Controlling Formality in Low-Resource NMT with Domain Adaptation and Re-Ranking: SLT-CDT-UoS at IWSLT2022

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 341 - 350 May 26-27, 2022 c⃝2022 Association for Computational Linguistics Controlling Formality in Low-Resource NMT with Domain Adaptation and Re-Ranking: SLT-CDT-UoS at IWSLT2022 Sebastian T. Vincent, Loïc Barrault, Carolina Scarton Department of Computer Science, University of Sheffield Regent Court, 211 Portobello, Sheffield, S1 4DP, UK {stvincent1, l.barrault, c.scarton}@shef.ac.uk Abstract This paper describes the SLT-CDT-UoS group’s submission to the first Special Task on Formal- ity Control for Spoken Language Translation, part of the IWSLT 2022 Evaluation Campaign. Our efforts were split between two fronts: data engineering and altering the objective func- tion for best hypothesis selection. We used language-independent methods to extract for- mal and informal sentence pairs from the pro- vided corpora; using English as a pivot lan- guage, we propagated formality annotations to languages treated as zero-shot in the task; we also further improved formality controlling with a hypothesis re-ranking approach. On the test sets for English-to-German and English- to-Spanish, we achieved an average accuracy of .935 within the constrained setting and .995 within unconstrained setting. In a zero-shot setting for English-to-Russian and English-to- Italian, we scored average accuracy of .590 for constrained setting and .659 for unconstrained. 1 Introduction Formality-controlled machine translation enables the system user to specify the desired formality level at input so that the produced hypothesis is expressed in a formal or informal style. Due to discrepancies between different languages in for- mality expression, it is often the case that the same source sentence has several plausible hypotheses, each aimed at a different audience; leaving this choice to the model may result in an inappropriate translation. This paper describes our team’s submission to the first Special Task on Formality Control in SLT at IWSLT 2022 (Anastasopoulos et al., 2022), where the objective was to achieve control over bi- nary expression of formality in translation (enable the translation pipeline to generate formal or infor- mal translations depending on user input). The task evaluated translations from English (EN) into Ger- man (DE), Spanish (ES), Russian (RU), Italian (IT), Japanese (JA) and Hindi (HI). Among these, EN- {RU,IT} were considered zero-shot; for other pairs, small paired formality-annotated corpora were pro- vided. The task ran in two settings: constrained (limited data and pre-trained model resources) and unconstrained (no limitations on either resource). Submissions within both the constrained and un- constrained track were additionally considered in two categories: full supervision and zero-shot. Our submission consisted of four primary sys- tems, one for each track/subtrack combination, and we focused on the EN-{DE,ES,RU,IT} lan- guage directions. We were interested in lever- aging the provided formality-annotated triplets (src, tgtformal, tgtinformal) to extract sufficiently large annotated datasets from the permitted training corpora, without using language-specific resources or tools. We built a multilingual translation model in the given translation directions and fine-tuned it on our collected data. Our zero-shot submis- sions used fine-tuning data only for the non-zero- shot pairs. To boost the formality control (espe- cially within the constrained track), …

Content truncated to 500 words