Graph neural networks, as powerful tools for graph representation learning, have been widely applied in graph tasks such as node classification, link prediction, and recommendation. However, most existing graph neural network models focus on homogeneous graphs, overlooking the rich heterogeneity in real-world scenarios. With the increasing attention of researchers on heterogeneous information networks and heterogeneous graphs, designing heterogeneous graph neural network models that can effectively aggregate diverse node types and connection patterns from heterogeneous graphs, and integrate both neighbor node information and semantic information, has become a challenge. Inspired by attention mechanisms, this paper proposes the HAM-HGNN model. This model utilizes hierarchical attention aggregation, where the first layer employs self-attention aggregation to assign weights to neighbor nodes, and the second layer uses semantic-level attention to learn the importance of different meta-paths. Through hierarchical aggregation, the model generates node embeddings based on meta-paths and can be applied to heterogeneous graph data. To validate the effectiveness of this model, experiments were conducted on the real-world heterogeneous graph dataset ACM. The proposed HAM-HGNN model was compared with eight common graph neural network models. The results indicate that the HAM-HGNN model performs well in classification and clustering tasks.