Improving Document Ranking with Relevance-based Entity Embeddings

Abstract

Document ranking aims to rank and return a group of documents in accordance to their relevance to queries. Traditional document ranking solutions utilize sparse vectors to represent query and documents, and then rank documents in accordance to the similarity between sparse vectors. Over recent years, with the advance of big data technology, such as the emergence of knowledge graph (KG), entities are considered as more essential pivots to connect queries and documents, which are used to improve the document ranking process. However, state-of-the-art entity embedding methods usually place entities with close proximity or similar contexts into the same area in the entity embedding space, which does not meet the goal of document ranking, in which entity relevance plays a more important role. As thus, we propose to enhance document ranking with relevance-based entity embedding. In particular, we first introduce the neural network for training such embeddings, and set the objective that given information needs, i.e., query entities, the frequently occurring entities in the top retrieved documents should be predicted. Then, we trained the model based on Wikipedia articles, and used it to improve the baseline document ranking framework. Empirical experiments validate the superiority of the proposed method.