Improving MT post-editing productivity with adaptive confidence estimation for document-specific translation model

We present an adaptive translation confidence estimation method to predict the human-targeted translation error rate for a document-specific machine translation model. We show the effectiveness of our method that uses features derived from the internal translation decoding process and from the source sentence analysis, in both classification and regression estimate of MT quality. By dynamically training the confidence model for the document-specific MT model, we are able to achieve consistency and prediction quality across multiple documents, demonstrated by the higher correlation coefficient and F scores in finding good sentences. Furthermore, the proposed method is applied to an English–Japanese MT post-editing field study. A strong correlation between our prediction and human selection is observed with a 10 % increase in the productivity of human translators.