Multi-source Transformer for AutomaticPost-Editing of Machine Translation Output

Automatic post-editing (APE) of machine translation (MT) is the task of automatically fixing errors in a machine-translated text by learning from human corrections. Recent APE approaches have shown that best results are obtained by neural multi-source models that correct the raw MT output by also considering information from the corresponding source sentence. In this paper, we pursue this objective by exploiting Transformer (Vaswani et al. 2017), the state-of-the-art architecture in MT. Our approach presents several advantages over previous APEsolutions, both from the performance perspective and from an industrial deployment standpoint. Indeed, besides competitive results, our Transformer-based architecture is faster to train (thanks to parallelization) and easier to maintain (thanks to the reliance on a single model rather thana complex, multi-component architecture). These advantages make our approach particularly appealing for the industrial sector, where scalability and cost-efficiency are important factors, complementary to pure performance. Besides introducing a novel architecture, we also validate the common assumption that training neural APE systems with more data always results in stronger models. Along this direction, we show that this assumption does not always hold, and that fine-tuning the system only on small in-domain data can yield higher performance. Furthermore, we try different strategies to better exploit the in-domain data. In particular, we adapt reinforcement learning (RL) techniques to optimize our models by considering task-specific metrics (i.e. BLEU and TER) in addition to maximum likelihood. Our experiments show that,alone, the multi-source approach achieves slight improvements over a competitive APE system based on a recurrent neural network architecture. Further gains are obtained by the full-fledged system, fine-tuned on in-domain data and enhanced with RL optimization techniques. Our best results (with a single multi-source model) significantly improve the performance of the best (and much more complex) system submitted to the WMT 2017 APE shared task.