plausibly be related tothe predicate. Consider the examples in (11) and (12):

(11) Both nouns plausibleThe boy near the dogs was running away.

(12) Subject only plausibleThe boy near the trees was running away.

In (11), both nouns could plausibly act as subjects for the sentence predi-cate, while in (12) only the subject noun could. If the extent of interactionwithin the production system extends beyond simple lexical level effectsand, in fact, allows the overall conceptual sentential representation to exertcontrol over the agreement process, then this plausibility variable shouldaffect agreement rates such that sentences like (11) should have higher errorrates than sentences like (12). Thornton and MacDonald (1999) have reportedpreliminary data suggesting just such an effect; however, the preliminarywork on these results are difficult to interpret at present. In addition, Greenslitand Badecker (2000) found a similar effect in comprehension, but it is notclear whether these results would also be obtained in production. We turn toan examination of this variable in Experiment 3.

EXPERIMENT 3

The purpose of this experiment was to investigate whether the agree-ment process is affected by the degree to which each of the nouns in thecomplex noun phrase bears a plausible relation to the sentence predicate.This is of immediate importance because this factor could have affected theresults of Experiment 1 and is of more general importance for demonstratingthat sentence level semantic representations can interact with the grammati-cal encoding operations such as agreement. Because we were specificallyinterested in whether this variable affected the results of Experiment 1, wechose to run this study as a posthoc examination of the items in that exper-iment. Although it is probably true that plausibility covaries with semanticoverlap and animacy (e.g., two nouns with a high degree of semantic over-lap are more likely to be able to take the same predicate than two nouns withstrongly disparate semantic features), this covariance is not at all absolute.High semantic overlap nouns can often differ strongly in terms of how plau-sibly they relate to a predicate (e.g., The woman near the girl was teaching.and low semantic overlap nouns can often be equally plausibly related to thepredicate (e.g., The boy near the tree was tall.) Because of this, our itemsfrom Experiment 1 afforded a wide range of items across this plausibilityscale.


In this experiment, we had independent raters judge the plausibility withwhich each noun in our sentence preambles from Experiment 1 fit with theprovided target predicates and then used these ratings to reevaluate the results.

Method

Participants

Twenty-five University of Arizona undergraduates participated forcourse credit.

Method

The nouns and target predicates from the items in Experiment 1 wereextracted from their sentence contexts and presented as simple word pairs ona questionnaire sheet. The first word in the pair was always a noun and thesecond was always the target predicate. Subjects were asked to rate how wellthey felt the first word in the pair “went with” the second word on a five-point scale where (1) meant “they don’t go together at all” and (5) meant“they go perfectly well together.” In order to make it clear to the subjectswhat kind of judgment we were asking them to make, they were given spe-cific examples of how we wanted them to approach the task. As an example,they were given the word pairs “dog/hungry” and “car/hungry” and told thatthey were to ask themselves, “Can a dog be hungry?”, “ Can a car be hun-gry?” and to rate how plausibly they felt they could answer YES to thosequestions. For these examples, the ratings should be (5) and (1), respectively.Ratings for each noun were then averaged across all 25 subjects.

Once all of the nouns were assigned a rating, we could then calculatea plausibility index for each item in Experiment 1. By subtracting the aver-age plausibility rating of the distractor noun from the average plausibilityrating of the subject noun, we arrived at a “plausibility difference” score,which expressed the degree to which both nouns in the sentence bore plau-sible relations to the target predicate. An example pair of sentences withplausibility ratings is provided in Figure 1.

This creates a plausibility difference scale, which can theoretically rangefrom −4 to +4, with positive scores indicating the head is most plausible andnegative scores indicating the distractor noun is most plausible. In practice,the actual range is from about −1 to +4, since our items were, of course, allconstructed such that the predicate bore at least a moderately plausible rela-tion to the head. Items with scores close to zero represent those in which bothnouns are plausibly related to the predicate, while items closer to +4 representthose items in which only the subject noun bears a plausible relation to thepredicate. If this plausibility variable is relevant to agreement, then items with


scores closer to zero should be more prone to error than items with scorescloser to +4. After assigning each item its plausibility rating, we could thencorrelate this with the proportion of errors made on each item.

Results

A scatterplot of the data can be seen in Figure 2. Because there wereessentially zero errors in the ss (singular head, singular distractor noun) sen-tence conditions, we restricted our analysis to the sp items only. As can beseen, our items provided a wide range of plausibility values, with a fairlyequal distribution of ratings across the scale. Overall, 10% of the items hada score between −1 and 0; 34% fell between .1 and 1, 13% fell between 1.1and 2, 22% fell between 2.1 and 3, and 21% fell between 3.1 and 4. As canbe seen in the figure, there was almost no relationship between the plausi-bility difference score and the percentage of agreement errors on a givenitem. A correlation analysis revealed a near zero relationship between thetwo variables (Spearman r = −.052).

