As can be seen from the data in the table, PEGCN has the highest classification accuracy in each dataset when λ is 0.3, that is, when the weight ratio of GCN to BERT is 0.3/0.7. The accuracy of the four datasets is 98.22%, 96.65%, 73.26%, and 89.59%, respectively. Compared with the case of a λ- value of 0.5, the accuracy is 0.04%,1.63%, 0.96%, and 2.08% higher, so λ is set to 0.3 herein because the large-scale pre-training model can significantly improve the classification effect, and assigning a larger weight is conducive to improving the accuracy. For specific datasets, GCN is also required for further feature extraction, so the best effect arises when the weight ratio is 0.3/0.7.
Based on TextGCN, the proposed method delivers an improvement. A position graph convolutional network (PGCN) is a GCN model that adds position information. A position and Bert graph convolutional network (PBGCN) is a network model that combines BERT based on the addition of position information. Meanwhile, experiments on GAT were conducted to verify the effectiveness of the proposed method. A position and Bert graph attention network (PBGAT) is a GAT network with position information and a model trained together with BERT. Finally, the method of PEGCN is GCN model using position information and edge features. The improved experimental results are listed in Table 3.
Table 3. Module validity experiment accuracy comparison. metric: accuracy (%)