Fujitsu Laboratories Trec9 Report 1 System Description 2 Common Processing 2.1 Indexing/query Processing 2.1.1 Indexing Vocabulary 2.1.2 Stemmer 2.1.4 Stop Word List for Query Processing

This year a Fujitsu Laboratory team participated in web tracks. For TREC9 we experimented passage retrieval which is expected to be e ective for Web pages which contain more than one topic. To split document into passages, we used NLP based paragrah detecting program, not by xed (variable) window size. But it did not produce better result for TREC9 Web data. For indexing large web data faster, we developped two techiniques. One is multi-partional selective sorting for inversion which is about 10-30% faster than normal quick sorting in sorting term-number, text-number pair. The other is compressed trie dictionary based stemming. 1 System Description Except reranking by passage retrieval, and passage segementing program for index preprocessing, the frame work we used, is same as that of TREC8[1].