Language and Cognitive Processes Priming effects ...

This article was downloaded by: [University of Oklahoma Libraries] On: 08 January 2015, At: 06:50 Publisher: Routledge Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Language and Cognitive Processes Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/plcp20

Priming effects from gaps to antecedents Maryellen C. Macdonald

a

a

Department of Psychology , Northeastern University , Boston, Massachusetts, USA Published online: 13 Dec 2007.

To cite this article: Maryellen C. Macdonald (1989) Priming effects from gaps to antecedents, Language and Cognitive Processes, 4:1, 35-56, DOI: 10.1080/01690968908406356 To link to this article: http://dx.doi.org/10.1080/01690968908406356

PLEASE SCROLL DOWN FOR ARTICLE Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and should be independently verified with primary sources of information. Taylor and Francis shall not be liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of the Content. This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http:// www.tandfonline.com/page/terms-and-conditions

LANGUAGE AND COGNITIVE PROCESSES, 1989, 4 (l), 35-56

Priming Effects from Gaps To Antecedents Maryellen C. MacDonald Downloaded by [University of Oklahoma Libraries] at 06:50 08 January 2015

Department of Psychology, Northeastern University, Boston, Massachusetts, U . S . A . Recent research has suggested that null anaphors o r gaps can prime their antecedents analogous to facilitation from pronouns. These results are open to alternative interpretations and may reflect other factors such as processing load rather than actual priming effects from gaps. An experiment designed to eliminate alternative interpretations was conducted using passive sentences, where gap and no-gap conditions differed by only one word. In a probe recognition task, subjects responded faster to a probe word when that word corresponded to the antecedent of a gap than when it did not. Responses in three control measures indicated that facilitation effects were due to the gap and not to extraneous factors. The stimuli were evaluated in Experiment 2. Ratings o n measures critical for the presence of a gap correctly predicted Experiment 1 response times to the gap’s antecedent; additional ratings indicated that the priming effect could not be attributed to discourse focus shifts. The results reveal that syntactic differences in three superficially similar grammatical constructions affect the processing of these constructions, and they reaffirm that gaps d o prime their antecedents during on-line language processing.

INTRODUCT10N There is a growing body of evidence that a pronominal or nominal anaphor can prime its antecedent during on-line language processing (Chang, 1980; Cloitre & Bever, 1984; Dell, McKoon, & Ratcliff, 1983; Leiman, 1982; MacDonald & MacWhinney, submitted; McKoon & Ratcliff, 1980). This priming effect is typically assessed using stimuli such as the sentences in ( l ) , where a probe word is presented at some point in the second clause of the sentence: Requests for reprints should be addressed to Maryellen MacDonald. Department of Psychology, 125 Nightingale Hall, Northeastern University, Boston, Mass. 02115. U.S.A. The author was supported by a Sloan Foundation Grant to the Department of Psychology at Carnergie Mellon University during the period in which this research was Conducted. Thanks are due to Robin Clark for many valuable discussions throughout this project. and to Tom Bever, Janet Dean Fodor, Nancy Hidlebrandt. Howard Kurtzman. Joanne Miller, and an anonymous reviewer for helpful comments on an earlier draft of this paper.

0 1989 Lawrence Erlbaurn Associates Limited and VSP Publications

36

MACDONALD

Downloaded by [University of Oklahoma Libraries] at 06:50 08 January 2015

la. John went to the store, and he talked to the cashiers. b. John went to the store, and the cashiers were friendly. Subjects’ naming or recognition responses to the probe word JOHN are generally faster following presentation of (la) than (lb), a result that is interpreted to indicate that the pronoun he in (la) primes its antecedent, John. Recently, a number of researchers have investigated whether similar priming effects could be obtained for antecedents of ‘‘null’’ anaphors or gaps (Bever & McElree, 1988; Swinney et al., 1987). In non-technical terms, gaps can be considered place-holders in a syntactic representation, marking the point of origin of a constituent (such as a nounphrase) that has been moved to some other point in the sentence.’ The gap is said to be bound to this moved constituent; binding is a syntactically determined coreference relationship (Chomsky, 1981). A number of grammatical constructions have been hypothesised to be produced through movement of some constituent, and thus they are hypothesised to contain a gap. Some examples can be seen in the sentences in (2): 2a. [The girlIi was kissed [eli. b. [Which bookIi did Mary read [el, c. Billi is easy to fool [el,. The presence of a gap is indicated by [el in these sentences, and the binding relationship is indicated by subscripts. In (2a). for example, the nounphrase [he girl is the antecedent of the gap. This binding relationship captures English speakers’ intuitions that although the girl is the grammatical subject of this sentence, it is the logical patient of the verb kiss. Similar intutions apply for the sentences in (2b and c)-which book and Bill are the subjects of these sentences, yet they are the patients of the verbs read and fool respectively. Because binding entails a co-reference relationship, it is logical to propose the binding hypothesis. i.e. binding between a gap and its antecedent could produce priming effects analogous to those found with pronominal reference. Investigations of priming effects from a binding relationship have supported this hypothesis (Bever & McElree, 1988; ‘Government Binding Theory (Chomsky. 1981) makes a distinction between several kinds of gaps. These distinctions will not be discussed here. The issue for this paper is whether gaps can be shown to have any effect on their antecedents. Given findings that gaps do affect processing. additional research could compare the effects of different types of gaps. Bever and McElree (1988) have found some evidence that gaps produced through movement of a nounphrase produce stronger priming effects than the null pronominal PRO.

