Exploiting Background Information Networks to Enhance Bilingual Event Extraction Through Topic Modeling

In this paper we describe a novel approach of biased propagation based topic modeling to exploit global background knowledge for enhancing both the quality and portability of event extraction. The distributions of event triggers and arguments in topically-related documents are much more focused than those in a heterogeneous corpus. Based on this intuition, we apply topic modeling to automatically select training documents for annotation, and demonstrate it can significantly reduce annotation cost in order to achieve comparable performance for two different languages and two different genres. In addition, we conduct cross-document inference within each topic cluster and show that our approach advances state-of-the-art.