|
Do Speakers Help Listeners? Karen Mims and John C. Trueswell, University of Pennsylvania | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
What must a speaker consider when forming an utterance? Of course, he must consider his choice of words; if he wants to convey the idea of a dog he says dog, and if he wants to convey the idea of a cat he says cat. In addition, he mustalso consider syntactic options; to use the active or passive voice, a pronoun or a full noun phrase, or a reduced or full embedded clause. What influences a speaker to choose the active or the passive, or to use a full or reduced embedded clause? One popular hypothesis is ambiguity, proposed by many researchers including Grice (1975), who suggested that speakers may "avoid ambiguity" in their linguistic choices. Because listeners interpret speech incrementally, as they hear it (Tyler & Marslen-Wilson, 1981; Trueswell, Sekerina, Hill, & Logrip, in press; Tanenhaus, Spivey-Knowlton, Eberhard, & Sedivy, 1996), there are points in almost every sentence where the meaning is temporarily unclear. For example, if a speaker says Pick up the red ball, there are several points of temporary (local) ambiguity. At pick, the speaker could say Pick the flower, or Pick up [something]. Once the speaker says Pick up, the Pick the flower is out, but there is still plenty of ambiguity about what the listener should pick up-the cup, the ball, a shoe... There are many possibilities. Once the listener has heard Pick up the red, the possible things that could be picked up are reduced to those things that are red, but if there are several red things, the sentence is still unclear. Only after hearing the entire utterance does the ambiguity resolve itself. Local ambiguity extends beyond the problem of picking out unique referents in the world. Some sentences contain syntactic local ambiguities. For example, in The secretary remembered you ..., when the listener hears you, he may take you to be the direct object (DO) of remembered (e.g. The secretary remembered you from your previous visits) or the subject of an embedded sentential complement (e.g. The secretary remembered you were out for the day). If a listener is expecting you to be the DO of remembered, and it turns out to be the subject of a sentence complement (SC), the listener will get confused at were and will be forced to reevaluate his initial prediction. This is called being led down the "garden-path" and causes processing difficulty for listeners (Frazier & Rayner, 1982). Indeed, there is good evidence that readers and listeners do commit early to a single interpretation of the DO/SC ambiguity and experience a garden-path when the commitment turns out to be incorrect (Ferreira & Henderson, 1991; Trueswell, Tanenhaus, & Kello, 1993; Garnsey, Pearlmutter, Myers, & Lotocky, 1997). When choosing which path to take (DO or SC), readers and listeners weigh factors such as verb bias and plausibility. Verb bias refers to the structures a verb usually prefers to take, and is determined by examining the frequency with which a particular verb occurs with a particular construction in a language corpus. Some verbs are more likely to take a SC, others are biased to take a DO, and others are equibiased, and will take either. Plausibility is the idea that some nouns are more or less plausible as subjects or objects of particular verbs. For example, Tuesday is an implausible DO for the verb knew. Therefore, if a listener hears the fragment The student knew Tuesday..., he is likely to expect a SC continuation, not a DO continuation, because it is implausible to know Tuesday. On the other hand, the answer is a plausible DO for the verb knew (especially when the subject is student), so a listener who hears The student knew the answer... is likely to expect a DO continuation. Speakers may choose their words in part to help their listeners avoid the garden-path. For example, if a speaker simply inserted that after the verb in The secretary remembered (that) you were out for the day, listeners would avoid being led down the garden path because that you unambiguously introduces the embedded SC. In an exhaustive study, Ferreira and Dell (1998) investigated whether speakers would insert that into ambiguous SC sentences. In their study, subjects memorized reduced SC sentences, then recalled them after a delay. Ferreira and Dell hoped this quasi-spontaneous production would reveal speakers' tendency to insert that when a listener would potentially be garden-pathed. Subjects were presented with sets of three sentences and asked to memorize them. Each set of three sentences included one target sentence embedded in two filler sentences. After subjects had memorized the three sentences, they were asked to recall them, one at a time, into a microphone. Ferreira and Dell indicated to the subject which sentence was to be recalled by prompting them with the main subject and verb of the sentence (the cued recall paradigm). The materials were SC sentences that were either locally ambiguous (The secretary remembered you were out for the day) or unambiguous (The secretary remembered I was out for the day-here, I is case-marked to be a subject, and therefore unambiguously signals a SC). To prevent a speaker bias either for or against inserting that, Ferreira and Dell manipulated whether the stimuli included that. Ferreira and Dell predicted that subjects would insert that more often in the ambiguous sentences (...you...) than the unambiguous sentences (...I...) to help a listener avoid the garden-path. They found, unsurprisingly, that speakers inserted that more when it had been present in the stimulus. They did not, however, find any significant differences in that-insertion between the ambiguous and unambiguous conditions. These findings led Ferreira and Dell to concludethat the production system is 'ego-centric'-that speakers do not take a listener's possible garden-path into account. If speakers do not insert that to avoid ambiguity, then why do they? Earlier studies (Bock 1986, 1987; Bock & Irwin, 1980) have shown that speakers have a retrieval-based algorithm for speech production. This algorithm allows speakers to mention easily retrievable words earlier in a sentence and delay the mention of less accessible words until later, when they have had time to retrieve them from memory. Easily retrievable words include those that have been primed (Bock, 1986) and those that have been already mentioned, especially when they are lexically identical to previously mentioned material (Bock & Irwin, 1980). Ferreira and Dell hypothesized that that-insertion might be dependent on how accessible the next word in the sentence was. They predicted that in reduced embedded clause sentences, like the ones they had been studying, if the word following the verb was less accessible, speakers would insert that as a filler word, to give themselves time to retrieve the next wordfrom memory. Ferreira and Dell manipulated the accessibility of the word following that by controlling whether it was the same as the main subject. In the accessible condition, the embedded subject following that was the same as the main subject of the sentence (e.g. I remembered I had an appointment... or You remembered you had an appointment...). In the inaccessible condition, the main subject and the embedded subject were different (e.g. I remembered you had an appointment or You remembered I had an appointment). Ferreira and Dell predicted that speakers would insert that less when the embedded subject is the same as the main subject, and show just that. After studying both the ambiguity avoidancemodel and the retrieval-based model of speech production, Ferreira and Dell found no evidence that speakers insert that to avoid ambiguity to help the listener. Rather, that insertion appeared to depend on low-level memory demands on the production system. That is, speakers seemed to be using that only to fill a pause, analogous to um. Though speakers do not seem to use lexical that-insertion to avoid ambiguity, they do have another way to help listeners avoid the garden-path-prosody. Prosody is the tone and meter of speech, its duration, pitch, and pauses. Speakers may useprosodic cues to signal phrase and clause boundaries, and to group constituents. We are all familiar with the clear, unambiguous way of speaking common to television and radio announcers, and studies have shown that trained speakers effectively produce disambiguating prosodic cues for their listeners (Price, Ostendorf, Shattuck-Hufnagel, & Veilleux, 1988; Nespor & Vogel, 1983). But the debate is still on regarding whether untrained speakers also use disambiguating prosody. Furthermore, even if speakers are producing disambiguating prosody, are listeners using it? Though speakers in Ferreira and Dell's study did not use lexical means of disambiguation, they could have been using prosody. Earlier research by Beach (1991) suggests that prosodic cues exist which could signal a direct object or sentence complement, exactly the ambiguity studied by Ferreira and Dell. Beach (1991) investigated whether listeners could use prosody to predict the syntactic structure of an utterance. She used synthesized speech to precisely manipulate the prosody of sentence fragments. Listeners heard sentences that were spliced off after the main verb or the noun phrase following the main verb (e.g. The scientist found... or The scientist found a combination...). The sentence fragments had either DO or SC prosody imposed on them; prosody was manipulated on the main verb that immediately preceded the ambiguous part of the sentence. DO prosody was indicated with short duration (75 ms) and small pitch drops (2%); SC prosody was indicated with long duration (275 ms) and large pitch drops (30%). Pitch drop on the verb was defined as follows: the pitch value at the top of the first peak minus the pitch value at the bottom of the following trough. After hearing a fragment, the listeners' task was to choose one of two possible continuations, either DO or SC. The continuations were provided for the subjects; they simply had to circle the one they thought was correct. Subjects were not instructed in any way about the grammatical structure of the sentences or any cues theymight use to make their choice. Beach found that subjects reliably chose the DO continuation when the fragment had DO prosody, and the SC continuation when the fragment had SC prosody, regardless of whether the subjects heard the long or short version of the sentence fragment (with or without the post-verbal noun phrase). Because subjects were able to predict sentence structure from even the short fragments, Beach concluded that listeners only require a small