Abstract

Entity disambiguation with a knowledge base becomes increasingly popular in the NLP community. In real life application, not all entities in the knowledge base have rich text content or sufficient labeled data. So it is important to solve entity disambiguation problem when lacking of sufficient labeled training data. In this paper, we leverage the labeled data via the type taxonomy. We trained type classifiers based on the labeled data, then annotate the unlabeled data with these type classifiers. We iteratively re-train the entity classifiers with the newly labeled data, and update the annotations. Experiments on 33743 real life entities show the advantages of considering all distinguishable types and the ambiguity of different entity mentioning.

Paper Details