HSIE: Improving Named Entity Disambiguation with Hidden Semantic Information Extractor

Abstract

Named Entity Disambiguation (NED) is a fundamental task in natural language processing. However, existing methods pay more attention to the acquisition of global features, but ignore hidden semantic information in local features. In this paper, we propose a Hidden Semantic Information Extractor (HSIE) to capture hidden semantic features. Specifically, the HSIE is composed of multi-layer neural networks with multi-head attention and a transition layer. Vectors of candidate entities are iteratively updated in the HSIE to capture more elaborated semantic features, which are beneficial for Neural Attention Module (NAM) to compute precise feature scores of candidate entities. An extensible vector space module aims to connect the HSIE and the NAM effectively, which enhances the performance of the HSIE. Such an HSIE can be embedded into local model effectively and extended to global and global (neighbor) models. Moreover, we develop a disambiguation system by orchestrating local, global and global (neighbor) models, each of which is equipped with their own HSIE, the extensible vector space module and the NAM respectively. Experimental results show that our disambiguation system achieves the best performance on AIDA-B, MSNBC and ACE2004 datasets.