A Rational Model of Eye Movement Control in Reading

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 1168–1178, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics A Rational Model of Eye Movement Control in Reading 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 A number of results in the study of real- time sentence comprehension have been explained by computational models as re- sulting from the rational use of probabilis- tic linguistic information. Many times, these hypotheses have been tested in read- ing by linking predictions about relative word difficulty to word-aggregated eye tracking measures such as go-past time. In this paper, we extend these results by ask- ing to what extent reading is well-modeled as rational behavior at a finer level of anal- ysis, predicting not aggregate measures, but the duration and location of each fix- ation. We present a new rational model of eye movement control in reading, the cen- tral assumption of which is that eye move- ment decisions are made to obtain noisy visual information as the reader performs Bayesian inference on the identities of the words in the sentence. As a case study, we present two simulations demonstrating that the model gives a rational explanation for between-word regressions. 1 Introduction The language processing tasks of reading, listen- ing, and even speaking are remarkably difficult. Good performance at each one requires integrat- ing a range of types of probabilistic information and making incremental predictions on the ba- sis of noisy, incomplete input. Despite these re- quirements, empirical work has shown that hu- mans perform very well (e.g., Tanenhaus, Spivey- Knowlton, Eberhard, & Sedivy, 1995). Sophisti- cated models have been developed that explain many of these effects using the tools of com- putational linguistics and large-scale corpora to make normative predictions for optimal perfor- mance in these tasks (Genzel & Charniak, 2002, 2003; Keller, 2004; Levy & Jaeger, 2007; Jaeger, 2010). To the extent that the behavior of these models looks like human behavior, it suggests that humans are making rational use of all the infor- mation available to them in language processing. In the domain of incremental language compre- hension, especially, there is a substantial amount of computational work suggesting that humans be- have rationally (e.g., Jurafsky, 1996; Narayanan & Jurafsky, 2001; Levy, 2008; Levy, Reali, & Grif- fiths, 2009). Most of this work has taken as its task predicting the difficulty of each word in a sen- tence, a major result being that a large component of the difficulty of a word appears to be a function of its probability in context (Hale, 2001; Smith & Levy, 2008). Much of the empirical basis for this work comes from studying reading, where word difficulty can be related to the amount of time that a reader spends on a particular word. To re- late these predictions about word difficulty to the data obtained in eye tracking experiments, the eye movement record has been summarized through …

Content truncated to 500 words