Improving Formality-Sensitive Machine Translation Using Data-Centric Approaches and Prompt Engineering

Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 420–432 July 13-14, 2023 c⃝2023 Association for Computational Linguistics Improving Formality-Sensitive Machine Translation using Data-Centric Approaches and Prompt Engineering Seungjun Lee1, Hyeonseok Moon1, Chanjun Park1,2, Heuiseok Lim1∗ 1Korea University, South Korea 2Upstage, South Korea {dzzy6505, glee889, bcj1210, limhseok}@korea.ac.kr chanjun.park@upstage.ai Abstract In this paper, we present the KU x Up- stage team’s submission for the Special Task on Formality Control on Spoken Language Translation, which involves translating En- glish into four languages with diverse gram- matical formality markers. Our methodology comprises two primary components: 1) a language-specific data-driven approach, and 2) the generation of synthetic data through the employment of large-scale language mod- els and empirically-grounded prompt engineer- ing. By adapting methodologies and models to accommodate the unique linguistic prop- erties of each language, we observe a no- table enhancement in performance relative to the baseline, substantiating the heightened effi- cacy of data-driven approaches. Moreover, our devised prompt engineering strategy yields su- perior synthetic translation instances. 1 Introduction Neural machine translation (NMT) models have achieved remarkable progress in recent years, as evidenced by their high BLEU scores (Britz et al., 2017; Stahlberg, 2020). Nonetheless, these models generally rely on generic parallel corpora and as- sume a single target translation for a given source sentence, often overlooking the significance of style and pragmatic aspects in translation, such as formality or politeness (Li et al., 2022). To address this issue, formality-sensitive machine translation (FSMT) has emerged as a research area, aiming to control grammatical formality in translated text across languages (Niu et al., 2017). The Special Task on Formality Control on Spo- ken Language Translation introduces a new bench- mark with high-quality training datasets for di- verse languages, encompassing both supervised and zero-shot language pairs. Despite these new datasets (N˘adejde et al., 2022), controlling formal- ity in MT remains a challenging problem due to the Source: It did, many people liked his show so yeah, do you like Chris Pratt? Korean Formal: 그랬어요, 많은 사람들이 그의 쇼를 좋아했죠. 그래서 당신 크리스 프랫 좋아해요? Korean Informal: 그랬어, 많은 사람들이 그의 쇼를 좋아했지. 그래서 너 크리스 프랫 좋아해? Table 1: Contrastive translations in formal and informal styles into Korean are presented. Grammatical formal- ity markers, which are bolded, can be aligned through colors. absence of gold translations with alternate formal- ity levels and the extensive variation in grammatical formality markers across languages. In the 2023 shared task, an English source seg- ment is paired with two references that are mini- mally contrastive in grammatical formality, repre- senting both formal and informal levels as shown in Table 1. Training and test samples are provided in the domains of “telephony data” and “topical chat” (Gopalakrishnan et al., 2019) for two super- vised language pairs, English-Korean (EN-KO) and English-Vietnamese (EN-VI), and two zero-shot language pairs, English-Portuguese (EN-PT) and English-Russian (EN-RU). Grammatical formality markers differ across these languages. Personal pro- nouns and verb agreement signal formality in many Indo-European languages (e.g., PT, RU), while in Korean, …

Content truncated to 500 words