T. Florian Jaeger (Department of Brain and Cognitive Sciences, University of Rochester)
Language processing seems to rely on prediction at all levels of language processing studied so far. However, prediction is only a truly effective means of dealing with noisy and incomplete input if comprehenders’ implicit probabilistic knowledge provides sufficiently veridical representations of the actual statistics of the input (the speech or visual signal). However, at least at the level of phonetics, the mapping of linguistic categories to acoustical properties (or even percepts) is highly invariant (e.g., due to talker-specific differences in the realization of sounds).
Research in speech perception suggests that listeners overcome this problem by recognizing familiar talkers, generalizing based on previously experienced talkers, and adapting to novel talkers (for a review and a proposed computational framework, see Kleinschmidt and Jaeger, 2015-Psychological Review). In recent work, we have found that similar adaptation is also observed in lexical and syntactic processing (Fine, Jaeger, Farmer, and Qian, 2013-PLoS One; Fine and Jaeger, 2013-Cognitive Science; Jaeger and Snider, 2013-Cognition; Yildirim, Degen, Tanenhaus, and Jaeger, under revision). I discuss these findings and the questions they raise for discourse processing.