UMass Amherst and UT Austin @ the TREC 2009 Relevance Feedback Track

We present a new supervised method for estimating term-based retrieval models and apply it to weight expansion terms from relevance feedback. While previous work on supervised feedback [Cao et al., 2008] demonstrated significantly improved retrieval accuracy over standard unsupervised approaches [Lavrenko and Croft, 2001, Zhai and Lafferty, 2001], feedback terms were assumed to be independent in order to reduce training time. In contrast, we adapt the AdaRank learning algorithm [Xu and Li, 2007] to simultaneously estimate parameterization of all feedback terms. While not evaluated here, the method can be more generally applied for joint estimation of both query and feedback terms. To apply our method to a large web collection, we also investigate use of sampling to reduce feature extraction time while maintaining robust learning.