Interactive Post-Editing for Verbosity Controlled Translation

Proceedings of the 29th International Conference on Computational Linguistics, pages 5119–5128 October 12–17, 2022. 5119 Interactive Post-Editing for Verbosity Controlled Translation Prabhakar Gupta1, Anil Nelakanti1, Grant M. Berry2, and Abhishek Sharma1 1Amazon Prime Video 1{prabhgup,annelaka,naabhiss}@amazon.com 2Department of Spanish, Villanova University 2grant.berry@villanova.edu Abstract We explore Interactive Post-Editing (IPE) models for human-in-loop translation to help correct translation errors and rephrase it with a desired style variation. We specifically study verbosity for style variations and build on top of multi-source transformers that can read source and hypothesis to improve the latter with user inputs. Token-level interaction in- puts for error corrections and length interac- tion inputs for verbosity control are used by the model to generate a suitable translation. We report BERTScore to evaluate semantic quality with other relevant metrics for trans- lations from English to German, French and Spanish languages. Our model achieves supe- rior BERTScore over state-of-the-art machine translation models while maintaining the de- sired token-level and verbosity preference. 1 Introduction Recent machine translation (MT) mod- els (Sutskever et al., 2014; Bahdanau et al., 2015) have shown to excel with aspects of translation quality like adequacy and fluency but these models still suffer notable shortcomings like out-of-domain data, low-resource languages, rare words and longer sentences (Koehn and Knowles, 2017). Hence, MT systems are often supplemented by human translators for editing and correcting MT outputs to achieve the desired quality bar for various use-cases (Peris et al., 2017). Broadly, MT employing human input can be classified as one of two types: manual post-editing (MPE) and interactive post-editing (IPE) (Escribe and Mitkov, 2021). MPE relies on humans to make all necessary edits on top of MT output to deliver the final translation. Whereas, IPE uses a human-in-the-loop approach: the model offers human translators various cues like auto-complete suggestions, word look-ups, etc until the human arrives at a translation. Both approaches have their trade-offs and it is observed that while MPE is slower, it delivers a higher quality output relative to the faster IPE (Green et al., 2014). Our work is a human-in-the-loop model that aims to make efficient use of human effort in delivering improved translations. Translation quality, largely encompassing ade- quacy and fluency, has been the primary focus of most MT studies. Some recent work has explored other aspects of MT models like translation diver- sity, word choice (rare vs frequent words), transla- tion style, etc (Wang et al., 2021; Niu and Carpuat, 2020; Agrawal and Carpuat, 2019; Marchisio et al., 2019; Lakew et al., 2019; Niu et al., 2018; Yam- agishi et al., 2016). This style-aware modelling further broadens the scope and usability of MT, even more so when users can control levers to achieve the desired style variation. Our work iden- tifies one of the most important style features – verbosity or translation length. Controlling length of translation output has been studied before for NMT models (Lakew et al., 2019) but to the best of our knowledge our work is the first to propose a solution for controlling length of …

Content truncated to 500 words