Narrative Schema as World Knowledge for Coreference Resolution

JOSEPH HOWARD IRWIN (1051134)

State-of-the-art coreference resolution systems use supervised machine learning methods with various linguistic features. Recent research has focused on incorporating semantic knowledge through ontologies such as WordNet. Despite almost two decades of research, however, high-accuracy systems remain elusive. In part this is because current approaches lack knowledge bases which would allow them to make the kind of commonsense inferences which humans are able to do.

In this work, we investigate the feasibility and effectiveness of incorporating a form of knowledge similar to Schank and Abelson's scripts, which has been made possible by recent work in aquiring this kind of knowledge structure in an unsupervised manner. We develop a baseline coreference resolution system and evaluate the effect of including features based on these narrative schema. In its current implementation the effectiveness of the schema-based features seems to be mostly equivalent to incorporating selectional restriction information. I analyse the behavior of these features, and develop some suggestions aimed at making better use of the unique aspects of the schema structures.