Taxonomy-Enhanced Document Retrieval with Dense Representations

Document retrieval is a task that powers several downstream applications such as search and question answering. One way to approach this task is to take embeddings of the documents to be retrieved, and of the query, and use a similarity function to rank results. In this work, we extend this approach by incorporating knowledge about entities mentioned in either the document or the query, in the form of taxonomic relations and canonical labels of said entities. The method, when applied to a domain-specific corpus, improves retrieval recall over a state of the art method trained on a general domain corpus. It does so without requiring any further retraining of the machine learning models involved, thus making it applicable for use cases where training is not feasible because of data or infrastructure limitations.