Lexical surprisal as a general predictor of reading time

Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 398–408, Avignon, France, April 23 - 27 2012. c⃝2012 Association for Computational Linguistics Lexical surprisal as a general predictor of reading time Irene Fernandez Monsalve, Stefan L. Frank and Gabriella Vigliocco Division of Psychology and Language Sciences University College London {ucjtife, s.frank, g.vigliocco}@ucl.ac.uk Abstract Probabilistic accounts of language process- ing can be psychologically tested by com- paring word-reading times (RT) to the con- ditional word probabilities estimated by language models. Using surprisal as a link- ing function, a significant correlation be- tween unlexicalized surprisal and RT has been reported (e.g., Demberg and Keller, 2008), but success using lexicalized models has been limited. In this study, phrase struc- ture grammars and recurrent neural net- works estimated both lexicalized and unlex- icalized surprisal for words of independent sentences from narrative sources. These same sentences were used as stimuli in a self-paced reading experiment to obtain RTs. The results show that lexicalized sur- prisal according to both models is a signif- icant predictor of RT, outperforming its un- lexicalized counterparts. 1 Introduction Context-sensitive, prediction-based processing has been proposed as a fundamental mechanism of cognition (Bar, 2007): Faced with the prob- lem of responding in real-time to complex stim- uli, the human brain would use basic information from the environment, in conjunction with previ- ous experience, in order to extract meaning and anticipate the immediate future. Such a cognitive style is a well-established finding in low level sen- sory processing (e.g., Kveraga et al., 2007), but has also been proposed as a relevant mechanism in higher order processes, such as language. In- deed, there is ample evidence to show that human language comprehension is both incremental and predictive. For example, on-line detection of se- mantic or syntactic anomalies can be observed in the brain’s EEG signal (Hagoort et al., 2004) and eye gaze is directed in anticipation at depictions of plausible sentence completions (Kamide et al., 2003). Moreover, probabilistic accounts of lan- guage processing have identified unpredictability as a major cause of processing difficulty in lan- guage comprehension. In such incremental pro- cessing, parsing would entail a pre-allocation of resources to expected interpretations, so that ef- fort would be related to the suitability of such an allocation to the actually encountered stimulus (Levy, 2008). Possible sentence interpretations can be con- strained by both linguistic and extra-linguistic context, but while the latter is difficult to evalu- ate, the former can be easily modeled: The pre- dictability of a word for the human parser can be expressed as the conditional probability of a word given the sentence so far, which can in turn be es- timated by language models trained on text cor- pora. These probabilistic accounts of language processing difficulty can then be validated against empirical data, by taking reading time (RT) on a word as a measure of the effort involved in its pro- cessing. Recently, several studies have followed this ap- proach, using “surprisal” (see Section 1.1) …

Content truncated to 500 words