PRIMING EFFECTS FROM GAPS

37


Swinney et al., 1987). These studies have used the basic methodology from the pronominal reference experiments, in which subjects were presented with sentences either with or without a gap, and they responded to a probe word. In one experiment, Bever and McElree (1988) presented sentences such as those in (3), where only the (a) version of the sentence contains a gap: 3a. [The astute lawyerIi, who faced the female judge, was certain [eli to argue during the trial. b. The astute lawyer, who faced the female judge, hated the long speeches during the trial. If the gap in (3a) does prime its antecedent, rhe m u r e lawyer, then responses to the probe word ASTUTE should be faster following the gap sentence (3a) than following the no-gap sentence (3b). These predictions were supported and also replicated in experiments testing several other syntactic constructions, as well as in work by Swinney et al. (1987). Such results are clearly consistent with the hypothesis that binding produces priming, but the data are also open to other interpretations. Because it is necessary to vary the syntactic structure of the stimuli to manipulate the presence or absence of a gap, it is possible that processing load also varies. Responses could thus be faster in the gap condition simply because sentences in that condition are easier to process than the no-gap sentences. The fact that Bever and McElree (1988) found similar priming effects from gaps in passive sentences using both subject-controlled and experimenter-controlled sentence presentations, suggests that processing load may not be a serious concern. However, it is clear that independent assessments of processing load would go further to dismiss these issues. A somewhat more troubling problem is that because manipulating the presence of a gap requries changes in the syntactic structure of the sentence, it is possible that discourse focus shifts could be confounded with the manipulations of interest. The probe task has been shown to be especially sensitive to shifts in discourse focus (Gernsbacher & Hargreaves, 1988; MacDonald & Just, 1989), and so it is a very real possibility that focus shifts, rather than gaps, are the real source of the priming effects in these experiments. As there are now a number of gaps priming studies (Bever & McElree, 1988; Swinney et al., 1987), it is probably unlikely that discourse focus shifted to produce faster responses in the gap condition in every one of the experiments reported to date. None the less, direct evidence that discourse focus is not responsible for the priming effects is clearly essential for such a counterintuitive argument as this one, where “invisible” anaphors prime their antecedents. These problems of interpretation do not indicate that the situation is


38

MACDONALD

hopeless for studying priming effects with gaps, but rather that additional controls are needed to clarify the earlier work. If the role of gaps in language processing is to be explicated, then subsequent experiments must (1) minimise changes in the stimulus sentences across the gap and no-gap conditions, (2) use additional dependent measures to monitor for other, unwanted differences in processing of the stimuli. and (3) use additional probe words to check for discourse-level changes across gap and no-gap conditions. The first experiment presented below meets all three of these criteria, providing a much more rigorous test of binding effects than has yet been available. The first step in developing a stronger test of the binding predictions is to choose a syntactic construction where the binding manipulation produces the fewest changes to the stimulus materials. Although there are a number of sentence constructions that are hypothesised to contain a gap, the passive construction makes an excellent choice for a test of the priming effect because only one word must be varied in a sentence to manipulate the presence or absence of a gap. As will become apparent in the discussion of passives, an added benefit of the passive construction is that it allows for a third sentence type to be added to the gap vs no-gap comparison. Clearly, a comparison of responses when there are predictions for performance in three conditions permits a stronger test of the binding hypotheses than is possible with only two conditions. A number of researchers working within a variety of linguistic theories have noted that there seem to be two types of passive constructions (Bresnan, 1982; Levin 8i Rappaport, 1986; Wasow, 1977). One type, called adjectival or lexical passives, is illustrated in sentence (4). The second type, verbal passives, is illustrated in sentence (2a), which is repeated here as ( 5 ) : 4. The girl was surprised. (Adjectival passive) 5 . The girl was kissed. (Verbal passive)

Adjectival and verbal passives can be distinguished first by the subtle semantic differences between them. The sentence in (4) is most naturally interpreted as describing a state or quality of the subject of the sentence, in this case that the girl is in a state of surprise. There is the potential for an “event” interpretation of (4) as well. so that the sentence describes the event of the girl getting surprised by something, but speakers’ intuitions suggest that interpretation is less favoured. In contrast, sentence ( 5 ) unambiguously describes an event happening to the subject of the sentence; it does not attribute some quality to the subject. The two types of passives can also be distinguished by the syntactic environments in which they can appear. Adjectival passives (in their


39

adjectival or “state” interpretation) can appear in exactly those environments where adjectives can appear, but the appearance of verbal passives is more restricted (Levin & Rappaport, 1986; Wasow, 1977). As the phrases in (6) demonstrate, adjectives and adjectival passives can appear as prenominal modifiers, whereas verbal passives cannot:


6a. the happy girl (Adjective) b. the surprised girl (Adjectival passive) c. *the kissed girl (Verbal passive) Similarly, adjectival passives can appear as the complements of verbs such as seem and appear that select for adjectival rather than verbal complements. Verbal passives cannot appear in this environment: 7a. the girl looked happy (Adjective) b. the girl appeared surprised (Adjectival passive) c. *the girl seemed kissed (Verbal Passive) Whereas some adjectives and adjectival passives can be prefixed by a negative un, verbal passives cannot take this prefix, as evidenced by (8): 8a. the girl was unhappy (Adjective) b. the girl was unsurprised (Adjectival passive) c. *the girl was unkissed (Verbal passive) Finally, as the examples in (9) indicate, verbal passives cannot be modified by intensifiers such as very and rather, whereas adjectival passives and adjectives can: 9a. the girl was very happy (Adjective) b. the girl was quite surprised (Adjectival passive) c. *the girl was rather kissed (Verbal passive) Using the evidence illustrated in (6-9) and other examples of similarities between adjectival passives and adjectives (Levin & Rappaport, 1986; Wasow, 1977), syntacticians working within Government Binding Theory have argued that adjectival passives in their “state” interpretation are adjectives, despite their superficial similarity to verbal passives (Chomsky, 1981). For reasons well outside the scope of this paper, constructions containing adjectives (and thus adjectival passives) are assumed to be represented without a gap, whereas verbal passives are formed through movement of a nounphrase, producing a gap. Other analyses have been proposed, such as in Lexical Functional Grammar, where both adjectival

40

MACDONALD


and verbal passives are formed by lexical rules, and neither contains a gap (Bresnan, 1982). Such an analysis makes no predictions of any priming effects, and the Government Binding approach will be pursued here. Given the Government Binding analyses of adjectival and verbal passives, an experiment including all three sentence types illustrated in (10) could provide a strong test of the binding predictions: 10a. The girl was happy. (Adjective) b. The girl was surprised. (Adjectival passive) c. [The girl], was kissed [eli. (Verbal passive) When sentences such as those in (10) are followed by the probe GIRL, the verbal passives like (1Oc) should produce faster probe responses, compared to responses following the adjective condition in (10a) or the adjectival passive in (lob), because girl is the antecedent of a gap only in the verbal passive. As neither (loa) nor (lob) contain a gap, probe responses in these two sentences should not differ. This work extends an experiment by Bever and McElree (1988), who found priming effects in the verbal passive condition compared to the adjective condition, but did not test the adjectival passive condition.

