notesum.ai
Published at November 18Transcending Language Boundaries: Harnessing LLMs for Low-Resource Language Translation
cs.CL
cs.AI
Released Date: November 18, 2024
Authors: Peng Shu1, Junhao Chen1, Zhengliang Liu1, Hui Wang2, Zihao Wu1, Tianyang Zhong3, Yiwei Li1, Huaqin Zhao1, Hanqi Jiang1, Yi Pan1, Yifan Zhou1, Constance Owl4, Xiaoming Zhai5, Ninghao Liu6, Claudio Saunt4, Tianming Liu6
Aff.: 1School of Computing, The University of Georgia, Athens 30602, USA; 2Second Language Acquisition and Teaching, University of Arizona, Tucson, 85721, USA; 3Department of Mathematical and Statistical Sciences, University of Alberta, Edmonton, Canada; 4Department of History, The University of Georgia, Athens 30602, USA; 5Department of Mathematics, Science, and Social Studies Education, University of Georgia, Athens, GA, USA; 6National GENIUS Center, Athens, GA, USA

| Language | Model | BLEU | ROUGE-L |
|
|
|
Human Evaluation | ||||||
| Llama 3.1 405B | 0.0 | 0.0 | 0.931 | 0.927 | 0.929 | 0.0 | |||||||
| Cherokee | GPT-4o | 0.003 | 0.0 | 0.938 | 0.938 | 0.938 | 0.0 | ||||||
| GPT-4o + RAG | 0.115 | 0.117 | 0.962 | 0.964 | 0.963 | 0.0 | |||||||
| Llama 3.1 405B | 0.0 | 0.0 | 0.879 | 0.859 | 0.869 | 0.067 | |||||||
| Tibetan | GPT-4o | 0.0 | 0.0 | 0.833 | 0.851 | 0.842 | 0.147 | ||||||
| GPT-4o + RAG | 0.108 | 0.123 | 0.802 | 0.810 | 0.806 | 0.293 | |||||||
| Llama 3.1 405B | 0.0 | 0.104 | 0.693 | 0.663 | 0.678 | 0.040 | |||||||
| Manchu | GPT-4o | 0.0 | 0.125 | 0.726 | 0.703 | 0.714 | 0.173 | ||||||
| GPT-4o + RAG | 0.077 | 0.188 | 0.716 | 0.696 | 0.706 | 0.333 |