Fujitsu Laboratories TREC2001 Report

This year a Fujitsu Laboratory team participated in web tracks. Both for ad hoc task, and entry point search task, we combined the score of normal ranking search and that of page ranking techniques. For ad hoc style task, the eect of page ranking was very limitted. We only got very little improvement for title eld search, and the page rank was not eective for description, and narrative eld search. For entry point search task, we compared three heuristics. The rst heuristics supposed that entry point page contains key word and had high page rank score. The second heuristics supposed that entry point page contains key word in its head part and had high page ran score. The third heuristics supposes that entry point is pointed by the pages whose anchor string contains key word, and has high page rank score. The page rank improved the result of entry point search about 20-30% in rather small VLC10 test set, and the rst heuristics got the best result because of its high recall. For TREC2001, we added the new functions to trec exec for entry point search. The functions includes score merging, evaluation of reciprocal rank and so on. We u s e d W eb Recommener Agent to get page ranking score. Except above modications, the framework is same as that of TREC9[1]. 1.0.1 Tera Tera[2] is a fulltext search library, designed to provide an adequate number of ecient functions for commercial service, and to provide parameter combination testing and easy extension for experiments in IR. 1.0.2 trec exec trec exec is designed for automatic processing of TREC. It contains a procedure controller, evaluation module , logging module, and all non-searching units such as query generation, query expansion and so on. trec exec can execute all the TREC processing for one run in a few minutes, and it can be used for system tuning by hill-climing. The new functions added for TREC2001 Web track are heuristics for entry point search, evaluation of reciprocal rank, and accepting non-digit query number. We used web recommeder agent tool developed for automatic domain specic web directory Tsuda et al[3] to get page ranking score. The page rank score is put into Tera, and it is marged with normal ranking score.