EXPERIMENT 1 This experiment presented the three sentence types shown in (10) and measured responses in a probe recognition task to probes corresponding to the antecedent of the gap. Three additional response measures were taken to ensure that any priming effects were not due to extraneous differences across the conditions. Processing load measures are of particular concern with passive sentences, because there is some evidence that passives are more difficult to process than are active transitive sentences (Forster & Olbrei, 1973). It is not clear that passives are also more difficult than sentences with adjective predicates as in (10a). To assess these possibilities, the reading times for the three sentence types were measured, as were response times and the accuracy of answering questions about the three types of sentences. There was one additional control-the inclusion of another probe condition in which the probe corresponded to a word that was not involved in a binding relationship in any of the three sentence types. The addition of these controls permits a very precise specification of the predictions of the binding hypothesis. If gaps can prime their antecedents and the experiment succeeds in holding extraneous factors such as discourse focus and processing load constant, then the only difference that should appear in the data is that responses to a probe of the antecedent [e.g.



41

GIRL in (lo)] should be faster in the verbal passive condition than in the other two conditions. That is, the binding hypothesis predicts that of the six conditions produced from all combinations of sentences like (loa-c) and probes (antecedent and control probe conditions), only one condition (antecedent probes in the verbal passive condition) should differ from the other five. Thus although the experimental design is a standard 2 x 3 factorial, it is not clear that the standard two-way interaction will adequately test the hypotheses. It is easy to imagine that responses to the antecedent probe in the verbal passive condition could be significantly faster than in the other five conditions when the interaction was not significant, precisely because five of the six conditions did not differ. All comparisons of individual conditions needed to test the binding hypothesis will therefore be performed independent of the presence of the interaction. Clearly, this decision to relax statistical criteria necessitates additional care and scrutiny of all the additional control measures.

Method Subjects. The sample consisted of 30 undergraduates enrolled in psychology classes at Carnegie Mellon University who participated as part of a course requirement. All subjects were native speakers of English. An additional five subjects were tested but were not used in the final analyses because of error rates of over 20% on comprehension questions. Materials. The stimuli were two-sentence passages. The passage type variable was manipulated at three levels (adjective, adjectival passive, and verbal passive) by changing only the last word of the passage. Fourteen word triples were constructed such that one member of the triple was an adjective, one an adjectival passive, and one a verbal passive, such as furio~lsurprisedlshof.The words were matched across the three conditions for length and frequency of occurrence in English (KuCera & Francis, 1967). Three passages were written for each triple, so that any member of the triple could sensibly appear as the last word of the passage. The topics of the passages were unrelated. An example passage for the triple furiowlsurprisedlshot is shown in Table 1(all stimuli are contained in the Appendix). The length of the first sentence of each passage varied between 7 and 11 words. The second sentence was always nine words in length. The last word of this sentence was always one member of the word triple for that passage. The two sentences in the passage were constructed so that the subject of the first sentence could be a plausible agent of the action

42

MACDONALD TABLE 1 Example Stimuli

Adjective Passage The Irish terrorists wanted to disrupt the ceremonies. The new mayor at the centre podium was furious.


Adjectival Passive Passage The Irish terrorists wanted to disrupt the ceremonies. The new mayor at the centre podium was surprised. Verbal Passive Passage The Irish terrorists wanted to disrupt the ceremonies. The new mayor at the centre podium was shot. Agent probe: TERRORISTS Antecedent probe: MAYOR

described by the verbal passive. For example, in the passage in Table 1, terrorists is the most plausible agent of the shooting in the version of the passage ending in shot. The presence of the first sentence thus created a more coherent discourse than would be possible with one-sentence stimuli. Inclusion of the first sentence in the passage also allowed for a second position to be probed. The two probe words for each passage were the subject of the first sentence (the second or third word), and the subject of the second sentence (the second word of the sentence). For ease of exposition, the probe from the second sentence (e.g. MAYOR) will be termed the anfecedent probe, as this probe corresponds to the antecedent of the gap in the verbal passive construction. The first sentence probe (e.g. TERRORISTS) is called the agent probe, by virtue of the fact that this probe is a logical agent of the action in the verbal passive condition. Each passage contained many words that were not in a binding relationship that could have been used in a control probe condition; the choice of the agent probe for the control condition was designed to provide additional information. If the experimental manipulations do produce discourse shifts for these passages, then the subject of the first sentence of the passage (a likely discourse topic) might be especially sensitive to these shifts. The binding predictions suggest that there should be no differences in responses to the agent probe, but, if such differences do emerge, they should be useful in determining how syntactic changes affect discourse shifts in these experiments. A total of 66 practice and filler passages were prepared, containing a variety of syntactic constructions. The majority of probes for filler passages did not occur in the passage; each of these false probes was an associate of



43

a word that had occurred in the passage. This control forced subjects to pay close attention to the passages and refrain from making the probe recognition judgement merely on the basis of familiarity of the concept being probed. All m e probes for filler passages were contained in the passage in some position other than the subject position of either sentence. A yesho comprehension question was prepared for each passage, with yes as the correct answer for half of the questions. The questions for the experimental passages avoided reference to the last word in the passage, so that the same question could be used for a passage regardless of whether the passage ended in an adjective, adjectival passive, or verbal passive.

Procedure. Stimuli were presented on a CRT in a modified “moving window” display (Just, Carpenter, & Woolley, 1982), where subjects were able to read one sentence at a time. At the start of each trial, the display contained three lines of dashes indicating all non-space characters of the passage, probe word, and comprehension question. When subjects pressed a key with the left hand, the first sentence of the passage was displayed. A subsequent keypress replaced the first sentence with dashes and revealed the second sentence. A third keypress removed the second sentence and displayed the probe word in upper-case letters four spaces to the right of the end of the second sentence. Prior to this keypress, the probe word was always indicated by 11 dashes regardless of its actual length, so that subjects would not have early information about the length of the word to be probed. Subjects indicated whether the probe had occurred anywhere in the passage by pressing a YES or a NO key with the right hand. This response removed the probe and displayed a comprehension question, which was answered with the YES or NO key. Subjects were encouraged to read at a normal rate, and speed and accuracy were stressed for responses to both the probes and the comprehension questions. Response times to the probe and comprehension question were recorded, as were reading times for each sentence. Following 10 practice trials, the experimental and filler trials were presented in random order. Every subject saw all 42 experimental passages once, with an equal number of passages in each of the six combinations of probe and passage type conditions. Subjects were tested individually and completed the experiment without a break in one 30-min session.

