Hybrid LETCNN-P Transformer Architecture for Enhanced Translation of Low-Resource Languages

This research introduces an innovative method for neural machine translation (NMT) of Low-Resource Languages (LRLs) using the LETCNN-P Transformer model. The proposed model achieves significant improvements in translation accuracy and efficiency by combining Log Exponential Tanh CNN in the encoder phase and P Transformer in the decoder phase. This architecture is designed to effectively capture word relationships and contextual nuances, resulting in precise and contextually accurate translations. Comparative analysis with existing NMT approaches demonstrates superior performance, evidenced by higher scores in bilingual evaluation understudy (BLEU), F measures, and shorter training times. Additionally, qualitative assessments highlight the model's ability to accurately translate  complex sentences across multiple languages, underscoring its practical utility. Developed using the TensorFlow framework, the model is trained on a dataset ('samanantar') comprising English and Kannada sentence pairs. This research significantly advances MT technology, promising enhanced global communication and cross-cultural interaction.