Structural Change and Reanalysis Difficulty in Language Comprehension

Journal of Memory and Language 40, 136 –150 (1999) Article ID jmla.1998.2606, available online at http://www.idealibrary.com on

Structural Change and Reanalysis Difficulty in Language Comprehension Patrick Sturt Human Communication Research Centre, University of Edinburgh, Scotland

Martin J. Pickering Human Communication Research Centre, Department of Psychology, University of Glasgow, Scotland

and Matthew W. Crocker Human Communication Research Centre, University of Edinburgh, Scotland Many theories of parsing predict that the difficulty of syntactic reanalysis depends on the type of structural change involved. However, most existing experimental data show that reanalysis difficulty is affected by nonstructural factors like plausibility and verb bias, whereas claims about structural change are typically based on intuition alone. We report two self-paced reading experiments which demonstrate clear differences in the magnitude of garden path effects associated with different types of structural change. However, difficulty of reanalysis was not affected by the position of the head noun within the ambiguous phrase. We interpret these results in terms of theories of structural change such as Sturt and Crocker (1996). © 1999 Academic Press

A great deal of research in the past 20 years has been directed at the question of how people deal with ambiguity in natural language. Most of this work has been concerned with uncovering the mechanisms underlying people’s initial decisions when faced with an ambiguity. For example, in sentences like (1) and (2) below, much evidence suggests that the human parser initially interprets the famous doctor as the direct object of the immediately preceding verb. Address correspondence and reprint requests to Patrick Sturt, Human Communication Research Centre, University of Edinburgh, 2 Buccleuch Place, Edinburgh EH8 9LW, Scotland, U.K. Fax: 144 131 650 4587. E-mail: [email protected]. We thank Mike Burton, Don Mitchell, Keith Rayner, Tony Sanford and members of the Human Communication Research Centre sentence processing group at the universities of Edinburgh and Glasgow for comments on previous drafts of this paper. This research was supported by ESRC Grant No. R000222286 and was presented at the CUNY Conference on Human Sentence Processing, New Brunswick, NJ, March, 1998. 0749-596X/99 $30.00 Copyright © 1999 by Academic Press All rights of reproduction in any form reserved.

(1) The Australian woman saw the famous doctor had been drinking quite a lot. (2) Before the woman visited the famous doctor had been drinking quite a lot.

These findings may reflect general structural preferences such as Minimal Attachment and Late Closure (Frazier, 1978; Frazier & Rayner, 1982) or individual statistical preferences associated with the verbs saw and visited (Garnsey, Pearlmutter, Myers, & Lotocky, 1997; Trueswell, Tanenhaus, & Kello, 1993). Attachment decisions such as these sometimes have to be revised, or reanalyzed, in order to interpret the sentences correctly. For example, in both (1) and (2), the presence of had indicates that the initial analysis cannot be correct. Many studies have shown processing difficulty (relative to controls) soon after had is encountered (e.g., Clifton, 1993; Ferreira & Henderson, 1990; Frazier & Rayner, 1982; Garnsey et al., 1997; Pickering & Traxler, 1998; Trueswell et al.,

136

REANALYSIS IN LANGUAGE COMPREHENSION

1993).1 Henceforth, we refer to sentences like (1) as NP/S ambiguities because they involve verbs which can take either a noun phrase (NP) or a sentence (S) as a complement and to sentences like (2) as NP/Z ambiguities, because they involve verbs which can take either a noun phrase as a complement or no complement at all (“Z” stands for “zero”). Likewise, we refer to the NP, S, and Z analyses for such sentences. In experimental psycholinguistics, processing difficulty associated with disambiguation in certain constructions, including NP/S and NP/Z ambiguities, has typically been used as evidence for proposed theories of initial attachment preferences. However, recent years have seen increased interest in the study of the reanalysis processes themselves. This paper is concerned with such issues. In particular, we address the question of what factors affect the difficulty of reanalysis. Many accounts claim that reanalysis difficulty is affected by the type of structural change involved. However, the evidence for this claim is currently very weak, whereas, in contrast, the evidence that other factors affect reanalysis difficulty is strong. The main goal of our experiments is to determine whether structural change has a major effect on reanalysis difficulty. First we review current structural change models and then we discuss experimental evidence about reanalysis difficulty. STRUCTURAL CHANGE THEORIES OF REANALYSIS The first structural change theory was due to Pritchett (1988, 1992), who proposed that reanalysis is difficult when it results in a major rearrangement of thematic structure, but easy if it does not. For example, he claimed that NP/Z 1

Ranked parallel models (e.g., Gibson, 1991; Jurafsky, 1996), assume that preferences for different structural hypotheses are reordered in many cases where serial models assume that reanalysis occurs. In Gibson’s model, (1) would involve reordering, but only (2) would involve reanalysis. Models based on the satisfaction of multiple constraints (e.g., MacDonald, Pearlmutter, & Seidenberg, 1994) may also reorder analyses through shifts in the relative activation values of alternative hypotheses. However, some constraint satisfaction models still recognize the existence of true garden path effects (i.e., reanalysis effects) (Trueswell et al., 1993).

137

reanalysis (as in 2) requires a constituent (here, the famous doctor) to be moved out of its thematic domain (roughly corresponding to the argument of a verb or other argument assigner; see Pritchett (1988) for details). In contrast NP/S reanalysis (as in 1) does not require a constituent to be moved out of its thematic domain.2 Hence, Pritchett predicted that NP/Z reanalysis should be hard, whereas NP/S reanalysis should be easy. However, his evidence for this claim was based on intuition alone. Other theories make different claims about the types of structural change that causes difficulty. Such theories claim that reanalysis is difficult if it involves the destruction of the structural relations which describe tree structure (Gorrell, 1995; Sturt & Crocker, 1996; Weinberg, 1993) (cf. Marcus, Hindle, & Fleck, 1983), the rearrangement of intonational phrases (Bader, 1998), or the destruction of certain aspects of thematic structure (Sturt & Crocker, 1997). These theories can be seen as following a guiding strategy of representation preservation (Sturt, 1997; Sturt & Crocker, 1998) in which the processor prefers to reanalyze in a manner which is minimally destructive to some representation type. In these theories, the representation is often described in terms of transitive structural relations. A transitive relation is one such that if the relation holds between X and Y and between Y and Z then it also holds between X and Z. One such relation is dominance, which holds between nodes in the tree representations typically assumed in syntactic theory. Most of the above theories employ this relation (see Partee, ter Meulen, & Wall, 1993). If node X dominates node Y at a point during parsing, then any number of new nodes may be inserted into the path between X and Y at subsequent points without falsifying the original statement that X dominates Y. Let us see how such theories explain the 2 Technically speaking, in (2) the noun phrase the famous doctor has to be removed from the domain in which it receives the theme thematic role from visited, while in (1) this noun phrase remains within the relevant thematic domain. The theme role is reassigned to the complement clause, but the famous doctor remains within this domain after reanalysis.