Discussion

It does not appear then that “sentence-level” semantic information affectsthe agreement process. While semantic features at the lexical level demon-strated an impact in both Experiments 1 and 2, the plausibility of the NP-predicate relation did not. We can certainly be confident that this variable didnot affect our positive results in Experiment 1 and, based on the near zerocorrelation found here, it is likely that this type of information is simply notavailable to the agreement mechanism.

GENERAL DISCUSSION

First let us recap the results of the three experiments presented here.Experiment 1 demonstrated that animacy can, indeed, play a role in the agree-


Fig. 1. Sample sentences for Experiment 3 showing plausibility ratings for each noun and thecomposite plausibility difference score for each sentence.

ment process in at least two ways. First, sentence preambles with animate sub-ject nouns were less prone to agreement errors than sentences with inanimatesubjects. Second, sentence preambles in which the head and distractor nounsmatch in animacy are more prone to errors than those in which the nouns donot match in animacy. Experiment 2 demonstrated that the latter finding wasa specific example of a more general effect of semantic overlap: sentence pre-ambles in which the subject and distractor nouns have a large amount of over-lap of semantic features are more prone to error than preambles with moresemantically distinct nouns. Pairs of animate or inanimate nouns will, onaverage, have more semantic overlap than pairs of nouns that mismatch inanimacy. Taken together, these results support past research which has indi-cated that the agreement process is sensitive to semantic factors and is not a“syntax-only” process (e.g. Eberhard, 1997; Vigliocco et al.,1995) and placesthe agreement process at the functional level of processing. Further, theseresults extend past research in demonstrating that semantic factors, which donot actually involve number per se (e.g., conceptual number), can affect theprocess. The results of Experiment 3 indicate that there are constraints on thetypes of semantic factors that the agreement process is sensitive to. While lex-ical level semantic features on and between nouns in the complex NP hadreliable effects on error rates, the degree to which each noun in the NP borea plausible semantic relationship to the sentence predicate did not. This couldindicate either that this type of “sentence-level” semantic information is notvisible to the agreement mechanism, or it could indicate that the time courseof the flow of information through the production system precludes this


Fig. 2. Scatterplot for data in Experiment 3. Data shown for sp condition only.

information from affecting the agreement process. These possibilities will bediscussed in more detail below.

The results of Experiments 1 and 2 fit well within the activation frame-work for agreement proposed by Eberhard (1997). Recall that the two mostrelevant parts of that proposal here are (1) that features associated with headsare assumed to be more highly activated than those associated with nonheadsand (2) that any features associated with nonheads are assumed to create somedegree of noise in the system. Any factor which serves to increase the activa-tion of features on the distractor noun and, hence, the amount of noise presentin the system, should work to increase error rates. If the head and distractornouns in the complex NP have substantial overlap of semantic features, itcould be the case that any number features that may be present on the dis-tractor noun could be given undue weight. Similarly, if the head noun in thephrase is animate, then this may help to decrease error rates because of theincreased activation of the head noun relative to the distractor noun. If headsare typically more highly activated to begin with, the addition of a featuresuch as animacy, which is highly correlated with subjecthood, may work toboost activation rates. Note that, like the original Bock and Miller (1991)study, we did not find an effect of distractor noun animacy per se. Having ananimate noun in the distractor position did not increase error rates, as mightbe predicted if the system relies, at least in part, on correlational cues to deter-mine subjecthood. Simply being animate does not bias a noun toward being asubject, however, if a noun has been assigned the syntactic role of “subject”and it is also animate, it may be more highly activated than a subject that isinanimate. In other words, as stated in the introduction, it is not that animatenouns prefer to be subjects, but that subject nouns prefer to be animate.Thisinterpretation is supported by the results of Experiment 3 in the original Bockand Miller (1991) study. They found that in sentence preambles that containedtwo actual subject nouns (e.g., The king that the islands . . .; The song thatthe composers . . .),error rates were higher when the noun that should notgovern the first verb was animate, indicating that when faced with two subjectnouns, subjects were inclined to give more weight to the one that was animate.

While this interpretation fits the present data well, it is problematicwhen viewed alongside certain other results in the language productionerror literature. If the account offered above is correct, one might also pre-dict that word exchange errors should be more common between nouns thatshare substantial overlap of semantic features than between nouns that aremore semantically distinct. If the system can become confused as to whichnoun should be governing the number-agreement process when the nounsare semantically similar, it may also be more likely to confuse the positionsof those two nouns within the utterance. However, the data on naturally


occurring speech errors does not support this prediction. While word exchangeerrors are highly constrained by syntactic category (e.g., nouns exchange withnouns; verbs exchange with verbs) they are not constrained by semantics(e.g., highly dissimilar nouns are as likely to exchange as highly similarones) (e.g., Garrett, 1975, 1984; Bock & Levelt, 1994).

