UCSC on REC 2006 Blog Opinion Mining

The University of California Santa Cruz team submitted three runs for the TREC Blog Track opinion mining task. We developed a two stage retrieval system. We started with retrieving relevant documents from the corpus for each topic, and then ran each retrieved document through a classifier to estimate the probability that the document contains opinion expressions. The documents were ranked according to the product of the retrieval score and the estimated probability. The Lemur search engine, which is based on the language modeling approach, was used for retrieval. A Bayesian Logistic Regression classifier was trained using a noisy training data set from other domains, which include news articles, product reviews and movie reviews. All runs are automatic.