Adapting Neural Machine Translation for Automatic Post-Editing

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 315–319 November 10–11, 2021. ©2021 Association for Computational Linguistics 315 Adapting Neural Machine Translation for Automatic Post-Editing Abhishek Sharma∗ Prabhakar Gupta Anil Nelakanti Amazon Prime Video {naabhiss, prabhgup, annelaka}@amazon.com Abstract Automatic post-editing (APE) models are used to correct machine translation (MT) system outputs by learning from human post-editing patterns. We present the system used in our submission to the WMT’21 Automatic Post- Editing (APE) English-German (En-De) shared task. We leverage the state-of-the-art MT sys- tem (Ng et al., 2019) for this task. For further improvements, we adapt the MT model to the task domain by using WikiMatrix (Schwenk et al., 2021) followed by fine-tuning with ad- ditional APE samples from previous editions of the shared task (WMT-16,17,18) and ensem- bling the models. Our systems beat the baseline on TER scores on the WMT’21 test set. 1 Introduction Automatic Post-Editing (APE) is the task of au- tomatically correcting machine translation (MT) outputs. Along with fixing systematic errors in MT outputs, APE models can adapt general pur- pose MT systems to new domains and provide bet- ter translations to reduce the human post-editing effort (Chatterjee et al., 2015). APE has seen significant progress with Transformer based mod- els (Yang et al., 2020; Lopes et al., 2019; Chatter- jee et al., 2019, 2020) dominating the landscape as opposed to the earlier Statistical Machine Trans- lation (SMT) based models (Simard et al., 2007; Béchara et al., 2012) and RNN based sequence-to- sequence models (Junczys-Dowmunt and Grund- kiewicz, 2017). To track this progress, WMT has been conducting APE shared tasks since 2015 on different data domains and language pairs (Bojar et al., 2015, 2016, 2017; Chatterjee et al., 2018, 2019, 2020). WMT 2021’s shared task focused on English- German and English-Chinese language pairs. We participated in the English-German sub-task and describe our submission in this paper. Participants ∗Work done as intern at Amazon Prime Video were provided a training set with 7000 instances and a development set with 1000 instances. Each dataset consisted of source, machine-translation, post-edit triplets. The source sentences came from the English Wikipedia, the MT outputs were gener- ated with a black-box state-of-the-art MT system and the post-edits were created by professional translators correcting MT outputs. The test set con- sisted of 1000 pairs of source and MT outputs for which the participants had to submit the post-edits generated by their systems. The task organisers pro- vided two additional synthetic post-editing datasets – ‘artificial training data’ (Junczys-Dowmunt and Grundkiewicz, 2016) and ‘eSCAPE corpus’ (Negri et al., 2018) and permitted using additional data to train the model. TER scores (Snover et al., 2006) and BLEU (Papineni et al., 2002) scores were used as primary and secondary evaluation metrics re- spectively. Last year’s entries primarily focused on transfer learning (Yang et al., 2020; Lee, 2020; Wang et al., 2020) and novel data augmentation techniques (Lee et al., 2020b,a; Wang et al., 2020). The winning submission (Yang et al., 2020) was based on fine- tuning …

Content truncated to 500 words