Modeling Bilingual Conversational Characteristics for Neural Chat Translation

TL;DR

This paper enhances neural chat translation by modeling conversational properties with latent variables, significantly improving translation quality on bilingual dialogue datasets.

Contribution

Introduces three latent variational modules to capture role preference, dialogue coherence, and translation consistency in neural chat translation.

Findings

Significant performance boost over strong baselines

Outperforms state-of-the-art context-aware NMT models

Provides a new bilingual dialogue dataset BMELD

Abstract

Neural chat translation aims to translate bilingual conversational text, which has a broad application in international exchanges and cooperation. 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 translation consistency. In this paper, we aim to promote the translation quality of conversational text by modeling the above properties. Specifically, we design three latent variational modules to learn the distributions of bilingual conversational characteristics. Through sampling from these learned distributions, the latent variables, tailored for role preference, dialogue coherence, and translation consistency, are incorporated into the NMT model for better translation. We evaluate our approach on the benchmark dataset BConTrasT (English-German) and a self-collected bilingual dialogue corpus, named BMELD (English-Chinese). Extensive 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. Additionally, we make the BMELD dataset publicly available for the research community.