Entity disambiguation leveraging multi-perspective attention

Abstract

Entity disambiguation aims to map mentions in text to the corresponding entities in a knowledge base, which is a basic task in natural language processing (NLP). The major challenge in entity disambiguation is how to extract key imformation in mention context and entity description that is discriminative for disambiguation. State-of-the-art entity disambiguation systems apply attention mechanism to identify the imformative components, but most of the methods only focus on mention context, and neglect entity side. Besides, attention mechanism is employed in a single aspect, which may not be effective in difficult circumstances. In this work, we propose a neural network with multi-perspective attention to enrich the representation of mentions and entities in different perspectives. Specifically, we utilize intra-attention to aggregate internal pivotal information in mention context and entity description separately, and utilize interattention to interact their latent semantics in multiple directions and highlight the interrelated information, so as to capture more informative features and improve the disambiguation performance. The experimental results show that our proposed model outperforms other state-of-the-art entity disambiguation models and attain more improvements on hard datasets, which validates the effectiveness and superiority of our model, especially in complex situations.