Utilizing structured knowledge bases in open IE based event template extraction

摘要

Automatic template extraction including event template has been studied intensively in recent years. Researchers study the topic in order to solve the problem of manually defining a template that is required in most information extraction systems. Several studies of event template extraction rely on the documents characteristics to discover the pattern. Although there exist some structured knowledge bases, such as: FrameNet, Predicate Matrix, ACE (Automatic Content Extraction) event type keywords seeds, and FrameNet-ACE event type mapping, no previous researchers have studied combining this information for event template extraction. This paper presents an event template extraction approach that incorporates structured knowledge bases. We propose event template extraction from Open Information Extraction (Open IE) results (relation tuples) in two stages: relation tuple clustering and relation tuple filtering. Both processes utilize structured knowledge bases, as constraint sources in the clustering process and as the basis for the filtering process. The filtering process employs the word embedding representation to capture the semantic relatedness between words. We argue that by involving structured knowledge bases, the relation tuple semantic information can be enriched. Therefore, we can get groups of relation tuples with a similar event sense that represent event templates. The empirical experiment was based on an event argument extraction task and showed that our proposed approach outperforms similar methods that do not use structured knowledge bases. We also compare our proposed system performance to the performance of state-of-the-art systems. The comparison result shows that our proposed system outperforms other state-of-the-art systems, in terms of precision.