Working Memory Load in Sentence Parsing

This thesis presents a computational model of working memory load in human sentence processing that attempts to explain a number of well-established linguistic performance effects that are associated with processing difficulty of certain syntactic structures. The computational model presented here is based on recent assumptions of working memory in computational cognitive modeling. In particular the assumption that there are working memory demands in on-line sentence comprehension associated with both storage of incomplete syntactic structure and integration of new lexical input into the structure built thus far. The model has been realized as an implemented system for parsing a fragment of English into dependency representations. The implementation provides detailed quantitative predictions of the time course of sentence comprehension and processing difficulty. These predictions afford comparison with empirical evidence such as reading time data. In contrast to current sentence processing theories which rely on phrase structure representations, we implement an algorithm for parsing where the basic syntactic relations necessary for interpretation are established directly as asymmetrical dependency relations between lexical nodes. In addition to accounting for well-established processing contrasts, the dependency-based parsing model detailed here reflects important principles of real-time sentence processing, such as incrementality and predictivity.