ResuI ts Analyses were performed on three response time measures-responses to the probe word, reading times for the second sentence of the passage, and responses to the comprehension question. Accuracy rates for responses to the probe and comprehension question were also analysed. For all analyses


44

MACDONALD

presented below, the min F’ statistic is reported except where effects are significant in Fl and/or R analyses, but not in min F‘. Prior to analysis of the response time data, all incorrect probe responses (the error rate was 4%) and incorrect comprehension question responses (5.8% error rate) were removed. Extremely long responses in each of the three response time measures (reading, probe recognition, question comprehension) were also removed, using a two-step procedure. First, the grand mean for each of the measures was calculated, and all responses over three times the appropriate grand mean were removed. Next, means and standard deviations were calculated for each subject for reading, probe judgement, and question comprehension, and any responses that were more than 3.5 S.D. over the subject’s mean for that measure were removed. This procedure removed less than 3% of the responses in each of the three measures. Response times and accuracy rates (where appropriate) for all response measures are shown in Table 2. Analysis of the probe response times showed that the passage type x probe interaction was not significant (min F’ < 1). As discussed above, this result was not unexpected and does not necessarily run counter to the binding hypothesis. Given the nature of that hypothesis, other comparisons are still of interest. The additional comparisons are complicated by another finding, however. There was a trend for overall responses to antecedent probes (971 msec) to be faster than responses to agent probes (1026 msec)--Fl(l, 29) = 2.73, P > 0.10; R(1,41) = 4.23, P C 0.05. This result is probably due to the fact that at the point of the probe, the antecedent word in the passage had been presented more recently than the agent. It is thus not useful to compare

TABLE 2 Effect of Passage Type on Response Times (msec) and Percentage Correct (in Parentheses) for Three Response Measures

Passage Type Response Measure Probe judgement Agent probe Antecedent probe Second sentence reading time Question comprehension

Adjective

Adjectival Passive

Verbal Passive

1041 (95.8) 1006 (98.1)

1020 (96.1) 975 (96.7)

1015 (94.4) 931 (98.9)

295 1

3005

294 1

1717 (93.4)

1748 (94.7)

1666 (94.4)



45

responses to the critical condition (the antecedent probe in the verbal passive condition) to the mean of the other five conditions, because any differences that were obtained could be the result of this probe distance effect, and not the result of priming from the binding relationship. The solution to this problem is to compare the effects of passage type at each level of probe. The binding hypothesis predicts no effects of passage type for the agent probe, but it does predict that for the antecedent probe, the verbal passive condition should produce faster responses than the other two conditions. There was a main effect of passage type for the antecedent probes [F1(2, 58) = 4.58, P < 0.02; F2(2, 82) = 3.20, P < 0.051. Additional analyses of this effect indicated that, as predicted, responses to this probe were faster in the verbal passive condition than in the adjective condition [min F'(1,69) = 4.19, P < 0.0251. In the analysis with subjects as a random effect, there was a trend for antecedent probe responses to be faster in the verbal passive condition compared to the adjectival passive condition [Fl(l, 29) = 3.46, P < 0.081, but this effect was not reliable in the items analysis [F2(1, 41) = 1.67, P > 0.201. Antecedent probe responses did not differ in the adjective and adjectival passive conditions (min F' < l), a result that is entirely consistent with the hypothesis that neither of these conditions contains a gap that could prime the antecedent. As expected, there were no differences in response times to the agent probe across the three passage types (min F' < 1). The lack of an effect of passage type for the agent probes was not due to higher variance in agent probe response times than in antecedent probe response times. Standard deviations for agent probe responses were actually smaller than those for antecedent probe responses at each level of passage type (across passage type, agent and antecedent probe standard deviations were 240 and 267 msec, respectively). Accuracy rates in the probe judgement task, shown in Table 2, clearly indicate that there were n< speedlaccuracy trade-offs in responses to the probes. As Table 2 indicates, subjects were somewhat more accurate responding to the antecedent probe (97.9% correct) than to the agent probe (95.4%)--F1(1, 29) = 6.59, P < 0.05; n ( l , 41) = 3.90, P < 0.06. This result is probably also due to the more recent presentation of the antecedent than the agent. There was no effect of the passage type manipulation on accuracy, nor was there any interaction of passage type and probe (min F's < 1). The other control measures also indicated that processing load did not vary across passage type. Reading times in the second sentence did not differ across the three levels of this factor (min F' < 1). This result is not inconsistent with previous findings that subjects take longer to read passive sentences than active transitive sentences (Forster & Olbrei, 1973)-active transitive sentences were not used in this study, and the passage endings

46

MACDONALD

were matched for length and frequency. Response times and response accuracy to the comprehension questions also did not differ across passage type (min F‘s < 1).