138

STURT, PICKERING, AND CROCKER

difference between NP/S and NP/Z ambiguities. First, consider Fig. 1, which shows how the tree representation changes in the reanalysis of the NP/S ambiguity (1). The top panel of Fig. 1 represents the state of the parse at which the famous doctor has been attached as the direct object of saw. Notice that there is a dominance relation between the verb phrase (VP) node headed by saw and the noun phrase (NP) node dominating the famous doctor. The bottom panel represents the state of the parse after reanalysis. Notice also that this structural relation continues to hold even though there is now a sentence (S) node between the VP node and the NP node. This means that the set of dominance relations can be updated simply by adding new relations rather than by deleting any existing relations. We call such operations monotonic (i.e., nondestructive). In contrast, consider Fig. 2, which shows how the tree representation changes in the reanalysis of the NP/Z ambiguity (2). The upper panel represents the state of the parse immediately following the attachment of the famous doctor

FIG. 2. Nonmonotonic update of structural relations in the reanalysis of an NP/Z ambiguity.

as the object of visited. There is a dominance relation between the VP node headed by visited and the NP node dominating the famous doctor. Again, the bottom panel represents the state of the parse after reanalysis. Here, the VP node no longer dominates the NP node. This means that existing dominance relations have been deleted in this process. We call such operations nonmonotonic (i.e., destructive). These theories assume that processing difficulty accompanies parsing behavior that is nonmonotonic. In contrast, monotonic reanalysis is assumed to be comparatively easy. See Sturt and Crocker (1996) for a processing account which implements this distinction. REANALYSIS AND EXPERIMENTAL EVIDENCE

FIG. 1. Monotonic update of structural relations in the reanalysis of an NP/S ambiguity.

A serious weakness of most structural change theories is that they have not been tested experimentally and thus rely on intuitive data alone. For instance, Pritchett (1988, 1992) divided reanalysis into two discrete categories, defined in terms of whether the processing difficulty involved is available to conscious introspection.

REANALYSIS IN LANGUAGE COMPREHENSION

He assumed that NP/Z ambiguities like (2) cause noticeable difficulty, whereas NP/S ambiguities like (1) do not. The use of intuitive data is clearly unsatisfactory. For example, purportedly easy types of reanalysis can be made difficult to the extent that many readers will report some noticeable difficulty. Pritchett (1988, 1992) predicts that the following NP/S sentence should be easy to process: (3) Last week, I read a fascinating article by a famous historian about the American War of Independence had been published in a magazine.

However, (3) appears to cause a noticeable garden path effect for many people. If it does, it falsifies Pritchett’s claims; it is not even certain that it causes less difficulty than some NP/Z ambiguities. Whatever the status of this particular example, it is clear that basing theories of reanalysis on intuition is no more satisfactory than basing any other theory of processing on intuition. In contrast, experimental work has revealed a number of ways in which reanalysis difficulty is affected by nonstructural factors. First, Pickering and Traxler (1998) demonstrated that the difficulty of reanalysis is affected by the plausibility of the initial analysis, using sentences like (4a) and (4b): (4a) The criminal confessed his sins which upset kids harmed too many people. (4b) The criminal confessed his gang which upset kids harmed too many people.

Readers experienced more difficulty after disambiguation in (4a), where the initial NP analysis is plausible, than in (4b), where the initial NP analysis is implausible. Experiments using NP/Z ambiguities produced similar results. Pickering and Traxler argued that readers semantically commit more strongly to a plausible sentence than to an implausible sentence (cf. Clifton, 1993). Other studies have found that the degree of verb bias affects processing difficulty (Garnsey et al., 1997; Mitchell & Holmes, 1985; Trueswell et al., 1993). Garnsey et al. considered sentences like the following:

139

(5a) The photographer accepted the money could not be sent yet. (5b) The job applicant believed the interviewer had been dishonest.

They found that sentences like (5a), which employ verbs such as accepted that take a noun phrase direct object more often than a clausal complement, were more difficult to process in the disambiguating region than sentences like (5b), which employ verbs such as believed that take clausal complements more often than noun phrase objects. Another factor that may affect reanalysis difficulty concerns the syntactic characteristics of the ambiguous phrase. Ferreira and Henderson (1991) used the Speeded Grammaticality Judgement task in which participants judge the grammaticality of a string presented one word at a time using rapid serial visual presentation to investigate sentences like (6a) and (6b). In (6a) the head noun of the ambiguous phrase, dog, is adjacent to the disambiguating word yawned, while in (6b), it is four words distant. (6a) After the boy scratched the big and hairy dog yawned loudly. (6b) After the boy scratched the dog that is hairy yawned loudly.

Ferreira and Henderson (1991) found that participants judged (6a) to be grammatical more often than (6b). It is unclear whether such effects occur during normal reading; we return to this issue below. In summary, experimental evidence suggests that at least some nonstructural factors affect reanalysis difficulty, but there is little clear evidence to back up the intuitive claim that structural factors play a role. Indeed, it is quite possible that all apparent structural effects might in fact be due to nonstructural factors such as plausibility and verb bias in the examples involved. For example, Pritchett (1992, p. 77) cites the following sentence as an example of easy reanalysis: (7) Ron believes Rex is a threat.

However, believe is a verb which is biased toward the S reading. If verb bias influences processing difficulty, Pritchett’s observation may not be due to structural factors at all.

