Bayesian Pronoun Interpretation In Mandarin Chinese

2y ago
29 Views
4 Downloads
242.34 KB
6 Pages
Last View : 16d ago
Last Download : 2m ago
Upload by : Annika Witter
Transcription

Bayesian Pronoun Interpretation in Mandarin ChineseMeilin Zhan (mezhan@ucsd.edu)1 , Roger P. Levy (rplevy@mit.edu)1,2 , Andrew Kehler (akehler@ucsd.edu)11 Departmentof LinguisticsUniversity of California, San DiegoSan Diego, CA 92093 USA2 DepartmentAbstractKehler and Rohde (2013) proposed a Bayesian theory of pronoun interpretation where the influence of world knowledgeemerges as effects on the prior and the influence of informationstructure as effects on the likelihood: P(referent pronoun) P(pronoun referent)P(referent). Here we present two experiments on Mandarin Chinese that allow us to test the generalityof the theory for a language with different syntactic-semanticassociations than English. Manipulations involving two different classes of implicit-causality verbs and passive vs. activevoice confirmed key predictions of the Bayesian theory: effectsof these manipulations on the prior and likelihood in production were consistently reflected in pronoun interpretation preferences. Quantitative analysis shows that the Bayesian modelis the best fit for Mandarin compared to two competing analyses. These results lend both qualitative and quantitative support to a cross linguistically general Bayesian theory of pronoun interpretation.Keywords: Bayesian modeling; pronoun interpretation; Mandarin ChineseIntroductionSuccessful language understanding requires comprehendersto resolve uncertainty in language. One source of potentialuncertainty emerges from pronouns (e.g. he, she, they, it)since pronouns carry little information and usually do notfully specify the intended referent semantically (e.g. Janeinvited Anne to her house and she had a great time). Nevertheless, humans generally interpret pronouns rapidly andaccurately. Previous work has proposed a variety of factorsthat influence how comprehenders resolve pronouns. Oneset of approaches has focused on grammatical factors such assubjecthood preferences (Crawley et al., 1990), first-mentionpreferences (Gernsbacher & Hargreaves, 1988), and parallelgrammatical role preferences (Smyth, 1994). Others havefocused on information structural features such as topicality. Centering Theory (Grosz et al., 1995), for instance,uses information-structural relationships within and betweenutterances and the grammatical roles of potential referentsto guide pronoun interpretation. Another approach has argued for the role of world knowledge in referent assignment.The coherence-driven approach (Hobbs, 1979), for example,models pronoun interpretation as a side-effect of the inferenceprocesses that underpin general discourse processing.Recent studies on pronoun interpretation (Kehler et al.,2008; Kehler & Rohde, 2013) integrate aspects of the information structural approach and coherence-driven approachby way of a probabilistic Bayesian framework. Kehler andRohde (2013) proposed that comprehenders reverse-engineerspeakers’ intended referents in terms of Bayes’ rule, as shownin (M1). The posterior P(referent pronoun) represents the interpretation bias: upon hearing a pronoun, the probability ofof Brain & Cognitive SciencesMassachusetts Institute of TechnologyCambridge, MA 02139 USAthe pronoun referring to a particular referent. On the otherhand, the likelihood P(pronoun referent) represents the production bias: the probability of the speaker using a pronounto refer to an intended referent. The prior P(referent) denotesthe next-mention bias: the probability of a specific referentgets mentioned next in the context, independent of the formof referring expression used.P(referent pronoun) P(pronoun referent)P(referent) P(pronoun referent)P(referent)referent referents(M1)Equation M1 says that the relationship between pronouninterpretation and pronoun production follows Bayesian principles, without further specifying which factors affect eachterm in the numerator. Here we refer to this as the WEAKform of the Bayesian hypothesis. Kehler & Rohde (2013)further suggest that the factors conditioning the prior and thelikelihood are different: the influence from world knowledgeemerges as effects on prior next-mention biases, and the influence from information structure as effects on likelihood. Werefer to this specification as the STRONG form of the Bayesianhypothesis. This Bayesian model successfully accounts for arange of English data. However, little work has investigatedwhether the model generalizes across different languages.The Present StudyThis paper has two objectives. The first is to test the generality of the Bayesian pronoun interpretation theory crosslinguistically, specifically in the case of Mandarin Chinese. Experiment 1 serves as a base-line test case, where we use a passage completion paradigm that allows us to tease apart the influences of world knowledge from those of grammatical andinformation structural factors. We provide both qualitativeand quantitative model evaluations for the experimental datato test the Bayesian account.The second objective is to further test the theory by wayof a syntactic manipulation, specifically the voice used in acontext sentence (passive or active). Here we follow up ontwo results found by Rohde & Kehler (2014) for English; tofacilitate our descriptions, for the passive voice we use the“first noun phrase” NP1 to refer to the clause’s logical object(the syntactic subject), and the “second noun phrase” NP2 torefer to the logical subject (e.g., in Jane was amazed by Sue,the logical subject and object are Sue and Jane respectively).First, on the assumption that being the syntactic subject of apassive clause is a stronger indicator of topichood than beingthe syntactic subject of an active clause in English (becausePre-final draft of 11 April 2016, accepted to CogSci Conference 2016 -- comments welcome!

