Attention-Based Bilinear Joint Learning Framework for Entity Linking

Abstract

Entity Linking (EL) is a task that links entity mentions in the text to corresponding entities in a knowledge base. The key to building a high-quality EL system involves accurate representations of word and entity. In this paper, we propose an attention-based bilinear joint learning framework for entity linking. First, a novel encoding method is employed for coding EL. This method jointly learns words and entities using an attention mechanism. Next, for ranking features, a weighted summation model is introduced to model the textual context and coherence. Then, we employ a pairwise boosting regression tree (PBRT) to rank candidate entities. As input, PBRT takes both features constructed with a weighted summation model and conventional EL features. Finally, through the experiment, we demonstrate that the proposed model learns embedding efficiently and improves the EL performance compared with other state-of-the-art methods. Our approach achieves superior result on two standard EL datasets: CoNLL and TAC 2010.