Computational Linguistics of Automatic Post-Editing

Abstract

An automatic post-editing (APE) system is an automatic system that proofreads machine translation (MT) results as human post-editors do. It is generally more practical and cost-effective compared to a domain-specific MT system, particularly in the perspective of system development. In the last decade, researchers have made great progress in improving the effectiveness of existing APE systems. However, as time passed, it was found that the effectiveness of an APE system is significantly affected by some external factors such as the target language pair, the target domain, and the quality of given MT results. Until now, many studies have explored solutions to this problem, but not one approach has attained a remarkable success yet. This dissertation classifies such efforts into three categories: utilizing the knowledge of pre-trained artificial neural networks, using various kinds of synthetic training data, and making alterations to the typical system design. In summary, utilizing the knowledge of pre-trained artificial neural networks requires further studies, particularly in the direction of tuning them with delicacy, because no meaningful progress has been made since 2019, when a pioneering work, which was later shown to be only effective only in certain situations, was published. Next, using various kinds of synthetic training data seems helpful at first glance, but it seems that the ultimate effectiveness relies on one specific synthetic training data set, and the blending of different kinds of synthetic training data should be very delicate. In contrast, not only has the effectiveness of making alterations to the typical system design been verified through controlled experiments, it also does not require such a delicate touch. In addition, small but meaningful progress has been made in this direction recently. Thus, this dissertation concludes that making alterations to the typical system design is the most promising approach at the moment.