Medical Entity Recognition with Few-Shot Based on Chinese Character Radicals

In medical text entity recognition tasks, Chinese character radicals are often closely related to the semantics of the characters. Based on this insight, we proposed the CSR-ProtoLERT model to integrate Chinese character radical information into few-shot entity recognition to enhance the contextual representation of the text. We optimized the pre-training embeddings, extracted radicals corresponding to Chinese characters from an online Chinese dictionary for the extensive collection of medical texts we acquired, and stored these radicals as key-value pairs. Concurrently, we employed CNN to optimize the radical embedding representation. We input the static embedding vectors of multiple Chinese characters sharing the same radical into the CNN network, extracting common feature representations for Chinese characters, ultimately obtaining the embedding representation of the Chinese character radicals. The Cross Star-Transformer model we proposed employs two Star-Transformers to model the embeddings of the input medical text character sequence and the corresponding radical sequence embedding. It fuses the Chinese character radical features with the character features, enabling the few-shot entity recognition model to learn more about medical Chinese character entity features. In the CMF 5-way 1-shot and 5-way 5-shot scenarios of the Chinese medical text few-shot entity recognition dataset we constructed, we achieved F1 values of 54.07% and 57.01%, respectively.