LAUSR

Entity linking is the task of mapping mentions in text to target knowledge base, which is crucial to knowledge-base-related tasks such as knowledge fusion and knowledge base construction. Although English-oriented entity linking task has undergone continuing advancement, the entity linking systems targeted at Chinese language still suffer from lagged development. State-of-the-art Chinese entity linking systems devise multiple handcrafted features to measure similarity between mention and entity, whereas fail to mine semantic relations underneath the surface forms. In this paper, we propose to take the advantage of latent text features and generate representations of mention and entity via double-attention-based long short term memory network, which are further utilized to calculate mention-entity similarity. Furthermore, joint word and entity embedding training and well-designed candidate entities generation strategies are put forward to facilitate the implementation of neural network. The experimental results validate the superiority of our method Celan. Our proposal not only offers an improved deep neural network for generating mention and entity representation, but also enhances the performance of entity linking on Chinese microblogs.