Improving a Multi-Source Neural Machine Translation Model with Corpus Extension for Low-Resource Languages

Improving a Multi-Source Neural Machine Translation Model with Corpus Extension for Low-Resource Languages Gyu-Hyeon Choi1, Jong-Hun Shin2, Young-Kil Kim3 1Korea University of Science and Technology (UST), South Korea 2Electronics and Telecommunication Research Institute (ETRI), South Korea 3Electronics and Telecommunication Research Institute (ETRI), South Korea choko93@ust.ac.kr1, jhshin82@etri.re.kr2, kimyk@etri.re.kr3 Abstract In machine translation, we often try to collect resources to improve performance. However, most of the language pairs, such as Korean-Arabic and Korean-Vietnamese, do not have enough resources to train machine translation systems. In this paper, we propose the use of synthetic methods for extending a low-resource corpus and apply it to a multi-source neural machine translation model. We showed the improvement of machine translation performance through corpus extension using the synthetic method. We specifically focused on how to create source sentences that can make better target sentences, including the use of synthetic methods. We found that the corpus extension could also improve the performance of multi-source neural machine translation. We showed the corpus extension and multi-source model to be efficient methods for a low-resource language pair. Furthermore, when both methods were used together, we found better machine translation performance. Keywords: Neural Machine Translation, Multi-Source Translation, Synthetic, Corpus Extension, Low-Resource 1. Introduction We often try to collect resources to improve machine translation performance. Using the large size of a parallel corpus, it is possible to achieve high-quality machine translation performance. However, there are many cases where resources of language pairs are insufficient. Except for major European languages and some Asian languages, most of the language pairs do not have sufficient resources to develop a neural machine translation (NMT) system. It is also difficult to obtain parallel corpora for some language pairs such as Korean to Arabic or Korean to Vietnamese. Since the machine translation performance largely depends on the size of a parallel corpus, it is important to find an efficient way to extend the corpus. Although it is difficult to find a proper parallel corpus, we can create an artificial parallel corpus by translating the source or target of a language pair. Some researchers have studied the extension of a parallel corpus using the pivot method (Cohn and Lapata, 2007; Utiyama and Isahara, 2007; Wu and Wang, 2007). This method introduces another language referred to as the pivot language which is a third language that is different from the source and target languages. There are many different pivot strategies. The first is the transfer method which translates a source sentence to a pivot sentence and then to a target sentence (Cohn and Lapata, 2007; Wu and Wang, 2007). The second is the triangulation method which multiplies corresponding translation probabilities and lexical weights to create a new source-target phrase table (Utiyama and Isahara, 2007). The third is the synthetic method, which uses existing translation models to build a synthetic parallel source-target corpus from source-pivot or pivot-target (Bertoldi et al., 2008). There are other approaches that have been proposed for multilingual training with low-resource parallel corpora. Among the approaches, there is a multi-source …

Content truncated to 500 words