The quantity of data increases exponentially with the advent of artificial intelligence which resulted in the emergence of a heterogeneous network data characterized by diverse structural complexity. The most models tend to demonstrate deficiencies in parsing intricate semantic relations within heterogeneous graphs, which consequently impact the effectiveness of classification tasks. To resolve this issue, this paper develops a deep Semantic Interaction-based Heterogeneous Graph Neural Network for Node Classification (SIHG-NC), with the objective of efficiently generating high-precision node representations to optimize the node classification task. Firstly, the heterogeneous network is initially deconstructed through a metapath strategy, with the objective of subdividing it into multiple subgraphs. Subsequently, a dynamic graph attention network GATv2 is introduced to implement graph convolution operations for each semantic subgraph, to extract node representations that contain specific semantic information. The model's most notable feature is the design of a semantic interaction scoring mechanism, which is capable of accurately quantifying the weights among different semantic subgraphs, and promoting the effective integration and fusion of features by dynamically adjusting the attention weights. Finally, the efficacy of the node classification of the model is evaluated on the basis of three authentic, heterogeneous web datasets.