Discussion This experiment set up an extremely rigorous test of the hypothesis that gaps can prime their antecedents, so that a very precise pattern of data was necessary to confirm that hypothesis. Antecedent probe responses had to be faster in the verbal passive condition than in the adjective and adjectival passive conditions and, crucially, all the control measures-agent probe responses, reading times, and question-answering times and accuracycould not differ across the three conditions. The results obtained are extremely close to this ideal pattern of data. Antecedent probe responses were significantly faster in the verbal passive condition than in the adjective condition, and they were non-significantly faster than responses in the adjectival passive condition. Antecedent probe responses did not differ in the adjective and adjectival passive conditions, and all other control measures-responses to the agent probe, reading times, and questionanswering times and accuracy-also showed no differences across passage type. The one point where the data do not clearly fit the criteria outlined here was in responses to antecedent probes in the adjectival passive condition. The binding hypothesis predicts that responses to the antecedent probe in the adjectival passive condition should not differ from the adjective condition, and should be slower than antecedent probe responses in the verbal passive condition. In the research reported here, response times to the antecedent probe in the adjectival passive condition were intermediate between response times in the adjective and verbal passive conditions. One possible explanation for this finding may be that subjects sometimes posited a verbal passive structure (with a gap) for some of the adjectival passives. While it was assumed that the adjectival “state” interpretation was favoured for the adjectival passive passages in Experiment 1, all adjectival passives are ambiguous-they can always have the verbal passive “event” interpretation. For example, in the passages presented in Table 1, surprised was intended to be taken as an adjectival passive, giving a description of the mayor being in a state of surprise. Some subjects may have taken surprised to be a verbal passive, however, so that the sentence described the event of the mayor getting surprised by someone. If subjects treated surprised as a verbal passive, they would have posited a gap following surprised, resulting in facilitation effects for the antecedent probe in this adjectival passive trial. Enough such trials could have produced the intermediate response times in the adjectival passive condition. Experiment 2 addresses the possibility that some adjectival passives


47

were treated as verbal passives and further explores the possibility that discourse focus shifts could account for some of the priming effects found in Experiment 1.


EXPERIMENT 2 The first purpose of Experiment 2 was to assess the degree to which subjects interpreted the experimental passages as being descriptions of states or events. Verbal passives should be uniformly perceived as event descriptions, whereas adjectives should be state descriptions. By examining the interpretations subjects give to the adjectival passive passages, it is possible to determine whether some passages tended to produce an “event” (verbal passive) reading. The second purpose of this experiment was to investigate further whether discourse focus shifts affected the response times obtained in Experiment 1. Such shifts were assumed not to be producing the differences obtained in that experiment, as it would be expected that responses to the agent probe would also be affected by discourse focus shifts. No differences in response times to the agent probe were found, but it is valuable to have an additional, independent assessment of discourse focus in the experimental passages. The stimuli were assessed with two questionnaires. The binding hypothesis provides several predictions about the relationship between the rating data and the reaction time data in Experiment 1. First, discourse focus should not shift in the passages as a function of the passage type manipulation, but the degree to which the passage described a state or event should shift with passage type. If these statelevent questionnaire ratings reflect how perceivers treat the passages during on-line processing, then they should provide an assessment of how consistently subjects posited gaps while comprehending the passages-passages with strong “event” ratings should most often have been treated as containing a verbal passive, and thus a gap. The binding hypothesis predicts that these ratings should be a good predictor of response times to the antecedent probe, bound to the gap, but they should not predict agent probe response times. Moreover, if the ratings reveal that some passages in the adjectival passive condition were frequently given a verbal p,assive interpretation, removal of response times in these passages from the Experiment 1 data set should produce slower response times to the antecedent probe (more like the adjective condition), whereas agent probe responses should not be affected.

Method Subjects. The sample consisted of 36 native English-speaking college students who were paid for their participation in this and an unrelated

48

MACDONALD

experiment. Of these, 18 responded to the discourse focus questionnaire, and 18 received the state/event questionnaire. None of the subjects had participated in Experiment 1.


Materials and Procedure. Each subject who answered the discourse focus questionnaire received a booklet containing the 42 experimental passages, 14 each with the adjective, adjectival passive, and verbal passive endings. Three different booklets were prepared to fully counterbalance passage endings across items. Seven passages appeared on each page of the booklet; each subject’s booklet had a different ordering of the six pages of materials. Each passage in the discourse focus questionnaire was followed on the next line by its two probe words, separated by the numbers 1-5. For example, the passage shown in Table 1 was followed by terrorists 1 2 3 4 5 mayor. The numbers were used by the subjects to rate which of the two words the passage was “more about”. Subjects were instructed to circle I if the passage was about the first word (e.g. terrorists), or 5 for the second word (e.g. mayor). “Mostly about” judgements were coded with 2 and 4 , and 3 was used for passages that were equally about the two words on either side of the scale. Subjects were encouraged to make their judgements quickly after reading the passage. The procedure for the statelevent questionnaire was identical except that subjects were instructed that they would be rating whether the passages were more a description of a state of some entity mentioned in the passage or more a description of some event happening to that entity. The passages in these booklets were followed by the antecedent probe word, followed by the words STATE and EVENT, separated by the numbers 1-5. as in mayor: STATE 1 2 3 4 5 EVENT. Subjects circled 1 if the passage described the state of the mayor, 5 if it described an event happening to the mayor, and the numbers 2-4 were used for intermediate judgements.

Results and Discussion

’

Ratings on both questionnaires are shown in Table 3. As can be seen from Table 3, ratings on the state/event questionnaire showed much larger differences across passage type than did the discourse focus ratings. The data were analysed in two ways; first, in ANOVAs assessing the effects of passage type on ratings, and then in a multiple regression to determine the value of the two ratings as predictors of reaction times in Experiment 1. The discourse focus ratings showed no reliable differences across passage type (min F‘ < 1). In contrast, there was a substantial effect of passage type on subjects’ ratings in the state/event questionnaire [min F’(2, 57) = 83.56, P < O.OOOl]. Adjectival passive passages were rated as being


49

TABLE 3 M e a n Focus and Event Ratings and Standard Deviations (in Parentheses) Across Passage Type

Passage Type

Questionnaire


Discourse focus" State/eventb

Adjective

Adjectival Pnssive

Verbal Passive

3.04 (0.59) 1.70 (0.81)

3.06 (0.5Y)

2.33 (1.09)

2.91 (0.62) 4.45 (0.62)

"Rating of 1 = agent focus; 5 = antecedent focus. "Rating of I = state reading; 5 = event reading.

