Recently, contrastive learning has shown promising performance in multi-view clustering. However, existing methods suffer from the generation of false negative pairs due to a negative sample construction mechanism that overlooks semantic consistency, leading to conflicts with the clustering objective. To address this limitation, we propose a novel framework called Deep Contrastive Multi-view Clustering under Semantic Feature Guidance (DCMCS). Our framework extracts view-specific features from raw data and fuses them to create a fusion view. Considering the consistency of instance cluster labels among views, specific view and fusion view semantic features are learned by cluster-level contrastive learning and concatenated to obtain instance pair weights measuring the semantic similarity of instances. By adopting instance pair weights, DCMCS adaptively weakens the impact of false negative pairs in instance-level contrastive learning. Additionally, DCMCS utilizes the fusion views as the anchor to alleviate the influence of view differences. Extensive experiments on multiple public datasets demonstrate that DCMCS significantly outperforms state-of-the-art methods, showcasing its effectiveness and robustness in multi-view clustering tasks.