Prompt-Driven Neural Machine Translation

Findings of the Association for Computational Linguistics: ACL 2022, pages 2579 - 2590 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Prompt-Driven Neural Machine Translation Yafu Li♠♡ , Yongjing Yin♠♡ , Jing Li♣ , Yue Zhang♡♢ ♠ Zhejiang University ♡ School of Engineering, Westlake University ♢ Institute of Advanced Technology, Westlake Institute for Advanced Study ♣ Sichuan Lan-bridge Information Technology Co., Ltd. yafuly@gmail.com yinyongjing@westlake.edu.cn judyli.5266@gmail.com yue.zhang@wias.org.cn Abstract Neural machine translation (NMT) has ob- tained significant performance improvement over the recent years. However, NMT models still face various challenges including fragility and lack of style flexibility. Moreover, current methods for instance-level constraints are lim- ited in that they are either constraint-specific or model-specific. To this end, we propose prompt-driven neural machine translation to incorporate prompts for enhancing translation control and enriching flexibility. Empirical results demonstrate the effectiveness of our method in both prompt responding and trans- lation quality. Through human evaluation, we further show the flexibility of prompt control and the efficiency in human-in-the-loop trans- lation. 1 Introduction Neural machine translation (NMT) has achieved much performance improvement over the recent years (Vaswani et al., 2017; Edunov et al., 2018; Hassan et al., 2018; Liu et al., 2020), yet still faces various challenges such as low cross-domain ro- bustness (Müller et al., 2020), fragility (Li et al., 2021) and lack of style flexibility (Li and Jurafsky, 2016; Shu et al., 2019). To address these issues, a line of work considers introducing constraints to the translation outputs, typically in the form of lexi- cal constraints (Song et al., 2019; Chen et al., 2020) and style control (Sennrich et al., 2016a; Michel and Neubig, 2018; Shu et al., 2019). For example, Song et al. (2019) ensure that polysemous words are translated to their domain-specific senses in eCommerce. Such instance-level constraint has been shown useful for improving both the translation adequacy and readability in practical applications (Song et al., 2019; Chen et al., 2020; Jwalapuram et al., 2020; Konieczny, 2021; Chen et al., 2021a). However, they are limited in being (1) model-specific and (2) Prompt: the translation should include “on the desk” Prompt: “���” should be translated before “�” Translation: The apple pie on the table was eaten by me yes -terday. Translation: Yesterday, I ate the apple pie on the desk. Prompt: “���” should be translated into “Apple Pie” Translation: Yesterday, I ate the Apple Pie on the table. Prompt: the translation should begin with “I” Translation: I ate the apple pie on the table yesterday. (a) (b) (c) (d) Figure 1: A Prompt-driven NMT model outputs differ- ent translations for the sentence “昨天, 我吃了桌上的 苹果派。” (English: Yesterday, I ate the apple pie on the table.) based on the given prompts. One can specify phrase translations, guarantee translation positions or alter word order by feeding the system with different prompts. constraint-specific. For instance, lexical constraints are typically integrated into a model by either modi- fying the decoding process (Hokamp and Liu, 2017; Post and Vilar, 2018; Chen et al., 2021a) or intro- ducing special …

Content truncated to 500 words