Modeling Bilingual Conversational Characteristics for Neural Chat Translation

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 5711–5724 August 1–6, 2021. ©2021 Association for Computational Linguistics 5711 Modeling Bilingual Conversational Characteristics for Neural Chat Translation Yunlong Liang1∗, Fandong Meng2, Yufeng Chen1, Jinan Xu1† and Jie Zhou2 1Beijing Key Lab of Traffic Data Analysis and Mining, Beijing Jiaotong University, Beijing, China 2Pattern Recognition Center, WeChat AI, Tencent Inc, China {yunlongliang,chenyf,jaxu}@bjtu.edu.cn {fandongmeng,withtomzhou}@tencent.com Abstract Neural chat translation aims to translate bilin- gual conversational text, which has a broad ap- plication in international exchanges and coop- eration. Despite the impressive performance of sentence-level and context-aware Neural Machine Translation (NMT), there still remain challenges to translate bilingual conversational text due to its inherent characteristics such as role preference, dialogue coherence, and trans- lation consistency. In this paper, we aim to pro- mote the translation quality of conversational text by modeling the above properties. Specif- ically, we design three latent variational mod- ules to learn the distributions of bilingual con- versational characteristics. Through sampling from these learned distributions, the latent vari- ables, tailored for role preference, dialogue co- herence, and translation consistency, are incor- porated into the NMT model for better transla- tion. We evaluate our approach on the bench- mark dataset BConTrasT (English⇔German) and a self-collected bilingual dialogue cor- pus, named BMELD (English⇔Chinese). Ex- tensive experiments show that our approach notably boosts the performance over strong baselines by a large margin and significantly surpasses some state-of-the-art context-aware NMT models in terms of BLEU and TER. Ad- ditionally, we make the BMELD dataset pub- licly available for the research community.1 1 Introduction A conversation may involve participants that speak in different languages (e.g., one speaking in En- glish and another in Chinese). Fig. 1 shows an example, where the English role R1 and the Chi- nese role R2 are talking about the “boat”. The ∗Work was done when Yunlong Liang was interning at Pattern Recognition Center, WeChat AI, Tencent Inc, China. †Jinan Xu is the corresponding author. 1Code and data are publicly available at: https:// github.com/XL2248/CPCC Y2: nǐ huì jiàshǐ fānchuán？ X1: You know, Joey, I could teach you to sail, if you want.? Y3: duìa, wǒ zhè bèizǐ dōu zài jiàchuán, wǒ shíwǔ suì shí, wǒ bà song wǒ yì sōu chuán。 Y4: nǐ yǒu yì sōu fānchuán？ X2: You could? X3: Yeah! I've been sailing my whole life. When I was fifteen, my dad bought me my own boat. X4: Your own boat? X5: What? What? He was trying to cheer me up! My pony was sick. Ref: zěnme? bù xìn? tā song wǒ yì sōu chuán lái ānwèi wǒ, wǒ de xiǎomǎ bìng le R1 R2 R2 R1 R1 Y5: S-NMT: shénme? shénme? tā xiǎng rang wǒ gāoxìng qǐlái,wǒ de xiǎomǎ bìng le。 。 Y1: qiáoyī, rúguǒ nǐ xiǎng, wǒ kěyǐ jiào nǐ jiàchuán。？ Figure 1: An ongoing bilingual conversation example (English⇔Chinese), where the Chinese utterances are presented in pinyin style. Ri: Role i. The dashed ar- …

Content truncated to 500 words