the speaker chose a syntactically marked construction to getNP1 into the syntactic subject position), their theory predictsthat the rate of pronominalization of the syntactic subject inthe passive will be higher than that in the corresponding activecase, a prediction that was confirmed by their study. Second,they also found an unexpected effect whereby the voice manipulation influenced next-mention biases. Specifically, Rohde et al. found that passivization increased the rate of nextmention for NP2 , the logical subject. This result was unexpected on the STRONG form of the Bayesian theory. However,Rohde et al. also found that pronoun interpretation reflectedthis effect of passivization on next-mention preference, aspredicted by the WEAK form of the theory.Here we examine whether similar effects are found forMandarin Chinese. The syntactic-semantic properties of Chinese passives make these constructions a good test case. Passive voice in Mandarin Chinese is generally realized via thebei construction, with linear arrangement NP1 bei NP2 verb(e.g., Li & Thompson, 1981). In this construction, NP1 isthe logical object followed by the passive signaling wordbei, which introduces the logical subject NP2 . The bei construction conveys the notion of affectedness (LaPolla, 1988);it describes an event in which the logical object is affectedphysically or psychologically in some way (Li & Thompson,1981). We expect that this affectedness introduced by the beiconstruction may increase the probability of mentioning thelogical object in the next sentence. If this expectation is borneout in the data, it affords the opportunity for an additional testof the WEAK Bayesian theory, which predicts that any effectof passivization on next-mention preferences should have acorresponding effect on pronoun interpretation preferences.We test these predictions in Experiment 2.Experiment 1Experiment 1 provides a first test case for investigating thegenerality of the Bayesian pronoun interpretation theory inMandarin Chinese. We used a passage completion paradigmwith a factorial design that allowed us to tease apart the influence of world knowledge-based inference that emerges fromverb semantics, and the influence of information structuraland grammatical factors by conditioning on the grammaticalrole of the referent in the analysis.MethodsParticipants We recruited 50 self-reported native Mandarin speakers over Witmart (a China-based online crowdsourcing platform). All of them were paid 4 for their participation in the experiment.Materials and procedures Participants completed twosentence passages, writing a second sentence after atransitive-verb context sentence with two like-gender animate arguments. The Free condition (1) included no material in the second sentence prompt, allowing us to estimateprior next-mention preferences P(referent) by analyzing thefirst-mentioned referent in each condition, and the likelihoodP(pronoun referent) that a pronoun is produced to refer tothat referent by analyzing the referential forms that participants used. The Pronoun condition (2) included an overt pronoun in the second sentence, allowing us to measure empirical pronoun interpretation preferences P(referent pronoun).The verbs in the first sentence were taken from one of twoimplicit causality (IC) classes (Garvey & Caramazza, 1974;Brown & Fish, 1983). The use of IC verbs allowed us to manipulate the prior, with IC-1 and IC-2 favoring the syntacticsubject (NP1 ) and the non-subject (NP2 ) next-mentions respectively. A norming study (N 45) was conducted prior toExperiment 1 to ensure the verbs we selected have a clear biastowards re-mentioning either the subject or the non-subjectin the following explanations. We selected sixteen subjectbiased IC-1 verbs and twenty object-biased IC-2 verbs for themain experiments based on the results of the norming study.(1)[NP1 Meihui] dadong-leIC-1 [NP2 Jieyi]. . . .Meihui impressedJieyi. . . .(2)[NP1 Meihui] dadong-leIC-1 [NP2 Jieyi]. Ta . . .Jieyi. She . . .Meihui impressed(3)[NP1 Meihui] jiegu-leIC-2 [NP2 Jieyi]. . . .Meihui firedJieyi. . . .(4)[NP1 Meihui] jiegu-leIC-2 [NP2 Jieyi]. Ta . . .Meihui firedJieyi. She . . .Because IC verbs exhibit the most polarized effects onnext-mention biases when the follow-on sentence providesan explanation of (i.e., a cause or reason for) the event described by the prompt sentence (Rohde, 2008), we instructedparticipants to limit their continuations to such explanations.Each participant completed 36 target items and 36 filler itemswith pseudo-randomization. Prompt Type varied within participants and within items; Verb Type varied only within participants. Each item was presented via the web interface ona separate page with a text box where the participants wereinstructed to write their continuations.Coding The responses were coded by two trained nativespeakers, the first author and a UC San Diego graduate student who was blind to the hypothesis of the study. Each coderwent through the responses independently to code the firstmentioned referent (or the assignment of the pronoun in thecase of pronoun prompt condition) in each continuation asone of the five categories: NP1 , NP2 , both, unclear, and other.If the two coders did not agree on a reference, or there wasnot enough information available to identify the intended referent, the response was coded as unclear. For continuationsin the Free prompt conditions, choice of referring expressionswere coded as name, overt pronoun, null pronoun, and other.Results and discussionThe analysis only included continuations for which the firstmentioned entity was NP1 or NP2 ; hence continuations wereexcluded if the referent was coded as unclear (3.8%), both(0.8%), or other (2.7%). Continuations were also excludedPre-final draft of 11 April 2016, accepted to CogSci Conference 2016 -- comments welcome!