140

STURT, PICKERING, AND CROCKER

Current data are therefore consistent with two different accounts. On one account, reanalysis difficulty is due to nonstructural factors alone, as suggested by Frazier and Clifton (in press). In their terminology, reanalysis may be based on sentence token (i.e., non-structural factors, such as plausibility or verb bias) rather than sentence type (i.e., structural factors). On the other account, reanalysis difficulty is due to a combination of structural and nonstructural factors. The two experiments reported below attempt to distinguish these positions by directly considering the effect of the type of structural change on processing difficulty. We employed a standard psycholinguistic technique, self-paced reading, to make a comparison between two types of ambiguous sentence and either controlled or directly manipulated other factors that might affect reanalysis difficulty. EXPERIMENT 1 Experiment 1 compared NP/S and NP/Z ambiguities and included disambiguated controls. Structural change theories predict that processing difficulty at disambiguation will be greater in NP/Z ambiguities like (2) than NP/S ambiguities like (1), when other factors have been controlled. These factors include plausibility, degree of verb bias, and characteristics of the ambiguous phrase (as discussed above). However, in addition to the nonstructural factors discussed above, it is also possible that the processing of such sentences differs for reasons that are unrelated to ambiguity. For example, it may be harder to process a main clause after an initial subordinate clause (as in 2) than to process a complement clause after a main verb (as in 1), perhaps because of a delay associated with the computation of the semantic relation between the initial subordinate clause and the main verb. This might be the case even for unambiguous constructions. The experiment therefore compared locally ambiguous items similar to (1) and (2) above with unambiguous baseline controls. We decided to use segmentby-segment noncumulative self-paced reading, which is known to be sensitive to garden path effects (e.g., Mitchell, Corley, & Garnham,

1992). We are interested in the process of reanalysis given that misanalysis has already occurred. Therefore, we are not concerned with the possibility that this technique might exaggerate the extent of misanalysis. Method Participants. Twenty-four participants from the University of Glasgow were paid to take part in the experiment. No participants took part in more than one experiment or pretest reported in this paper. Items. We constructed 32 sets of items similar to (8a) through (8d) below (see the Appendix): (8a) NP/S ambiguous: The Australian woman saw the famous doctor had been drinking quite a lot. (8b) NP/S unambiguous: The Australian woman saw that the famous doctor had been drinking quite a lot. (8c) NP/Z ambiguous: Before the woman visited the famous doctor had been drinking quite a lot. (8d) NP/Z unambiguous: Before the woman visited, the famous doctor had been drinking quite a lot.

These were constructed on the basis of pretests that equated plausibility of the NP analysis and degree of verb bias across conditions (see below). Additionally, the sentences did not differ from the ambiguous noun phrase onward. The unambiguous conditions were created by adding the complementizer that after the first verb in the NP/S construction and a comma after the first verb in the NP/Z construction (cf. Pickering & Traxler, 1998). The ambiguous conditions had identical or nearly identical numbers of characters. The experimental design treated ambiguity (ambiguous vs unambiguous) and construction type (NP/S vs NP/Z) as within-items factors. Verb bias pretest. Verb bias was important for two reasons. First, we wished it to be uncontroversial that the verbs used in the experiment were biased toward the NP reading. Current theories all assume that both NP/S and NP/Z ambiguities with NP-biased verbs are initially misanalyzed (e.g., Frazier, 1987; Trueswell et al., 1993), though we return to this issue

141

REANALYSIS IN LANGUAGE COMPREHENSION

in the discussion below. Second, we wanted each item to use a verb pair which was as closely balanced as possible in terms of the degree to which the NP analysis was preferred over the alternative analysis. To check this, we ran the CORSET corpus search tools (Corley, Corley, & Crocker, 1997) on the 100-millionword British National Corpus3 to obtain a random sample of sentences containing one of a preselected set of NP/S and NP/Z verbs in the past tense. We removed sentences in which the complements of the verb appeared in noncanonical positions due to extraposition, heavy NP shift, or long-distance dependencies. The sentences were then categorized according to the complements appearing with the verb. If the verb appeared with an NP direct object and no other complement, the sentence was classified as NP. If it appeared with a tensed clausal complement and no overt complementizer, it was classified as S. If it appeared with no overt complement at all, it was classified as Z. For each NP/S verb, the NP bias was calculated as the following ratio, where uNPu represents the number of NP sentences in the sample and uSu represents the number of S sentences in the sample: ?NP? ?S? 1 ?SP? The NP bias for the NP/Z verbs was calculated in a completely analogous way except that uNPu was compared with uZu (i.e., the number of Z sentences in the sample). The mean NP bias was 0.89 for both the NP/Z and the NP/S verbs (F , 1). Table 1 shows the the NP bias of each verb pair used in the experiment. Plausibility pretest. As we have seen above, reanalysis difficulty can be affected by the plausibility of the misanalysis. To control for this, we conducted a norming study in which 20 further participants rated the plausibility of sentences created from the misanalysis of the NP/S and NP/Z items (e.g., The Australian woman saw the famous doctor and The woman visited 3

See http://info.ox.ac.uk/bnc/ for documentation of the British National Corpus.

TABLE 1 NP Bias of Each Verb Pair in Experiment 1 NP/S verb

NP bias

NP/Z verb

NP bias

understood accepted recalled heard confirmed maintained forgot mentioned found announced discovered noticed saw acknowledged remembered remembered read revealed revealed doubted

.92 .93 .87 .89 .81 .98 .89 .94 .94 .91 .71 .65 .97 .97 .97 .97 .99 .79 .79 .83

negotiated polished scratched packed typed built painted debated lost investigated watched knitted visited questioned attacked invaded edited washed followed typed

.94 .93 .91 .89 .86 .97 .94 .90 .90 .93 .68 .67 .98 .99 .97 .95 .98 .77 .78 .86

the famous doctor) on a scale from 1 (least plausible) to 7 (most plausible). The pretest included 40 fillers, including extremely implausible sentences and mildly implausible sentences. There was no significant difference in the mean ratings for the NP/S and NP/Z sentences (NP/S: 6.66; NP/Z: 6.69; both Fs , 1). Procedure The sentences were divided into four regions. The first region boundary was placed immediately before the first verb. The second region boundary was placed immediately before the first disambiguating word. The critical third region contained 2– 4 words, and the fourth region contained a final phrase, which was also 2– 4 words in length (see the Appendix for region breaks). The items were divided into four lists each of which contained exactly one condition of each experimental item so that each list had an equal number of items in each condition. In addition to the experimental items, we included 96 filler sentences. The fillers included eight NP/S ambiguities and eight NP/Z ambiguities, which

