Combining Structural Information and the Use of Priors in Mixed Named-Page and Homepage Finding

This paper presents Carnegie Mellon University’s experiments on the mixed named-page and homepage finding task of the TREC 12 Web Track. Our results were strong; we achieved the success using language models estimated from combining information from document text, in-link text, and information present in the structure of the documents. We also present experiments using expectations about posterior distributions to create class-based prior probabilities. We find that priors do provide a large gain for our official runs, but we do further experiments that show the priors do not always help. Some preliminary analysis shows that the prior probabilities are not providing the desired posterior distributions. In cases where applying the priors harm performance, the observed posterior distributions in the rankings are far off of the desired posterior distributions.