significantly more "event-like" than were adjective passages [min F'(1,45) = 12.82, P < 0.0011, and verbal passive passages were in turn rated as describing events significantly more than adjectival passive passages were [min F'(1, 28) = 68.60, P < O.OOOl]. Because verbal passives inherently describe events whereas adjectives and adjectival passives describe states, the state/event ratings should reflect how often the subjects in Experiment 1 interpreted each passage as containing a verbal passive. The hypothesis that gaps can prime their antecedents suggests that these ratings should thus be a good predictor of response times to the antecedent probes, but that they should not predict response times to the agent probes. Moreover, if the task in Experiment 1 was truly measuring binding and not discourse focus shifts, then the discourse focus ratings should not be a good predictor of probe response times. These predictions were tested in a multiple regression analysis. The mean discourse and state/event ratings for each stimulus item in the three passage types were entered as predictors of Experiment 1 antecedent and agent probe response times at each level of passage type. As it collapses across subjects and thus cannot take into account subject variability in either response times or ratings, this procedure is a conservative one (Knight, 1984). The partial correlations of the two ratings on antecedent and agent response times supported the binding hypothesis predictions. The state/ event questionnaire ratings were a good predictor of antecedent probe response times [t(123) = 1.85, P < 0.071, but they did not predict agent probe response times ( t < 1). Given that the ratings were collected in an off-line task using different subjects than those tested in Experiment 1, the degree of correspondence between the state/event ratings and the antecedent probe responses is quite impressive. Discourse focus ratings were not a good predictor of response times to either probe (ts < 1).


50

MACDONALD

From these results it is clear that when a passage is interpreted as containing a verbal passive, response times to the antecedent probe are faster. The state/event rating data can thus be used in another way, to test the possibility that some of the adjectival passive stimuli were treated as verbal passives in Experiment 1. Table 3 shows that the standard deviation for the adjectival passive statelevent ratings is larger than that of the adjectives and verbal passives. This fact suggests that the more “eventlike” ratings for adjectival passives than for adjectives is not the result of a general tendency for all adjectival passives to be more “event-like” than adjectives, but is rather due to a small number of “event-like” adjectival passive stimulus items. If so, then the binding hypothesis would predict that removal of these outlier adjectival passives would produce a slower mean response time for antecedent probes, but would not affect agent probe response times. Ratings of the adjectival passive passages were examined for evidence of outliers in the distribution, and one cluster of “event-like’’ outliers was found. The three passages containing the adjectival passive folded (items 22-24 in the Appendix) all had ratings over 3.0, with a mean rating of 3.67. Folded was the only 1 of the 14 adjectival passives with a rating of more than 2 S.D. above the mean rating for the adjectives, and it also had the only rating within 2 S.D. of the mean rating for the verbal passives.2 Removal of probe responses on trials containing folded in Experiment 1 (one trial per subject) had exactly the effect predicted by the binding hypothesis: Response times to the antecedent probe in the adjectival passive condition slowed to 990 msec (an increase of 15 msec), becoming more similar to the antecedent probe response times in the adjective condition (1006 msec). Moreover, removal of these items had a negligible effect on agent probe response times in the adjectival passive conditionthey decreased by 4 msec to 1016 msec. Reanalysis of the response times in Experiment 1 with the folded items replaced by these new cell means revealed that this small change in the stimuli had a major impact. Response times to antecedent probes in the adjectival passive condition were now significantly slower than in the verbal passive condition (990 vs 931 msec)-Fl(1, 29) = 4.57, P < 0.05; n ( l . 41) = 3.29, P < 0.08. With the removal of the folded items, the response times to the antecedent probes now correspond exactly to the predictions of the binding ’The opposite hypothesis, that removing especially “state-like” outliers should make the adjectival passive data more resemble the verbal passive data. could in principle be tested. However, as would be expected if most adjectival passives were being interpreted as adjectives. there were no outliers at the “state” end of the scale. While the rating for folded was 0.62 points higher than that of its nearest neighbour, there were five adjectival passives within 0.62 points of the most “state-like” adjectival passive.



51

hypothesis. Response times to the antecedent are reliably faster in the verbal passive condition than in the adjectival passive and adjective conditions, and these latter two conditions do not differ. This result was obscured in Experiment 1 by the fact that the adjectival passive stimuli are inherently ambiguous, and some items were occasionally given their verbal passive interpretation. The state/event ratings provide a measure of which items tended to have this interpretation; if it were possible to know for each subject in Experiment 1 how each item was interpreted, so that the items could be divided into adjectival and verbal passives for each subject according to that subject’s interpretation, then the data would be expected to fit the predictions of the binding hypothesis even better. This insight into how each subject will interpret the adjectival passives is of course impossible, and so future research will have to try to disambiguate the adjectival passives. The grammatical constructions shown in (6-9) could provide disambiguating information, as when adjectival passives appear in these constructions they must have an adjectival (i.e. “state”) interpretation. Alternatively, unambiguous adjectival passives could be tested in languages such as Hebrew where adjectival and verbal passives are morphologically distinct (Borer & Wexler, 1987). In such languages the perceiver has immediate unambiguous information about whether the passive is adjectival or verbal, and thus the adjectival passives should pattern with the adjectives in such languages, without the ambiguity found in Experiment 1.

GENERAL DISCUSSION Taken together, the results of Experiments 1 and 2 indicate that when a sentence contains a passive construction that is interpreted as an event, the activation level of the subject of the sentence increases (evidenced by faster response times to the antecedent probe), but the activation level of the agent of this action does not increase. The data are at the very least curious without a theoretical framework with which to interpret them, but they follow naturally from the binding hypothesis. This hypothesis assumes a particular grammatical analysis for passives: Verbal passive constructions are generated through movement of the patient of the action to subject position in the sentence, leaving behind a gap or “null” pronominal that is bound (i.e. co-referential) with the subject. Adjectival passives, despite the ed ending that they share with verbal passives, are assumed to be adjectives, generated without movement and thus without gaps. The hypothesis that gaps, like overt pronouns, can prime their antecedents predicts that when a sentence contains a verbal passive (an event), the presence of a gap will prime its antecedent. If this priming effect is due only to the binding relationship between the gap and its antecedent, then