142

STURT, PICKERING, AND CROCKER

were all resolved in favor of the NP analysis. These were included to increase the likelihood of misanalysis in the experimental items and therefore to encourage reanalysis. None of the remaining fillers exhibited the ambiguities being studied. Eight fillers included commas and eight included complementizers. The experiment employed the PsyScope experimental package (Cohen, MacWhinney, Flatt, & Provost, 1993) and was run on Motorola Starmax machines running the Macintosh operating system, connected to a PsyScope button box. Each participant was seated in front of the computer monitor in a soundproof booth and was instructed to press the middle button of the button box with the index finger of his or her dominant hand to see each subsequent region of the sentence. As each region was presented, all preceding and subsequent regions were displayed as a series of underscore characters, with spaces in the positions between the words of the actual text. The first three regions appeared on one line and the fourth region on the next line. A yes/no question followed 50% of the experimental and filler items, which the participants had to answer by pressing the right (yes) or left (no) button as appropriate. The appropriate answer for the questions was balanced equally between “yes” and “no.” The experiment was preceded by five practice trials, and the experiment proper began with at least 10 filler items. The experimental items and fillers were randomized independently for each participant so that no two experimental items appeared adjacent to each other. An equal number of participants was assigned to each of the four items lists. Results and Discussion We removed all trials including a region with a reading time that was either extremely long (8000 ms or over) or extremely short (250 ms or under). For the remaining data, all reading times over 2.5 standard deviations either side of the mean for each participant and each region were replaced with the cut-off value. Together, these two procedures affected less than 5% of the data. Separate 2 3 2 analyses of variance were computed on the reading times for each region,

TABLE 2 Reading Times for Experiment 1 Region Region Region Region 1 2 3 4 NP/S (ambiguous) NP/S (unambiguous) NP/Z (ambiguous) NP/Z (unambiguous)

990 981 914 998

1183 1282 1269 1384

877 790 1335 935

771 768 848 832

treating participants (F1) and items (F2) as the random factor. Mean reading times for all regions can be found in Table 2. For completeness, we report data for all regions, although the first and second regions are not crucial to our argument. No significant main effects or interactions were found in the first region (all Fs , 2). Analysis of the second region revealed a main effect of ambiguity, with ambiguous sentences being read more quickly than their unambiguous counterparts, for both NP/S and NP/Z constructions [F1(1,23) 5 11.86, MSe 5 22858, p , .01; F2(1,31) 5 7.74, MSe 5 48093, p , .01]. This effect is probably due to the increased number of characters associated with the presence of the complementizer and comma in these conditions and, possibly, wrap-up effects associated with the comma (cf. Rayner, Sereno, Morris, Schmauder, & Clifton, 1989). There was a nonsignificant tendency for NP/S constructions to be read more quickly than NP/Z constructions [F1(1,23) 5 2.27, MSe 5 93810, p . .1; F2(1,31) 5 2.93, MSe 5 95448, p , .1]. There was no interaction between ambiguity and construction type (both Fs , 1). In the third region there were main effects of ambiguity [F1(1,23) 5 26.74, MSe 5 53266, p , .001; F2(1,31) 5 59.33, MSe 5 33285, p , .001] and construction type [F1(1,23) 5 22.36, MSe 5 97488, p , .001; F2(1,31) 5 28.96, MSe 5 104302, p , .001]. The two factors also interacted significantly [F1(1,23) 5 15.29, MSe 5 38360, p , .001; F2(1,31) 5 11.60, MSe 5 68979, p , .01]. Simple effects analysis revealed that unambiguous sentences were read more quickly than their ambiguous counterparts both for the NP/S constructions [F1(1,23) 5 6.97, MSe 5 13122, p , .02; F2(1,31) 5 4.19,

REANALYSIS IN LANGUAGE COMPREHENSION

MSe 5 31101, p , .05] and for the NP/Z constructions [F1(1,23) 5 24.45, MSe 5 78503, p , .001; F2(1,31) 5 37.16, MSe 5 71163, p , .001]. Additionally, the NP/S constructions were read more quickly than the NP/Z constructions, not only in the ambiguous conditions [F1(1,23) 5 23.24, MSe 5 108156, p , .001; F2(1,31) 5 32.92, MSe 5 105266, p , .001], but, interestingly, also in the unambiguous conditions [F1(1,23) 5 9.11, MSe 5 27692, p , .01; F2(1,31) 5 5.23, MSe 5 68016, p , .05]. Note that the critical third region used identical words in all conditions, eliminating the problems associated with analyzing conditions with different lengths (Trueswell, Tanenhaus, & Garnsey, 1994). In the final region, there was a nonsignificant tendency for NP/S constructions to be read more quickly than NP/Z constructions [F1(1,23) 5 2.74, MSe 5 44292, p . .1; F2(1,31) 5 1.91, MSe 5 85254, p . .1]. There was no effect of ambiguity or interaction between the two factors (all Fs , 1). Overall comprehension accuracy for the 16 items which included questions was 92%, and there were no significant differences in accuracy amongst the conditions. The data from Experiment 1 show clearly that both NP/S and NP/Z ambiguous sentences are associated with difficulty in the disambiguating region, but that the garden path effect in the NP/Z sentences is considerably larger than in the NP/S sentences. If we define “garden path effect” as the difference in reading time between ambiguous and unambiguous conditions in the disambiguating region, then the garden path effect for the NP/Z sentences is 400 ms, which is more than four-and-a-half times the size of that of the NP/S conditions (87 ms). The interaction between the two factors of construction type and ambiguity shows us that this difference in magnitude of garden path is highly reliable. One interesting aspect of the data is the fact that the unambiguous NP/Z sentences were also read more slowly than the corresponding unambiguous NP/S sentences. As we have mentioned above, this may be due to the fact that the order of main and subordinate clauses differs. Alternatively, it may be that the disambiguating

143

comma is missed in a proportion of trials. The resulting garden paths for these trials would have the effect of increasing the mean reading time of the third region in the NP/Z unambiguous cases. Whatever the reason for the relative difficulty of the unambiguous NP/Z sentences, the statistically reliable interaction shows us that the difficulty associated with the ambiguous NP/Z sentences is more than just an additive effect associated with the difficulty of NP/Z sentences in general. However, there is an alternative explanation of the interaction, which we will call the probability of misanalysis account. It is possible that the cost of reanalysis does not differ between the two construction types we have investigated in this experiment, but that there is a difference in the probability with which participants adopt the (globally incorrect) NP analysis. If so, readers might have experienced particular difficulty with the NP/Z ambiguous sentences because they nearly always initially misanalyzed them, but misanalyzed the NP/S ambiguous sentences considerably less often. Notice that this explanation requires that readers often adopt the S analysis for NP/S ambiguous sentences even when the verb is NP biased. There are two main reasons to doubt this explanation for our findings. First, it makes wrong predictions about the distributions of reading times within each condition. If people misanalyze ambiguous sentences less than 100% of the time, then reading times in the disambiguating region will be drawn from two separate populations and thus be approximately bimodal in their distribution. More specifically, there should be one set of slower reading times, reflecting trials on which misanalysis occurred, and another set of faster reading times, reflecting trials on which no misanalysis occurred. We would expect the standard deviation of such an approximately bimodal distribution to be greater than that of both a unimodal distribution reflecting misanalysis alone and a unimodal distribution reflecting no misanalysis at all. If the NP/S ambiguous sentences are often not misanalyzed, and the NP/Z sentences are always or nearly always misanalyzed, then the standard deviation for the NP/S ambiguous condition

