Improved Chinese Few-Shot relation extraction using Large Language Model for Data Augmentation and Prototypical Network

Chinese few-shot relation extraction aims to effectively identify relationships between entities in text using limited data, which is crucial for information extraction and knowledge construction and assists in extracting critical information from medical cases to support personalized rehabilitation treatment plans. However, due to the limited number of samples, existing methods struggle to capture sufficient relational features from limited data, resulting in poor extraction performance. Therefore, we propose an Improved Chinese Few-Shot Relation Extraction Using Large Language Model for Data Augmentation and Prototypical Network to address this issue. Specifically, we establish the Chinese few-shot relation extraction datasets DUIE Few and SanWen Few. Notably, we introduce a framework based on large language models for dataset augmentation, which effectively alleviates the problem of feature extraction due to insufficient data and improves task performance. Finally, we present baselines for prototype networks, Siamese networks, and the CFSRE model based on relational category information. Experimental results show that the CFSRE model improves accuracy, recall, and F1 score under few-shot conditions, particularly as the sample size decreases. In summary, the method we propose demonstrates promising results in Chinese few-shot relation extraction tasks and holds the potential to advance medical rehabilitation research.