Abstract: Term extraction methods in middle school mathematics do not consider the dependency structure information of the sentence, which leads to a limited semantic understanding of the sentence. In addition, the term extraction method of dependency structure information exists some errors in dependency structures, resulting in poor accuracy and completeness. This paper proposes a robust extraction model for middle school mathematical terms based on dependency structure learning. The dependency structure information and context information of text is captured by the graph neural network model with a denoising attention mechanism layer and bidirectional recurrent neural network, respectively. The attention mechanism is further used to integrate text structure information and context information to integrate the dependency structure information while alleviate the influence of incorrect word segmentation or dependency structure. The proposed model achieves 83.82% precision and 82.51% and F₁ value in the experiment, indicating 2.21% and 1.22% improvements compared with the baseline model, respectively.