Dissecting Human Pre-Editing toward Better Use of Off-the-Shelf Machine Translation Systems

Machine translation (MT) systems are not able to always produce translations of human-level quality. As a practical means of such MT systems, we investigated the potential of pre-editing strategy, by collecting actual pre-edit instances using a human-in-the-loop protocol. In our study, targeting Japanese-to-English translation on four different datasets and using an offthe-shelf MT system, we collected a total of 12,287 pre-edit instances for 400 source sentences and showed promising results; more than 85% of source sentences turned out to be accurately translated by the MT system. We also found that the pre-edited Japanese source sentences were better translated into Chinese and Korean, confirming the usefulness of pre-editing strategy in a multilingual setting. Through decomposing the collected pre-edit instances, we built a typology of primitive edit operations comprising 53 types, which unveils the subjects for further research.