Transformer-based Neural Machine Translation System for Hindi – Marathi: WMT20 Shared Task

Proceedings of the 5th Conference on Machine Translation (WMT), pages 393–395 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 393 Transformer-based Neural Machine Translation System for Hindi – Marathi: WMT20 Shared Task Amit Kumar, Rupjyoti Baruah, Rajesh Kumar Mundotiya, Anil Kumar Singh Department of Computer Science & Engineering Indian Institute of Technology (B.H.U.) Varanasi, India {amitkumar.rs.cse17, rupjyotibaruah.rs.cse18, rajeshkm.rs.cse16, aksingh.cse}@iitbhu.ac.in Abstract This paper reports the results for the Machine Translation (MT) system submitted by the NL- PRL team for the Hindi – Marathi Similar Translation Task at WMT 2020. We apply the Transformer-based Neural Machine Trans- lation (NMT) approach on both translation di- rections for this language pair. The trained model is evaluated on the corpus provided by shared task organizers, using BLEU, RIBES, and TER scores. There were a total of 23 systems submitted for Marathi to Hindi and 21 systems submitted for Hindi to Marathi in the shared task. Out of these, our submission ranked 6th and 9th, respectively. 1 Introduction In the last decade and a half, neural machine trans- lation (NMT) (Sutskever et al., 2014) has achieved great success in automatically translating human language text, outperforming statistical machine translation (SMT) (Koehn et al., 2003). Both the system require very large corpus sizes to train and evaluate the results. They, however, don’t work very well for low resource data (He et al., 2016; Koehn and Knowles, 2017; Dowling et al., 2018). Translation from or to low resource languages is the major challenges faced by today’s NMT systems. Different methods have been proposed to over- come the data sparsity problem for low resource languages by researchers around the world. These include using monolingual data (Wu et al., 2019), fine-tuning (Miceli Barone et al., 2017) the high resource monolingual and parallel data on low re- source data, back translation (Hoang et al., 2018), etc. They succeed up to some extent, but the suc- cess is limited, as the reported results show when compared to those for resource rich languages. In this paper, we use the Transformer network- based NMT system (Vaswani et al., 2017) because it is among the state of the art models for machine translation. The work reported for this shared task is an extension of the work done by (Kumar and Singh, 2019) for similar languages task for 2019, which had also used a transformer based NMT system. 2 Similar Languages Two languages are considered similar or closely related if they are close relatives in terms of the linguistic family of the linguistic family tree (or forest), or if the speakers of the two languages are in close contact over a long period of time. Contact over a long period leads to the exchange of cognates and loanwords between the speakers, sometimes even grammatical constructs. Leveraging the close similarity of languages is one way to overcome the problem of data scarcity. Using similar features between such languages and improving translation is one of the directions for research for low resource machine translation. For this submission, the …

Content truncated to 500 words