IR-Specific Searches at TREC 2007: Genomics & Blog Experiments

This paper describes our participation in the TREC 2007 Genomics and Blog evaluation campaigns. Within these two tracks, our main intent is to go beyond simple document retrieval, using different search and filtering strategies to obtain more specific answers to user information needs. In the Genomics track, the dedicated IR system has to extract relevant text passages in support of precise user questions. This task may also be viewed as the first stage of a Question/Answering system. In the Blog track we explore various strategies for retrieving opinions from the blogsphere, which in this case involves subjective opinions about various targets entities (e.g., person, location, organization, event, product or technology). This task can be subdivided in two parts: 1) retrieve relevant information (facts) and 2) extract positive, negative or mixed opinions about the specific entity being targeted. To achieve these objectives we evaluate retrieval effectiveness using the Okapi (BM25) and various other models derived from the Divergence from Randomness (DFR) paradigm, as well as a language model (LM). Through our experiments with the Genomics corpus we find that the DFR models perform clearly better than the Okapi model (relative difference of 70%) in terms of mean average precision (MAP). Using the blog corpus, we found the opposite; the Okapi model performs slightly better than both DFR models (relative difference around 5%) and LM (relative difference 7%) model.