A Dual Reinforcement Method for Data Augmentation using Middle Sentences for Machine Translation

Proceedings of Machine Translation Summit XIX, Vol. 1: Research Track, pages 48–58 September 4–8, 2023, Macau SAR, China. ©2023 The authors. This article is licensed under a Creative Commons Attribution 4.0 International License (CC BY 4.0). 48 A Dual Reinforcement Method for Data Augmentation using Middle Sentences for Machine Translation Wenyi Tang tang@akane.waseda.jp Yves Lepage yves.lepage@waseda.jp Graduate School of IPS, Waseda University, Kitakyushu, Japan Abstract This paper presents an approach to enhance the quality of machine translation by leveraging middle sentences as pivot points and employing dual reinforcement learning. Conventional methods for generating parallel sentence pairs for machine translation rely on parallel corpora, which may be scarce, resulting in limitations in translation quality. In contrast, our proposed method entails training two machine translation models in opposite directions, utilizing the middle sentence as a bridge for a virtuous feedback loop between the two models. This feed- back loop resembles reinforcement learning, facilitating the models to make informed deci- sions based on mutual feedback. Experimental results substantiate that our proposed method significantly improves machine translation quality. 1 Introduction The accuracy of neural machine translation is limited by the quantity of available training data (Wang et al., 2022; Sennrich et al., 2016), leading to the development of various techniques for data augmentation. In this paper, we propose a novel method that leverages middle sentences (Wang et al., 2021) as pivot points and uses dual reinforcement learning (Zhou et al., 2019) for data augmentation in machine translation. Dual learning (He et al., 2016; Yi et al., 2017; Zhou et al., 2019) entails the concurrent training of two neural networks, to enhance translation accuracy by leveraging the reconstruc- tion model’s ability to generate synthetic parallel sentence pairs. Data augmentation involves ar- tificially augmenting the size of the training data by generating additional sentence pairs through diverse techniques, such as back-translation (Brislin, 1970; Douglas and Craig, 2007; Edunov et al., 2018). These techniques offer potential solutions to mitigate the scarcity of parallel corpora and improve the quality of machine translation models by providing supplementary training data. In our proposal, we aim to combine the strengths of dual learning and data augmentation with the use of middle sentences as pivot points to reinforce the training process and further enhance the accuracy of the machine translation model. We start by presenting our dual rein- forcement method in Section 2. We present our experiment setup in Section 3 and results in Section 4. 49 2 Methods Our method combines the use of a dual learning framework with data augmentation techniques, leveraging the middle sentences of parallel sentence pairs as pivot points. The general process involves generating additional parallel sentence pairs through middle sentence generation, using the middle sentences to create new sentence pairs and refining the translations using a machine translation model. This process is repeated iteratively, forming a reinforcement loop that en- hances the quality of the translation model through synthetic data, i.e., middle sentences. In the following subsections, we provide a detailed explanation of each step in …

Content truncated to 500 words