This paper proposes the SAD-SP model, which improves open-world compositional zero-shot learning by capturing contextuality and feasibility dependencies between states and objects. Using semantic attention and knowledge disentanglement, the approach …