Controlling Translation Formality Using Pre-trained Multilingual Language Models

Proceedings of the 19th International Conference on Spoken Language Translation (IWSLT 2022), pages 327 - 340 May 26-27, 2022 c⃝2022 Association for Computational Linguistics Controlling Translation Formality Using Pre-trained Multilingual Language Models Elijah Rippeth⇤and Sweta Agrawal* and Marine Carpuat Department of Computer Science University of Maryland {erip,sweagraw,marine}@cs.umd.edu Abstract This paper describes the University of Mary- land’s submission to the Special Task on For- mality Control for Spoken Language Transla- tion at IWSLT, which evaluates translation from English into 6 languages with diverse grammatical formality markers. We investigate to what extent this problem can be addressed with a single multilingual model, simultane- ously controlling its output for target language and formality. Results show that this strategy can approach the translation quality and formal- ity control achieved by dedicated translation models. However, the nature of the underlying pre-trained language model and of the finetun- ing samples greatly impact results. 1 Introduction While machine translation (MT) research has pri- marily focused on preserving meaning across lan- guages, linguists and lay-users alike have long known that pragmatic-preserving communication is an important aspect of the problem (Hovy, 1987). To address one dimension of this, several works have attempted to control aspects of formality in MT (Sennrich et al., 2016; Feely et al., 2019; Schioppa et al., 2021). Indeed, this research area was formalized as formality-sensitive machine translation (FSMT) by Niu et al. (2017), where the translation is not only a function of the source segment but also the desired target formality. The lack of gold translation with alternate formality for supervised training and evaluation has lead re- searchers to rely on manual evaluation and syn- thetic supervision in past work (Niu and Carpuat, 2020). Additionally, these works broadly adapt to formal and informal registers as opposed to specifi- cally controlling grammatical formality. The Special Task on Formality Control on Spo- ken Language Translation provides a new bench- mark by contributing high-quality training datasets ⇤ equal contribution. Source: Do you like1 Legos? did you2 ever play with them as a child or even later? German Informal: Magst du1 Legos? Hast du2 jemals als Kind mit ihnen gespielt oder sogar später? German Formal: Mögen Sie1 Legos? Haben Sie2 jemals als Kind mit ihnen gespielt oder sogar später? Table 1: Contrastive formal and informal translations into German. Grammatical formality markers are bolded and aligned via indices. for diverse languages (N˘adejde et al., 2022). In this task, a source segment in English is paired with two references which are minimally contrastive in gram- matical formality, one for each formality level (for- mal and informal; Table 1). Training and test sam- ples are provided in the domains of “telephony data” and “topical chat” (Gopalakrishnan et al., 2019) for four language pairs (English-{German (DE), Span- ish (ES), Hindi (HI), Japanese(JA)}) and a test dataset for two additional “zero-shot” (ZS) lan- guage pairs (EN-{Russian (RU), Italian (IT)}). Markers of grammatical formality vary across these languages. Personal pronouns and verb agreement mark formality in many Indo-European languages (e.g., DE, HI, IT, RU, ES), …

Content truncated to 500 words