amount of prosodic information to make a syntactic commitment. In a second experiment, Beach tested the robustness of the effect by manipulating the prosody-lengthening and shortening the verb duration and changing the magnitude of the pitch drop on the verb. She found evidence for cue-trading, that is, when the pitch cues are ambiguous, the duration cues have strong influence, and vice versa. Beach's (1991) results suggest that speakers in the Ferreira and Dell study may have had another way to disambiguate potentially ambiguous material, namely prosodic markings on the verb. The current study explored this hypothesis by investigating the use of disambiguating cues, both lexical and prosodic, from the point of view of the speaker and the listener. Experiment 1 explores Ferreira and Dell's findings that speakers do not use that to disambiguate locally ambiguous SC sentences, and investigates speakers' use of prosodic cues for disambiguation. A paradigm similar to Ferreira and Dell's, manipulates the syntactic ambiguity of the sentence and the plausibility that listeners will be led down the garden-path. There are two predictions associated with this experiment. The first is that speakers will follow the garden-path model. Speakers will only disambiguate when their listeners will otherwise be led down the garden-path. A second possibility is that speakers will follow the ambiguity avoidance model, and use disambiguating prosody whenever the sentence is locally ambiguous, regardless of the likelihood that their listeners will be led down the garden-path. Experiment 2 investigates the listener's ability to predict, based on
the prosody produced by the speakers in Experiment 1, if a sentence will
continue as a DO sentence or a SC sentence. A paradigm similar to Beach's
requires listeners to predict sentence continuations from spliced
fragments. The recordings made from subjects in Experiment 1 contain only
SC sentences. Based on Beach's findings, I predict that listeners will be
able to use the stronger prosody of the ambiguous (...you...)
sentences to correctly predict SC continuations, and that they will match
DO continuations to the weaker prosody of the unambiguous (...I...)
sentences.
Experiment 1: Production Experiment 1 examines the question of lexical and prosodic disambiguation from the production side. Do untrained speakers (college students) use lexical or prosodic means to disambiguate local ambiguities? Ferreira and Dell's work suggests that speakers do not use lexical means (that insertion). Beach's work proposes that listeners can use prosodic information to predict which syntactic structure will occur. Neither investigation explored speakers' use of prosody. This experiment examinesthis problem further. Experiment 1 uses a modified version of the Ferreira and Dell task, recalling aloud written sentences. Target sentences are like the following example: 1. a. The secretary remembered you were out for the day. Items 1. and 2. differ in the plausibility that the listener will be led down the garden-path. The verb/pronoun combination in 1. (...remembered you...) is plausibly from a DO sentence. Thus, a listener is likely to be garden-pathed. In 2., o n the other hand, the verb/pronoun combination (...learned you...) is not plausibly from a DO sentence, so the listener is unlikely to expect a DO and is therefore unlikely to be led down the garden-path. The a. and b. items differ with regard to ambiguity. The a. items have the verb followed by an ambiguous pronoun (...you...); one cannot tell from the form of the word whether it is an object or a subject. This leads to ambiguity because a listener cannot tell whether you is the direct object of the verb or is the subject of a sentence complement until he hears the rest of the sentence. The b. items are control items-the verb is followed by the unambiguous pronoun I. I is unambiguously marked to be a subject (its object-marked counterpart is me), and therefore signals the listener that a sentence complement follows. The crossing
of the two experimental factors (plausibility and ambiguity) led to four
conditions: ambiguous-plausible (AP), ambiguous-implausible (AI),
unambiguous-plausible (UP), and unambiguous-implausible (UI). Because
Ferreira and Dell did not find any differences in that insertion
between conditions, none are predicted here. However, prosodic
differences between conditions are predicted, based on Beach's findings.
The garden-path hypothesis predicts that speakers will use disambiguating
prosody only in the AP condition, where garden-paths are most likely. In
the other three conditions, either the unambiguous I or the
implausibility of the verb taking you as a DO would encourage
listeners to discount the DO continuation and therefore avoid the
garden-path. The ambiguity avoidance hypothesis predicts that speakers
will use disambiguating prosody in any situation where the listener might
choose the wrong interpretation, that is, in the two ambiguous conditions,
AP and AI. Essentially, the ambiguity avoidance hypothesis predicts that
speakers will use disambiguating prosody in cases of syntactic ambiguity.
Method Subjects Subjects were 33 University of Pennsylvania undergraduates, 21 women and 12 men, who were either paid or received course credit for participating. All subjects were native speakers of English with normal hearing and normal vision (corrected or uncorrected). Four subjects (three women and one man) were excluded from analysis for failing to follow instructions (e.g. speaking before the 10 second delay was over or changing the sentences entirely). The data from 29 subjects, 18 women and 11 men, were retained for analysis. Materials and Design Materials were 60 pairs of sentences, including 20 target items and 40 fillers. A full list of all the stimuli can be found in Appendix A. Target sentence pairs began with the sentence of interest (like 1a - 2b above), then continued with a second sentence designed to continue the story in a natural way. Filler pairs consisted of two filler sentences, also constructed to form a natural story. Sentences were presented to subjects on white 6x4 inch cards. The experimental design crossed two factors: the syntactic ambiguity of the post-verbal pronoun, and the semantic plausibility of you as a direct object for the individual verbs. Ambiguity was manipulated via the post-verbal pronoun, either you or I. You is syntactically ambiguous-it can be a subject or a direct object. I is unambiguously case-marked to be a subject, and cannot be a direct object. The you sentences are locally ambiguous at you between DO or SC completions, and therefore will cause a garden-path if the listener commits to the wrong interpretation. The I sentences, though, are unambiguously SC sentences, and do not cause a garden-path. Because of concerns of biasing speakers outright toward either the DO or the SC interpretation, verbs were chosen to be roughly equibiased, as normed by Garnsey et al. (1997). Garnsey et al. conducted a sentence completion study which presented subje cts with sentence beginnings (e.g. The psychic warned...) and required them to complete the sentence in a natural way. Garnsey et al. classified each completion as a DO, a SC, or some other type of completion. They also recorded (for SC completions) whether the subject had produced a full SC (with a that) or a reduced SC (without a that). The verbs were then classified as being DO-biased, SC-biased, equibiased, or none of the above based on what percentage of their completions were DOs, SCs, or something else. For this study, I chose only equibiased verbs (those that took DO and SC completions roughly equally), to avoid biasing subjects toward one interpretation or the other. In fact, verbs in the plausible condition were, on average, slightly more biased toward SC structure than verbs in the implausible condition. Another motivating factor behind choosing these particular verbs was that Ferreira and Dell had also used verbs from the Garnsey et al. norms; this study was modeledas closely as possible after Ferreira and Dell's to facilitate comparison between the two studies. Among the equibiased verbs, some took you as a DO more plausibly than others (e.g. ...remembered you... is more plausibly from a DO sentencethan ...learned you...). This plausibility factor was manipulated to determine if the implausibility of the post-verbal noun phrase as a DO would mitigate the garden-path by leading listeners toward a SC interpretation. The verbs in the plausiblecondition were balanced with the verbs from the implausible condition for number of phonemes (an average of 6.2 phonemes per word in the plausible condition and 6.4 phonemes per word in the implausible condition). Four lists of stimuli were generated. Each list had 20 target items and 40 filler items. Filler items were the same for each list. Target items for each list were divided among the four conditions (AP, AI, UP, UI), with five items per condition. List 2 was constructed from List 1 by switching the ambiguity for each verb. That is, if learned was followed by you in List 1 (the ambiguous condition), it would be followed by I in List 2 (the unambiguous condition). List 3 was a reverse order of List 1 and List 4 was a reverse order of List 2. Each list was balanced in a number of ways. Half (five) of the ambiguous target items occurred in the first half of the list, and half in the second. Likewise, five of the unambiguous target items occurred in the first half of the list, and five in the second. Plausible and implausible target items were balanced similarly. Ten of the forty fillers were sentences that included that, to guard against a subject bias against that. Fillers were also balanced to include ten items with you and ten with I to ensure subjects did not pick out items with you or I as being special in any way. Finally, the lists were pseudo-randomized (at least one filler appeared before each target, three practice fillers began the experiment, and blocks of target sentences with you or I were broken up). Procedure For each trial, subjects were given a card containing a sentence
pair. They had ten seconds to memorize this pair, then the card was taken
away. The subject sat through a ten second delay, then was cued to recall
the sentences to a confederate liste ner and into a microphone. Subjects'
voices were digitally recorded directly into SoundEdit 16. A confederate
was used instead of a real second subject to control for the listener's
behavior. The confederate's task was to rate the speaker's "naturalness
and clarity." This task was included to encourage subjects to speak with
natural prosody, not in a monotone or in an overly dramatic fashion.
