Deriving implicit user feedback from partial URLs for effective web page retrieval

Abstract

User click-throughs provide a search context for understanding the user need of complex information. This paper re-examines the effectiveness of this approach when based on partial clicked data using the language modeling framework. We expand the original query by topical terms derived from clicked Web pages and enhance early precision via a more compact document representation. Since our URLs of Web pages are stripped, we first reconstruct them at different levels based on different collections. Our experimental results on the GOV2 test collection and AOL query log show improvement by 31.7% and 28.3% significantly in statMAP for two sources of reconstruction and 153 ad-hoc queries. Our model also outperforms pseudo relevance feedback.