UTCMKG: A Lightweight Universal Model for Constructing Traditional Chinese Medicine Case Knowledge Graphs

This study introduces UTCMKG, a lightweight universal model designed to automate the extraction and structuring of unstructured Traditional Chinese Medicine (TCM) case records into knowledge graphs. By integrating ALBERT for context-aware semantic encoding, BiLSTM for sequential feature extraction, self-attention mechanisms to capture long-range dependencies, and CRF for structured prediction, UTCMKG addresses the complexity of TCM terminology (e.g., syndrome patterns, herbal prescriptions) and non-standardized documentation. Hyperparameter optimization (batch size=128, Adam optimizer, learning rate=0.0001) yielded a 81.3% reduction in training time compared to conventional BERT-based models, achieving a precision of 0.8472, recall of 0.8397, and F1-score of 0.8434 across nine entity types (e.g., symptoms, treatments) and eight relation types (e.g., treatment-syndrome associations). Experimental validation against baseline models (e.g., Word2vec-BiLSTM-CRF, Bert-BiLSTM-CRF) confirmed superior performance, particularly in recognizing rare or context-sensitive entities critical for TCM diagnosis. The model’s scalability and efficiency enable systematic digitization of TCM knowledge, supporting clinical decision-making, data mining, and standardization efforts.