Multilingual Multi-Domain NMT for Indian Languages

Proceedings of Recent Advances in Natural Language Processing, pages 727–733 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_083 727 Multilingual Multi-Domain NMT for Indian Languages Sourav Kumar IIIT Hyderabad sourav.kumar@ research.iiit.ac.in Salil Aggarwal IIIT Hyderabad salil.aggarwal@ research.iiit.ac.in Dipt Misra IIIT Hyderabad dipti@ iiit.ac.in Abstract India is known as the land of many tongues and dialects. Neural machine translation (NMT) is the current state-of-the-art approach for machine translation (MT) but performs bet- ter only with large datasets which Indian lan- guages usually lack, making this approach in- feasible. So, in this paper, we address the prob- lem of data scarcity by efficiently training mul- tilingual and multilingual multi domain NMT systems involving languages of the Indian subcontinent. We are proposing the technique for using the joint domain and language tags in a multilingual setup. We draw three major conclusions from our experiments: (i) Train- ing a multilingual system via exploiting lexi- cal similarity based on language family helps in achieving an overall average improvement of 3.25 BLEU points over bilingual baselines, (ii) Technique of incorporating domain infor- mation into the language tokens helps multilin- gual multi-domain system in getting a signifi- cant average improvement of 6 BLEU points over the baselines, (iii) Multistage fine-tuning further helps in getting an improvement of 1- 1.5 BLEU points for the language pair of in- terest. 1 Introduction Good translation systems are an important require- ment due to substantial government, business and social communication among people speaking different languages. Neural machine translation (Vaswani et al., 2017) is the current state-of-the-art approach for machine translation in both academia and industry (Bahdanau et al., 2014). The success of NMT heavily relies on a huge amount of parallel sentences as training data. But using the traditional approaches (Vaswani et al., 2017), one would still need to train a separate model for each translation direction and also a lot of parallel human translated corpora which is again expensive to generate. But on the other hand, multilingual neural machine translation (Johnson et al., 2017) enables training a single model that supports translation from multiple source languages to a single target language or from a single source language to multiple target languages. In addition to this, another benefit of a training a single model for multiple translation directions is the ability to learn not just from the training data of the language pair of interest, but also from other language pairs. But this learning is hindered in case of language pairs that do not show any kind of relatedness among themselves. But on the other hand, Indian languages exhibit a lot of lexical and structural similarities on account of sharing a common ancestry. It is therefore important to utilize the lexical similarity (Kunchukuttan and Bhattacharyya, 2020) of these languages to build efficient systems by combining all the related languages. Also, in a typical in-domain MT scenario, the amount of parallel texts from a single domain is not enough to train a good translation system, even for multilingual systems. Apart from this, one has to train …

Content truncated to 500 words