0.750.500.250.00IC 1(subj biased)IC 2(obj biased)Verb TypeFigure 1: Proportion of continuationsabout the syntactic subject NP1 , byVerb Type and Prompt Type1.001.0ReferentNP1NP20.75 Bayes. MSE .03 Expect. MSE .07 Mirr. MSE .05 0.50 0.5 NP1 obervedPrompt TypefreepronounRate of pronominal next mentionProportion of continuations about NP11.00 0.25 0.00IC 1(subj biased)0.0IC 2(obj biased)0.0Verb TypeNext-mention biases Figure 1 shows the proportion ofcontinuations about NP1 in both Free prompt and Pronounprompt conditions. We first evaluate the next-mention biases (dark blue bars) in the Free prompt data, which servesas the prior in the Bayesian model. Analyses showed amain effect of Verb Type (p 0.001): subject-biased IC-1items prompted significantly more continuations about NP1(77.7%) than object-biased IC-2 items (11.7%). This indicates that verb semantics had a strong effect on next-mentionbiases, an expected result on both the STRONG and WEAKforms of the Bayesian hypothesis.Production biases Figure 2 shows the rate of pronominalization conditioning on Re-mentioned NP collapsing null andovert pronouns, which serves as the likelihood in the Bayesianmodel. Analyses showed a large main effect of Re-mentioned0.51.0Model PredictionFigure 2: Rate of pronominalization byVerb Type and Re-mentioned NPif the choice of referring expression was other than a name,an overt pronoun, or a null pronoun (1.0%). These restrictions left 1651 out of 1800 continuations in the dataset. Allstatistical analyses in this paper report results from mixedeffect logistic regression models with the maximal randomeffect structure justified by the design (Barr et al., 2013), conducted using the lme4 R package (Bates et al., 2015; R CoreTeam, 2015); we report significance levels based on the Waldz statistic. In cases where we encountered convergence failurewith lme4, we report analyses carried out using the MCMCglmmR package (Hadfield, 2010), which we indicate by pMCMC inreporting statistical significance. Error bars in all figures arestandard errors over by-participant means.The WEAK form of the Bayesian hypothesis predicts thatpronoun interpretation biases and production biases followBayesian principles, without fully specifying what factors affect each component. The STRONG form of the hypothesisadditionally predicts that semantic factors (Verb Type) affectonly next-mention biases P(referent) and not pronoun production biases P(pronoun referent); the primary factor affecting pronoun production biases is instead predicted to be thegrammatical role of the referent (Re-mentioned NP). We testthese predictions in the following analyses. Figure 3: Quantitative model comparisonsNP (β 1.37, p 0.001), with NP1 re-mentions much morelikely to be pronominalized than NP2 re-mentions. There wasno significant interaction between Re-mentioned NP and VerbType (β 0.053, p 0.716). Both results align with thepredictions of both the STRONG and WEAK hypotheses, indicating a clear disassociation between next-mention biasesand production biases. There was also a smaller, unanticipated main effect of Verb Type (β 0.43, p 0.05), withpronominalization rates higher in the IC-2 context than in theIC-1 context. The reasons for this effect remain unclear.Interpretation biases Now we examine the interpretationbiases (the posterior) and compare them to next-mention biases (the prior). As shown in Figure 1, there was a main effect of Verb Type (β 2.021, p 0.001), with IC-1 verbseliciting more NP1 mentions than IC-2 verbs. There was amain effect of Prompt Type (β 0.896, p 0.001), withthe proportion of continuations about NP1 in the Pronounprompt condition higher than that in the Free prompt condition. Analyses also showed a interaction between PromptType and Verb Type (β 0.408, p 0.001), whereby theeffect of Pronoun prompts was larger for IC-2 verbs than IC1 verbs. Given next-mention and production biases, the maineffects of Verb Type and Prompt Type are predicted by theBayesian theory (the interaction’s presence was neither predicted nor precluded).Quantitative model comparisons Following Rohde &Kehler (2014), we further evaluate the Bayesian account bycomparing its predictions with those of two competing models. The first competing model is the Expectancy model, according to which the interpretation bias towards a referentis equal to the probability the referent gets mentioned next(Arnold, 2001). For this model, the predicted interpretationbias is estimated to be the prior next-mention bias. We express this model in Equation (M2) below, using the assignment operator to emphasize that this model does not follownormative probability theory.P(referent pronoun) P(referent)(M2)The second competing model is called the Mirror model,Pre-final draft of 11 April 2016, accepted to CogSci Conference 2016 -- comments welcome!

