Entity Linking in Queries Using Word, Mention and Entity Joint Embedding

Entity linking in queries is an important task for connecting search engines and knowledge bases. This task is very challenging because queries are usually very short and there is very limited context information for entity disambiguation. This paper proposes a new accurate and efficient entity linking approach for search queries. The proposed approach first jointly learns word, mention and entity embeddings in a unified space, and then computes a set of features for entity disambiguation based on the learned embeddings. The entity linking problem is solved as a ranking problem in our approach, a ranking SVM is trained to accurately predict entity links. Experiments on real data show that our proposed approach achieves better performance than comparison approaches.