Neural Machine Translation for Tamil–Telugu Pair

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 284–287 November 10–11, 2021. ©2021 Association for Computational Linguistics 284 Neural Machine Translation for Tamil–Telugu Pair Sahinur Rahman Laskar, Bishwaraj Paul, Prottay Kumar Adhikary Partha Pakray, Sivaji Bandyopadhyay Department of Computer Science and Engineering National Institute of Technology Silchar Assam, India {sahinur_rs, bishwaraj_ug, prottay_ug, partha}@cse.nits.ac.in sivaji.cse.ju@gmail.com Abstract The neural machine translation approach has gained popularity in machine translation be- cause of its context analysing ability and its handling of long-term dependency issues. We have participated in the WMT21 shared task of similar language translation on a Tamil-Telugu pair with the team name: CNLP-NITS. In this task, we utilized monolingual data via pre- train word embeddings in transformer model based neural machine translation to tackle the limitation of parallel corpus. Our model has achieved a bilingual evaluation understudy (BLEU) score of 4.05, rank-based intuitive bilingual evaluation score (RIBES) score of 24.80 and translation edit rate (TER) score of 97.24 for both Tamil-to-Telugu and Telugu-to- Tamil translations respectively. 1 Introduction Machine translation (MT) works as an interface that handles language ambiguity concerns via auto- matic translation between two different languages. Neural machine translation (NMT) attains state-of- the-art results for both high and low-resource lan- guage pairs translation (Cho et al., 2014; Sutskever et al., 2014; Bahdanau et al., 2015; Luong et al., 2015; Pathak et al., 2018; Pathak and Pakray, 2018; Laskar et al., 2019; Laskar et al., 2020a). The NMT utilizes an artificial neural network to pre- dicts the likelihood of a sequence of words. But NMT requires a sizeable parallel corpus to get ef- fective MT output, challenging for low-resource pair translation. In this WMT21 shared task, we have participated on a similar language pair transla- tion task of Tamil–Telugu pair using NMT. We aim to utilize similarity features among such a similar language pair and monolingual data to overcome the less availability of parallel corpus. The trans- former model (Vaswani et al., 2017) based NMT is considered in this work, since it outperforms RNN based NMT. Moreover, NMT performance can be enhanced utilizing monolingual data (Weng et al., 2019; Wu et al., 2019; Ramachandran et al., 2017; Variš and Bojar, 2019; Qi et al., 2018). To evaluate the performance of our system’s output, WMT21 organizer used standard evaluation met- rics, namely, BLEU (Papineni et al., 2002), RIBES (Isozaki et al., 2010) and TER (Snover et al., 2006) which are reported in Section 4. 2 Related Work There are limited works on the Tamil–Telugu pair (Chakravarthi et al., 2021). The literature sur- vey found similar works on Indian similar lan- guage pairs, such as Hindi–Nepali (Laskar et al., 2019) and Hindi–Marathi (Laskar et al., 2020b) at WMT19 and WMT20. Both (Laskar et al., 2020b, 2019) used transformer model based NMT. More- over, Ramachandran et al. (2017); Variš and Bojar (2019); Qi et al. (2018) pre-trained methods are in- corporated in NMT to utilize advantage of mono- lingual data for low-resource pairs translation. In this work, GloVe (Pennington et al., 2014) …

Content truncated to 500 words