A Research On Distantly Supervised Entity Linking With Selection Consistency Constraint

The goal of entity linking(EL)task is to find out the corresponding entities of textual entity mentions from a knowledge base(KB).At present,under the condition of low-quality candidate generation,the performance of distantly supervised EL methods is not satisfactory due to the fact that supervision is weak.Therefore,in this paper,we consider using extra KBs to improve the performance when there are multiple KBs available.It is difficult to combine multiple KBs directly into a single KB through entity alignment(EA)technology,but our method,which utilizes multiple KBs through selection consistency constraint(SCC),is more feasible.First,for each KB,there is an EL model corresponding to it.Then,we propose the SCC,that is,for one sample,the entities selected from multiple KBs should be consistent if these entities are all correct.We aim to utilize the SCC to improve the performance of each EL model.Specifically,we define the SCC models to introduce the SCC into the training of multiple EL models,so that multiple EL models can supervise each other through SCC models.The experimental results show that our method,jointly training multiple EL models with the SCC models,outperforms the method which trains multiple EL models separately,and it has low cost.