BUPT at TREC 2006: Enterprise Track

This year, the expert search task requires a list of support documents provided for each expert. The change implied that support documents for the potential experts should be found before getting the experts themselves, which is one of the natural ways for expert search. The two-stage ranking method we used last year was just following this way. We develop an expert experience model using window-based method this year, in which our efforts were focused on the combination of using local content for evidence and quoting entire document for support. We also tried to treat some important types of data particularly both in the corpus and in a document. Finally the headings in every page were given a high weight. Each email author was given an additional weight for the confidence of their relationship with the email content. All our experiments were based on the 4.2version of Lemur Toolkit, in which language model with Bayesian smoothing was used for relevance computing. For candidate location, the candidate list and the name disambiguation rules[1] used last year were still working this time. But we found there were some problems in encoding which would cause missing match for a few candidates. We accepted several encoding representation in our system. The detail of the expert experience model and some improvements are in the following analysis.