Bilingual Low-Resource Neural Machine Translation with Round-Tripping: The Case of Persian-Spanish

Proceedings of Recent Advances in Natural Language Processing, pages 18–24, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_003 18 Bilingual Low-Resource Neural Machine Translation with Round-Tripping: The Case of Persian-Spanish Benyamin Ahmadnia1 and Bonnie J. Dorr2 1Department of Computer Science, Tulane University, New Orleans, LA, USA 2Institute for Human and Machine Cognition (IHMC), Ocala, FL, USA ahmadnia@tulane.edu , bdorr@ihmc.us Abstract The quality of Neural Machine Translation (NMT), as a data-driven approach, mas- sively depends on quantity, quality and relevance of the training dataset. Such approaches have achieved promising re- sults for bilingually high-resource scenar- ios but are inadequate for low-resource conditions. This paper describes a round- trip training approach to bilingual low- resource NMT that takes advantage of monolingual datasets to address training data scarcity, thus augmenting transla- tion quality. We conduct detailed experi- ments on Persian-Spanish as a bilingually low-resource scenario. Experimental re- sults demonstrate that this competitive ap- proach outperforms the baselines. 1 Introduction Neural Machine Translation (NMT) has made considerable progress in recent years. However, to achieve acceptable translation output, large sets of aligned parallel sentences are required for the training phase. Thus, as a data-driven paradigm, the quality of NMT output strongly depends on the quality as well as quantity of the provided train- ing data (Bahdanau et al., 2015). Unfortunately, in practice, collecting such parallel text by human labeling is very costly and time consuming (Ah- madnia and Serrano, 2017). Low-resource languages are those that have fewer technologies and datasets relative to some measure of their international importance. The biggest issue with low-resource languages is the extreme difficulty of obtaining sufficient re- sources. Natural Language Processing (NLP) methods that have been created for analysis of low-resource languages are likely to encounter similar issues to those faced by documentary and descriptive linguists whose primary endeavor is the study of minority languages. Lessons learned from such studies are highly informative to NLP researchers who seek to overcome analogous chal- lenges in the computational processing of these types of languages. Assuming that large monolingual texts are available, an obvious next step is to leverage these texts to augment the NMT systems’ per- formance. Various approaches have been devel- oped for this purpose. In some approaches, target monolingual texts are employed to train a Lan- guage Model (LM) that is then integrated with MT models trained from parallel texts to enhance translation quality (Brants et al., 2007; G¨ulc¸ehre et al., 2015). Although these approaches utilize monolingual text to train a LM, they do not ad- dress the shortage of bilingual training datasets. In other approaches, bilingual datasets are au- tomatically generated from monolingual texts by utilizing the Translation Model (TM) trained on aligned bilingual text; the resulting sentence pairs are used to enlarge the initial training dataset for subsequent learning iterations (Ueffing et al., 2008; Sennrich et al., 2016). Although these ap- proaches enlarge the bilingual training dataset, there is no quality control and, thus, the accuracy of the generated bilingual dataset cannot be guar- anteed (Ahmadnia et al., 2018). To …

Content truncated to 500 words