Automatically extracting unknown drug-drug interactions from biomedical literature can update the drug database quickly,which is of great importance and medical value in application. Existing neural network mod? els often can only learn a single one-sided feature in a certain aspect from sentence sequences or other external infor? mation,but it is difficult to fully mine the potential long-distance dependency features from sentences to obtain a comprehensive feature representation. This paper proposes a novel drug-drug interaction extraction method combin? ing semantics and dependency. In this method,we not only use the Bi-GRU network to learn the semantic feature representation from the sentence sequence and the shortest dependency path of the target drug entities,but also combine the multi-head self-attention mechanism to further capture the potential dependencies between words. Finally, these multi-source features are fully fused to effectively improve the performance of drug-drug interaction extraction. The experimental results on the DDIExtraction-2013 dataset show that our method outperforms other existing meth? ods and obtains an F1 value of 75.82%.