1994 年 9 巻 4 号 p. 569-579
This paper describes the design of a set of 6200 Japanese-English sentence pairs for testing Japanese to English MT systems. Japanese expressions are organized into 600 test items with consideration for the characteristics of both Japanese and English, an average of 10 sentence pairs were then made for each item. In machine translation, translations between two very different languages (e.g. Japanese and English) are more difficult than between two similiar languages (e.g. Japanese and Korean). This is believed to be due to differences in morphology and in how things are conceived in different language groups. Therefore, we have focused our attention beyond morphological differences to cover differences in perception and concepts. The process of constructing Japanese texts is examined at 5 levels (part of speech, phrase, expression, sentence and text). Based on these 5 levels and considering the differences between the two languages, about 600 test items were chosen to test the basic functions of Japanese to English machine translation systems. Finally examples of these test items were extracted from various documents and publications and combined with specially constructed sentences to make the test set of 6200 sentence pairs.