A3-108 Machine Translation System for Similar Language Translation Shared Task 2020

Proceedings of the 5th Conference on Machine Translation (WMT), pages 451–455 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 451 A3-108 Machine Translation System for Similar Language Translation Shared Task 2020 Saumitra Yadav, Manish Shrivastava Machine Translation - Natural Lanugae Processing Lab Language Technologies Research Centre Kohli Center on Intelligent Systems International Institute of Information Technology - Hyderabad saumitra.yadav@research.iiit.ac.in m.shrivastava@iiit.ac.in Abstract In this paper, we describe our submissions for Similar Language Translation Shared Task 2020. We built 12 systems in each direction for Hindi ⇐⇒ Marathi language pair. This paper outlines initial baseline experiments with var- ious tokenization schemes to train statistical models. Using optimal tokenization scheme among these we created synthetic source side text with back translation. And prune syn- thetic text with language model scores. This synthetic data was then used along with train- ing data in various settings to build transla- tion models. We also report configuration of the submitted systems and results produced by them. 1 Introduction Machine Translation systems are models which aim to translate text from one language into an- other. There are multiple ways of building such a model (Rule Based, Data driven, Hybrid etc.). In this system description paper, we use data driven techniques to build MT systems. As the name sug- gests, data driven MT systems make use of parallel sentences (i.e. xth sentence in two languages have same meaning). We make use of statistical (Koehn et al., 2003) and neural (Bahdanau et al., 2014) methods to build systems for Hindi Marathi pair. Hindi Marathi language pair comes under purview of similar languages. Similar Languages are languages which exhibit lexical and structural similarities (Kunchukuttan et al., 2014a). This can be due to common ancestry or being in close prox- imity for long time. In current digital age commu- nication, translation between similar language is a justifiable requirement. But there is a scarcity of good quality bitext for many language pairs, as is the case of Hindi Marathi. Hence, we used characteristics displayed by similar languages (in this case Hindi and Marathi) like similar form of spelling, pronunciation etc. Following Kunchukut- tan and Bhattacharyya (2017) and Kunchukuttan et al. (2014b) we made use of byte pair encod- ing (Sennrich et al., 2016b) and morfessor toolkit (Virpioja et al., 2013) respectively as part of pre- processing step before training. Using cues from Koehn and Knowles (2017) and looking at the size of training data provided, we use statistical method to build initial models. To further salvage simi- larity between this language pair we made use of backtranslation (Sennrich et al., 2016a) to generate more synthetic data for further training using both neural and statistical methods. For this shared task we developed 12 translation systems in each direction (Hindi ⇐⇒ Marathi). To rank systems, we went through some test in- stances subjectively and also compared our BLEU scores with another Translation system. And chose top 2 systems in both direction using both sub- jective examination and detoknized BLEU scores. Subsequent sections give more …

Content truncated to 500 words