Composition and Decomposition of Japanese Katakana and Kanji Morphemes for Decision Rule Induction from Patent Documents

We propose a new method to construct a word list for rule induction from Japanese patent documents. For word segmentation in Japanese, statistical morphological analyzers have been used in many applications. However, the output of these morphological analyzers presents defects when analyzing unknown words, specifically words that contain Kanji/Katakana morphemes. Some words are overly segmented, and their original meanings are obscured. Furthermore, boundaries between compound nouns are uncertain, which impedes investigation in the initial stages of the application. In our method, we first perform morphological analysis to segment sentences into morphemes. Second, segmented compound words are filtered by character types and Katakana/Kanji morphemes in the compound words are concatenated. Third, the concatenated morphemes are truncated to reduce verbosity. Then, words comprising Katakana/Kanji are retained for use in a word list for rule induction. The experiment results show that our method is effective for extracting decision rules for patent classification.