Subjects were advised to try to get the meaning of the sentence across as
clearly as possible. Complete instructions for both the subject and the
confederate can be found in Appendix B. Each utterance was recorded
digitally. After all 60 sentence pairs had been read, the subject and
confederate filled out a questionnaire intended to ascertain what the
subjectthought the experiment was about (no subject guessed the real
intent of the experiment). The subject was then debriefed and
compensated.
Results Lexical insertion Subjects inserted that 33% of the time, on average. As expected, there were no differences in that-insertion across conditions. There is a marginally significant effect of ambiguity; subjects inserted that more in the unambiguous conditions. This effect can be seen in Figure 1. No other effects or interactions approach significance. It is interesting to note that the trend evident in the data is in the opposite direction than the one predicted by the garden-path model. Subjects inserted that most in the UI condition, where listeners are expected to correctly predict the syntactic structure without being garden-pathed. Also, subjects inserted that least in the AP condition, where listeners are expected to experience the most garden-paths. Separate analyses of variance (ANOVAs) were computed for subjects and items, with ambiguity (ambiguous, unambiguous) and plausibility (plausible, implausible) as factors. In the subject analysis, there was a marginally significant main effect of ambiguity in the direction opposite the one predicted by the garden-path model (ambiguous: 30%, unambiguous: 36%, F1(1,29) = 3.8, p = .06; F2(1,29) = 1.7). There were no other effects or interactions. INSERTFIG. 1 HERE To further investigate speakers' motivation for inserting that, the proportion of that insertion in this study was regressed against the proportion of that insertion in Garnsey et al. (1997). In the same sentence completion study from which the stimulus verbs were obtained, Garnsey et al. recorded the proportion of all SC completions that included that. These proportions served as the independent variable against which the proportions of that insertion from this study were regressed. After unequal variance was adjusted for by taking the log of each variable, the regression yielded a r2 of .72 (see Figure 2). This regression suggests that speakers' that insertion is not based on any high-level interactions with ambiguity or plausibility, but rather on some low-level verb-that co-occurrence information. That is, that insertion depends on the frequency with which a particular verb co-occurs with that. INSERT FIG. 2 HERE Prosodic disambiguation Prosodic cues for disambiguation were examined by analyzing duration and pitch data for the target verbs. Prosodic analysis was done only on those trials for which subjects did not insert that or change the verb in any way. After such trials were removed, some subjects had cells with no data. Those subjects plus subjects with one or more cells with only one observation were dropped from further analysis. This left 20 subjects, 14 women and 6 men, for analysis. Duration of the verb was measured in milliseconds from the onset of the verb to the onset of the word following the verb (you or I), including any pause following the verb (pauses after the verb occurred on 10% of all trials). Typical durations were between 300 and 400 ms, though the range was much bigger. Pauses following the verb were included in the duration analysis because pauses are considered an additional cue of a major clause boundary, such as the one between a verb and its sentential complement (Lehiste, Olive, & Streeter, 1976; Price, Ostendorf, Shattuck-Hufnagel, & Fong, 1991; Scott, 1982). Raw duration data for each subject was converted to Z-scores to remove any effects of speakers' varied speech rates. Positive Z-scores indicate a longer duration than average, and therefore a cue toward a SC interpretation (following Beach (1991)). Negative Z-scores indicate shorter duration than average and therefore a cue away from a SC interpretation. Results show an effect of plausibility, with longer verb duration for the implausible verbs and shorter verb duration for the plausible verbs. This effect is in the opposite direction from that predicted by the garden-path model of speech production. Separate ANOVAs of the Z-scores were computed for subjects and items, with ambiguity (ambiguous, unambiguous) and plausibility (plausible, implausible) as factors. The analysis revealed a main effect of plausibility (in the unexpected direction) in the subject analysis (plausible: -0.18, implausible: 0.15, F1(1,19) = 12.3, p < .01; F2(1,18) = 1.5). There were no other effects or interactions. See Figure 3 for a graphical summary of the data. INSERT FIG. 3 HERE Pitch was analyzed using the Xwaves program for Unix, which shows the pitch contours of an utterance in Hertz. Pitch change over the verb was computed as the value of the pitch contour (f0) at the end of the verb subtracted from the value of the pitch contour at the beginning of the word (pitch change = f0beginning ( f0end). Thus, pitch drop is a positive number (the bigger the number, the bigger the pitch drop), and pitch rise is a negative number (the smaller the number, the bigger the pitch rise). Raw pitch data for each subject was standardized to Z-scores to eliminate any effects due to variation in subjects' pitch range. Once standardized, positive numbers indicate a bigger pitch drop and therefore a stronger cue towards a SC interpretation (based on Beach (1991)). Negative numbers indicate a smaller pitch drop (though pitch rises exist in the data, all means show a pitch drop) and a cue away from a SC interpretation. There were no significant effects or interactions for the Z-scores of the pitch data, but the effect of plausibility was marginally significant, in the expected direction (pitch was dropped more on the plausible verbs than the implausible verbs). The raw pitch data (before Z-score standardization) was significant for a main effect of plausibility, in the expected direction. Separate ANOVAs of the pitch Z-scores for subjects and items were computed, with ambiguity (ambiguous, unambiguous) and plausibility (plausible, implausible) as factors. The analyses revealed a marginally significant effect of plausibility in the subject analysis, in the expected direction (plausible: 0.10, implausible: -0.11, F1(1,19) = 2.3, p = .14; F2(1,18) = 1.1). Analysis of the raw pitch data showed a main effect of plausibility in the subject analysis (plausible: 13.9 Hz, implausible: 5.3 Hz, F1(1,18) = 6.0, p < .05; F2(1,18) = 1.7), with no other effects or interactions. See Figure 4 for illustration. INSERT FIG. 4 HERE Because the duration and pitch data had been standardized, the Z-scores could be added together (duration + pitch drop) to give a combined measure of the speakers' prosody. The combined prosody analysis was significant for a main effect of ambiguity, in the expected direction. Taken together, the prosodic cues of duration and pitch disambiguated the syntactic ambiguity. That is, speakers used "more" prosody in the ambiguous conditions (where a listener would have no syntactic cues indicating which structure, DO or SC, was coming up). ANOVAs were computed separately for subject and item means, with ambiguity (ambiguous, unambiguous) and plausibility (plausible, implausible) as factors. The subject analysis of the combined Z-score data shows a main effect of ambiguity in the direction predicted by the ambiguity avoidance hypothesis (ambiguous: 0.14, unambiguous: -0.17, F1(1,19) = 4.7, p < .05; F2(1,18) = 1.1) and no other effects or interactions. See Figure 5 for a representation of the data. INSERT FIG. 5 HERE Interactions with sex When ANOVAs were recomputed, adding sex of the speaker (female, male) as a third factor, some of the effects interacted with sex. Both lexical and prosodic data had some interaction with the sex. These effects all involved the female speakers showing the relevant effect, and the male speakers showing no effect. For all of the analyses including sex as a factor, only the subject analyses were computed. Sex was unable to be included in the item analysis due to unbalanced cell design. The unbalanced cell design was due to the fact that some items were usable in only a few subjects. Usable items were those that were produced by the subject without changing the verb and (for the prosody analyses) without inserting that. For example, few subjects produced the sentences with reported, estimated, felt, and acknowledged correctly. Subjects generally changed the verb (e.g. they said The jury convicted you of murder instead of The jury felt you were