Neural Models for Measuring Confidence on Interactive Machine Translation Systems

Reducing the human effort performed with use of interactive-predictive neural machine translation (IPNMT) systems is one main goals in this sub-field (MT). Prior works have focused on changing human–machine interaction method and simplifying feedback performed. Applying confidence measures (CM) to an IPNMT system helps decrease number words that user has check through session, reducing needed, although supposes losing a few points quality translations. The reduction comes from decreasing translator review—it only ones score lower than threshold set. In paper, we studied performance four based most used metrics MT. We trained recurrent network (RNN) models approximate scores metrics: Bleu, Meteor, Chr-f, TER. experiments, simulated obtain compare translations generated reduction. also between them see which obtains best results. results achieved showed 48% Bleu 70 points—a significant almost perfect.