Enhanced BERT-based Multi-Head Self-Attention Transformer for Transformation of Marathi Text to Marathi Sign Language Gloss

Abstract

One recent advancement in the field of machine learning is the translation of text into sign language gloss, which is a form of natural language for the deaf community. The work presented is a new method to translate Marathi text to Marathi sign language gloss by combining salient features of Bidirectional Encoder Representation from Transformer (BERT) for tokenization and complementing the tokenized frame with Encoder with Attention mechanism and decoding with the LSTM decoder. The work conducted experiments on the created corpora of Marathi text and Marathi sign language gloss sentence pairs. The experiments that employed three models show that the suggested model performs better than the existing approaches. The results show that the translation of Marathi text to sign gloss achieves improved performances with an accuracy of 91.5% and the BLEU scores BLEU-1: 85, BLEU-2: 75, BLEU-3: 65, and BLEU-4: 57.