Multi-task entity linking with supervision from a taxonomy

Entity linking is the task of resolving ambiguous mentions in documents to their referent entities in a knowledge graph (KG). Existing solutions mainly rely on three kinds of information: local contextual similarity, global coherence, and prior probability. But the information of the mentions’ types is rarely utilized, which is helpful for precise entity linking. That is to say, if the type information of a mention is obtained from a mention classifier, we can exclude candidate entities with different types. However, the key challenge of realizing it lies in obtaining the type labels with appropriate granularity and performing entity linking with the error propagated from the mention classifier. To solve the challenges, we propose a model named type-oriented multi-task entity linking (TMTEL). First, we select types with appropriate granularity from the taxonomy of a KG, which is modeled as a nonlinear integer programming problem. Second, we use a multi-task learning framework to incorporate the selected types into entity linking. The type information is used to enhance the representation of the mentions’ context, which is more robust to the errors of the mention classifier. Experimental results show that our model outperforms multiple existing solutions.