Event Related Document Retrieval Based on Bipartite Graph

Abstract

Given a short event name, event retrieval is a process of retrieving event related documents from a document collection. The existing approaches employ the-state-of-art retrieval models to retrieve relevant documents, however, these methods only regard the input query as several keywords instead of an event, thus the special aspects of the event are not considered in the models. Aiming at this problem, we first propose a novel bipartite graph model to describe an event, where one bipartition represents event type and the other represents the event specific information. Each edge between two bipartitions issues co-occurrence relationship. Then we model an event with a unigram language model estimated through the corresponding bipartite graph. Based on KL-divergence retrieval framework, event model is integrated into the query model for more accurate query representation. Experiments on publicly available TREC datasets show that our method can improve the precision@N metric of event retrieval.