An expectation model of referring expressions

Abstract

This thesis introduces EMRE, an expectation-based model of referring expressions. EMRE is proposed as a model of non-syntactic dependencies - in particular, discourse-level semantic dependencies that bridge sentence gaps. These include but are not limited to anaphora (references to noun phrases in previous sentences) and coherence predicates such as causality, temporal ordering and resemblance -- two domains that have typically been treated as entirely distinct aspects of language. EMRE is a computational-level model, and is agnostic about any particular algorithms, cognitive faculties, or neurological substrates that might be applied to the problem of semantic reference. Instead, it describes reference as a computational problem framed in terms of expectation and inference, and describes a solution to the problem based on rational top-down expectations about the likely targets of referring expressions, and on bottom-up feature-based matching that occurs when a referring expression is encountered. EMRE is used to derive novel empirical predictions about how people will construe particular discourse constructions involving NP anaphora and coherence predicates. These predictions are tested in controlled behavioral experiments, in which participants read and answer questions about short texts. The results of these experiments are shown to be consistent with a model of reference as an expectation-based computational structure with different underlying rules than those governing syntactic processing.