Meet Changes with Constancy: Learning Invariance in Multi-Source Translation

Proceedings of the 28th International Conference on Computational Linguistics, pages 1122–1132 Barcelona, Spain (Online), December 8-13, 2020 1122 Meet Changes with Constancy: Learning Invariance in Multi-Source Translation Jianfeng Liu♦△♣, Ling Luo♦△♦, Xiang Ao♦△♦†, Yan Song♠♥, Haoran Xu♦△, Jian Ye♦♣† ♦Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China △University of Chinese Academy of Sciences ♣Beijing Key Laboratory of Mobile Computing and Pervasive Device ♦Key Laboratory of Intelligent Information Processing of Chinese Academy of Sciences ♠The Chinese University of Hong Kong (Shenzhen) ♥Shenzhen Research Institute of Big Data {liujianfeng18s;luoling18s;aoxiang;xuhaoran18s;jye}@ict.ac.cn songyan@cuhk.edu.cn Abstract Multi-source neural machine translation aims to translate from parallel sources of informa- tion (e.g. languages, images, etc.) to a single target language, which has shown better per- formance than most one-to-one systems. Despite the remarkable success of existing models, they usually neglect the fact that multiple source inputs may have inconsistencies. Such dif- ferences might bring noise to the task and limit the performance of existing multi-source NMT approaches due to their indiscriminate usage of input sources for target word predictions. In this paper, we attempt to leverage the potential complementary information among distinct sources and alleviate the occasional conflicts of them. To accomplish that, we propose a source invari- ance network to learn the invariant information of parallel sources. Such network can be easily integrated with multi-encoder based multi-source NMT methods (e.g. multi-encoder RNN and transformer) to enhance the translation results. Extensive experiments on two multi-source trans- lation tasks demonstrate that the proposed approach not only achieves clear gains in translation quality but also captures implicit invariance between different sources. 1 Introduction Neural machine translation (NMT) systems in general translate one source language to a target lan- guage. Various one-to-one attentional encoder-decoder architectures (Bahdanau et al., 2014; Luong et al., 2015b; Vaswani et al., 2017) were designed to learn word and structure mappings including forma- tions, grammatical correspondences between source and target languages. Recently, multi-source NMT (Zoph and Knight, 2016), which simultaneously takes multiple different languages (Dabre et al., 2017; Libovick`y et al., 2018; Currey and Heafield, 2018) as input when trans- lating to another one, is emerging. Its intuitive idea is ambiguity between one source language and the target language could be reduced by another source language via the “triangulation” proposed by (Kay, 2000). For example, it is hard to tell who was with a telescope by the ambiguous English sentence in Figure 1 (a). But it could be more easily to be translated to Spanish if provided with the corresponding Chinese phrase “我通过望远镜”, which corresponds to “I see through a telescope” in English. Besides, the complementary source could be extended to non-language input such as images, also known as multi- modal NMT (Libovick`y and Helcl, 2017; Elliott et al., 2017). It takes an image with description in source languages, that is then translated into a target language. The images are expected to provide additional signals for better translations, which abides by similar assumptions in multi-source NMT with different languages. A specific instance is illustrated in Figure …

Content truncated to 500 words