TREC-8 Ad-Hoc, Query and Filtering Track Experiments using PIRCS

In TREC-8, we participated in automatic ad-hoc retrieval as well as the query and filtering tracks. The theme of our participation is ‘retrieval lists combination’, and the technique is applied throughout our experiments to various degree. It is pointed out that our PIRCS system may be considered as a combination of probabilistic retrieval model and a language model approach. For adhoc, three types of experiments were done with short, medium and long queries as before. General approach is similar to TREC-7, but combination of retrieval lists from different query types were used to boost effectiveness. For query track, we submitted one shortquery set, and performed retrieval for twenty one natural language query vairants. For filtering track, experiments for adaptive, batch filtering, and routing were performed. For adaptive, historical selected document list was used to train profile term weights and dynamically vary retrieval status value (rsv) threshold for deciding document selection during the course of filtering. For batch filtering, Financial Times FT92 data was used to define 6 retrieval profiles whose results were combined based on coefficients trained via a genetic algorithm. Logistic regression transforms rsv`s to probabilities. Routing was similarly done with additional training data obtained from non-FT collections and two additional profiles were defined and combined