NTCIR-4 PATENT Experiments at Osaka Kyoiku University —Gram-based Passage Index and Essential Words—
Takashi Sato, Nao Hatta · NTCIR · 2004
Long gram-based indices are experimented at NTCIR-4 patent task. No morphological analyses are required to make gram-based indices. The ABJ and DEJ tag fields are extracted and indexed from NTCIR-4 patent corpus. Passages are extracted and indexed also. The total index size is 240Gbyte and time to make indices is about 86 hours. By merging the result of passage retrieval with the result of document retrieval by ABJ and DEJ field, we aimed at improvement in accuracy. Ranking algorithm used is based on a traditional probabilistic model. We also tried to set essential words in a query.