Research on Fine-Grained Intelligent Classification and Grading Techniques for Data in the Traditional Chinese Medicine Industry

By constructing a terminology database covering 150,000 geo-authentic medicinal materials and 4,618 natural medicines, the thesis overcomes the adaptation bottleneck of general-purpose segmentation tools in professional terminology boundary splitting.Finally, a closed-loop multimodal intelligent annotation model of “collection—cleaning—pre-annotation—manual review—fine-tuning—application” is designed.

The Segment Anything Model (SAM) is integrated to achieve zero-shot image segmentation. The Self-Instruct method is utilized to synthesize low-resource annotation data. Cross-modal knowledge communities are constructed through Graph Retrieval-Augmented Generation (Graph RAG) to support hybrid retrieval and intelligent question answering.

In large-scale practice at Yunnan Baiyao Group, this system improved data annotation efficiency by 30 times, enhanced medicinal material quality inspection efficiency by 50%, and saved annual operating costs exceeding ten million yuan. The diagnosis, treatment, and marketing assistant based on Graph RAG has been promoted and applied in thousands of pharmacies. This thesis provides a systematic technical solution and engineering paradigm for the secure circulation and value transformation of data elements in the TCM industry.