This study aimed to compare the performance of large language models (LLMs) like GPT with traditional deep learning models (LSTM and BioBERT) in extracting relations related to acupuncture point locations. The researchers utilized the World Health Organization Standard Acupuncture Point Locations in the Western Pacific Region (WHO Standard) as the data source and manually annotated five types of relations: 'direction_of,' 'distance_of,' 'part_of,' 'near_acupoint,' and 'located_near.'
The results showed that the fine-tuned GPT-3.5 model consistently outperformed the other models, achieving the highest micro-average F1 score of 0.92. The pre-trained GPT-4 model, however, exhibited lower performance compared to the fine-tuned GPT-3.5. The error analysis revealed that the fine-tuned GPT-3.5 model struggled the most with accurately identifying the 'near_acupoint' relation, suggesting challenges in capturing complex spatial relationships between acupuncture points.
The study highlights the effectiveness of LLMs, particularly fine-tuned GPT models, in extracting relations related to acupuncture point locations. This has implications for accurately modeling acupuncture knowledge, promoting standard implementation in acupuncture training and practice, and advancing informatics applications in traditional and complementary medicine.
他の言語に翻訳
原文コンテンツから
arxiv.org
深掘り質問