Ensembling Factored Neural Machine Translation Models for Automatic Post-Editing and Quality Estimation

摘要

This work presents a novel approach to Automatic Post-Editing (APE) andWord-Level Quality Estimation (QE) using ensembles of specialized NeuralMachine Translation (NMT) systems. Word-level features that have proveneffective for QE are included as input factors, expanding the representation ofthe original source and the machine translation hypothesis, which are used togenerate an automatically post-edited hypothesis. We train a suite of NMTmodels that use different input representations, but share the same outputspace. These models are then ensembled together, and tuned for both the APE andthe QE task. We thus attempt to connect the state-of-the-art approaches to APEand QE within a single framework. Our models achieve state-of-the-art resultsin both tasks, with the only difference in the tuning step which learns weightsfor each component of the ensemble.