Exploring Distinctive Features in Distant Supervision for Relation Extraction.

Distant supervision (DS) for relation extraction suffers from the noisy labeling problem. Most solutions try to model the noisy instances in the form of multi-instance learning. However, in the non-noisy instances, there may be noisy features which would harm the extraction model. In this paper, we employ a novel approach to address this problem by exploring distinctive features and assigning distinctive features more weight than the noisy ones. We make use of all the training data (both the labeled part that satisfies the DS assumption and the part that does not), and then employ an unsupervised method by topic model to discover the distribution of features to latent relations. At last, we compute the distinctiveness of features by using the obtained feature-relation distribution, and assign features weights based on their distinctiveness to train the extractor. Experiments show that the approach outperforms the baseline methods in both the held-out evaluation and the manual evaluation significantly. © 2013 Springer-Verlag.