CSIR at TREC 2007 Expert Search Task

This is the second year for the participation of Center for Studies of Information Resources (CSIR) in the TREC Expert Search Task. Rather than using the candidate profile based approach, a simplified two stage approach is used in our experiment, that is, documents are ranked based on topics, and each expert is scored intuitively by the weights of documents the expert appeared. Instead of the modeling of expert search, we have mainly focused on the effect of document filtering in the expert search. In our experiment, only the top n ranked topic-relevant documents where the expert also appeared are calculated into the expert score. The tuned value of n under W3C corpus for a best performance is 10, which is proved to be stable under CERC corpus.