CMF-NERD: Chinese Medical Few-Shot Named Entity Recognition Dataset with State-of-the-Art Evaluation

Current works about medical few-shot named entity recognition (NER) predominantly focuses on English texts. There are also some supervised Chinese medical NER datasets available. The difficulty to share private data and varying specifications presented pose a challenge to this research. In this paper, We merged and cleaned multiple sources of Chinese medical NER dataset, then restructured these data into few-shot settings. CMF-NERD was constructed by weighted random sampling algorithm, containing 8,891 sentences and comprising 16 entity types. We adapted the most recent state-of-the-art few-shot learning methods and large language model for NER and conducted systematic experiments. The results indicate that the Chinese medical small-sample NER task is challenging and requires further research. Our further analysis provides promising directions for future studies.