guilty of murder) or inserted that, or both. Thus, there were only a few usable observations for those items, and it sometimes happened that data in at least one cell was completely missing. So, for example, there were no instances of ...reported I... for any of the female subjects. Because of the amount of missing data, sex could not be included as a factor in the item analysis. The marginal ambiguity effect in the that-insertion data reached significance once sex was included as a factor. Speakers inserted that significantly more often in the unambiguous conditions, which is the unexpected result. The expected result was either to replicate Ferreira and Dell (1998), who found that speakers did not insert that more often in either the ambiguous or the unambiguous condition, or to find that speakers inserted that more often in the ambiguous conditions, in order to avoid ambiguity. Once sex was added as a factor, not only did the main ambiguity effect reach significance, but there was also a sex by ambiguity interaction. When the data were split by sex, the female subjects showed the ambiguity effect, inserting that more often in the unambiguous conditions, but the male speakers did not. An ANOVA of the that-insertion data for subjects was computed, with ambiguity (ambiguous, unambiguous), plausibility (plausible, implausible), and sex (female, male) as factors. The analyses revealed a main effect of ambiguity (ambiguous: 30%, unambiguous: 36%, F1(1,28) = 4.2, p = .05). There was also an ambiguity by sex interaction (F1(1,28) = 4.4, p < .05). Females reliably inserted that more often in the unambiguous conditions (ambiguous: 26%, unambiguous: 38%, F1(1,17) = 9.7, p < .01). Males showed no difference (ambiguous: 36%, unambiguous: 34%, F1(1,10) < 1). See Figure 6 for a graphical representation of the results. INSERT FIG. 6 HERE The sex of the speaker also had an effect on the duration data. The main effect of plausibility (with duration longer for the implausible verbs) stayed the same, but there was also a sex by ambiguity interaction. For female speakers only, verbs had longer duration in the ambiguous condition. This effect is in the expected direction, based on the predictions of the ambiguity avoidance model, which predicts that speakers will lengthen the main verb to cue an upcoming SC when the syntax is temporarily ambiguous. An ANOVA of the duration data with ambiguity (ambiguous, unambiguous), plausibility (plausible, implausible), and sex (female, male) as factors was computed. Again, this analysis was only computed for the subject means. The analysis showed a main effect of plausibility, in the unexpected direction (plausible: -0.18, implausible: 0.15, F1(1,18) = 11.8, p < .01). There was also a significant ambiguity by sex interaction (F1(1,18) = 8.7, p < .01). When the data were split by sex, female speakers showed a main effect of ambiguity (ambiguous: 402.0 ms, unambiguous: 368.6 ms, F1(1,13) = 9.8, p < .01). Male speakers showed no effect (ambiguous: 338.6 ms, unambiguous: 356.6 ms, F1(1,5) = 2.0). Figure 7 is a graphical representation of the sex by ambiguity interaction. INSERT FIG. 7 HERE When sex was added as a factor to the analysis of the pitch data, there were no differences in the type of effects found. The pitch data was still marginally significant for plausibility, in the expected direction (plausible: 0.10, implausible: -0.11, F1(1,19) = 2.2, p = .15) When the combined prosody data was reanalyzed with sex as a factor, the main effect of ambiguity remained, with subjects producing more disambiguating prosody in the ambiguous conditions. In addition, adding sex as a factor revealed a marginally significant sex by ambiguity interaction. Splitting the data by sex revealed that the female speakers were showing the effect of ambiguity, while the male speakers were not. An ANOVA, with factors for ambiguity (ambiguous, unambiguous), plausibility (plausible, implausible), and sex (female, male) was computed for the combined prosody data. The results showed a main effect of ambiguity, in the expected direction (ambiguous: 0.14, unambiguous: -0.17, F1(1,18) = 5.4, p < .05), and a marginal sex by ambiguity interaction, in the expected direction for the female speakers (female-ambiguous: 0.23, female-unambiguous: -0.24, male-ambiguous: -0.07, male-unambiguous: 0.01, F1(1,18) = 3.7, p = .07). See Figure 8 for graphical illustration. INSERT FIG. 8 HERE
Discussion Lexical insertion When the data for that-insertion were analyzed as they were in Ferreira and Dell's (1998) study (without sex included as a factor), Ferreira and Dell's results were replicated. That is, there were no significant effects of either ambiguity or plausibility nor was there an ambiguity by plausibility interaction. There was, however, a marginal effect of ambiguity. This effect was troubling because it went in the unpredicted direction (unpredicted by either Ferreira and Dell's results or by the ambiguity avoidance or garden-path models). This effect reached significance when sex was added to the analysis as a factor. Though there was an effect of ambiguity on that-insertion, it was in the opposite direction than the ambiguity avoidance model would predict. This effect was absent in Ferreira and Dell's study. Because the effect is in the unexpected direction, it is unclear what the effect means. One possible explanation is that I is more difficult to retrieve from memory or harder to articulate than you. If this is the case, then Ferreira and Dell's retrieval-based model of that insertion would predict more that insertion in the unambiguous I conditions because subjects would need to fill the pause created as they searched for I or programmed the motion needed to articulate it. However, it is unlikely that I is less easily retrieved or harder to articulate than you for several reasons. First, the fact that I refers to the speaker suggests that it would be easier to retrieve than you, since speakers are likely to be at least somewhat 'ego-centric,' and should therefore find it easy to retrieve the pronoun which refers to themselves. Second, the experimental situation made it awkward to produce the you sentences. In those cases, the speaker was telling the listener something about the listener (e.g. The secretary remembered you were out for the day), which is more awkward than the speaker telling the listener something about himself. Finally, I may be more difficult to articulate than you, because I begins with a glottal stop, and you begins with a glide. However, when I is produced immediately after the last sound of the previous word, as it is in normal pronunciation of sentences like The secretary remembered I was out for the day, the glottal stop is deleted and I is no more difficult to articulate than you. One way to find out if I is more difficult to process than you is to conduct a reading study on these sentences. If there is some sort of processing difficulty with I that is not present with you, it should show up in a reading study as slower reading times for I than for you. One thing regarding the that data seems clear-speakers are not using that insertion to help their listeners avoid the garden-path, and any differences in that insertion between conditions are probably due to some low-level production factors. The possibility of that-insertion depending on low-level production factors was examined in the regression depicted in Figure 2. The rate of that-insertion from the Garnsey et al. (1997) study served as the predictor variable for the rate of that-insertion in this study. Though the methods of the two studies were different (written sentence completion vs. oral production), the prediction was very strong. Such a tight fit in the regression is good evidence that that-insertion is based on a low-level speech process, such as verb-that co-occurrence. It has been shown that humans are capable of using statistical information about language in sophisticated ways (Trueswell, 1996; Trueswell, Tanenhaus, & Kello, 1993; Kelly & Martin, 1994). Therefore, it is unsurprising that speakers are sensitive to the frequency with which a verb co-occurs with that, and insert that accordingly. Prosodic disambiguation The combined prosody data (duration + pitch) support the ambiguity avoidance model of speech production. Speakers used disambiguating cues of longer verb duration1 and greater pitch drop when the verb was followed by the syntactically ambiguous you . This suggests that speakers do use disambiguating prosody (some combination of duration and pitch) when producing a syntactically ambiguous sentence. The lack of a plausibility effect or an ambiguity by plausibility interaction may be due to the weakness of the plausibility manipulation. The