Sarcasm, a prevalent linguistic style in social media and online commentary, employs exaggerated expression is somewhat agitating and easily leads to public opinion. Existing deep learning-based methods do not perform satisfactorily when dealing with sarcastic texts with complex expressions or that require specific background knowledge to understand. To meet this challenge, the present paper introduces a detection model that incorporates semantic dependency and syntactic dependency analysis. For specific background problems, TextCNN and BiLSTM are used in parallel to obtain local key features and global features of speech. For the problems of context dependency and complex language structure, the UniMP model is introduced to effectively extract and fuse semantic dependency and syntactic dependency information. Then, the multi-head self-attention mechanism is implemented to integrate these three parts to extract the intrinsic features of the speech. Ultimately, the output layer facilitates the completion of sarcasm detection. Experimental results indicate that this approach achieves an accuracy of 80.9%, surpassing the performance of the baseline model.