ESearch: Incorporating Text Corpus and Structured Knowledge for Open Domain Entity Search

The paper introduces an open domain entity search system called ESearch, which aims at finding a list of relevant entities to an open domain entity search query (a natural language question). The system is built on top of a Wikipedia text corpus, as well as the structured DBPedia knowledge base. Entities are initially ranked by a model which effectively associates context matching (based on the contexts of entities in the unstructured text corpus) and category matching (based on the types of entities in the structured knowledge base). They are ranked further by a re-ranking component supported by blind feedback or user feedback on entities. We show that category matching is critical for the search performance and the re-ranking component can boost the performance largely. Category matching therefore needs some query entity types (especially specific entity types) as input. However, it is often hard for systems to detect specific entity types because users may not be familiar with how the types of desired entities are defined in the structured knowledge base. In ESearch, we design an effective ranking model of entity types to facilitate blind feedback and user feedback on desired entity types for category matching, so that users can effectively perform entity search without the need of explicitly providing any query entity types as inputs.