verbs used in this experiment were all roughly equally biased in term of how plausibly they took a DO or a SC construction. However, they did not all plausibly take you as a DO. For example, remembered plausibly takes you as a DO, but learned does not. As a simple illustration, consider the sentences The man remembered you well and The man learned you well. Clearly, the first sentence is much more likely to occur. This difference in plausibility with you was the basis of the plausibility manipulation. However, some cases were not as clear as remembered and learned. For example, acknowledged you, announced you, and demanded you are somewhere in the middle in terms of the plausibility of you as a DO. Since the plausibility of verbs is more of a continuum than a clear division, the verbs in the middle of the continuum could be interacting in unexpected ways with the prosodic cues of duration and pitch drop. In addition, because the plausibility manipulation was a "between items" manipulation (a verb could not be both plausible and implausible), the effects that were found could be a product of having different verbs in the two conditions. The two sets of verbs could be causing effects based on some features of the verbs which are uncontrolled for in this experiment. The separate data for duration and pitch do not support either the garden-path model or the ambiguity avoidance model. The duration data indicates that speakers lengthen a verb's duration when it is implausible that the listener will be led down the garden-path. This is the unexpected result. However, the pitch data indicates that speakers drop their pitch more on verbs that plausibly led listeners down the garden-path, the expected result. These two cues seem to lead listeners in opposite directions. A possible explanation is that speakers have an easier time anticipating the embedded clause when the verb implausibly takes a DO, and are therefore able to signal the approaching embedded clause with a lengthened verb. However, this explanation is subject to criticism for two reasons. First, these verbs are not actually SC-biased. They were chosen to be equibiased, and only implausibly lead toward a DO interpretation when the word following the verb is you. So speakers could not be expected to anticipate a SC continuation based on the verb alone. Second, speakers used the other prosodic cue, pitch drop, in the expected way: to cue a SC. Another possibility is that pitch drop is indeed a cue toward a SC interpretation, but that long duration is a cue away from a SC interpretation. However, long duration has been shown (by Beach (1991), for example) to be a cue toward a SC interpretation. See Table 1 for a summary of the data.
Another puzzling fact about the data is that an effect of ambiguity shows up in the combined prosody analysis, but not in either of the duration or pitch analyses, and a plausibility effect shows up in both the duration and pitch analyses, but not in the combined analysis. This can be explained by the way the ambiguity and plausibility effects interact in the combined analysis. Notice that duration shows an effect of plausibility in the unexpected direction, and pitch shows an effect of plausibility in the expected direction. The Z-scores for duration are: plausible = -0.18, implausible = 0.15. The Z-scores for pitch are: plausible = 0.10, implausible = -0.11. When the Z-scores are added together in the combined prosody analysis, then, plausible = -0.18 + 0.10 = -0.08 and implausible = 0.15 + (-0.11) = 0.04, both of which are near zero. There is such a small difference between -0.08 and 0.04 that the plausibility effect is no longer significant. For ambiguity, much the same thing happens, but in the opposite direction. The duration values are: ambiguous = 0.07, unambiguous = -0.09. The pitch values are: ambiguous = 0.07, unambiguous = -0.08. When the values are added together in the combined analysis, the values are: ambiguous = 0.07 + 0.07 = 0.14, unambiguous = -0.09 - 0.08 = -0.17. The difference between 0.14 and -0.17 is large enough to reach significance. Though there are no significant ambiguity effects in either the duration or pitch analysis alone, their non-significant effects are in the same direction, so when the Z-scores are added, the effect reaches significance. Though prosodic effects were found in Experiment 1, the methods used
to quantify them were rough. It is likely that these measurements missed
some information present in the sound stream that listeners can use. For
example, pitch drop was measured as a simple function: the pitch value at
the end of the verb subtracted from the pitch value at the beginning of
the verb. While this method has the advantage of being objective and
unambiguous, it does not take into account variation in the shape of the
pitch contours (e.g. monotonic descending, monotonic ascending, ascending
then descending, descending then ascending). Also, the combined measure
of prosody was the simple addition of the duration data and the pitch
data. Listeners, however, probably interpret the combined prosody in a
more sophisticated way. Though science has not determined the formula
listeners use to interpret prosody, prosodic effects can be tested by
experimentally asking listeners to make judgements based on them.
Experiment 2 serves as just such a test, asking listeners to use prosody
to predict syntactic structure.
Experiment 2: Comprehension In the second experiment, the effectiveness of prosody is explored from the listener's point of view. Knowing that speakers produce disambiguating prosody is only one part of the story. It is also important to know if listeners can use that prosody. If listeners cannot make use of the prosodic effects found in Experiment 1, this suggests that those effects are not meaningful in terms of syntactic structure. On the other hand, if listeners can predict syntactic structure from prosodic cues, then either the cues found are real or there are other cues present that were not found. This experiment used a forced-choice sentence completion
paradigm similar to Beach's to investigate whether prosody produced by
speakers in Experiment 1 could be used by listeners in Experiment 2 to
predict the syntactic structure of an utterance. Listeners were forced to
choose between DO and SC completions of fragments from utterances produced
in Experiment 1. Though none of the stimuli were actually from DO
sentences, the more obvious prosody from the ambiguous you
sentences is predicted to lead listeners to choose the SC completion,
whereas the less obvious prosody from the unambiguous I sentences
is predicted to lead listeners to choose the DO completion.
Method Subjects Subjects were 10 University of Pennsylvania undergraduates taking Psychology 1. There were 4 women and 6 men. Subjects participated for course credit. All subjects were native English speakers with normal hearing and vision (corrected or uncorrected). Materials and Design Materials were spliced fragments of the utterances recorded during Experiment 1. The utterances from eight subjects (one woman and one man from each of the four lists) from Experiment 1 were chosen based on the criterion of having at least 3 "no-that" target utterances per condition. Each qualifying target item was spliced off after the verb (e.g. The secretary remembered...). Also, 3 random filler items were selected for each speaker. For each fragment, two complete sentences were created, one continuing with a direct object and the other with a sentence complement. The sentences were designed to be as similar as possible to each other. Because you was implausible as a direct object for half of the verbs (the implausible verbs), which made writing natural you DO completions nearly impossible, the noun phrase following the verb in the complete sentences was a definite noun phrase. For example, two of the completed sentences were: The secretary remembered the appointment in the morning, (DO completion) and The secretary remembered the appointment was that afternoon, (SC completion). To make the task as easy as possible for the listeners, all the fragments from a particular speaker were heard in a block, in the order that the speaker had said them. Target items were heard one after the other, with no fillers interspersed. Also, the answer sheets were constructed so that either the SC completion or the DO completion came first for every trial. Finally, preceding each speaker were three filler items from that speaker, so listeners would have a chance to acclimate to the speaker's voice. Speakers alternated between male and female and between lists 1-4. All completions can be found in Appendix C. Procedure Listeners heard three practice items to introduce the task.
