A model of local coherence effects in human sentence processing as consequences of updates from bottom-up prior to posterior beliefs

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 665–673, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics A model of local coherence effects in human sentence processing as consequences of updates from bottom-up prior to posterior beliefs Klinton Bicknell and Roger Levy Department of Linguistics University of California, San Diego 9500 Gilman Dr, La Jolla, CA 92093-0108 {kbicknell,rlevy}@ling.ucsd.edu Abstract Human sentence processing involves integrat- ing probabilistic knowledge from a variety of sources in order to incrementally determine the hierarchical structure for the serial input stream. While a large number of sentence pro- cessing effects have been explained in terms of comprehenders’ rational use of probabilistic information, effects of local coherences have not. We present here a new model of local coherences, viewing them as resulting from a belief-update process, and show that the rele- vant probabilities in our model are calculable from a probabilistic Earley parser. Finally, we demonstrate empirically that an implemented version of the model makes the correct predic- tions for the materials from the original exper- iment demonstrating local coherence effects. 1 Introduction The task of human sentence processing, recovering a hierarchical structure from a serial input fraught with local ambiguities, is a complex and difficult problem. There is ample evidence that comprehen- ders understand sentences incrementally, construct- ing interpretations of partial structure and expecta- tions for future input (Tanenhaus et al., 1995; Alt- mann and Kamide, 1999). Many of the main behav- ioral findings in the study of human sentence pro- cessing have now been explained computationally. Using probabilistic models trained on large-scale corpora, effects such as global and incremental dis- ambiguation preferences have been shown to be a result of the rational use of syntactic probabilities (Jurafsky, 1996; Hale, 2001; Narayanan and Juraf- sky, 2001; Levy, 2008b; Levy et al., 2009). Simi- larly, a number of other effects in both comprehen- sion and production have been modeled as resulting from rational strategies of languages users that take into account all the probabilistic information present in the linguistic signal (Genzel and Charniak, 2002; Genzel and Charniak, 2003; Keller, 2004; Levy and Jaeger, 2007). One class of results from the literature that has not yet been explained in terms of a rational com- prehender strategy is that of local coherence effects (Tabor et al., 2004; Gibson, 2006; Konieczny and M¨uller, 2007), cases in which it appears that the parser is systematically ignoring contextual infor- mation about possible syntactic structures and pur- suing analyses that are probable only locally. These effects are problematic for rational models, because of the apparent failure to use all of the available in- formation. This paper describes a new model of lo- cal coherence effects under rational syntactic com- prehension, which proposes that they arise as a re- sult of updating prior beliefs about the structures that a given string of words is likely to have to pos- terior beliefs about the likelihoods of those struc- tures in context. The critical intuition embodied in …

Content truncated to 500 words