144

STURT, PICKERING, AND CROCKER TABLE 3

Mean Standard Deviations for Participants and Items Data in Region 3 in Experiment 1

NP/S (ambiguous) NP/S (unambiguous) NP/Z (ambiguous) NP/Z (unambiguous)

Participants

Items

360 260 561 413

394 281 765 450

should be larger in comparison to its unambiguous control condition than the NP/Z ambiguous condition in comparison to its unambiguous control condition. To check this possibility, we ran analyses of variance on the standard deviations of reading times for the third region, for each participant (F1), and each item (F2). The mean standard deviations are given in Table 3. The analyses of variance yielded a main effect of ambiguity such that ambiguous conditions had a greater standard deviation than unambiguous conditions [F1(1,23) 5 14.00, MSe 5 26222, p , .01; F2(1,31) 5 19.45, MSe 5 75003, p , .001] and a main effect of construction type such that the NP/Z conditions had a greater standard deviation than the NP/S conditions [F1(1,23) 5 10.30, MSe 5 26222, p , .005; F2(1,31) 5 17.86, MSe 5 132582, p , .001]. The interaction between the two factors approached significance on the items analysis, but not on the participants analysis [F1 , 1; F2(1,31) 5 2.97, MSe 5 109628, p , .1]. Note that this weak trend toward an interaction is actually the opposite predicted by the probability of misanalysis account. Hence, the evidence from the distribution of reading times does not support the probability of misanalysis account. A second argument against the probability of misanalysis account is based on degree of verb bias for individual items. The most plausible reason why participants might have adopted the NP analysis more often in the NP/Z ambiguous condition than the NP/S ambiguous condition is that probability of misanalysis is closely related to degree of verb bias. Although our pretests showed that the NP/S verbs and the NP/Z verbs did not differ in their degree of NP bias (both

means .89), it is possible that we did not pay attention to the appropriate level of granularity in determining these biases and that our NP/Z sentences are more strongly biased than our NP/S sentences at the appropriate level of granularity. Specifically, the probability of the NP analysis may be particularly high in the syntactic contexts employed in the NP/Z ambiguous condition, where the verb occurs in a preposed subordinate clause and is immediately followed by a noun phrase with no intervening comma. To investigate whether this is the case, we performed another search of the British National Corpus, which was restricted to NP/Z verbs appearing in the relevant context. Specifically, we searched for sentences which began with a subordinating conjunction (e.g., while, because, after) and continued with a noun phrase, an NP/Z verb used in our experimental items, and another noun phrase with no preceding comma. This search yielded a total of 162 sentences, which we classified either as NP (the noun phrase following the NP/Z verb was attached as its object and the verb took no other complement) or Z (this noun phrase was attached as the subject of the following main clause verb). Four verbs never occurred as either NP or Z (edited, investigated, knitted, scratched). The remaining 15 verbs were used to calculate revised NP biases, using the same formula as in the verb bias pretest above. The mean NP bias for these verbs was .95, as opposed to .90 for the same 15 verbs in the verb bias pretest. Because this estimate of the bias could be particularly unreliable on very small samples, we performed the same analysis on the four verbs which appeared at least 10 times in this context (followed, lost, visited, watched). The mean NP bias for these four verbs was .91 as opposed to .83 for the same four verbs in the verb bias pretest. It is therefore true that the overall NP bias of our NP/Z verbs is slightly elevated when we consider them in the contexts found in the experiment. In theory, this small difference might affect the proportion of trials on which misanalysis occurred in the two ambiguous conditions. However, for this to be the case, degree of bias, within the much wider range employed in our

REANALYSIS IN LANGUAGE COMPREHENSION

items (.65–.99), should have affected processing difficulty. In fact, there was no relation between processing difficulty and degree of verb bias. We performed correlational analyses between the garden path effect of each item (i.e., ambiguous reading time minus unambiguous reading time) and degree of NP-bias of the verb for both NP/S and NP/Z conditions, using the original measures of bias reported in the verb bias pretest. No correlation was found either for the NP/S conditions (r 5 20.12, n.s.) or for the NP/Z conditions (r 5 20.04, n.s.). This suggests that the magnitude of difficulty was not affected by verb bias and so the question of whether the relevant bias for our NP/Z ambiguous items was .89 or slightly higher cannot explain our effects. In conclusion, the particular difficulty with the NP/Z ambiguous items must have been due to difficulty of reanalysis rather than probability of misanalysis. EXPERIMENT 2 Experiment 1 showed that type of structural change has an effect on reanalysis difficulty. Experiment 2 tested the relationship between structural change and head position, a factor which can affect reanalysis difficulty in certain tasks (Ferreira & Henderson, 1991). Consider (6) again: (6a) After the boy scratched the big and hairy dog yawned loudly. (6b) After the boy scratched the dog that is hairy yawned loudly.

Ferreira and Henderson interpreted their finding that (6b) is judged grammatical far less often than (6a) in terms of a model in which a thematic role is assigned at the head of the phrase receiving the role in question. They proposed that multiple subcategorization frames are activated when a verb with ambiguous argument structure, such as scratched, is read and that the assignment of a thematic role causes alternative argument structures to decay rapidly. Therefore, the relative difficulty of (6b) can be explained by assuming that the transitive argument structure associated with scratched has been assigned earlier than in (6a) due to the difference in the position of the word dog. This means that