These three practice items were spliced from three filler sentences from
the first speaker. They did not contain DO/SC ambiguities. After
completing the three practice sentences, subjects heard the three filler
items for the first speaker, followed by the target items for the first
speaker. They then heard the three filler items for the second speaker,
followed by the target items for the second speaker. They proceeded
similarly through the rest of the speakers. Each target fragment was
played twice, with a two-second pause in between. There was a five-second
break, then the next item was played. After hearing the fragment twice,
subjects circled the complete sentence with the completion they thought
the speaker said. Subjects were told that they must choose one and only
one completion for each item. The subjects were given no instruction
about the syntactic structure of the sentences, nor were they given any
suggestions about how they could decide between the two alternatives.
Results There were no significant effects or interactions in the results of Experiment 2. However, the trends in the means for ambiguity and plausibility do go in the expected direction. Listeners chose more SC completions when the sentence fragment came from a syntactically ambiguous sentence (where the ambiguity avoidance hypothesis predicts stronger prosody). Listeners also chose more SC completions for sentence fragments that were originally from plausible sentences (where the garden-path model predicts stronger prosody). Sex of the speaker and sex of the listener interacted in interesting ways with the effects of ambiguity and plausibility. In the subject analysis, a sex of the speaker by plausibility interaction suggests that the listeners were better at using plausibility to decide between DO or SC completions when the fragments were from male speakers. In the item analysis, a sex of the listener by ambiguity interaction suggests female listeners were better than male listeners at using ambiguity to decide between DO or SC completions. An ANOVA with factors for sex of the listener (female, male), sex of the speaker (female, male), ambiguity of the sentence (ambiguous, unambiguous), and plausibility of the verb (plausible, implausible) produced no significant effects or interactions. However, in the subject analysis, there was one marginally significant interaction for sex of the speaker by plausibility of the verb (female-plausible: 55.8%, female-implausible: 58.1%, male-plausible: 65.2%, male-implausible: 55.7%, F1(1,8) = 3.8, p = .09). Sex of the speaker could not be included in the item analysis because of unbalanced cell design. There was one marginally significant interaction in the item analysis, sex of the listener by ambiguity (female-ambiguous: 65.1%, female-unambiguous: 55.6%, male-ambiguous: 57.8%, male-unambiguous: 58.6%, F2(1,17) = 3.0, p = .10). Though none of the other effects or interactions were even marginally significant, the cell means were in the direction predicted. Subjects chose the SC completion more often in the ambiguous conditions (ambiguous: 60.4%, unambiguous: 57.0%, F1(1,8) = 1.3; F2(1,17) = 1.6) They also chose the SC completion more often in the plausible conditions (plausible: 60.5%, implausible: 56.9%, F1(1,8) < 1; F2(1,17) < 1). There were interesting small effects involving the sex of the listeners and the speakers. Female listeners did a better job of picking the correct completion than male listeners did, for both plausibility and ambiguity. Both female and male listeners did a better job choosing the correct completion for male speakers than for female speakers (see Table 2 for a summary of the interactions with sex).
Discussion Though there are interesting effects of plausibility and ambiguity and interactions with sex of both the speaker and the listener, none of the effects reach significance. There are several reasons for this. First, this is a small, preliminary sample of only ten subjects. The small sample size combined with the small size of the effects makes insignificant results unsurprising. Second, though the subjects' task was to decide between SC and DO completions for the fragments, none of the fragments actually came from DO sentences. Every fragment heard by the subjects was originally from a SC sentence. The inclusion of fragments spliced from DO sentences is expected to strengthen the effects, and a revised version of this experiment would therefore benefit by having DO fragments included in the stimuli. Though the effects are small, they are in the direction indicated by the ambiguity avoidance hypothesis. Speakers used disambiguating prosody (duration and pitch drop) to differentiate the ambiguous you sentences from the unambiguous I sentences. Listeners showed a slight tendency to assign the prosody from ambiguous fragments to SC completions and the prosody from the unambiguous fragments to DO completions. It is a little puzzling that listeners chose SC completions more for the plausible conditions than the implausible conditions, since analysis of the speakers' utterances showed conflicting cues in those conditions. That is, listeners are doing the expected thing, by choosing more SC completions when the sentence plausibly leads them toward the garden-path, but they are doing it with prosody that did not clearly lead toward the SC interpretation. Remember that the plausibility effect in Experiment 1 was toward SC structure for the pitch cues, but away from SC structure for the duration cues. That listeners were able to make the correct choices in spite of the conflicting cues may mean that listeners are only using pitch drop in the plausible conditions as a cue toward a SC. Listeners' ability to overcome the misleading duration cues and still choose the correct completion could mean that the duration cue is less important to listeners than the pitch cue, or that the speakers in Experiment 2 had stronger pitch cues than duration cues. Interactions with sex Although small, the sex interactions in
the data are interesting. Keysar & Henly (1998) found that female
listeners were better able than male listeners to predict sentence
structure based on prosody (though male listeners were more
confident o f their ability to use prosody). In addition, Keysar
found that female speakers were better at using disambiguating prosody
than male speakers (though male speakers were more confident that they had
gotten the correct meaning across). In Experiment 2, there is a trend
similar to Keysar's effect for the sex of the listener-the female
listeners in my study were better at using the prosody to correctly
predict the completion. The effect for the sex of the speaker, however,
is in the opposite direction from Keysar's. In this experiment, listeners
were better at predicting sentence structure from male speakers.
Since none of the effects were reliable, however, more experimentation
must be done before anything conclusive can be said about listeners'
abilities to use natural prosody from untrained speakers to predict
sentence structure.
