There is strong evidence that human sentence processing is
incremental, i.e., that structures are built word by word. Recent
experimental results show that the processor us also predictive, i.e.,
it can anticipate upcoming linguistic material on the basis of
previous input. However, the granularity of this prediction process is
currently unclear. We present evidence from two visual world
experiments that show that speakers can make fine-grained predictions
based on the frame and the frequency of a verb. We use this data to
motivate a computational model of human parsing that includes an
explicit mechanism for generating and verifying predictions. Based on
this mechanism, our model can capture both locality effects and
surprisal effects, and thus unify empirical results that have so far
been accounted for separately. We evaluate the model on the Dundee
eye-tracking corpus and on relative clause data that existing models
fail to capture adequately.
Joint work with Manabu Arai, Vera Demberg, and Roger Levy.