Confidence estimation for spoken language translation based on Round Trip Translation

In this paper we propose a Round Trip Translation (RTT) based approach to sentence-level confidence estimation (CE) for spoken language translation without the assistant of reference translations generated by human. A number of novel RTT based features are introduced to reflect the quality of spoken language translation in more detail. After combing various kinds of features together, support vector regression (SVR) method is employed to learn human´s assessment patterns of translation quality. Experimental results show that RTT based features could improve the accuracy of CE significantly and SVR method could model human´s assessment pattern accurately and robustly. In the final CE task of spoken language translation from Chinese to English, our system achieves comparable performance with that of BLEU, which needs the assistance of human´s reference, even with small training data.