Fujitsu Laboratories Trec7 Report 2 System Description 2.1 Overall 2.2 the Search System Teraa

1 Abstract In our rst participation in TREC, our focus was on improving the basic ranking systems and applying text clustering techniques for query expansion. We tested a variety of techiniques including reference measures, passage retrieval, and data fusion for the basic ranking systems. Some te-chiniques were used in the oocial run, others were not used because of time limitations. We applied the text clustering techiniques for query expansion with a text clustering engine. Clustering base query expansion uses the top N best text clusters from the top 1000 documents instead of just using the top N documents. Clustering base query expansion produces better results than simple query expansion based on passage retrieval. We submitted three runs, Flab7at , Flab7ad, and Flab7atE. Flab7at is combination of ranking and query expansion by clustering the top 1000 documents on the title eld, Flab7ad is combination of ranking and query expansion by clustering on the description eld, and Flab7atE is combination of ranking with Boolean (existence) operators and query expansion by passage retrieval. We particpated in TREC with two groups. One group was concerned with search engines, and included index construction, searching process, and normal query expansion. The other group was concerned with the Environment for Document Analysis (EDA), which is related to query expansion with text clustering. The two groups had diierent locations, and the two systems were developped in completely diierent environments. To combine the two systems , we constructed an experimental search procedure using perl script. We also wrote a TREC local procedure in perl script for such tasks as removing stop patterns from input query. Teraaa1],,2] is fulltext search system, designed to provide an adequate number of eecient functions for commercial service, and to provide parameter combination testing and easy extension for experiment of information retrieval. To satisfy both the commercial and experimental requirements, Teraa has many functions and extensibility as described below. 1. Basic search operations Boolean, Boolean Ranking (Ranking with Boolean operators for commercial use), Ranking (Accumulator Method) Near operators for phrase, Not operator, existing operator for term in ranking and parameter control, such as df limiter for ranking, term evaluation order control for ranking, etc. 2. Index type Inverted le index for fulltext search, number array index for range search, number array index for multiple occurences of number in a single text (eg. IPC code in a patent text) , and B-Tree index for item search. …