Multi-Source Neural Machine Translation with Data Augmentation

Multi-Source Neural Machine Translation with Data Augmentation Yuta Nishimura1 , Katsuhito Sudoh1 , Graham Neubig2,1, Satoshi Nakamura1 1Nara Institute of Science and Technology, 8916-5 Takayama-cho, Ikoma, Nara 630-0192, Japan 2Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA 15213, USA {nishimura.yuta.nn9, sudoh, s-nakamura}@is.naist.jp gneubig@cs.cmu.edu Abstract Multi-source translation systems translate from multiple lan- guages to a single target language. By using information from these multiple sources, these systems achieve large gains in accuracy. To train these systems, it is necessary to have corpora with parallel text in multiple sources and the target language. However, these corpora are rarely complete in practice due to the difficulty of providing human transla- tions in all of the relevant languages. In this paper, we pro- pose a data augmentation approach to fill such incomplete parts using multi-source neural machine translation (NMT). In our experiments, results varied over different language combinations but significant gains were observed when us- ing a source language similar to the target language. 1. Introduction Machine Translation (MT) systems usually translate one source language to one target language. However, in many real situations, there are multiple languages in the corpus of interest. Examples of this situation include the multilingual official document collections of the European parliament [1] and the United Nations [2]. These documents are manually translated into all official languages of the respective organi- zations. Many methods have been proposed to use these mul- tiple languages in translation systems to improve the trans- lation accuracy [3, 4, 5, 6]. In almost all cases, multilin- gual machine translation systems output better translations than one-to-one systems, as the system has access to multi- ple sources of information to reduce ambiguity in the target sentence structure or word choice. However, in contrast to the more official document col- lections mentioned above where it is mandated that all trans- lations in all languages, there are also more informal multi- lingual captions such as those of talks [7] and movies [8]. Be- cause these are based on voluntary translation efforts, large portions of them are not translated, especially into languages with a relatively small number of speakers. Nishimura et al. [9] have recently proposed a method for multi-source NMT that is able to deal with the case of miss- ing source data encountered in these corpora. The imple- mentation is simple: missing source translations are replaced with a special symbol ⟨NULL⟩ as shown in Figure 1(a). This How are you? <NULL> Ako sa máte? Cs En Sk Original Missing Original (a) Multi-source NMT with filling in a symbol [9] How are you? Jak se máte? Ako sa máte? Cs En Sk Original Pseudo Original How are you? Ako sa máte? Original Original Data Augmentation with trained multi-source NMT {En, Sk}-to-Cs En Sk (b) Proposed Method: Multi-source NMT with data augmentation Figure 1: Example of multi-source NMT with an incomplete corpus; The language pair is {English, Czech}-to-Slovak and the translation of Czech is missing. method allows us to use incomplete corpora both at training time and test time, …

Content truncated to 500 words