Recommendation systems rely on modeling user behavior to deliver personalized suggestions, with latent intention disentanglement being pivotal for performance enhancement. However, two challenges remain: 1) insufficient multidimensional semantic fusion for fine-grained intention representation; 2) implicit feedback sparsity limiting intention characterization. To address these issues, this paper proposes a dual-intent recommendation framework integrating knowledge graph and hypergraph modeling semantics (KHSDIR), comprising three modules: First, a hypergraph of user-item-group interactions is constructed to characterize social semantics, while linking item-side knowledge graphs to represent entity semantics; Second, social-semantic-enhanced user representations and knowledge-aware item representations are utilized to model collective intentions (capturing common preferences) and individual intentions (capturing personalized preferences), with a designed dual-intent fusion mechanism to achieve fine-grained characterizations; Finally, to mitigate the impact of data sparsity and enable better intention disentangling, graph contrastive regularization techniques are employed, along with a knowledge-based multi-view contrast and bidirectional intention-aware contrast mechanism to further constrain the consistency among users, items and intentions. Experiments on four datasets demonstrate that the KHSDIR has superiority over state-of-the-art methods, providing new insights for disentangled intention modeling.