A Comparative Analysis of Transformers for Multilingual Neural Machine Translation
R Prasanna Kumar, S Sudharson, Avuthu Avinash Reddy, B Siva Jyothi Natha Reddy, Vemireddy Anvitha · 2023
Multilingual neural machine translation (NMT) has emerged as a promising solution to break down language barriers and promote cross-lingual communication. Transformer-based NMT models have gained significant attention due to their ability to learn contextual information and their effectiveness in handling long-range dependencies. The transformer model, with its self-attention and cross-attention mechanisms, fundamentally alters machine translation when compared to other neural machine translation models and traditional machine translation models. Token alignments between source and target sentences are successfully modelled by these mechanisms. This paper deals with the performance of a transformer-based multilingual NMT model that uses a shared encoder and decoder. We conducted experiments on the publicly available Multi30k dataset for translation jobs involving German, French, and Czech into English. In comparison of models, the Transformer model has performed better against the Convolutional SeqtoSeq and Attention-Based Seq2Seq models. We have evaluated our model on the metric Bilingual Evaluation Understudy (BLEU), where the transformer achieved 36.67, 48.47, and 33.66 for DE, FR, and CZ to EN translations.