ANVITA-African: A Multilingual Neural Machine Translation System for African Languages

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1090 - 1097 December 7-8, 2022 ©2022 Association for Computational Linguistics ANVITA-African: A Multilingual Neural Machine Translation System for African Languages Pavanpankaj Vegi, Sivabhavani J, Biswajit Paul, Prasanna Kumar KR, Chitra Viswanathan pavanpankaj333@gmail.com, {jsbhavani.cair, biswajit.cair, prasanna.cair, chitrav.cair}@gov.in Centre for Artificial Intelligence and Robotics, CV Raman Nagar, Bangalore, India Abstract This paper describes ANVITA African NMT system submitted by team ANVITA for WMT 2022 shared task on Large-Scale Machine Translation Evaluation for African Languages under the constrained translation track. The team participated in 24 African languages to English MT directions. For better handling of relatively low resource language pairs and effective transfer learning, models are trained in multilingual setting. Heuristic based cor- pus filtering is applied and it improved per- formance by 0.04-2.06 BLEU across 22 out of 24 African→English directions and also improved training time by 5x. Use of deep transformer with 24 layers of encoder and 6 layers of decoder significantly improved per- formance by 1.1-7.7 BLEU across all the 24 African→English directions compared to base transformer. For effective selection of source vocabulary in multilingual setting, joint and language wise vocabulary selection strategies are explored at the source side. Use of lan- guage wise vocabulary selection however did not consistently improve performance of low resource languages in comparison to joint vo- cabulary selection. Empirical results indicate that training using deep transformer with fil- tered corpora seems to be a better choice than using base transformer on the whole corpora both in terms of accuracy and training time. 1 Introduction Africa is very rich in languages, and around 1200 to 2100 languages are spoken in African countries1, 24 African languages and 100 language pairs were selected for the WMT22 Large-Scale Machine Translation Evaluation for African Languages shared task Adelani et al. (2022b). Selected 24 African languages include Afrikaans(afr), Amharic(amh), Chichewa(nya), Hausa(hau), Igbo(ibo), Kamba(kam), Kinyarawanda(kin), 1https://en.wikipedia.org/wiki/Languages_of_Africa Lingala(lin), Luganda(lug), Luo(luo), Nige- rian Fulfulde(fuv), Northern Sotho(nso), Oromo(orm), shona(sna), Somali(som), Swahili(swh), Swati(ssw), Setswana(tsn), Umbundu(umb), Wolof(wol), Xhosa(xho), Xitsonga(tso), Yoruba(yor) and Zulu(zul) and language pairs include African-English, selective African-French, and African-African pairs, where many of the pairs fall under the low resource category. In this task, organizers permitted two submissions, Best scoring submission is considered as Primary model and other one being the Contrastive model. This paper describes our submission to WMT 2022 Large-Scale Machine Translation Evaluation for African Languages shared task where we participated for translation of 24 African languages to English. We are not officially given a rank as we didn’t participate in all African MT directions. 2 Related Work Developing quality machine translation system for low resource languages still remains a major chal- lenge and many of the world languages fall under this category. Some of the recent developments do show that multilingual NMT is a promising di- rection. In massively multilingual neural machine translation, the authors have shown to train a single model for translating 102 languages to and from English and the results outperformed the strong bilingual baseline MT system especially …

Content truncated to 500 words