Abstract
User dictionaries are important for practical machine translation. However it is difficult for users to enter the detailed semantic attributes that a system may require. This paper proposes a method of automatically determining semantic attributes for noun pairs entered by users. The method compares the Japanese and English words with words already in the system dictionary. When this method was applied to words in two user dictionaries (for newspaper articles and software manuals) it generated more attributes than trained humans did, with a recall rate of 40-80%. Evaluation showed that translations using the machine-made dictionary were similar in quality to translations using the human-made dictionary. Thus automatic determination of semantic attributes removes the need for highly trained lexographers to make user dictionaries.
