The Integration of Syntax and Semantic Plausibility in a Wide-Coverage Model of Human Sentence Processing

Models of human sentence processing have paid much attention to three key characteristics of the sentence processor: Its robust and accurate processing of unseen input (wide coverage), its immediate, incremental interpretation of partial input and its sensitivity to structural frequencies in previous language experience. In this thesis, we propose a model of human sentence processing that accounts for these three characteristics and also models a fourth key characteristic, namely the influence of semantic plausibility on sentence processing. The precondition for such a sentence processing model is a general model of human plausibility intuitions. We therefore begin by presenting a probabilistic model of the plausibility of verb-argument relations, which we estimate as the probability of encountering a verb-argument pair in the relation specified by a thematic role in a role-annotated training corpus. This model faces a significant sparse data problem, which we alleviate by combining two orthogonal smoothing methods. We show that the smoothed model’s predictions are significantly correlated to human plausibility judgements for a range of test sets. We also demonstrate that our semantic plausibility model outperforms selectional preference models and a standard role labeller, which solve tasks from computational linguistics that are related to the prediction of human judgements. We then integrate this semantic plausibility model with an incremental, widecoverage, probabilistic model of syntactic processing to form the Syntax/Semantics (SynSem) Integration model of sentence processing. The SynSem-Integration model combines preferences for candidate syntactic structures from two sources: Syntactic probability estimates from a probabilistic parser and our semantic plausibility model’s estimates of the verb-argument relations in each syntactic analysis. The model uses these preferences to determine a globally preferred structure and predicts difficulty in human sentence processing either if syntactic and semantic preferences conflict, or if the interpretation of the preferred analysis changes non-monotonically. In a thorough evaluation against the patterns of processing difficulty found for four ambiguity phenomena in eight reading-time studies, we demonstrate that the SynSem-Integration model reliably predicts human reading time behaviour.