General Discussion Taken together, the results of Experiments 1 and 2 provide some preliminary evidence that speakers produce and listeners use disambiguating prosody. The production results, while not as clear or robust as hoped, indicate that speakers are using some disambiguating prosodic cues. The pitch drop and the combined prosody both lead to a SC interpretation in the conditions that are the most misleading, that is, the cases in which a listener would plausibly be led down the garden-path (pitch drop), and the cases that are syntactically ambiguous (combined). The duration results that seem to lead away from SC interpretations are questionable. In all cases, the effects are small, and must be considered carefully and with the knowledge that the experiments did not include any DO sentences with which to compare the SC prosody. Because the production study was modeled on Ferreira and Dell's (1998) study, in an effort to replicate and extend their findings, speakers did not produce DO sentences. In a future version of this experiment, a DO/SC manipulation should be included. Likewise, listeners in a comprehension study should have the opportunity to hear fragments from DO sentences, to better make decisions about whether a fragment is from a DO sentence. Another problem with the original production study is all the missing data. Many trials were not usable in the prosody analysis because the speaker changed the verb or inserted a that. A future rendition of this experiment should attempt to m inimize the memory demands on the speaker, either by increasing the time allotted to memorize the sentences, decreasing the delay after memorization and before production, or simply having the subjects read the sentences. Not only would an increase in the usable trials benefit the power of the experiment, but it would allow the inclusion of sex as a factor in both the subject and the item analyses. Though this experiment attempted to answer the question of whether speakers help listeners avoid the garden-path, even a completely successful experiment with ideal results would have been unable to provide an answer. Even if speakers had provided strong, unambiguous prosodic cues in support of the garden-path hypothesis, we could not know whether those cues were produced to help the listener avoid the garden-path or as an artifact of the speaker's own confusion. That is, we could not tell whether the speaker actually had a model of his listener's comprehension system in mind, and was anticipating the listener's garden-path, or whether the speaker was simply acting in an ego-centric way, producing disambiguating prosody when he himself had been garden-pathed. Future studies might explore this question, leading to insight into the speaker's model of his listener's understanding. Future Directions A study that is currently in the planning stages could shed some light on the question of why speakers are using disambiguating prosody (for their listeners or for themselves). The study uses a different ambiguity than was researched in this paper-the early/late closure ambiguity. An example is While the man was drying the dog jumped into the pool. Here, the dog could locally be either the direct object of drying or the subject of a new clause. If the dog is taken to be the subject of a new clause, the sentence is said to have early closure because the first clause closes off earlier in the sentence. If the dog is taken as the direct object of drying, the sentence has late closure, because the first clause closes off later in the sentence. As in the DO/SC ambiguity, if the reader or listener chooses the wrong interpretation (here, late closure), he will experience a garden-path at jumped, when he realizes that he has chosen the wrong interpretation. Target sentences containing this type of local ambiguity will be paired with pictures depicting the action. Each target sentence will be paired with two pictures, one of a helpful scene, and one of an unhelpful scene. For example, While the man was drying the dog jumped into the pool will be paired with two pictures: one of a man toweling himself off and a dry dog standing near a pool (the helpful picture) and one of a dry man toweling off a wet dog (the unhelpful picture). If a listener sees the helpful picture while he is hearing the accompanying sentence, he is likely to commit to the early closure interpretation, and the garden-path should be mitigated. If, however, a listener sees the unhelpful picture while hearing the sentence, he will likely commit to the late closure interpretation, and experience a severe garden-path. The helpful picture/unhelpful picture manipulation will be crossed with a common ground/privileged ground manipulation. Common ground/privileged ground refers to what is shared knowledge in a conversation and what is known only to one of the participants. The four conditions would be: 1. a. Helpful picture in common ground. If speakers are indeed using prosody to avoid ambiguity and help their listeners avoid the garden-path, the results should show an interaction between the two factors. Speakers should use the most dramatic prosodic cues in condition 2a, where their listeners are most likely to be garden-pathed (since the picture encourages the listener to commit to the wrong interpretation). The least dramatic prosody is expected condition 1a, where the picture encourages the listener to commit to the correct interpretation. Mid-level prosody is expected in the two privileged ground conditions, since the listener is being neither helped nor hindered by the picture. If speakers are not using prosody to avoid ambiguity for their listener, but for some other reason (maybe to reinforce the correct interpretation for themselves), the results should show a main effect of how helpful the picture was, no effect of common/privileged ground, and no interactions. Speakers should use the most dramatic prosody in conditions 2a and 2b, where the picture encourages the incorrect interpretation, and should use significantly less disambiguating prosody in conditions 1a and 1b, where the picture encourages the correct interpretation. No effect of common/privileged ground or interactions are expected. Though not described above (in order to keep the explanation
simple), this experiment could also include a manipulation of the
ambiguity of the stimulus, and a manipulation of the structure (whether
the sentence is early or late closure). If a comma is inserted after the
main verb (While the man was drying, the dog jumped into the pool),
the sentence is unambiguous to a reader (but still ambiguous to a
listener). This manipulation will allow for replication of Experiment 1
with a different ambiguity. Including both early and late closure
sentences in the stimuli will enable this experiment to avoid one of the
pitfalls of Experiment 1-the lack of a comparison with the other possible
structure.
References Beach, C. M. (1991). The interpretation of Prosodic Patterns at Points of Syntactic Structure Ambiguity: Evidence for Cue Trading Relations. Journal of Memory and Language 30, 644-663. Bock, J. K. (1986). Meaning, sound, and syntax: Lexical priming in sentence production. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 575-586. Bock, J. K. (1987). An effect of the accessibility of word forms on sentence structures. Journal of Memory and Language, 26, 119-137. Bock, J. K. & Irwin, D. E. (1980). Syntactic effects of information availability in sentence production. Journal of Verbal Learning and Verbal Behavior, 19, 467-484. Ferreira, F. & Henderson, J. M. (1991). Recovery from misanalyses of garden-path sentences. Journal of Memory and Language, 30, 725-745. Ferreira, V. & Dell, G. (1998). Syntactic and Lexical Choices in Language Production What we can learn from "that". Paper presented at the Eleventh annual CUNY Conference on Human Sentence Processing in New Brunswick, NJ. 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 to temporarily ambiguous sentences. Journal of Memory and Language, 37, 58-93. Grice, H. P. (1975). Logic and Conversation. In P. Cole and J. L. Morgan (Eds.) Syntax and Semantics. Vol. 3 Speech Acts. New York: Academic Press. Kelly, M. K. & Martin, S. (1994). Domain-specific abilities applied to domain-specific tasks: Sensitivity to probabilities in perception, cognition, and language. Lingua, 92, 105-140. Keysar, B. & Henly, A. S. (1998). Speakers overestimate the transparency of their intention. Paper presented at the 39th Annual Meeting of the Psychonomic Society, Dallas TX. Lehiste, I., Olive, J. P., & Streeter, L. (1976). Role of duration in disambiguating syntactically ambiguous sentences. Journal of the Acoustical Society of America, 60, 1199-1202. Nespor, M., & Vogel, I. (1983). Prosodic structure above the word. In A. Cutler & D. R. Ladd (Eds.), Prosody: Models and measurements. New York: Springer-Verlag. Price, P., Ostendorf, M., Shattuck-Hufnagel, S., & Fong, C. (1991). The use of prosody in syntactic disambiguation. Journal of the Acoustical Society of America, 90, 2956-2970. Price, P., Ostendorf, M., Shattuck-Hufnagel, S., & Veilleux, N. (1988). A methodology for analyzing prosody. Journal of the Acoustical Society of America, 84, (Suppl. 1), S99. Scott, D. R. (1982). Duration as a cue to the perception of a phrase boundary. Journal of the Acoustical Society of America, 71, 996-1007. Tanenhaus, M. K., Spivey-Knowlton, M., Eberhard, K., & Sedivy, J. (1996). Using eye movements to study spoken language comprehension: Evidence for visually mediated incremental interpretation. In T. Inoui & J. McClelland (Eds.) Attention and Performance 16: Information integration in perception and communication. Attention and Performance. Cambridge, MA: MIT Press. Trueswell, J. C. (1996). The role of lexical frequency in syntactic ambiguity resolution. Journal of Memory and Language, 35, 566-585. Trueswell, J. C., Sekerina, I., Hill, N., & Logrip, M. L. (in press). The kindergarten-path effect: Studying on-line sentence processing in young children. Cognition. Trueswell, J. C., Tanenhaus, M. K., & Kello, C. (1993). Verb-specific constraints in sentence processing: Separating effects of lexical preference from garden-paths. Journal of Experimental Psychology: Learning, 19, 528-553. Tyler, L. K., & Marslen-Wilson, W. D. (1981). Children's processing of spoken language.
Journal of Verbal Learning and Verbal Behavior, 20, 400-416.
