Design and Analysis of a Large Corpus of Post-Edited Translations: Quality Estimation, Failure Analysis and the Variability of Post-Edition

Machine Translation (MT) is now often used to produce approximate translations that are then corrected by trained professional post-editors. As a result, more and more datasets of post-edited translations are being collected. These datasets are very useful for training, adapting or testing existing MT systems. In this work, we present the design and content of one such corpus of post-edited translations, and consider less studied possible uses of these data, notably the development of an automatic Quality Estimation (QE) system and the detection of frequent errors in automatic translations. Both applications require a careful assessment of the variability in post-editions, that we study here.