Person re-identification (ReID) is a challenging task in computer vision, aimed at recognizing and tracking individuals across different scenes. The traditional approach involves feature extraction, preliminary sorting, leading to potential mismatches with persons distinctly different in critical features like color, posture, or clothing. To address this issue, we propose a novel S1 &S2 framework that combines traditional feature extraction with an advanced re-ranking method based on semantic analysis. Initially, we utilize backbone models from TransReID, IICS, and SINet as S1 for feature extraction, generating a preliminary ranking list. Next, we introduce the multimodal model S2, which utilizes adaptive position interpolation encoding and relies on extensive manual annotations to enhance the accuracy of semantic analysis, thus improving the re-ranking performance. By integrating the S1 and S2 models, our approach significantly enhances matching accuracy while maintaining computational efficiency with an O(n) complexity for re-ranking. Our contributions are summarized as follows: 1. We design a S1 &S2 framework that deeply analyses the semantic information of persons for re-ranking, thereby enhancing re-identification accuracy. 2. We create a carefully annotated dataset based on real-world images to train S2. 3. Our method consistently outperforms state-of-the-art methods in both supervised and unsupervised environments on large-scale person re-identification datasets.