IIT Bombay’s WMT22 Automatic Post-Editing Shared Task Submission

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 682 - 688 December 7-8, 2022 ©2022 Association for Computational Linguistics IIT Bombay’s WMT22 Automatic Post-Editing Shared Task Submission Sourabh Deoghare and Pushpak Bhattacharyya Computation for Indian Langauge Technology (CFILT) IIT Bombay, India {sourabhdeoghare, pb}@cse.iitb.ac.in Abstract This paper describes IIT Bombay’s submission to the WMT-22 Automatic Post-Editing (APE) shared task for the English-Marathi (En-Mr) language pair. We follow the curriculum train- ing strategy to train our APE system. First, we train an encoder-decoder model to perform translation from English to Marathi. Next, we add another encoder to the model and train the resulting dual-encoder single-decoder model for the APE task. This involves training the model using the synthetic APE data in multi- ple training stages and then fine-tuning it using the real APE data. We use the LaBSE tech- nique to ensure the quality of the synthetic APE data. For data augmentation, along with using candidates obtained from an external ma- chine translation (MT) system, we also use the phrase-level APE triplets generated using phrase table injection. As APE systems are prone to the problem of ‘over-correction’, we use a sentence-level quality estimation (QE) system to select the final output between an original translation and the corresponding out- put generated by the APE model. Our approach improves the TER and BLEU scores on the de- velopment set by -3.92 and +4.36 points, re- spectively. Also, the final results on the test set show that our APE system outperforms the baseline system by -3.49 TER points and +5.37 BLEU points. 1 Introduction Automatic Post-Editing (APE) is a post-processing task in a Machine Translation (MT) workflow. It aims to automatically identify and correct errors in MT outputs (Chatterjee et al., 2020). Läubli et al. (2013) and Pal et al. (2016) show that APE systems have the potential to reduce human effort by auto- matically correcting repetitive translation errors. The initial years of the WMT APE shared task focused on correcting errors in Statistical Machine Translation (SMT) translations, where participants explored various statistical and neural APE ap- proaches (Bojar et al., 2017). Although neural APE approaches showed high potential for significantly improving the quality of SMT translations, these approaches faced challenges in improving transla- tions obtained from relatively-more-robust neural machine translation (NMT) systems (Chatterjee et al., 2018). A possible reason for this could be that correcting a high-quality translation requires fewer edits, and therefore APE approaches need to be precise in identifying and in correcting the errors. Also, the neural APE approaches use large neural networks that require significant training data. APE training data consists of ‘triplets’ in the form of source sentence (src), its translation gener- ated using an MT system (mt), and a human post- edited version of the translation (pe). Obtaining pe is an expensive task in terms of time and money; therefore, there is a lack of large APE datasets. To deal with this problem, various data augmen- tation techniques have been proposed (Junczys- Dowmunt and Grundkiewicz, 2016; Negri …

Content truncated to 500 words