Negative-aware Entity Set Expansion

Anonymous

Abstract: Entity Set Expansion (ESE) aims to find all entities of one target semantic class with a few seed entities describing it. However, existing ESE methods cannot express what entities we explicitly dislike, and thus hinder its application in real-world scenarios.In this paper, to endow models with the capability of understanding the ``dislike'' relationship among seed entities, we express the target semantic class with both positive and negative seed entities.To this end, we propose an efficient and learnable negative-aware entity set expansion framework, which is essentially a retrieval model. To facilitate this study, a large-scale Negative-aware ESE Dataset NED with more than 1M entities is further collected and annotated. Extensive experiments on NED show that the proposed framework can effectively understand the dislike relations expressed by the negative seeds and expand fewer dislike entities than baseline methods.