Event Group Similarity Analysis using Latent Dirichlet Allocation

An event, or a predicate and its arguments, is a semantic representation in the task of information extraction (IE) that provides little information beyond the sentence level. This paper describes event groups, sets of related events that encode semantic information across multiple sentences. We define an information retrieval task that uses our proposed event group representation as queries to create a more semantically oriented search. This task requires finding a method to measure event group similarity, which this paper addresses. Our approach approximates the similarity between pairs of event groups using distances between their topic distributions generated by a Latent Dirichlet Allocation (LDA) model. We use the ACE 2005 dataset to manually construct a gold standard corpus with annotated event groups and ground truth pairwise similarity scores. We evaluate the LDA similarity scores against the gold standard and show an average Spearman’s rank correlation coefficient of 0.120.