There are a number of ways to possibly reconcile this discrepancy.First, it is possible that the effects reported here are somewhat dependent onthe task involved. Experimental subjects are provided with partial sentencesand asked to complete them, rather than constructing sentences from scratchas they would normally, and this may bias the task toward creating or exag-gerating lexical effects, which may be more relevant to language compre-hension than language production. It is probably important to corroboratethese results with corpus analyses or other more naturalistic methods todetermine the robustness of this effect, considering the wealth of naturallyoccurring speech error data that exists for word exchanges.

Second, it is possible that word-exchange errors and agreement errorsare different enough phenomenon that they are not governed by exactly thesame factors. When implementing number agreement, the system is looking tothe subject noun of the head NP to govern the number marking on the verb.On occasion, for a variety of reasons, the system allows the number of the verbto be governed by the nonsubject noun (the distractor noun). It does not appearthat the system has actually misidentified the subject, since sentences in whichspeakers make an agreement error are rarely completed in such a way that theyare actually treating the distractor noun as the true subject of the sentence (thatis, people rarely say things, such as “The key to the cabinets are being refin-ished”). In the case of exchange errors, the system is not looking to one wordto govern the properties of another, but rather, it is simply trying to order theconstituents of the sentence properly. It could be the case that the process ofjuggling features between lexical items required by agreement is susceptibleto different kinds of interference than the process of ordering lexical itemswithin an utterance. The studies reported here do not allow us to explore thisquestion fully, but it will be important to reconcile this possible discrepancybetween experimentally elicited and naturally occurring speech errors.

Experiment 3 also fits well with existing research, and serves to demon-strate that there are constraints on the ways in which semantic informationcan effect the agreement process. While strictly lexical semantic overlapbetween nouns in the complex NP affected error rates in Experiments 1 and 2,the degree to which both nouns bore plausible relations to the sentence pred-icate did not. There could be at least two possible reasons for this. First, itcould be the case that this type of “sentence-level” semantic information isnot visible to the agreement process. The type of semantic overlap involved


in Experiments 1 and 2 can be quantified solely by reference to the semanticfeatures of the lexical items involved. In Experiment 3, this is not the case, asour “plausibility difference score” relied on some reference to the overall sen-tence, and not simply to the lexical items themselves. For example, the sametwo nouns may or may not bear similar relations to the predicate in differ-ent sentence contexts, as in “The boy near the tree was tall” and “The boynear the tree was smart.” While competition or interference may arise dueto simple semantic feature overlap, predicate plausibility may require refer-ence to information that is unavailable to the agreement mechanism.

Another possibility is that it is the time course of information through thesystem that is important and not the type of semantic information involvedper se. If the flow of information through the production system is incremen-tal, such that as soon as any given piece of an utterance has completed pro-cessing at one level it is sent on to the next (e.g., Kempen & Hoenkamp,1987; Kempen & Vosse, 1989), and if the agreement process is begun as soonas the subject noun reaches the functional level of processing, then it could bethe case that a factor such as plausibility simply does not typically becomeavailable during the window of time in which agreement is computed. Studiesby Nicol (1995) and Nicol and Barker (1996) provide support for this pro-posal. Both studies demonstrated that the likelihood of a plural feature on adistractor noun interfering with the correct implementation of agreement wasdependent on it being encoded closely following the head. Distractor nounswhich were distant from the head were less likely to cause errors. For instance,Nicol and Barker (1996) found that sentence . preambles such as (13) weresignificantly more prone to error than those such as (14):

(13) The pilot near the hangars. . .(14) The pilot at the airport near the hangars. . .

By this view, any information arriving at the functional level significantlyafter the head noun will be too late to interfere with the agreement processand the plausibility manipulation of Experiment 3 may be one such type ofinformation. It is worth noting that, if anything, our experimental procedureactually encouraged the use of this plausibility information, since partialpredicates were specifically provided to the subjects before they even sawthe sentence preamble. It is quite striking, then, that speakers did not appearto make any use of this information in terms of computing agreement.

CONCLUSION

Overall, these results support previous research that has indicated thatthe agreement process is sensitive to semantic factors and that agreement is


likely carried out at the functional level of processing. We extend theseresults in demonstrating that even semantic information that does not involvenumber information per se can affect agreement. The degree of semanticfeature overlap of nouns within the complex NP, in general, and the specificsemantic feature of animacy were both shown to affect error rates in a num-ber agreement error elicitation task. These results also support the proposalthat sentence production occurs incrementally and that interfering informa-tion must be available in close proximity to the subject noun in order for itto affect the agreement process, as evidenced by the lack of a plausibilityeffect in Experiment 3. These studies may also highlight the need to inte-grate the wealth of data now available on experimentally elicited agreementerrors with the existing data on naturally occurring speech errors.

