Joint embedding of hierarchical structure and context for entity disambiguation

Entity linking refers to the task of constructing links between the mentions of context and the description pages from knowledge base. Due to the polysemy phenomenon, the key issue of entity linking is entity disambiguation. To simplify the goal of entity disambiguation, the main problem is choosing the correct entity from candidates. In this paper, we propose a novel embedding method specifically designed for entity disambiguation. Existing distributed representations are limited in utilizing structured knowledge from knowledge bases such as Wikipedia. Our method jointly maps the information from hierarchical structure of knowledge and context words. We extend the continuous bags-of-words model by adding hierarchical categories and hyperlink structure. So far, we have trained a joint model which adds category information. We demonstrate the utility of our proposed approach on an entity relatedness dataset. The results show that our jointly embedding model is superior to the model simply using context words. In addition, we do disambiguation experiment on a dataset, and the results show slight superiority of the novel embedding model.