Deep neural architectures for Kashmiri-English machine translation.

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Artificial Intelligence

Background:

• Neural Machine Translation (NMT) has advanced significantly, yet low-resource languages like Kashmiri remain under-explored.
• Developing effective NMT systems requires substantial parallel data and robust model architectures.

Purpose of the Study:

• To establish the first comprehensive deep learning-based NMT framework for the Kashmiri-English language pair.
• To create and release a high-quality parallel corpus for Kashmiri-English translation research.
• To benchmark different NMT architectures and provide a foundation for future low-resource NMT development.

Main Methods:

• Development of a 270,000-sentence Kashmiri-English parallel corpus.
• Training and evaluation of three NMT architectures: encoder-decoder, attention-enhanced, and Transformer.
• Utilizing byte-pair encoding for vocabulary generation and standard metrics (BLEU, GLEU, ROUGE, ChrF++) for evaluation.

Main Results:

• The Transformer-based NMT model significantly outperformed RNN-based baselines.
• Achieved a BLEU-4 score of 0.2965, demonstrating effective handling of Kashmiri's complexity.
• Linguistic error analysis confirmed the Transformer's superiority in managing long-range dependencies and morphology.

Conclusions:

• The Transformer architecture represents a state-of-the-art approach for Kashmiri-English NMT.
• The released dataset, models, and methodology provide a crucial benchmark for low-resource NMT research.
• This work paves the way for improved machine translation capabilities for Kashmiri and similar languages.