Read, Retrospect, Select: An MRC Framework to Short Text Entity Linking

Entity linking (EL) for the rapidly growing short text (e.g. search queries and news titles) is critical to industrial applications. Most existing approaches relying on adequate context long EL are not effective concise sparse text. In this paper, we propose a novel framework called Multi-turn Multiple-choice Machine reading comprehension (M3) solve from new perspective: query generated each ambiguous mention exploiting its surrounding context, an option selection module employed identify golden entity candidates using query. way, M3 sufficiently interacts limited with candidate entities during encoding process, as well implicitly considers dissimilarities inside bunch in stage. addition, design two-stage verifier incorporated into address commonly existed unlinkable problem To further consider topical coherence interdependence among referred entities, leverages multi-turn fashion deal mentions sequence manner by retrospecting historical cues. Evaluation shows that our achieves state-of-the-art performance five Chinese English datasets real-world EL.