Distinguished Speakers in Language Science

Thursday, 11 December 2014, 16:15
Conference Room, Building C7.4

Language comprehension as rational probabilistic inference on perceptual input

Language comprehension is the process of identifying intended meaning from written or spoken language input. It is well established that comprehension proceeds incrementally: a comprehender’s interpretation of an utterance evolves continuously as the linguistic signal unfolds. In this talk, I present a view of incremental language comprehension across the written and spoken domains as the rational combination of rich, probabilistic language knowledge with noisy perceptual input. I discuss three consequences of this view: (1) It motivates a principled model of how we move our eyes as we read, in which eye movements are made to efficiently gather visual information. I show that this model captures a wide range of properties of how we read. (2) It suggests a linking hypothesis for why we often move our eyes backwards as we read, which is supported by the statistical analysis of a large reading corpus. (3) It makes precise predictions for the trade–off between contextual and acoustic information in spoken comprehension, which are well supported by new experimental data. The success of this unbounded view of comprehension suggests a novel understanding of the notion of ‘difficulty’ in language comprehension: the primary bottleneck in language processing may not be the inherent difficulty of particular language processing operations, but rather the amount of perceptual data available to the language system at each moment in time. More broadly, this work adds to the growing evidence suggesting that many aspects of cognition can be usefully understood as optimized to solve common problems in very efficient ways.

If you would like to meet with the speaker, please contact Vera Demberg.