Petr Babkin, RPI Graduate Student

 

Petr Babkin, RPI Graduate Student

Sage 4101

November 1, 2017 12:00 PM - 1:30 PM

   

Language understanding has traditionally been viewed in AI research as an input preprocessing step, which translates unstructured text into some formal representation ---prerequisite to downstream reasoning and action (Bar Hillel, 1970). By contrast, a substantial amount of psycholinguistic evidence suggests that reasoning and even action happen in tandem with the increments of understanding. First, the interpretation (and possibly reaction) begins before all input is perceived (Kilger and Finkler, 1995). Second, the interpretation is built on different levels of linguistic analysis in parallel rather than serially (Marslen-Wilson, 1975). Anticipation and predictability play a major role in this process and they have been studied from viewpoints ranging from simple co-occurence probabilities, to saccadic eye movements and contextual coherence (Kamide, 2008). On one hand, anticipation is beneficial as it allows the hearer to be more robust to noise and to respond proactively. On the other hand, it comes with a cognitive overhead and may incur additional cost due to possible misprediction (Gibson, 1998). We review some of the work on the subject and present an anticipatory model based on scripts (Schank and Abelson, 1977) and discourse coherence. We evaluate it against a non-anticipatory mode of meaning representation construction ultimately aiming to weigh the benefits against the costs.

Add to calendar
Share|