Entity-Based Relevance Feedback for Document Retrieval

Eilon Sheetrit, Fiana Raiber, Oren Kurland

Abstract: There is a long history of work on using relevance feedback for ad hoc document retrieval. The main types of relevance feedback studied thus far are for documents, passages and terms. We explore the merits of using relevance feedback provided for entities in an entity repository. We devise retrieval methods that can utilize relevance feedback provided for tokens whether entities or terms. Empirical evaluation shows that using entity relevance feedback falls short with respect to utilizing term feedback on average, but is much more effective for difficult queries. Furthermore, integrating term and entity relevance feedback is of clear merit; e.g., for augmenting minimal document feedback. We also contrast approaches to presenting entities and terms for soliciting relevance feedback.