Traditional Chinese medicine (TCM) patents, as significant achievements in TCM technological innovation, contain rich semantic information and complex structures. The effective extraction of entity relations in TCM patents is of great importance for the efficient utilization of TCM knowledge. To effectively address the issues of complex entity relations, diverse relation semantics, and insufficient semantic interaction in TCM patent texts, this paper proposes an improved model based on a large language model and a diversified semantic interaction strategy to accurately extract entity relations such as TCM preparation and pharmacological mechanism from TCM patent texts. The model uses the Qwen3-Embedding-8B large language model as an encoder, providing deep representations of subject-object entities and relations through deep semantic modeling. It also designs a dual-channel cross-attention mechanism to precisely capture and enhance the modeling of associations between entities and relations through diversified semantic interaction. Additionally, an adversarial learning strategy is introduced to solve noise and long-tail data distribution issues in TCM patent texts. The experimental results show that the proposed model surpasses current benchmarks when applied to TCM patent texts. It also aids in the construction of TCM knowledge databases and contributes to research and development decision-making.