In this chapter we provide the implementation of the proposed model which is based on hybrid BERT model using Asian languages for MNMT. Due to its benefits in streamlining the training process, lowering the cost of online maintenance, and boosting low-resource and zero-shot translation, multilingual neural machine translation (NMT) (Ni et al., 2022), which translates many languages using a single model, is of significant practical significance. It is exceedingly laborious to manage them all in a single model or use a new model for each language pair given that there are thousands of languages in the world, some of which are very diverse. As a result, multilingual NMT relies on the ability to choose which languages should be supported by a single model given a set resource budget, such as the number of models. Unfortunately, this issue has not been addressed by prior work (Singh et al., 2020). In this study, we create a framework for grouping languages into various categories and training a single multilingual model for each category. Moreover, we also provides two language clustering techniques: (1) language family-based language clustering utilizing prior information, and (2) language-based language embedding, where each language is represented by an embedding vector and grouped in the embedding space. In specifically, we train a universal neural machine translation model to extract the embedding vectors of every language (S. Yang et al., 2020). Our studies on 23 languages reveal that the first clustering approach is simple and understandable but results in subpar translation accuracy, whereas the second approach adequately captures the relationship between languages and enhances translation accuracy for almost all languages (Araújo et al., 2020). (Yang et al., 2021) proposed incorporation of many-to-one statistical machine translation is something else we have done (SMT). When compared to the outcome of the standard SMT, this novel technique gave results in terms of translation accuracy that were equivalent (Kituku et al., 2016).
|