Language Models for Expert Finding--UIUC TREC 2006 Enterprise Track Experiments

In this paper, we report our experiments in the TREC 2006 Enterprise Track. Our focus is to study a language model for expert finding. We extend an existing language model for expert retrieval in three aspects. First, we model the document-expert association using a mixure model instead of name matching heuristics as in the existing work; such a mixture model allows us to put different weights on email matching and name matching. Second, we propose to model the prior of an expert based on the counts of email matches in the supporting documents instead of using uniform prior as in the previous work. Finally, we perform topic expansion and generalize the model from computing the likelihood to computing the cross entropy. Our experiments show that the first two extensions are more effective than the third extension, though when optimized, they all seem to be effective.