Bitext Mining for Low-Resource Languages via Contrastive Learning
This addresses the problem of limited parallel data for low-resource language translation, though it appears incremental as it builds on existing contrastive learning techniques.
The paper tackled the challenge of mining high-quality bitexts for low-resource languages by fine-tuning language models with a contrastive learning objective, resulting in parallel data that substantially outperformed the previous state-of-the-art method on Khmer and Pashto.
Mining high-quality bitexts for low-resource languages is challenging. This paper shows that sentence representation of language models 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.