52

MACDONALD

responses should be faster to the antecedent probe only in the verbal passive condition, and responses to the agent probe and all other measures (e.g. reading time) should not vary across passage type. This is exactly the pattern of results that was obtained. Those who find it implausible that syntactic structures in the form of an “invisible” pronominal can produce priming effects may be tempted to argue that the real source of the priming is the “event” interpretation. Such an analysis might be favoured by proponents of syntactic theories that do not posit gaps in verbal passive constructions, such as Lexical Functional Grammar (Bresnan, 1982). According to this view, the binding hypothesis was valuable only as a guide to the fortuitous discovery that “event” interpretations of passives increase the activation level of the patient of the action. This interpretation of the data fails on several counts, however. First, it provides no mechanism for the priming effects; there is no a priori reason to expect that a description of an event is going to produce more priming than a description of a state. The most obvious choice for such a mechanism is that “event” descriptions are for some reason more likely to produce discourse focus shifts than are “state” descriptions, and discourse focus has been shown to influence probe response times. This suggestion immediately gets into difficulty explaining why it is the patient of the event (the antecedent probe) that is facilitated and not the agent of this action (the agent probe). Equally important, it is clearly refuted by the discourse focus ratings in Experiment 2. Shifts from state to event descriptions did not change discourse focus in the experimental passages, and thus discourse focus cannot be a mediating mechanism for the priming effects observed here. Abandoning the binding hypothesis in favour of some “event”-priming mechanism causes other problems, because other grammatical constructions contain gaps without producing “event” interpretations. The gaps in these constructions [e.g. the sentences in (2) and (3a)l have been shown to produce priming of their antecedents (Bever & McElree, 1988; Swinney et al., 1987) despite the absence of any “event” interpretation that is unique to the gap condition. The binding hypothesis predicts that facilitation effects should be found for all of these constructions containing gaps, whereas the event-priming approach has no explanation for why many seemingly unrelated constructions should produce priming effects. The power of the binding hypothesis to predict the wide range of effects explored here and in previous work comes from the unification of two independent research efforts, Although none of the motivations have been elaborated here, the theory of the binding relationship between a gap and its antecedent has been extensively studied within the Government Binding framework (Chomsky, 1981). The motivations underlying the binding theory and the analyses of various syntactic constructions as having gaps



53

are crucially internal to the syntactic theory and have not been tailored to any psycholinguistic data. The link between the syntactic theory and language processing performance is completed by the second body of research, the work on priming effects from anaphoric reference. Because the binding relationship between a gap and its antecedent is viewed as a syntactically determined co-reference relationship, the natural union of the two fields is the binding hypothesis developed here, i.e. that gaps, like overt anaphors, prime their antecedents. The work presented here demonstrates that priming effects can in fact be produced by binding, despite the fact that this syntactic relationship cannot be perceived in the same way an overt pronoun can. The research demonstrates that these effects cannot be attributed to differences in processing load or to shifts in discourse focus. The research thus supports the surprising prediction that was shot, a verbal passive, should be processed differently than superficially similar adjectival passives and adjectives, such as was surprised and was furious. The binding hypothesis further unifies the effects obtained here with those obtained with strikingly different constructions, such as easy to fool (2c) and was certain to (3a) (Bever & McElree, 1988). The binding hypothesis gives researchers the ability to look beyond these superficial similarities and differences, and it should prove a valuable framework with which to continue investigations into the effect of syntactic representations on language comprehension.

Manuscript received February 1988 Revised manuscript received November 1988

REFERENCES Bever. T. G . & McElree. B. (1988). Empty categories access their antecedents during comprehension. Linguisric Inquiry. 19. 35-45. Borer, H. & Wexler, K. (1987). The maturation of syntax. In T. Roeper & E. Williams. (Eds), Parameter serting. Dordrecht: Reidel. Bresnan, J. (1982). The passive in lexical theory. In J . Bresnan (Ed.), The mental represenration of grammatical relarions. Cambridge. Mass.: MIT Press. Chang. F. R. (1980). Active memory processes in visual sentence comprehension: Clause effects and pronominal reference. Memory and Cognirion, 8. 58-64, Chomsky. N. (1981). Lectures on government and binding. Dordrecht: Foris. Cloitre. M. & Bever. T. G. (1984). Less is more: The processing economy of pronominal anaphors. Paper presented at the meeting of the Eastern Psychological Association. Dell, G . S.. McKoon, G., & Ratcliff. R. (1983). The activation of antecedent information during the processing of anaphoric reference in reading. Journal of Verbal Learning and Verbal Behavior, 22. 121-132. Forster, K. I. & Olbrei. 1. (1973). Semantic heuristics and syntactic analysis. Cognition, 2. 319-347.


54

MACDONALD

Gernsbacher, M. A. & Hargreaves, D. (1988). Accessing sentence participants: The advantage of first mention. Journal of Memory and Language, 27,699-717. Just, M. A., Carpenter, P. A.. & Woolley. J . D. (1982). Paradigms and process in reading comprehension. Journal of Experimental Psychology: General. 111, 228-238. Knight. G. P. (1984). A survey of some important techniques and issues in multiple regression. In D. E. Kieras & M. A. Just (Eds), New methods in reading comprehension research. Hillsdale, N.J.: Lawrence Erlbaum Associates Inc. KuCera, H. & Francis, W. N. (1967). Computational analysis of present-day American English. Providence: Brown University Press. Leiman, J. M. (1982). A chronometric analysis of referent assignment to pronouns. Unpublished Ph.D dissertation. Wayne State University. Detroit. Michigan. Levin, B. & Rappaport, M. (1986). The formation of adjectival passives. Linquistic Inquiry, 17. 623-661. MacDonald, M. C. &Just, M. A. (1989). Changes in activation levels with negation. Journal of Experimental Psychology: Learning, Memory & Cognition, 15, 633-642. MacDonald, M. C. & MacWhinney, B. (submitted). Pronouns can inhibit and facilitate nouns. McKoon. G . & Ratcliff. R. (1980). The comprehension processes and memory structures involved in anaphoric reference. Journal of Verbal Learning and Verbal Behavior. 19. 668-682. Swinney, D.. Nicol, J.. Ford. M.. Frauenfelder. U.. & Bresnan, J . (1987). The time course of co-indexation during sentence comprehension. Paper presented at the Annual Meeting of the Psychonomics Society. Seattle. Wasow. T. (1977). Transformations and the lexicon. In P. Culicover, T. Wasow. & A. Akmajian. (Eds). Formal syntax. London and San Diego: Academic Press.

