'Bag of Events' Approach to Event Coreference Resolution. Supervised Classification of Event Templates

摘要

We propose a new robust two-step approach to cross-textual event coreference resolution on news articles. The approach makes explicit use of event and discourse structure thereby compensating for implications of the Gricean Maxim of quantity. News follows the principle of language economy. Information tends not to be repeated within discourse boarders. This phenomenon poses a challenge for models comparing information about event mentions (and their arguments) on the sentence level. Our approach addresses this challenge by building a knowledge representation per unit of discourse - for present purposes, a document. We collect event information from a single document filling in a "document template " and by that creating a "Bag of Events." We then use supervised Classification to determine if pairs of document templates contain corefering event mentions. Next we solve coreference between event mentions from the same document cluster by means of supervised classification of "sentence templates." The results indicate that the new approach is promising.