Data relationships and the impact of synthetic loss have not been concerned by previous sample generation methods, which lead to bias in model training. To address above problem, in this paper, we propose a relationship-aware hard negative generation (RHNG) method. First, we build a global minimum spanning tree for all categories to measure the data distribution, which is used to constrain hard sample generation. Second, we construct a dynamic weight parameter which reflects the convergence of the model to guide the synthetic loss to train the model. Experimental results show that the proposed method outperforms the state-of-the-art methods in terms of retrieval and clustering tasks.