Noising Scheme for Data Augmentation in Automatic Post-Editing

Proceedings of the 5th Conference on Machine Translation (WMT), pages 783–788 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 783 Noising Scheme for Data Augmentation in Automatic Post-Editing WonKee Lee1, Jaehun Shin1, Baikjin Jung1, Jihyung Lee1, Jong-Hyeok Lee1,2 1Departmet of Computer Science and Engineering, 2Graduate School of Artificial Intelligence, Pohang University of Science and Technology (POSTECH), Republic of Korea {wklee, jaehun.shin, bjjung, jihyung.lee, jhlee}@postech.ac.kr Abstract This paper describes POSTECH’s submission to WMT20 for the shared task on Automatic Post-Editing (APE). Our focus is on increasing the quantity of available APE data to overcome the shortage of human-crafted training data. In our experiment, we implemented a nois- ing module that simulates four types of post- editing errors, and we introduced this module into a Transformer-based multi-source APE model. Our noising module implants errors into texts on the target side of parallel cor- pora during the training phase to make syn- thetic MT outputs, increasing the entire num- ber of training samples. We also generated additional training data using the parallel cor- pora and NMT model that were released for the Quality Estimation task, and we used these data to train our APE model. Experimental results on the WMT20 English-German APE data set show improvements over the baseline in terms of both the TER and BLEU scores: our primary submission achieved an improve- ment of -3.15 TER and +4.01 BLEU, and our contrastive submission achieved an improve- ment of -3.34 TER and +4.30 BLEU. 1 Introduction There has been a surge of interest in developing Automatic Post-Editing (APE) models, which is ca- pable of automatically correcting errors produced by a machine-translation (MT) system, and thus is an attractive way to improve the quality of the MT output. Currently, sequence-to-sequence modeling has become a dominant approach to constructing APE models (Chatterjee et al., 2019, 2018), which requires a large quantity of training samples. How- ever, APE data1 — comprising triplets of three texts: source (src), a machine-translation (mt) of src, and a human-crafted post-edited sentence (pe) of mt — is too small and costly to acquire. Con- sequently, the lack of APE data becomes a great 1http://www.statmt.org/wmt20/ape-task.html obstacle to a satisfactory performance of sequence- to-sequence models. To reduce such data scarcity, there have been several attempts at constructing synthetic APE data (Negri et al., 2018; Junczys-Dowmunt and Grundkiewicz, 2016). Most notably, Negri et al. (2018) proposed a simple but effective way to con- struct a large-scale synthetic APE data set eSCAPE (src, mt, ref), of which src and ref is the source and target text of freely available parallel corpora, respectively, and mt is a translation of src produced by the MT system that had been trained on those parallel corpora. As eSCAPE has shown to be beneficial in train- ing APE models (Chatterjee et al., 2019), it has become feasible to train deep APE models and also what most recent works have been relying on so far. Nevertheless, the availability of a limited quantity of parallel corpora may not only be …

Content truncated to 500 words