Collocations as Word Co-ocurrence Restriction Data - An Application to Japanese Word Processor -
Kosho Shudo, Masahito Takahashi, Yasuo Koyama, Kenji Yoshimura · 2000
A bstractCol l ocat i ons, t he com bi nat i on of speci fi c w ords are qui t e useful l i ngui sti c resources for N LP i n general .The purpose of t hi s paper i s t o show t hei r useful ness, exem pl i fyi ng an appl i cat i on t o K anj i charact er deci sion processes for Japanese w ord processors.U nl i ke recent t ri al s of aut om at i c ext ract i on, our col l ocat i ons w ere col l ect ed m anual l y t hrough m any years of i nt ensi ve i nvest i gat i on of corpus.O ur col l ect i on procedure consi st s of (1) fi ndi ng a proper com bi nat i on of w ords i n a corpus and (2) recol l ect i ng sim i l ar com bi nat i ons of w ords, inci t ed by i t .Thi s procedure, w hi ch depends on hum an j udgm ent and t he enri chm ent of dat a by associat i on, i s effect i ve for rem edyi ng t he sparseness of dat a probl em , al t hough t he arbi t rari ness of hum an j udgm ent is inevi t abl e.A pproxi m at el y sevent y t w o t housand and four hundred col l ocat i ons w ere used as w ord co-occurrence rest ri ct i on dat a for deci di ng K anj i charact ers in t he processing of Japanese w ord processores.Experi m ent s have show n t hat the col l ocat i on dat a yi el d 8. 9% hi gher fract i on of K ana-t o-K anj i charact er conversion accuracy t han t he syst em w hi ch uses no col l ocat i on dat a and 7. 0% hi gher, t han a com m erci al w ord processor soft w are of average perform ance.