Appendix A: Stimuli for Experiment 1 Target items: The nurse reported you/I had not taken all my medication. A doctor explained to you/me why it was important to take it all. The architect noted you/I changed some of the blueprints. He said the changes were not structurally sound. The bartender noticed you/I had had too many drinks. He refused to make you/me another one. The terrorist demanded you/I give him a million dollars immediately. If you/I didn't, he would kill the hostages. The statement acknowledged you/I knew nothing of the coverup. You/I had been an unknowing participant in the crime. The magician guessed you/I had a pen in your/my pocket. He then pulled your/my pen out of his bag. The teacher knew you/I skipped class Monday. She reported me to the principal. The investigator learned you/I concealed evidence from the police. He blackmailed you/me in exchange for his silence. The artist remembered you/I posed for him in college. That was before he made it big. The coach promised you/I could play in the next game. It was the big homecoming game, and you/I were/was really looking forward to it. The driver feared you/I were/was about to hit his new car. He swerved away from you/me, but hit a signpost instead. The accountant estimated you/I would fall in the highest tax bracket. She told you/me that making some charitable contributions would help. The priest sensed you/I needed to talk about your/my marriage. You and your spouse (My spouse and I) had been having problems lately. The surgeon predicted you/I would walk again. He had confidence in your/my desire to succeed. The sportscaster mentioned you/I made the winning goal. The stadium erupted in a cheer. The principal announced you/I won the essay contest. All your/my friends congratulated you/me. The receptionist denied you/I had an appointment with the CEO. She said it wasn't written down on his schedule. The reporter doubted you/I had witnessed the incident. She checked your/my version of the story against the facts. The secretary remembered you/I were/was out for the day. She cancelled all of your/my appointments. The jury felt you/I were/was guilty of murder. You/I had failed to convince them of your/my innocence. Filler items: Near the end of the marathon, I got sick from dehydration and had to leave the race. But I was proud of myself for having made it 23 miles. Gus bought a cheaper bike because he heard bike thefts were common in the area. But he didn't buy a lock and the bike got stolen. The weather suddenly turned sour so you had to cancel the picnic. Everyone was disappointed, but there was nothing you could do. The dinner party went off without a hitch. The food was delicious, the guests were interesting, and the weather was beautiful. When I started the business ten years ago I had almost no money, but now I'm a millionaire. I've started teaching a course on entrepreneurship. John's brother got thousands of dollars in the will, but John didn't get anything. He tried to fight the will, but lost. My mother thought I should go to medical school. She always wanted a doctor in the family. You had no idea what to get your mother for her birthday. She refused to give you any clues. During the interview with the detective, you were nervous. He asked you to take a lie detector test. Each time you ate at the restaurant, you tried a different dessert. You liked the mango cheesecake the best. You saw the table had been moved and concluded that someone had been in the house while you were gone. You checked to see if all your valuables were still there. On the application, Cheryl checked "Native American" under "Race." Robert accused her of lying to get an advantage over other applicants, but Cheryl denied it. The instructor emphasized that the test would be difficult. Kevin decided to go out drinking anyway. I got caught in traffic on the way to my appointment. Luckily, when I explained what happened, my client understood. This new line of shoes is the store's biggest seller. The store has agreed to carry them for two more years. The storm last night blew down two tree limbs. Luckily, it didn't blow down the whole tree. The cook admitted it was his fault the cake burned. He said he was preoccupied with the salmon. If I use the toaster oven and the microwave at the same time, it usually blows a fuse. One day I'm going to move the microwave to a different part of the kitchen. Mrs. Smith worried about getting sick, so she wore a surgical mask during the flu season. She also ate a clove of garlic every morning with breakfast. The building was insured for millions and all the evidence pointed toward arson. The police arrested the building's owner. Just last week, three students were suspended. They had been caught selling heroin. The psychic warned you that the trip would end badly. You took her advice and stayed home. Lisa regretted not remarrying after her husband died. She became lonely as her family grew up and moved away. The band's destruction of the hotel room was complete and utter. The hotel banned them from the premises. The archeologist discovered that the bones were from a previously unknown species. He became one of the most renowned scientists of his time. The farmer was tired because he spent the afternoon painting the fences. But he got it all done, so took it easy the next day. You decided to go to Florida instead of Mexico because Florida is cheaper. You had to save some money for later. It rained the whole time I was on vacation. All the pictures of me look like I've been swimming. When she called, Amanda told me that she wouldn't be able to come to the party. She had suddenly been called out of town. I waited until the children had left for school, then set up the decorations for the surprise party. It was Billy's 10th birthday, and all his friends were coming. Someone at the office had to be fired. The boss told you not to worry, though-it wasn't going to be you. Susie was afraid of the dark, so her parents always left the hallway light on. As she got older, she still slept with a nightlite. My physics professor told me that I was going to fail unless I did well on the final. I studied hard and aced the exam. You went to the gym and spent two hours running on the track. When you got home, you took a nap. I went to the park to meet Sarah, but she didn't show up. I called her later to find out what happened. I almost forgot that I needed to go to the store. I had to buy some wine for dinner. The little boy picked a mud puddle rather than a water puddle to splash in, much to the distress of his parents. All his clothes were ruined. The supervisor forgot to tie up the helium canister. It fell over and exploded and injured three people. Yesterday you took the day off and went to the beach. It was a great stress reliever. After her son was born the empress did a lot of travelling. Her son grew up in the care of a
nurse.
Appendix B: Subject Instructions for
Experiment 1 Subject instructions: In this study you will be reading sentences from cards, and trying to produce those sentences as you would if you were trying to get the meaning of the sentence across to a friend. Here's what we'll do for each set of sentences. The experimenter will give you a card with two sentences on it. You will have 10 seconds to memorize the sentences. Then the experimenter will take back the card and there will be a 10 second delay. After the delay, the experimenter will ask you if you if you are ready to produce the sentences. When you say "ready," the experimenter will start the computer program that records your voice. You should stay about 4 inches away from the microphone to get a clear recording. Please avoid stopping in t he middle of your speech as well as stopping and starting over. Say the sentences as naturally as you can, and try to make the meaning clear. The other subject will be listening to your sentences and rating how naturally you produce the sentence. Try t o pretend as if you were in the middle of a conversation with a friend. After you finish, the experimenter will give you the next card. There are 60 cards, and the whole experiment should take less than an hour. If you have any questions, please ask the experimenter. Confederate instructions: You will hear a series of sentences from the other subject. He or she will be saying the sentences into a microphone attached to a computer. After you hear each of the sentences, please rate the clarity of the sentences. You should consider how eas y the sentences were to understand as well as how naturally they were produced. You will rate each set of sentences on a scale of 1 - 5, 5 being the best. A rating of 5 means the speaker spoke in a natural way and was easy to understand. A rating of 1 means the speaker spoke in an unnatural way and was difficult to understand. You are encouraged to sue the entire scale. The most important thing to remember is that the speaker should sound natural, as if speaking to a friend in normal conversation. There are 60 sets of sentences, and the whole experiment should take less than an hour. If
you have any questions, please ask the experimenter.
Appendix C: Sentence Completions for
Experiment 2 The nurse reported the patient's status to the doctor. The architect noted the changes and studied them. The artist remembered the painting during a dream. The terrorist demanded the money in unmarked bills. The statement acknowledged the error of the official. The magician guessed the correct card in 10 seconds. The teacher knew the student when she was little. The investigator learned the description of the suspect. The bartender noticed the customer alone in the corner. The coach promised the team a day off after the game. The driver feared the snow because of his accident. The accountant estimated the taxes correctly, to the nickel. The receptionist denied the allegations at the meeting. The surgeon predicted the recovery time perfectly. The sportscaster mentioned the race three times. The principal announced the winner of the prize. The priest sensed the woman's distress acutely. The reporter doubted the story from the witness. The secretary remembered the appointment in the morning. The jury felt the witness's fear as she testified. | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||