Estimating Grammar Correctness for a Priori Estimation of Machine Translation Post-Editing Effort
Nicholas Hubert Kirk, Guchun Zhang, Georg Groh · 2014
We present a supervised learning pilot application for estimating Machine Translation (MT) output reusability, in view of supporting a human post-editor of MT content.We train our model on typed dependencies (labeled grammar relationships) extracted from human reference and raw MT data, to then predict grammar relationship correctness values that we aggregate to provide a binary segmentlevel evaluation.In view of scaling up to larger data, we provide implemented Naïve Bayes and Stochastic Gradient Descent with Support Vector Machine loss function approaches and their evaluation, and verify the correlation of predicted values with human judgement.