What to count and when to decide: Open issues in frequency‐based models of human parsing*

Abstract

Structural ambiguities like the PP‐attachment ambiguity in Sonja saw the man with the binoculars abound in many languages. It is the objective of psycholinguistic modeling to predict which of the alternative readings is preferred during on‐line processing, and to explain why. Frequency‐based models of human sentence processing claim that humans keep an internal statistics about ambiguous material in their linguistic input. This statistic is accessed during processing and guides the interpretation process in so far as frequent readings of ambiguous constructions are preferred. In order to generate processing predictions, these models count the number of alternative interpretations of ambiguous structures in large corpora, which are regarded as an approximation to the previous exposure of human subjects to ambiguous language data. Empirical testing of frequency‐based models requires both the specification of the grain size used for counting frequencies in corpora, and the modeling of temporal aspects of on‐line processing. It will be shown that current frequency‐based models are still underspecified with respect to at least one of these requirements. Empirical evidence will be presented that indicates the need for counting frequencies on multiple grain levels.