Empirical Analysis of Noising Scheme based Synthetic Data Generation for Automatic Post-editing

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 883–891 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 883 Empirical Analysis of Noising Scheme based Synthetic Data Generation for Automatic Post-editing Hyeonseok Moon1, Chanjun Park1,2, Seolhwa Lee3, Jaehyung Seo1, Jungseob Lee1, Sugyeong Eo1, Heuiseok Lim1,† 1Korea University Seoul, South Korea 2Upstage Gyeonggi-do, South Korea 3University of Copenhagen Copenhagen, Denmark 1{glee889, bcj1210, seojae777, omanma1928, djtnrud, limhseok}@korea.ac.kr 2chanjun.park@upstage.ai 3sele@di.ku.dk Abstract Automatic post-editing (APE) refers to a research field that aims to automatically correct errors included in the translation sentences derived by the machine translation system. This study has several limitations, considering the data acquisition, because there is no official dataset for most language pairs. Moreover, the amount of data is restricted even for language pairs in which official data has been released, such as WMT. To solve this problem and promote universal APE research regardless of APE data existence, this study proposes a method for automatically generating APE data based on a noising scheme from a parallel corpus. Particularly, we propose a human mimicking errors-based noising scheme that considers a practical correction process at the human level. We propose a precise inspection to attain high performance, and we derived the optimal noising schemes that show substantial effectiveness. Through these, we also demonstrate that depending on the type of noise, the noising scheme-based APE data generation may lead to inferior performance. In addition, we propose a dynamic noise injection strategy that enables the acquisition of a robust error correction capability and demonstrated its effectiveness by comparative analysis. This study enables obtaining a high performance APE model without human-generated data and can promote universal APE research for all language pairs targeting English. Keywords: Machine Translation, Automatic Post-Editing, Noise Injection, Data Generation 1. Introduction Automatic Post-editing (APE) aims at automatically correcting errors of translated sentences generated by machine translation systems (Chatterjee et al., 2020). The APE research field shows practical effectiveness as it can correct errors in translated sentence, and im- prove sentence quality regardless of the utilized ma- chine translation system. Currently, it is being actively studied at various conferences such as Conference on Machine Translation (WMT) (Chatterjee et al., 2018). With the development of Neural Machine Translation (NMT) technology, APE has effectively reduced the human effort required to correct errors and has im- proved the quality of the translation significantly. However, this study has a substantial limitation: time consumption and human resources that are required to generate the APE data (Negri et al., 2018). Generally, APE data comprises source sentences (SRC), its ma- chine translation results (MT), and human post-edited sentences (PE) of the translation results. Through these data, the APE model is trained to generate PE by feeding SRC and MT as input (Chatterjee et al., 2020). The most challenging part is the generation of † This author is the corresponding author. PE, which requires expert level human labor. More than simply translating a source sentence, deep in- spection about the machine-like errors …

Content truncated to 500 words