MAPS: A Multi-task Framework with Anchor Point Sampling for Zero-shot Entity Linking

Entity linking (EL) plays a crucial role in NLP tasks by linking ambiguous entity mentionsto relevant entities in a knowledge base. Due to the inconsistency in data distributionacross diverse domains, it is difficult to accurately estimate the overall data distributionof target domain, resulting in the zero-shot scenarios with a significant decrease ingeneralization performance. Currently, existing works primarily focus on sampling andincorporating fine-grained information to deal with above issue. Unfortunately, they mayface to either significant computational cost of negative samples for sampling strategy,or shortcomings in interaction between coarse and fine-grained information. To tacklethese challenges, in this paper, we propose a Multi-Task Framework with Anchor PointSampling (MAPS). Specifically, for the anchor point sampling (APS) part, with consideringfine-grained information, we pre-bind mention-entity pairs based on prior conditions(e.g., entity type) to introduce challenging negative samples and modifies the conditionaldistribution. In this way, the optimal trade-off between computational effectiveness andefficiency will be reached. Moreover, we propose a novel multi-task framework that sharescoarse-grained information at a lower level, and utilizes multiple extractors to extract finegrainedinformation at a higher level. By combining the multi-task framework and variousAPS approaches, comprehensive fusion of coarse and fine-grained information will befinally achieved. Experimental results on the benchmark dataset ZESHEL demonstrate thatMAPS significantly outperforms the competitive baselines.