according to which the interpretation bias towards an entityis proportional to the likelihood that a reference to that entitywould have been pronominalized by the speaker. This modelcaptures the intuition that pronoun interpretation and production are effectively mirror images of each other, as expressedby the assignment in (M3).P(referent pronoun) P(pronoun referent) P(pronoun referent)(M3)referent referentsTo determine the quantitative predictions of each modelM1–M3, we used the Free prompt data to estimate[condition item]-specific prior and likelihood probabilitiesP(referent) and P(pronoun referent), using add-one smoothing to avoid zero probabilities, and then comparing thesepredictions to [condition item]-specific human interpretationpreferences from our Pronoun-prompt data. Figure 3 plotsobserved NP1 interpretation rates against item-specific predictions of the three models. The x y dotted line would beperfect model fit. The Bayesian model had the least meansquared error (0.03), indicating the Bayesian model is a better fit than either of the competing models in predicting pronoun interpretation. In comparison, the Mirror model dramatically underpredicts cross-item/condition variability in interpretation preference, because it lacks the influence of theworld-knowledge-derived prior; and the Expectancy modelsystematically underpredicts the rate of NP1 pronoun interpretation, because it lacks the likelihood-derived bias towardNP1 obtaining from the speaker’s choice of pronominal form.Experiment 2Experiment 2 further tests the generality of the Bayesian pronoun theory by pursuing a set of findings regarding the effectof voice on pronoun behavior identified for English by Rohde& Kehler (2014). Recall that Rohde and Kehler hypothesizedthat a difference in syntactic form — specifically active vs.passive voice — would have an effect on production biases.This expectation is based on the STRONG Bayesian hypothesis, which posits that the likelihood of a speaker pronominalizing a mention of a referent is based on the referent’sdegree of topicality. Hypothesizing that being the subject ofa passive clause is a stronger indicator of topichood than being the subject of an active clause, Rohde & Kehler (2014)thus predicted that the syntactic subject of a passive clause ismore likely to be pronominalized than that of an active clause.The results of their study confirmed this prediction. However,the results also revealed a separate effect, unexpected underthe STRONG hypothesis, whereby passivization not only affected the next-mention bias as well, but did so in the opposite direction that one would expect: Passivization increasedthe next-mention rate of the logical subject, that is, the entitymentioned from within a by- adjunct.Here we evaluate the predictions of the models by employing a similar voice manipulation in Mandarin Chinese. TheSTRONG hypothesis predicts that the difference in information structure between active and passive voice will affectproduction biases. A finding similar to Rohde and Kehler’swhereby passivization increases the next-mention bias for thelogical subject is not predicted by the STRONG hypothesis,however. The WEAK hypothesis makes no commitment tothe either of these predictions. Instead, the WEAK hypothesis simply predicts that any change in the pronoun productionbiases or t

