Graph neural entity disambiguation

Abstract Entity Disambiguation (ED) aims to automatically resolve mentions of entities in a document to corresponding entries in a given knowledge base. State-of-the-art ED methods typically utilize local contextual information for obtaining mention embeddings which will be compared to candidate entity embeddings and then apply Conditional Random Field (CRF) for collective ED, considering global coherence. An inherent drawback of these methods is that, the global semantic relationships among the candidate entities in the same document are not encoded in the embedding process. As such, the resultant embeddings may not be sufficient to capture the global coherence effect. In this paper, to address the issue, we propose a novel end-to-end graph neural entity disambiguation model which fully exploits the global semantic information. In particular, a heterogeneous entity-word graph is first constructed for each document to model the global semantic relationships among candidate entities in a same document. Then graph convolutional network (GCN) is applied on the entity-word graph to generate enhanced entity embeddings encoding global semantics, which are fed to a CRF for collective ED. Extensive experiments have demonstrated the efficiency and effectiveness of our method over a few state-of-the-art ED methods.