English-to-Prakrit Machine Translation via Multilingual Transfer Learning
Researchers have developed a method for English-to-Prakrit machine translation, a low-resource language pair not supported by existing models like IndicTrans2. By mapping Prakrit to the Hindi language tag within a multilingual model, they achieved improved BLEU scores using a 1,474-pair Maharashtri Prakrit corpus. This approach demonstrates the potential for script-compatible language routing to enable translation for unsupported classical languages, though data scarcity and dialect differences present challenges. AI
IMPACT Demonstrates a technique for extending machine translation to unsupported classical languages, potentially opening new avenues for linguistic research and preservation.