Unsupervised Ultra-Fine Entity Typing with Distributionally Induced Word Senses

The lack of annotated data is one of the challenging issues in an ultra-fine entity typing, which is the task to assign semantic types for a given entity mention. Hence, automatic type generation is receiving increased interest, typically to be used as distant supervision data. In this study, we investigate an unsupervised way based on distributionally induced word senses. The types or labels are obtained by selecting the appropriate sense cluster for a mention. Experimental results on an ultra-fine entity typing task demonstrate that combining our predictions with the predictions of an existing neural model leads to a slight improvement over the ultra-fine types for mentions that are not pronouns.