145

in (6b), the alternative intransitive argument frame has decayed and become less accessible, giving rise to greater reanalysis difficulty. More recently, Ferreira and Henderson (1998) motivated a different account in terms of thematic processing domains. An alternative interpretation of these findings has been offered by Fodor and Inoue (1998), who propose a model of reanalysis in which NP/Z ambiguities like (6a) and (6b) are difficult because there is no direct link (in terms of grammatical dependencies) between the main verb (yawned) and the ambiguous noun phrase (the big and hairy dog/the dog that is hairy). In such cases, the parser cannot reanalyze straightforwardly and is therefore forced to use an emergency strategy called theft to perform the reanalysis. In theft, the parser does not consider tree structure, but instead removes previously analyzed words and incorporates them with incoming words. The head position effect is then simply explained in terms of the fact that the head noun dog is closer in the word string to the disambiguating word yawned in (6a) than it is in (6b). To date, head position effects have not been experimentally investigated in NP/S sentences. However, Fodor and Inoue make the prediction that head position should not affect the difficulty of NP/S reanalysis because NP/S reanalysis does not require theft. The main experimental data on head position comes from speeded grammaticality judgement. However, data from this task are not necessarily informative about on-line processing, since the grammaticality judgment is made after the sentence has been presented. An eyetracking study reported in Ferreira and Henderson (1993) failed to replicate the head-position effect in on-line measures, though the effect did occur in grammaticality judgments performed after the participants had read the sentences. We therefore decided to examine the effect of head position using self-paced reading to look for any on-line effects in our items. In order to compare any effect of head position with the effect of structural change which we found in Experiment 1, we manipulated head position and construction type within the same experiment. We can therefore investigate whether the manipula-

146

STURT, PICKERING, AND CROCKER

tion of head position differentially affects NP/S and NP/Z ambiguities as predicted by Fodor and Inoue (1998). The experiment also allows us to assess whether the difference between NP/S and NP/Z sentences found in Experiment 1 is dependent on the type of ambiguous region involved. Method Participants. Twenty-eight further participants took part in the experiment. Items. We employed 32 items like (9a) through (9d) below: (9a) NP/S recent: The Australian woman saw the famous doctor had been drinking quite a lot. (9b) NP/S nonrecent: The Australian woman saw the doctor who was famous had been drinking quite a lot. (9c) NP/Z recent: Before the woman visited the famous doctor had been drinking quite a lot. (9d) NP/Z nonrecent: Before the woman visited the doctor who was famous had been drinking quite a lot.

The recent head conditions were identical to the ambiguous conditions of Experiment 1, and the nonrecent head conditions were constructed by replacing the adjective in the second region (famous) with a relative clause (who was famous), thus separating the head (doctor) from the disambiguating material in the third region. Procedure As in Experiment 1, we included 96 filler sentences. The fillers included 16 NP/S ambiguities and 16 NP/Z ambiguities, which were all resolved in favor of the NP analysis. In half of these, the ambiguous NP included an adjective and in half it included a relative clause. The remaining fillers used a variety of constructions, none of which included the ambiguities being investigated in the experiment. Otherwise, the procedure was identical to Experiment 1. Results and Discussion Before conducting analyses, the data were trimmed according to the same procedure as Experiment 1, again affecting less than 5% of the data. Mean reading times for Experiment 2

TABLE 4 Reading Times for Experiment 2

NP/S (recent) NP/S (nonrecent) NP/Z (recent) NP/Z (nonrecent)

Region 1

Region 2

Region 3

Region 4

1008 989 974 944

1335 1645 1332 1702

1004 1037 1432 1411

860 899 1056 943

are given in Table 4. As in Experiment 1, we report data for all the regions, although the first and second regions are not central to our argument. No interactions or main effects approached significance on the first region (all Fs , 2). In the second region, there was a main effect of head position, with the distant head conditions being read more slowly than their recent head counterparts [F1(1,27) 5 37.28, MSe 5 86616, p , .001; F2(1,31) 5 33.89, MSe 5 98852, p , .001]. This effect can be attributed to length differences between the two conditions. There was no effect of construction type, and no interaction between the two factors (all Fs , 1). In the third and disambiguating region, there was a main effect of construction type, with NP/Z conditions being read more slowly than their NP/S counterparts [F1(1,27) 5 25.30, MSe 5 178303, p , .001; F2(1,31) 5 83.62, MSe 5 61303, p , .001]. This replicates the main effect of construction type found in Experiment 1. However, there was no effect of head position and no interaction between the two factors (all Fs , 1). Note that, as in Experiment 1, the third region includes identical words across all conditions. In the fourth region, there was a main effect of construction type [F1(1,27) 5 6.31, MSe 5 63985, p , .02; F2(1,31) 5 17.16, MSe 5 28692, p , .001], with the NP/Z conditions taking longer to process than the NP/S conditions, but no effect of head position (both Fs , 2). There was a marginal interaction between the two factors, which only approached significance on the participant analysis [F1(1,27) 5 3.56, MSe 5 45306, p 5 .07; F2(1,31) 5 1.55, MSe 5 89666, p . .2]. The tendency is for the

REANALYSIS IN LANGUAGE COMPREHENSION

NP/Z recent condition to be harder than the NP/Z nonrecent condition, contrary to the predictions of Ferreira and Henderson (1991, 1998) and Fodor and Inoue (1998). Overall comprehension accuracy for the 16 items that included questions was 91%, and there were no significant differences in accuracy among the conditions. These results replicate the effect of construction type seen in the previous experiment, with NP/Z sentences being harder in the disambiguating region than NP/S sentences. However, there is no evidence that head position had an effect on the difficulty of reanalysis. Nor is there any evidence for the interaction predicted by the theft-based account of Fodor and Inoue (1998). The lack of a head position effect is striking, given that the effect of Ferreira and Henderson (1991) was very large. We suggest that the difference is related to the different tasks used in the two experiments. Specifically, Ferreira and Henderson used a Speeded Grammaticality Judgement Task, where participants were required to judge the grammaticality of each stimulus sentence after it had been presented using rapid serial visual presentation. When judging the grammaticality of particularly difficult sentences, participants may rely on superficial clues such as whether the main verb is close to a word which may function as the head of its subject. These effects could purely reflect processes which are active after the participant has finished reading the sentence and may not reflect on-line behavior at all. This conclusion is suggested by our data from Experiment 2, as well as the eye-tracking study reported in Ferreira and Henderson (1993) [though cf. Konieczny, Hemforth, and Scheepers (in press)]. CONCLUSIONS Experiment 1 showed that there was a reliable difference in the processing of NP/S and NP/Z ambiguities over and above any differences in the processing of unambiguous sentences of this type. Experiment 2 replicated this difference, but showed no evidence for an online effect of head position. We have thus found very strong evidence that