APPENDIX Second sentence endings are given in the order adjectiveladjectival passive/verbal passive. followed by AGENT and ANTECEDENT probes. 1. The furious parents marched right into the classroom. The new principal with the modern methods was incoherent/embarrassed/assaulted. (PARENTS. PRINCIPAL) 2. The obnoxious reporters made the news conference a very rowdy affair. The famous politician from the tiny country was incoherent/embarrassed/assaulted. (REPORTERS. POLITICIAN) 3. The customers demanded an explanation for the milk shortage. The store owner with the wireless microphone was incoherent/embarrassed/assaulted. (CUSTOMERS. OWNER) 4. The dogs barked loudly from the other side of the yard. The grey squirrel in the oak tree was afraidfrightenedattacked. (DOGS, SQUIRREL) 5 . Some tough guys roam around in the park at night. The small child on the park bench was afraid/frightened/attacked. (GUYS. CHILD) 6. The gangsters will do anything to keep out of jail. The young witness to the bank robbery was afraid/frightened/attacked. (GANGSTERS. WITNESS) 7. The carpenter started work in the kitchen yesterday. The tile floor next to the stove was hideous/stained/removed. (CARPENTER. FLOOR) 8. The owner examined each article in the shop. The antique chair in the back closet was hideous/stained/removed. (OWNER. CHAIR) 9. The mechanics wanted to improve the condition of the car. The red upholstery in the back seat was hideous/stained/removed. (MECHANICS, UPHOLSTERY)



55

10. The proud mother watched the recital with rapt attention. The young ballerina in the pink costume was nervous/worried/photographed. (MOTHER. BALLERINA) 11. The game wardens stopped the jeeps near the clearing. The shy antelopes at the watering hole were nervous/worried/photographed. (WARDENS. ANTELOPES) 12. The journalism students were looking forward to the lecture. The guest speaker with the stained tie was nervoudworriedlphotographed. (STUDENTS, SPEAKER) 13. The researchers spent long hours in the library. The famous document on the American Revolution was complex/detailed/studied. (RESEARCHERS, DOCUMENT) 14. The accountants worked on the audit for several weeks. The sales books from the downtown stores were complex/detailed/studied. (ACCOUNTANTS, BOOKS) 15. The explorers had a new theory about the location of the treasure. The ancient map of the limestone cave was complex/detailed/studied. (EXPLORERS. MAP) 16. The peasants tried to storm the palace with crude knives and guns. The army general on the front steps was furious/surprised/shot. (PEASANTS, GENERAL) 17. The motorcycle gang demanded access to the private beach. The burly lifeguard in the blue trunks was furious/surprised/shot. (GANG. LIFEGUARD) 18. The Irish terrorists wanted to disrupt the ceremonies. The new mayor at the centre podium was furious/surprised/shot. (TERRORISTS, MAYOR) 19. The collector got a magnifying glass down from the shelf. The rare coin from the eighteenth century was beautiful/chipped/appraised. (COLLECTOR. COIN) 20. The experts conferred about a new shipment from England yesterday. The expensive plate from the estate sale was beautiful/chipped/appraised. (EXPERTS, PLATE) 21. The insurance agent surveyed the living room. The crystal vase on the black pedestal was beautifuUchipped/appraised. (AGENT. VASE) 22. The soldiers checked that everything was ready for Monday’s inspection. The blue flags in the front window were spotless/folded/cleaned. (SOLDIERS. FLAGS) 23. The decorator worked hard to finish the store window display. The yellow rugs on the green couches were spotless/folded/cleaned. (DECORATOR. RUGS) 24. The maid spent all day preparing the guest room. The satin sheets in the top drawer were spotless/folded/cleaned. (MAID. SHEETS) 25. The wpman examined the bake sale items on the table. The big bag full of oatmeal cookies was expensiveltornipurchased. (WOMAN, BAG) 26. The man bid for several prints and postcards at the auction. The colour photograph of the famous novelist was expensiveltornlpurchased. (MAN, PHOTOGRAPH) 27. The girl always selects clothes with great care in this store. The silk dress with the big belt was expensive/torn/purchased. (GIRL. DRESS) 28. The director was tired of all the arguments among the cast. The lead actress in the war film was upset/annoyed/replaced. (DIRECTOR, ACTRESS) 29. The band’s manager wanted to make several changes. The bass guitarist with the black van was upset/annoyed/replaced. (MANAGER. GUITARIST) 30. The boss blamed the staff for poor sales. The new clerk from the downtown office was upset/annoyed/replaced. (BOSS. CLERK) 31. The officers took measures to prevent a riot over bad food. The agitated prisoners in the east wing were unhappyldispleasedhransferred. (OFFICERS. PRISONERS) 32. The teacher demanded that the class remain quiet during the exam. The two boys with the cheat sheets were unhappy/displeased/transferred. (TEACHER. BOYS) 33. The foreman had some new ideas for improving production efficiency. The oldest workers in the machine room were unhappy/displeased/transferred. (FOREMAN, WORKERS) 34. The judges weren’t impressed with the poetry reading. The poor speaker with the soft voice was ineffective/humiliated/criticised. (JUDGES. SPEAKER)


56

MACDONALD

35. The restless audience wanted an inspiring speech. The hopeful candidate with the vague platform was ineffective/humiliated/criticised. (AUDIENCE. CANDIDATE) 36. The professors believed in a thorough oral examination. The history student with the complicated theory was ineffective/humiliated/criticised. (PROFESSORS. STUDENT) 37. The dangerous spies crept silently down the hall to the lab. The security guard in the monitoring room was desperate/concerned/spotted. (SPIES. G U A R D ) 38. The rescue team worried that the blizzard would soon strike. The stranded hunters on the mountain top were desperatelconcernedlspotted. (TEAM. HUNTERS) 39. The police bloodhounds followed the trail into the desert. The escaped convicts in the deep canyon were desperate/concerned/spotted. (BLOODHOUNDS, CONVICTS) 40. The governor had big plans for the downtown renovation. The experienced builder with the huge company was enthusiastic/pleased/hired. (GOVERNOR. BUILDER) 41. The executives needed some new managers in the sales office. The recent graduate from the business school was enthusiatic/pleased/hired. (EXECUTIVES. GRADUATE) 42. The producers finally decided on a set designer for the film. The talented sculptor with the innovative ideas was enthusiasticlpleasedIhired. (PRODUCERS. SCULPTOR)