Unsupervised and Very-Low Resource Supervised Translation on German and Sorbian Variant Languages

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1104 - 1110 December 7-8, 2022 ©2022 Association for Computational Linguistics PICT-NLP@WMT22-EMNLP2022: Unsupervised and Very-Low Resource Supervised Translation on German and Sorbian Variant Languages Aditya Vyawahare ∗ aditya.vyawahare07@gmail.com Rahul Tangsali ∗ rahuul2001@gmail.com Aditya Mandke † amandke@ucsd.edu Onkar Litake † olitake@ucsd.edu Dipali Kadam ‡ ddkadam@pict.edu Pune Institute of Computer Technology, India Abstract This paper presents the work of team PICT- NLP for the shared task on unsupervised and very low-resource supervised machine translation, organized by the Workshop on Machine Translation, a workshop in colloca- tion with the Conference on Empirical Meth- ods in Natural Language Processing (EMNLP 2022). The paper delineates the approaches we implemented for supervised and unsu- pervised translation between the following 6 language pairs: German-Lower Sorbian (de-dsb), Lower Sorbian-German (dsb-de), Lower Sorbian-Upper Sorbian (dsb-hsb), Up- per Sorbian-Lower Sorbian (hsb-dsb), German- Upper Sorbian (de-hsb), and Upper Sorbian- German (hsb-de). For supervised learning, we implemented the transformer architecture from scratch using the Fairseq library. Whereas for unsupervised learning, we implemented Face- book’s XLM masked language modeling ap- proach. We discuss the training details for the models we used, and the results obtained from our approaches. We used the BLEU and chrF metrics for evaluating the accuracies of the gen- erated translations on our systems. 1 Introduction Neural machine translation has witnessed signif- icant progress in the case of highly spoken lan- guages such as English (Bahdanau et al., 2015), Mandarin (Li et al., 2022), French (Emezue and Dossou, 2020), etc. However, in many cases, it becomes challenging to develop a robust bilingual machine translation system, especially with limited resources (Dong et al., 2015). There are big-tech companies such as Google1 and Bing2, which have taken initiatives to build translation systems for ∗ equal contribution † equal contribution ‡ equal contribution 1https://translate.google.com/ 2https://www.bing.com/translator multiple languages. Still, languages that are low- resource in nature, such as the Sorbian family of languages (Howson, 2017), have gotten lesser at- tention in terms of research. The paper focuses on the development of machine translation systems between pairs of languages from German, Lower Sorbian, and Upper Sorbian, using both supervised and unsupervised approaches. In the Indo-European language family, German (Deutsch) belongs to the western Germanic branch (Durrell, 2006). Approximately 95 million peo- ple speak it natively; 28 million speak it as a sec- ond language in more than 40 countries. Due to a phonetic mutation called High German Consonant Shift (Vennemann, 2008), German moved away from other Germanic languages. This shift in Ger- man consonants occurred between the 3rd and 5th centuries, and probably ended in the 9th century AD. Lower Sorbian (dolnoserbska rˇec) and Upper Sorbian (hornjoserbska rˇeˇc) are western Slavonic languages spoken in the region of Lower and Up- per Lusatia in the southeast of Germany respec- tively. They are closely related to other West Slavonic languages, including Polish, Czech, Slo- vak, and Kashubian. There are seven recognized autochthonous minorities and regional languages in Germany, including Danish, Saterfrisian, North Frisian, Romanes, and Lower German. …

Content truncated to 500 words