Mandarin Chinese. The syntactic-semantic properties of Chi-nese passives make these constructions a good test case. Pas-sive voice in Mandarin Chinese is generally realized via the bei construction, with linear arrangement NP 1 bei NP 2 verb (e.g., Li & Thompson, 1981). In this constructio

Related Documents:

Mandarin Chinese . Mandarin. Mandarin Chinese, also known as Standard Chinese or Modern Standard Mandarin, is the sole official language of China and Taiwan, and one of the four official languages of Singapore. Although there are eight major Chinese dialects, Mandarin is native to approxim

PowerPoint: Personal Pronouns Video: Personal Pronouns PowerPoint: Demonstrative Pronoun Video: Demonstrative Pronoun Video: Indefinite Pronoun PowerPoint: Indefinite Pronoun PowerPoint: Relative Pronoun Video: Relative Pronoun PowerPoint: Interrogative

Chinese students in Mandarin language classes. Twenty overseas Chinese students learning Mandarin participated in this stage. In stage 2, 24 overseas Chinese students were taught 3 learning units in Mandarin in SL. Analysis of the results showed that learning Mandarin in an SL environment

diction between Mandarin and written Chinese, together with the use of the colloquial Mandarin terms during lessons can boost the vocabulary bank of students. Second, the resemblance of syntax between Mandarin and written Chinese give students an edge to understand the regular structures and grammar of standard Chinese. Mandarin has great

Pronouns A Pronoun is a word used in place of a noun or of more than one noun. Example: The high school graduate accepted the diploma proudly. She had worked hard for it.-The pronoun she takes the place of the noun graduate. The pronoun it takes the place of the noun diploma. Antecedent- The word whose place a pronoun takes is called its a

Mandarin Chinese Grammar Modern Mandarin Chinese Grammar provides an innovative reference guide to Mandarin Chinese, combining traditional and function-based grammar in a single volume. The Grammar is divided into two parts. Part A covers traditional grammatical categories such as phrase order, nouns, verbs, and specifiers.

Mandarin Chinese 3. Nationalists in 1949, continued this policy, but they changed the name and coined the term . pu tong hua, or “common speech,” for “Mandarin.” This is the word for Mandarin used throughout mainland China. In Hong Kong, however, as in Taiwan and most overseas c

Annual Book of ASTM Standards, Vol. 04.02. 3 For referenced ASTM standards, visit the ASTM website, www.astm.org, or contact ASTM Customer Service at service@astm.org. For Annual Book of ASTM Standards volume information, refer to the standard’s Document Summary page on the ASTM website. 4 The boldface numbers in parentheses refer to a list of references at the end of this standard. 1 .