147

the type of structural change is an important determinant of reanalysis difficulty, suggesting that the intuition which has been the basis for a number of theories of reanalysis in the past is correct. This contrasts with head position, which does not appear to have a detectable effect. In conclusion, previous experimental work related to reanalysis has appeared to favor a theory of reanalysis which is based on nonstructural factors; that is, in Frazier and Clifton’s (1998) terms, a theory of reanalysis based on sentence token rather than sentence type. The experiments reported in the present paper demonstrate, however, that sentence type is also an important factor of reanalysis difficulty, as predicted by a wide range of structural change theories in the literature. APPENDIX Experimental Items The items shown below were used for the ambiguous conditions in Experiment 1 and for the recent head conditions in Experiment 2 (a 5 NP/S; b 5 NP/Z). To reconstruct the unambiguous conditions for Experiment 1, add that after the first verb for the NP/S sentences and a comma after the first verb for the NP/Z sentences. To reconstruct the distant head conditions for Experiment 2, remove the adjective from the second region and add the appropriate relative clause in the past tense, using which for inanimate nouns and who for animate nouns (e.g., the technical contract becomes the contract which was technical and the angry nurse becomes the nurse who was angry). Slashes correspond to region breaks. (1a) The faithful employees / understood the technical contract / would be changed / very soon. (1b) Because the employees / negotiated the technical contract / would be changed / very soon. (2a) The American diplomats / understood the unpopular treaty / would be altered / at the conference. (2b) Because the diplomats / negotiated

148

STURT, PICKERING, AND CROCKER

the unpopular treaty / would be altered / at the conference. (3a) The young woman / accepted the valuable trophy / had to be returned / to the committee. (3b) After the woman / polished the valuable trophy / had to be returned / to the committee. (4a) The young man / accepted the enormous car / looked more expensive / than before. (4b) After the man / polished the enormous car / looked more expensive / than before. (5a) The extremely old man / recalled the angry nurse / had complained / to the doctor. (5b) Whenever the old man / scratched the angry nurse / had complained / to the doctor. (6a) The foreign traveller / heard the loud clock / had woken everybody up / in the youth hostel. (6b) Before the traveller / packed the loud clock / had woken everybody up / in the youth hostel. (7a) The tabloid journalist / confirmed the sensational story / would be published / on the front page. (7b) Hours after the journalist / typed the sensational story / would be published / on the front page. (8a) The construction workers / maintained the rickety walls / fell down / in a heap. (8b) After the construction worker / built the rickety walls / fell down / in a heap. (9a) The clumsy apprentice / forgot the rusty bicycle / was standing / in the garage. (9b) While the apprentice / painted the rusty bicycle / was standing / in the garage. (10a) The industrial committee / mentioned the complex issue / would cause a problem / at the meeting. (10b) However much the committee / debated the complex issue / would cause a problem / at the meeting. (11a) The British army / found the expen-

sive medical supplies / saved many lives / during the conflict. (11b) Although the army / lost the expensive medical supplies / saved many lives / during the conflict. (12a) The Labour candidate / found the secret documents / could be very useful / to the press. (12b) If the Tory candidate / lost the secret documents / could be very useful / to the press. (13a) The chartered accountants / announced the impressive results / had appeared / in the report. (13b) Before the accountants / investigated the impressive results / had appeared / in the report. (14a) The corporate lawyers / announced the controversial proposals / would be sent / for approval. (14b) Before the lawyers / investigated the controversial proposals / would be sent / for approval. (15a) The tall football manager / discovered the skillful player / tried to show off / all the time. (15b) Because the football manager / watched the skillful player / tried to show off / all the time. (16a) The tall theatre director / discovered the young actor / got stage fright / very badly. (16b) Because the theatre director / watched the young actor / got stage fright / very badly. (17a) The young woman / noticed the fluffy sweater / changed shape / quite badly. (17b) While the woman / knitted the fluffy sweater / changed shape / quite badly. (18a) The younger children / noticed the woollen socks / were the wrong colour / for school. (18b) Because the children / knitted the woollen socks / were the wrong colour / for school. (19a) The Japanese coach party / saw the famous palace / was being restored / very carefully.

REANALYSIS IN LANGUAGE COMPREHENSION

(19b) When the coach party / visited the famous palace / was being restored / very carefully. (20a) The Australian woman / saw the famous doctor / had been drinking / quite a lot. (20b) Before the woman / visited the famous doctor / had been drinking / quite a lot. (21a) The old market researcher / acknowledged the clever assistant / worked very hard / on the project. (21b) While the market researcher / questioned the clever assistant / worked very hard / on the project. (22a) The tall lawyer / acknowledged the young witness / had been successful / in convincing the jury. (22b) Before the lawyer / questioned the young witness / had been successful / in convincing the jury. (23a) The tall knights / remembered the ancient castle / had been defended / very heavily. (23b) Before the knights / attacked the ancient castle / had been defended / very heavily. (24a) The tall fighter pilots / remembered the camouflaged airfield / was evacuated / very quickly. (24b) Before the fighter pilots / attacked the camouflaged airfield / was evacuated / very quickly. (25a) The old soldiers / remembered the industrial town / had been flattened / in the bombing raid. (25b) Before the soldiers / invaded the industrial town / had been flattened / in the bombing raid. (26a) The big policemen / remembered the remote peace camp / had been filmed / by the BBC. (26b) Before the policemen / invaded the remote peace camp / had been filmed / by the BBC. (27a) The academic publisher / read the controversial manuscript / had been reviewed / by a professor. (27b) Before the publisher / edited the

149

controversial manuscript / had been reviewed / by a professor. (28a) The contemporary fashion writer / read the famous magazine / was having trouble / with its finances. (28b) Even after the fashion writer / edited the famous magazine / was having trouble / with its finances. (29a) The tall archaeologist / revealed the valuable coins / had gone missing / from the site. (29b) Before the archaeologist / washed the valuable coins / had gone missing / from the site. (30a) The French male stripper / revealed his oily chest / felt uncomfortable / in the heat. (30b) Whenever the male stripper / washed his oily chest / felt uncomfortable / in the heat. (31a) The young journalist / revealed the corrupt politician / received some payments / from the millionaire. (31b) While the journalist / followed the corrupt politician / received some payments / from the millionaire. (32a) The old secretary / doubted the incriminating report / would be distributed / to many people. (32b) After the secretary / typed the incriminating report / would be distributed / to many people. REFERENCES Bader, M. (1998). Prosodic influences on reading syntactically ambiguous sentences. In J. D. Fodor & F. Ferreira (Eds.), Reanalysis in sentence processing (pp. 1–56). Dordrecht, The Netherlands: Kluwer. Clifton, C. (1993). Thematic roles in sentence processing. Canadian Journal of Psychology, 47, 222–246. Cohen, J. D., MacWhinney, B., Flatt, M., & Provost, J. (1993). Psyscope: A new graphic interactive environment for designing psychological experiments. Behavioral Research Methods, Instruments, and Computers, 25, 257–271. Corley, S., Corley, M. M. B., & Crocker, M. W. (1997). CORSET II User Manual. (Unpublished Technical Manual, Human Communication Research Centre, University of Edinburgh.) Ferreira, F., & Henderson, J. M. (1990). Use of verb information in syntactic parsing: Evidence from eye movements and word-by-word self-paced reading. Journal

