Comparison of similarity models for the relation discovery task

We present results on the relation discovery task, which addresses some of the shortcomings of supervised relation extraction by applying minimally supervised methods. We describe a detailed experimental design that compares various configurations of conceptual representations and similarity measures across six different subsets of the ACE relation extraction data. Previous work on relation discovery used a semantic space based on a term-by-document matrix. We find that representations based on term co-occurrence perform significantly better. We also observe further improvements when reducing the dimensionality of the term co-occurrence matrix using probabilistic topic models, though these are not significant.