A Model of Difficulty in Human Sentence Processing
Speaker: Ulrike Pado
Institution: Saarland University
Abstract:
While humans process most language input quickly and effortlessly, there is a number of phenomena that notoriously cause the language processing mechanism difficulty. We propose a model that predicts such difficulty given the same language input as humans. Our model evaluates the syntactic probability and semantic plausibility of candidate interpretations of the input, and predicts difficulty based on disagreement between syntactic and semantic evaluation and on incremental semantic discontinuity. The model is derived from corpus data and deals with unseen input, thereby also accounting for the processing of unproblematic language data.
We first introduce the semantic model component and then demonstrate how it is combined with the syntactic component in a way that overcomes problems with existing models of sentence processing. Finally, we verify that the model correctly predicts human data.