Constructing query-specific knowledge bases

Abstract Large general purpose knowledge bases (KB) support a variety of complex tasks because of their structured relationships. However, these KBs lack coverage for specialized topics or use cases. In these scenarios, users often use keyword search over large unstructured collections, such as the web. Instead, we propose constructing a 'knowledge sketch' that leverages existing KB data elements and relevant text documents to construct query-specific KB data. A knowledge sketch is a distribution over entities, documents, and relationships between entities, all for a specific information need. In our experiments we construct knowledge sketches for queries from the TREC 2004 Robust track, which emphasizes complex queries which perform poorly with existing text retrieval approaches.