Distantly Supervised Entity Linking with Selection Consistency Constraint.

Entity linking (EL) aims to find entities that the textual mentions refer to from a knowledge base (KB). The performance of current distantly supervised EL methods is not satisfactory under the condition of low-quality candidate generation. In this paper, we consider the scenario where multiple KBs are available, and for each KB, there is an EL model corresponding to it. We propose the selection consistency constraint (SCC), that is, for one sample, the entities selected from multiple KBs should be consistent if these selections are all correct. In this work, we aim to utilize the SCC to improve the performance of each EL model (not the combination of multiple EL models) under low-quality candidate generation. Specifically, we define an SCC model from two different aspects: minimizing probability and upper bound, which are used to introduce the SCC into the training of EL models. The experimental results show that our method, jointly training multiple EL models with the SCC model, outperforms the baseline which trains multiple EL models separately, and it has low cost.