Pagerank with text similarity and video near-duplicate constraints for news story re-ranking

Abstract

Pseudo-relevance feedback is a popular and widely accepted query reformulation strategy for document retrieval and re-ranking. However, problems arise in this task when assumed-to-be relevant documents are actually irrelevant which causes a drift in the focus of the reformulated query. This paper focuses on news story retrieval and re-ranking, and offers a new perspective through the exploration of the pair-wise constraints derived from video near-duplicates for constraint-driven re-ranking. We propose a novel application of PageRank, which is a pseudo-relevance feedback algorithm, and use the constraints built on top of text to improve the relevance quality. Real-time experiments were conducted using a large-scale broadcast video database that contains more than 34,000 news stories. © 2010 Springer-Verlag Berlin Heidelberg.