态靶辨治语料的深度学习标注方法

Deep Learning Annotation Methods for the Corpora of State-Target Differentiation and Treatment

摘要: 针对中医“态靶辨治”理论语料标注中存在的术语复杂性、表达灵活性及领域知识依赖等挑战，本研究提出一种融合深度学习的两阶段标注方法体系。首先，基于BERT + BiLSTM + CRF模型实现高精度命名实体识别，准确抽取疾病、证候、态、靶药等六类核心实体(F1值达72.06%)；其次，通过Attention + BiLSTM模型完成实体关系抽取，精准捕捉“疾病–证候–靶药”间的诊疗逻辑关系(F1值达98.23%)。实验表明，该方法显著提升标注效率与一致性，为构建态靶辨治知识图谱及开发智能诊疗系统提供了可靠的技术支撑。

Abstract: To address challenges in annotating Traditional Chinese Medicine (TCM) corpora for “State-Target Differentiation and Treatment” (STDT), including terminological complexity, expressive flexibility, and domain knowledge dependency, this study proposes a dual-stage deep learning annotation framework integrating deep learning. First, a BERT + BiLSTM + CRF model achieves high-precision named entity recognition, accurately extracting six core entities, such as diseases, syndromes, states, and target herbs (with an F1-score of 72.06%). Second, an Attention + BiLSTM model completes the extraction of entity relationships, accurately capturing diagnosis and treatment logical relationships between “disease-syndrome-target herb” (with an F1-score of 98.23%). Experiments demonstrate significant improvements in annotation efficiency and consistency, providing a robust technical foundation for constructing STDT knowledge graphs and developing intelligent diagnosis and treatment systems.