Event

Title:  Can surprisal explain syntactic disambiguation difficulty?

Abstract:
The disambiguation of a syntactically ambiguous sentence in favor of an unexpected parse can lead to slower reading at the disambiguation point. In the sentence "even though the girl phoned the instructor was very upset", for instance, the words "was very upset" typically cause considerable processing difficulty. Such garden path effects have motivated models in which readers only maintain a subset of the possible parses of the sentence. By contrast, surprisal theory -- a relatively new contender -- maintains that these effects arise as a special case of word predictability effects: quite generally, unpredictable words are read more slowly than predictable ones. We argue that as of yet this hypothesis has not been subjected to a rigorous quantitative evaluation, and use neural network language models to carry out such an evaluation. I will show that while the models show sensitivity to the relevant syntactic factors, the word predictability estimates they produce dramatically underestimate the magnitude of empirical garden path effects, and make incorrect predictions as to the detailed pattern of results. I will conclude by discussing prospects for addressing this discrepancy, as well as ongoing efforts to create a much larger empirical resource that can serve as a benchmark for more detailed computational models.