REFERENCES

Bock, K. (1995). Producing Agreement. Current Directions in Psychological Science, 4,56–61.Bock, K., & Cutting, J. C. (1992). Regulating mental energy: Performance units in language

production. Journal of Memory and Language, 31,99–127.Bock, K., & Eberhard, K. M. (1993) The meaning, sound, and syntax of English number agree-

ment. Language and Cognitive Processes, 8,57–99.Bock, K., & Levelt, W. J. (1994). Language production: Grammatical encoding. In M. A.

Gernsbacher (Ed.), Handbook of Psycholinguistics(pp. 945–984). San Diego, California:Academic Press.

Bock, K., & Miller, C. A. (1991). Broken agreement. Cognitive Psychology, 23,45–93.Clark, H. H. (1965). Some structural properties of simple active and passive sentences. Journal

of Verbal Learning and Verbal Behavior,pp. 365–370.Cutler, A. (1988). The perfect speech error. In L. M. Hyman & C. S. Li (Eds.), Language,

speech and mind: Studies in honor of Victoria A. Fromkin(pp. 209–223). London:Routledge.

Eberhard, K. M. (1997). The marked effect of number on subject-verb agreement. Journal ofMemory and Language, 36,147–164.

Eberhard, K. M. (1999). The accessibility of conceptual number to the processes of subject-verb agreement in English. Journal of Memory and Language, 41,560–578; Journal ofMemory and Language, 88,375–407.

Garrett, M. F. (1975). The analysis of sentence production. In G. H. Bower (Ed.), The psy-chology of learning and motivation(pp. 133–177). New York: Academic Press.

Garrett, M. F. (1984). Levels of processing in sentence production. In B. Butterworth (Ed.),Language Production(Vol. 1, pp. 177–220). London: Academic Press.

Garrett, M. F. (1988). Processes in language production. In F. J. Newmeyer (Ed.), Linguistics:The Cambridge survey, III: Language: Pyschological and biological aspects(pp. 69–96).Cambridge: Cambridge University Press.

Greenslit, N., & Badecker, W. (2000). The role of the lexical item in the comprehension ofsubject-verb agreement. Poster presented at the 13th Annual CUNY Conference on HumanSentence Processing,San Diego, California.

Kempen, G., & Hoenkamp, E. (1987). An incremental procedural grammar for sentence for-mulation. Cognitive Science, 11,201–258.


Kempen, G., & Vosse, T. (1989). Incremental syntactic tree formation in human sentence pro-cessing: A cognitive architecture based on activation decay and simulated annealing.Connection Science, 1,275–292.

Levelt, W. J. M. (1983). Monitoring and self-repair in speech. Cognition, 14,41–104.Levelt, W. J. M. (1989). Speaking: From intention to articulation.Cambridge, Massachusetts:

MIT Press.MacWhinney, B. (1977). Starting points. Language, 53,152–168.Nicol, J. L. (1995). Effects of clausal structure on subject-verb agreement errors. Journal of

Psycholinguistic Research, 24,507–516.Nicol, J. L., & Barker, J. (1996). Units of processing in language production. Poster Presented

at the 10th Annual CUNY conference on Human Sentence Processing.Santa Monica,California.

Nicol, J., Teller, M., Greth, D. (2000) The production of verb agreement in monolingual, bilin-gual, and second language speakers. In J. Nicol (Ed.), One mind, two languages:Language processing in the bilingualOxford: Blackwell Press.

Thornton, R., & MacDonald, M. (1999). Plausibility effects on subject-verb agreement errorsin English. Poster presented at the 12th Annual CUNY Conference on Human SentenceProcessing,New York, NY.

Vigliocco, G., Butterworth, B., & Semenza, C. (1995). Constructing subject-verb agreement inspeech: The role of semantic and morphological factors. Journal of Memory andLanguage, 34,186–215.

Vigliocco, G., Butterworth, B., & Garrett, M. F. (1996). Subject-verb agreement in Spanishand English: Differences in the role of conceptual constraints. Cognition, 61,261–298.

Vigliocco, G., Hartsuiker, R. J., Jarema, G., & Kolk, H. H. J. (1996b). One or more labels onthe bottles? Notional concord in Dutch and French. Language and Cognitive Processes,11, 407–442.

Vigliocco, G., & Nicol, J. (1998). Separating hierarchical relations and word order in languageproduction: Is proximity concord syntactic or linear? Cognition, 68,13–29.


Semantic Factors in the Production of Number Agreement

Documents

Transcript of Semantic Factors in the Production of Number Agreement