Learning from Knowledge Graphs: Neural Fine-Grained Entity Typing with Copy-Generation Networks

Abstract

Fine-grained entity typing (FET) aims to identify the semantic type of an entity in a plain text, which is a significant task for downstream natural language processing applications. However, most existing methods neglect rich known typing information about these entities in knowledge graphs. To address this issue, we take advantage of knowledge graphs to improve fine-grained entity typing through the use of a copy mechanism. Specifically, we propose a novel deep neural model called CopyFet for FET via a copy-generation mechanism. CopyFet can integrate two operations: (i) the regular way of making type inference from the whole type set in the generation model; (ii) the new copy mechanism which can identify the semantic type of a mention with reference to the type-copying vocabulary from a knowledge graph in the copy model. Despite its simplicity, this mechanism proves to be powerful since extensive experiments show that CopyFet outperforms state-of-the-art methods in FET on two benchmark datasets (FIGER (GOLD) and BBN). For example, CopyFet achieves the new state-of-the-art score of 76.4% and 83.6% on the accuracy metric in FIGER (GOLD) and BBN, respectively.