Bitext Mining for Low-Resource Languages via Contrastive Learning

Abstract: Mining high-quality bitexts for low-resource languages is challenging. This paper shows that sentence representation of LLMs fine-tuned with multiple negatives ranking loss, a contrastive objective, helps retrieve clean bitexts. Experiments show that parallel data mined from our approach substantially outperform the previous state-of-the-art method on low resource languages Khmer and Pashto.