150

STURT, PICKERING, AND CROCKER

of Experimental Psychology: Learning, Memory and Cognition, 16(4), 725–745. Ferreira, F., & Henderson, J. M. (1991). Recovery from misanalyses of garden-path sentences. Journal of Memory and Language, 30, 725–745. Ferreira, F., & Henderson, J. M. (1993). Reading processes during syntactic analysis and reanalysis. Canadian Journal of Psychology, 47(2), 247–275. Ferreira, F., & Henderson, J. M. (1998). Syntactic reanalysis, thematic processing, and sentence comprehension. In J. D. Fodor & F. Ferreira (Eds.), Reanalysis in sentence processing. Dordrecht, The Netherlands: Kluwer. Fodor, J. D., & Inoue, A. (1998). Attach Anyway. In J. D. Fodor & F. Ferreira (Eds.), Reanalysis in sentence processing. Dordrecht, The Netherlands: Kluwer. Frazier, L. (1978). On comprehending sentences: Syntactic parsing strategies. Unpublished doctoral dissertation, University of Connecticut. Frazier, L. (1987). Sentence processing: A tutorial review. In M. Coltheart (Ed.), Attention and performance (Vol. 12, pp. 559 –586). Hillsdale, NJ: Erlbaum. Frazier, L., & Clifton, C. (1998). Sentence reanalysis, and visibility. In J. D. Fodor & F. Ferreira (Eds.), Reanalysis in sentence processing. Dordrecht, The Netherlands: Kluwer. Frazier, L., & Rayner, K. (1982). Making and correcting errors during sentence comprehension: Eye movements in the analysis of structurally ambiguous sentences. Cognitive Psychology, 14, 178 –210. Garnsey, S. M., Pearlmutter, N. J., Myers, E., & Lotocky, M. A. (1997). The contributions of verb bias and plausibility to the comprehension of temporarily ambiguous sentences. Journal of Memory and Language, 37, 58 –93. Gibson, E. A. F. (1991). A computational theory of human linguistic processing: Memory limitations and processing breakdown. Unpublished doctoral dissertation, Carnegie Mellon University, Pittsburgh, PA. Gorrell, P. (1995). Syntax and parsing. Cambridge, England: Cambridge University Press. Jurafsky, D. (1996). A probabilistic model of lexical and syntactic access and disambiguation. Cognitive Science, 20, 137–194. Konieczny, L., Hemforth, B., & Scheepers, C. (in press). Head position and clause boundary effects in reanalysis. In L. Konieczny & B. Hemforth (Eds.), Aspects of German parsing. Dordrecht, The Netherlands: Kluwer. MacDonald, M. C., Pearlmutter, N. J., & Seidenberg, M. S. (1994). Lexical nature of syntactic ambiguity resolution. Psychological Review, 101(4), 676 –703. Marcus, M., Hindle, D., & Fleck, M. (1983). D-theory: Talking about talking about trees. In Proceedings of the 21st annual meeting of the Association for Computational Linguistics (pp. 129 –136). Cambridge, MA.

Mitchell, D. C., Corley, M. M. B., & Garnham, A. (1992). Effects of context in human sentence parsing: Evidence against a discourse-based proposal mechanism. Journal of Experimental Psychology: Learning, Memory and Cognition, 18(1), 69 – 88. Mitchell, D. C., & Holmes, V. M. (1985). The role of specific information about the verb in parsing sentences with local ambiguity. Journal of Memory and Language, 24, 542–559. Partee, B., ter Meulen, A., & Wall, R. E. (1993). Mathematical methods in linguistics. Dordrecht, The Netherlands: Kluwer. Pickering, M. J., & Traxler, M. J. (1998). Plausibility and recovery from garden paths: An eye-tracking study. Journal of Experimental Psychology: Learning, Memory, and Cognition, 24(4), 940 –961. Pritchett, B. L. (1988). Garden path phenomena and the grammatical basis of language processing. Language, 64(3), 539 –576. Pritchett, B. L. (1992). Grammatical competence and parsing performance. University of Chicago Press. Rayner, K., Sereno, S. C., Morris, R. K., Schmauder, A. R., & Clifton, C. (1989). Eye movements and on-line language comprehension processes. Language and Cognitive Processes, 4(3), 21– 49. Sturt, P. (1997). Syntactic reanalysis in human language processing. Unpublished doctoral dissertation, Centre for Cognitive Science, University of Edinburgh, Edinburgh, Scotland. Sturt, P., & Crocker, M. W. (1996). Monotonic syntactic processing: A cross-linguistic study of attachment and reanalysis. Language and Cognitive Processes, 11(5), 449 – 494. Sturt, P., & Crocker, M. W. (1997). Thematic monotonicity. Journal of Psycholinguistic Research, 26(3), 297–322. Sturt, P., & Crocker, M. W. (1998). Generalized monotonicity for reanalysis models. In J. D. Fodor & F. Ferreira (Eds.), Reanalysis in sentence processing. Dordrecht, The Netherlands: Kluwer. Trueswell, J. C., Tanenhaus, M. K., & Garnsey, S. M. (1994). Semantic influences on parsing: Use of thematic role information in syntactic ambiguity resolution. Journal of Memory and Language, 33, 285–318. Trueswell, J. C., Tanenhaus, M. K., & Kello, C. (1993). Verb-specific constraints on sentence processing: Separating effects of lexical preference from garden paths. Journal of Experimental Psychology: Learning, Memory, and Cognition, 19(3), 528 –553. Weinberg, A. (1993). Parameters in the theory of sentence processing: Minimal commitment theory goes East. Journal of Psycholinguistic Research, 22(3), 339 –364. (Received May 15, 1998) (Revision received July 21, 1998)