Incorporating Terminology Constraints in Automatic Post-Editing

Proceedings of the 5th Conference on Machine Translation (WMT), pages 1193–1204 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 1193 Incorporating Terminology Constraints in Automatic Post-Editing David Wan1, Chris Kedzie1, Faisal Ladhak1, Marine Carpuat2 and Kathleen McKeown1 1 Columbia University, 2 University of Maryland dw2735@columbia.edu, {kedzie,faisal,kathy}@cs.columbia.edu, marine@cs.umd.edu Abstract Users of machine translation (MT) may want to ensure the use of specific lexical terminolo- gies. While there exist techniques for incor- porating terminology constraints during infer- ence for MT, current APE approaches can- not ensure that they will appear in the final translation. In this paper, we present both autoregressive and non-autoregressive models for lexically constrained APE, demonstrating that our approach enables preservation of 95% of the terminologies and also improves trans- lation quality on English-German benchmarks. Even when applied to lexically constrained MT output, our approach is able to improve preservation of the terminologies. However, we show that our models do not learn to copy constraints systematically and suggest a sim- ple data augmentation technique that leads to improved performance and robustness. 1 Introduction Automatic post-editing (APE) aims to improve the quality of the output of an arbitrary machine trans- lation (MT) system by pruning systematic errors and adapting to a domain-specific style and vocab- ulary (Simard et al., 2007; Chatterjee et al., 2018). Although previous work has shown the usefulness of APE to prune errors by focusing on improving the translation error rate (TER), few have studied the effect of incorporating lexical constraints. There are several use cases where such a sys- tem would be beneficial. For example, content providers meticulously curate lists of terminolo- gies for their domains that indicate preferred trans- lations for technical terms. Lexically constrained APE would also be useful for cross-lingual infor- mation retrieval. When displaying snippets from retrieved documents, the query term should appear in the translation output (if it does in the source) as it can make relevance clear to the end user. Here, the query serves as the term. While recent approaches allow inference time adaptation of NMT systems using these terminolo- gies (Dinu et al., 2019; Post and Vilar, 2018), post- editing translations with a generic APE system may lead to dropped terms. A constraint-aware APE sys- tem would allow to fix systematic translation errors, while keeping the terminologies intact. Inspired by Dinu et al. (2019), we consider a range of representations which augment input se- quences with constraint tokens and factors for use in an autoregressive Transformer (AT) APE model. Using this approach, the constraints are explicitly represented in the encoder input sequence, and the model learns to prefer translations that contain the supplied terminologies during decoding. We also explore the use of the Levenshtein Transformer (LevT) (Gu et al., 2019), a non-autoregressive Transformer (NAT) model. The LevT model ap- plies neatly to the APE task since the decoder can be initialized with an incomplete sequence to be refined. Additionally, multiple corrections can be made simultaneously, yielding a decoding speedup over autoregressive models. We then show that constrained …

Content truncated to 500 words