This paper proposes a system for incremental English-Japanese spoken language translation, the system behaving like simultaneous interpretation. Since spontaneous speech appears continuously, each stage of a machine translation system for spoken languages should work incrementally. However, machine translation systems which have been proposed so far can not achieve high degree of incrementality because of the difference in word-order between the source language and the target one. In this paper, the system utilizing some ill-formed expressions characterizing Japanese spoken language such as repetitions, inversions, ellipses, repairs and hesitations, we have succeeded in the incremental transfer from English to Japanese. An experimental system on the basis of a chart parsing framework has been implemented. To evaluate the system we have made an experiment with 278 English dialogue sentences. 228 of them are translated correctly, providing a success rate of 82.0%, and 47msec on the average was taken to process one word. These results show our technique to be useful for spoken language translation with acceptable accuracy and high real-time nature.
J-STAGEがリニューアルされました! https://www.jstage.jst.go.jp/browse/-char/ja/