Confidence estimation for machine translation using context vectors
Salehi Marzieh, Shahram Khadivi, Noushin Riahi · 2014
Machine translation (MT) has been developed and has achieved wide successes over last years. But this technology is still not able to deliver high quality translation and therefore post-editing is needed. Since post-editing could be time consuming even more than the translation process, having a quality estimation of the translated parts can be very useful. It means we need to estimate the confidence of the output without having any references. Moreover, Confidence Estimation (CE) can be useful for some applications that their goal is to improve machine translation quality such as system combination, regenerating and pruning. But there is not yet any completely satisfactory method for CE task. We propose context vector-based features that are never used for CE task. We classify MT output at word level. We show that each proposed feature outperforms the baseline systems. The combination of proposed features outperforms the best baseline system 5.68% relative in CER, 3.88% relative in F-measure and 7.30% relative in negative class F-measure. Also combining proposed features with baseline features made noticeable improvement to the baseline systems.