A belief-updating model of adaptation and cue combination in syntactic comprehension

Abstract

A belief-updating model of adaptation and cue combination in syntactic comprehension Dave F. Kleinschmidt 1 , Alex B. Fine 1 , and T. Florian Jaeger 1,2 {dkleinschmidt, afine, fjeager} @ bcs.rochester.edu of Brain and Cognitive Sciences and 2 Department of Computer Science, University of Rochester, Rochester, NY, 14607 USA 1 Department Abstract Thus, behavioral evidence seems to suggest that adapta- tion, qualitatively speaking, is a very general feature of per- ception and cognition. A question that arises from all of this previous work is whether adaptation observed across all of these domains can be modeled within a single framework. The goal of this paper is to take a step in this direction. In particular, we model adaptation in syntactic comprehension in terms of Bayesian belief update. Modeling syntactic adap- tation in a Bayesian framework is appealing because the same basic computational approach has been successfully pursued in a variety of perceptual and motor domains (e.g. Koerd- ing et al., 2007) and, more recently, in speech perception (Kleinschmidt & Jaeger, 2011; Sonderegger & Yu, 2010). Moreover, Bayesian belief update is ideally suited to ex- plicitly model the fact that syntactic comprehension involves the combination of multiple cues. This offers the advantage of suggesting a single computational framework for adapta- tion and cue combination, since Bayesian approaches to cue combination have been successful in a number of domains including visually-guided grasping, audio-visual cue combi- nation, and the weighting of cues to phonetic category. In particular, Bayesian approaches to cue combination in per- ception have provided a formal means of capturing the fact that humans are able to weight multiple cues (e.g., multiple cues to object depth, such as shading and texture) according to how reliable those cues are. We return to the relationship between adaptation and cue combination in the discussion. The goal of the current study is to ask whether a rational model of adaptation–implemented in the form of Bayesian belief update–can account for behavioral evidence for adap- tation in the syntactic domain. Here we model behavioral data originally reported in Fine and Jaeger (2011), which concerns how subjects adjust their expectations about different syntac- tic structures conditioned on lexical information. Specifically, we exploit temporary syntactic ambiguities as a window onto syntactic expectations. In sentences such as (1), the syntac- tic assignment of the noun phrase the judge is temporarily ambiguous, since it can be parsed as either the subject of a sentence complement (SC) clause, as in (1a), or as the the direct object (DO) of acknowledged (as in 1b). We develop and evaluate a preliminary belief-updating model which links intermediate-term (i.e., over several days) syntac- tic adaptation to the joint statistics of syntactic structures and lexical cues to those structures. This model shows how sub- jects differentially depend on different cues to syntactic struc- ture following changes in the reliability of those cues, as shown by Fine and Jaeger (2011). By relating syntactic adaptation and cue combination to rational inference under uncertainty, this work links learning and adaptation in sentence processing with adaptation in speech perception and non-linguistic domains. Keywords: sentence processing, adaptation, Bayesian model- ing, cue combination, rational analysis Introduction Humans must maintain a stable representation of the envi- ronment despite the fact that available sensory input changes across time: for example, over the course of a day, we rec- ognize and grasp objects in a variety of lighting conditions; we execute accurate motor commands despite changes in our own motor systems due to fatigue, over-caffeination, etc.; and during linguistic communication, we process rapidly unfold- ing acoustic information that varies from talker to talker. Variability within each of these different modalities changes the correlation between cues—whether visual, hap- tic, or linguistic—and the things in the world we wish to make inferences about based on those cues. How do our brains make use of these cues in spite of variability in the environ- ment? One possibility, suggested by research across a number of domains, is that humans deal with variability in the envi- ronment by adapting to changes in the statistical properties of the environment (for examples from vision, motor plan- ning, and speech perception, see respectively: Blakemore & Campbell, 1969; Koerding, Tenenbaum, & Shadmehr, 2007; Norris, McQueen, & Cutler, 2003). While most work on adaptation has been concentrated in perception, the question of whether adaptation is operative in higher level cognition has recently received more attention, particularly in language processing research. For instance, a number of researchers have shown that, when given sufficient experience with a structure initially judged to be ungrammat- ical, listenders come to subsequently comprehend (Luka & Barsalou, 2005), generalize, and even produce (Kaschak & Glenberg, 2004) that structure. Similarly, recent work has shown that we fine-tune our expectations about which syn- tactic structures are likely to occur in a given context based on recent experience (Thothathiri & Snedeker, 2008; Farmer, Fine, & Jaeger, 2011). The lawyer acknowledged the judge . . . disambiguation z }| { a. . . . had been unfair to the defendant. b. . . . in the black robe. The sentence is disambiguated towards the latter reading at