Low Resource Similar Language Neural Machine Translation for Tamil-Telugu

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 288–291 November 10–11, 2021. ©2021 Association for Computational Linguistics 288 Low Resource Similar Language Neural Machine Translation for Tamil-Telugu Vandan Mujadia and Dipti Misra Sharma Machine Translation - Natural Language Processing Lab Language Technologies Research Centre Kohli Center on Intelligent Systems International Institute of Information Technology - Hyderabad vandan.mu@research.iiit.ac.in, dipti@iiit.ac.in Abstract This paper describes the participation of team oneNLP (LTRC, IIIT-Hyderabad) for the WMT 2021 task, similar language translation1. We experimented with transformer based Neu- ral Machine Translation and explored the use of language similarity for Tamil-Telugu and Telugu-Tamil. We incorporated use of differ- ent subword configurations, script conversion and single model training for both directions as exploratory experiments. 1 Introduction Machine Translation (MT) is a field of Natural Language Processing which aims to translate a text from one natural language to another. The meaning of the source text must be fully preserved in the resulting translated text in the target language. Recent years have seen significant quality advancements in machine translation with the advent of Neural Machine Translation. For the translation task, different types of machine translation systems have been developed and they are mainly categorized into Rule based Machine Translation (RBMT)(Forcada et al., 2011), Statistical Machine Translation (SMT) (Koehn, 2009) and Neural Machine Translation (NMT) (Bahdanau et al., 2014). Neural machine translation (NMT) shows high quality in terms of output fluency and translation quality, when large amounts of parallel data are available (Barrault et al., 2020). Unfortunately, for most language pairs, parallel data is either scare or non-existent. To overcome this, unsupervised MT (UMT) (Artetxe et al., 2020) focuses on util- ising monolingual data to generate synthetic par- allel training data. Other techniques like back- translation(Sennrich et al., 2015),(Hoang et al., 1https://www.statmt.org/wmt21/similar.html 2018), (Feldman and Coto-Solano, 2020) or denois- ing(Kim et al., 2019) also rely on parallel corpora of other language pairs and/or large quantities of monolingual data. This paper describes our experiments for very low resourced similar language translation. For our work, we focused only on Tamil-Telugu lan- guage pair (both directions) and participated in a constrained setting. We experimented only with Transformer (Vaswani et al., 2017) based Neural Machine Trans- lation throughout. Along with it, to tackle high agglutination of both languages, we explored the morph (Virpioja et al., 2013) induced sub-word seg- mentation with byte pair encoding (BPE)(Sennrich et al., 2016). Similar to Multilingual Neural Machine Transla- tion (MNMT), we explored the use of a tag trick, where a token like “< 2xx >” (xx is language code) is prefixed to each source sentence to indi- cate the desired target language(Dabre et al., 2020). Here, we trained a single model for both directions (Tamil-Telugu and Telugu-Tamil) on given parallel data and monolingual data under MNMT setting. The sections of the paper are organised as fol- lowing: Section 2 describes Data, Section 3 and 4 describe pre-processing and Training Configura- tion and in Section 5 we talk about results and we conclude in section 6. 2 …

Content truncated to 500 words