Learning Where to Look: Modeling Eye Movements in Reading

Proceedings of the Thirteenth Conference on Computational Natural Language Learning (CoNLL), pages 93–101, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Learning Where to Look: Modeling Eye Movements in Reading Mattias Nilsson Department of Linguistics and Philology Uppsala University mattias.nilsson@lingfil.uu.se Joakim Nivre Department of Linguistics and Philology Uppsala University joakim.nivre@lingfil.uu.se Abstract We propose a novel machine learning task that consists in learning to predict which words in a text are fixated by a reader. In a first pilot experiment, we show that it is possible to out- perform a majority baseline using a transition- based model with a logistic regression classi- fier and a very limited set of features. We also show that the model is capable of capturing frequency effects on eye movements observed in human readers. 1 Introduction Any person engaged in normal skilled reading pro- duces an alternating series of rapid eye movements and brief fixations that forms a rich and detailed be- havioral record of the reading process. In the last few decades a great deal of experimental evidence has accumulated to suggest that the eye movements of readers are reflective of ongoing language pro- cessing and thus provide a useful source of infor- mation for making inferences about the linguistic processes involved in reading (Clifton et al., 2007). In psycholinguistic research, eye movement data is now commonly used to study how experimental ma- nipulations of linguistic stimuli manifest themselves in the eye movement record. Another related strand of research primarily at- tempts to understand what determines when and where the eyes move during reading. This line of research has led to mathematically well specified ac- counts of eye movement control in reading being instantiated as computational models (Legge et al., 1997; Reichle et al., 1998; Salvucci, 2001; Engbert et al., 2002; McDonald et al., 2005; Feng, 2006; Reilly and Radach, 2006; Yang, 2006). (For a re- cent overview, see (Reichle, 2006).) These models receive text as input and produce predictions for the location and duration of eye fixations, in approxima- tion to human reading behavior. Although there are substantial differences between the various models, they typically combine both mechanisms of visuo- motor control and linguistic processing. Two impor- tant points of divergence concern the extent to which language processing influences eye movements and whether readers process information from more than one word at a time (Starr and Rayner, 2001). More generally, the models that have emerged to date are based on different sets of assumptions about the un- derlying perceptual and cognitive mechanisms that control eye movements. The most influential model so far, the E-Z Reader model (Reichle et al., 1998; Reichle et al., 2003; Pollatsek et al., 2006), rests on the assumptions that cognitive / lexical processing is the engine that drives the eyes through the text and that words are identified serially, one at a time. Although eye movement models typically have parameters that are fitted to empirical data sets, they are not based on machine learning in the standard sense and …

Content truncated to 500 words