Language comprehension is sensitive to changes in the reliability of lexical cues

Language comprehension is sensitive to changes in the reliability of lexical cues Alex B. Fine (afine@bcs.rochester.edu) Department of Brain and Cognitive Sciences and Department of Linguistics, University of Rochester Rochester, NY 14627 USA T. Florian Jaeger (fjaeger@bcs.rochester.edu) Department of Brain and Cognitive Sciences and Department of Computer Science, University of Rochester Rochester, NY 14627 USA Abstract This paper tests the hypothesis that language comprehen- ders update their beliefs about the statistics of their language throughout the lifespan, and that this belief update allows com- prehenders to combine probabilistic linguistic cues according to their reliability. We conduct a multi-day sentence compre- hension study in which the reliability of a probabilistic cue to syntactic structure is manipulated between subjects. We find that as the reliability of one cue to syntactic structure de- creases, comprehenders come to rely more on a second cue to syntactic structure. The results are consonant with ratio- nal models of cue integration in speech perception and in non- linguistic domains, thus suggesting a unifying computational principle governing the way humans use information across both perceptual and higher-level cognitive tasks. Keywords: psycholinguistics; adaptation; sentence process- ing; cue combination The lawyer acknowledged the judge had been unfair to the defendant. Probabilistic cues provide comprehenders with informa- tion that can guide inferences during incremental language processing, contributing to processing efficiency (see Smith & Levy, 2008 for an explicit proposal along these lines). How- ever, the cues relevant to comprehension are moving targets: probabilistic cues such as those mentioned above are context- dependent in that their validity (Bates & MacWhinney, 1987) may vary depending on speaker identity, context, and speaker dialect (see Tagliamonte, 2005 for a discussion of variability in syntax). Bates and MacWhinney (1987) define cue validity as the product of cue availability (how often a cue is present in the environment) and cue reliability (how often a cue leads to the correct inference, when present). How do comprehen- ders cope with this variability and maximize the usefulness of probabilistic cues? The current study addresses this ques- tion and tests a two-pronged hypothesis, framed in the spirit of rational analysis (Anderson, 1990): Introduction In order to process language, humans must make inferences about intended messages in the face of uncertainty arising from noisy perceptual data and ambiguity inherent in the lin- guistic signal. Research in psycholinguistics suggests that hu- mans accomplish this task partially by capitalizing on prob- abilistic cues in the linguistic as well as the non-linguistic context (Jurafsky, 1996; Tanenhaus, Spivey-Knowlton, Eber- hard, & Sedivy, 1995). For instance, in sentences such as (1), the judge is temporarily interpretable as both the direct object of acknowledged and the subject of an embedded sentence complement. By-word reading times at the point at which the sentence is disambiguated (had been) are correlated with the conditional probability of the structural representation as- signed to the incremental parse given a number of probabilis- tic cues. One such cue is the verb—the probability of a sen- tence complement occurring is greater given assert than ac- knowledge, based on corpus statistics and norming data (e.g., Trueswell, Tanenhaus, & Kello, 1993; MacDonald, Pearlmut- ter, & Seidenberg, 1994). A second cue is the post-verbal noun phrase—if a post-verbal noun phrase is unlikely to be a direct object of the verb, it is more likely to be the subject of an embedded clause, thereby increasing the probability of a sentence complement continuation (Garnsey, Pearlmutter, Myers, & Lotocky, 1997). Finally, if present, the comple- mentizer that (e.g., The lawyer acknowledged that the judge had been . . . ) also serves as a strong cue to syntactic struc- ture. Indeed, comprehenders have been shown to rely on all of these cues during the incremental processing of sentences such as (1) (MacDonald et al., 1994). • A: Lifelong implicit learning: Throughout adulthood, hu- mans continuously update and adjust estimates of proba- bilistic cues relevant to language comprehension. We will refer to the results of this process as adaptation (cf. also Chang, Dell, & Bock, 2006 and references therein). • B: Rational linguistic adaptation: Adaptation is rational in the sense that humans update the weight they assign to a particular cue based on changes in the validity of that cue. Preliminary evidence for (A) comes from language com- prehension studies at multiple levels of representation (at the phonetic level: Clayards, Tanenhaus, Aslin, & Jacobs, 2008; Kraljic & Samuel, 2007; at the syntactic level: Fine, Qian, Jaeger, & Jacobs, 2010; Wells, Christiansen, Race, Acheson, & MacDonald, 2009). Preliminary evidence for (B) comes primarily from speech perception (Clayards et al., 2008; Kraljic, Samuel, & Brennan, 2008), though these studies are not necessarily framed in terms of the hypotheses presented above. Of particular relevance is Clayards et al. (2008), who manipulated participants’ experience with voice-onset time (VOT), a probabilistic cue to phonetic category membership. For participants in one group, the distribution over VOT val- ues that emerged over the course of the experiment had a low variance; for participants in the other group, this distribution had a high variance. The rationale of the manipulation is that