Abstract: To address issues in relation extraction tasks, such as insufficient semantic representation, underutilization of structural knowledge, and prediction results prone to logical inconsistencies, this paper proposes a method that combines graph topology and semantics fusion with logical regularization. This method builds a dual-stream collaborative encoding framework composed of a text semantics branch, an R-GCN graph structure branch, and an adaptive gated fusion module. While modeling text context representations, it incorporates entity adjacency, relation types, and structural dependency information from domain relation graphs to enhance the ability to judge candidate entity pair relations. Furthermore, this paper transforms logical rules like relation type priors, entity type constraints, and relation exclusivity into differentiable regularization terms, which are jointly optimized with cross-entropy loss, allowing these rules to participate in parameter learning as soft constraints during training, thereby reducing the probability of candidate triples that violate the rules. Experimental results show that the proposed GTS-LR method achieved a precision of 81.30%, which is 3.90 percentage points higher than CasRel, with an F1 score of 72.60%. This method is particularly useful in domain knowledge base construction scenarios where tolerance for incorrect triples is low.