In order to improve the quality of machine translations, knowledge about words must be systematically collected. This paper describes (1) three semantic attribute systems (3,000 categories) for Japanese-to-English machine translation, (2) a semantic word dictionary (400,000 words) based on the semantic attribute systems, and (3) a semantic valency dictionary (16,000 patterns) based on the semantic attribute systems and the semantic word dictionary. It describes a Japanese-to-English transfer process which has four levels depending on abstraction and some active processes for element transfer.