Key points are not available for this paper at this time.
Communicating in one’s native language significantly improves comprehension across various aspects of life. The large linguistic diversity in India creates considerable communication challenges in the business, education, healthcare and creative industries, such as film, music, drama, etc. In this research, we propose a robust machine translation system that utilizes a multi-task, emotion and semantic aware neural architecture for preserving emotional nuances and semantic depth in literary translations between languages of Indo-Aryan and Dravidian language families. We employed a transfer learning approach by extending the pre-trained No Language Left Behind (NLLB-200) distilled transformer model trained with a multi-objective loss function optimized for literary translation quality. The proposed architecture contains two specialized components: an emotion recognition module classifying eight emotional states and a semantic consistency component enforcing meaning preservation through contrastive learning. Our model is trained on BHT25, a curated dataset of 25,000 parallel literary text corpora across Bengali, Hindi and Telugu language triplets, annotated with emotion labels. The model achieved Bilingual Evaluation Understudy scores of 42.66 for Bengali-Hindi, 36.74 for Bengali-Telugu, and 31.55 for Hindi-Telugu translation. The model demonstrated overall classification accuracy of 77.2% for 8 emotions across all language pairs, while maintaining semantic consistency with an average cosine similarity of 0.92. Our progressive three-phase training strategy with memory-efficient optimization techniques reduced GPU usage by 35% as compared to standard training. In addition, the ablation experiments confirmed the meaningful contribution of both emotion and semantic modules to translation quality. We deployed the trained model and dataset on Hugging Face, thereby providing, to the best of our knowledge, the first publicly available emotion-semantic-aware machine translation system for cross-family Indian languages.
Sani et al. (Thu,) studied this question.