Collocations as Word Co-ocurrence Restriction Data - An Application to Japanese Word Processor

Collocations, the combination of specific words are quite useful linguistic resources for NLP in general. The purpose of this paper is to show their usefulness, exemplifying an application to Kanji character decision processes for Japanese word processors. Unlike recent trials of automatic extraction, our collocations were collected manually through many years of intensive investigation of corpus. Our collection procedure consists of (1) finding a proper combination of words in a corpus and (2) recollecting similar combinations of words, incited by it. This procedure, which depends on human judgment and the enrichment of data by association, is effective for remedying the sparseness of data problem, although the arbitrariness of human judgment is inevitable. Approximately seventy two thousand and four hundred collocations were used as word co-occurrence restriction data for deciding Kanji characters in the processing of Japanese word processores. Experiments have shown that the collocation data yield 8.9% higher fraction of Kana-toKanji character conversion accuracy than the system which uses no collocation data and 7.0% higher, than a commercial word processor software of average performance.