Measuring Entity Relatedness via Entity and Text Joint Embedding

As unique identifiers of objects and basic components of knowledge graphs, entities are crucial to many natural language processing related works, such as entity linking and question answering, in which the estimation of entity relatedness is required. Current entity relatedness measures either consider entities as words, which neglects the rich semantics entities contain, or are integrated into extrinsic applications, which fail to evaluate the intrinsic effectiveness. In this work, we propose E5, an effective entity relatedness measure taking into account of entity description text in a neural embedding manner. We first jointly map words and entities to the same high-dimensional vector space, the output of which is utilized as the input for the following joint entity and text embedding training. The well-trained entity and text embedding network is then leveraged to measure similarity between entities and entity descriptions, which in combination with a graph structure based method, constitute the eventual entity relatedness measure. The experimental results validate the usefulness of E5.