Text classification model based on GNN and attention mechanism
Addressing the issue of low classification accuracy raised by the poor performance of the model, which is caused by the difficulty in learning from dynamic aggregation unknown neighboring nodes of graph data and insufficient fusion of semantic features, a model named graph attention text classificat...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | zho |
| Published: |
Beijing Xintong Media Co., Ltd
2025-05-01
|
| Series: | Dianxin kexue |
| Subjects: | |
| Online Access: | http://www.telecomsci.com/thesisDetails#10.11959/j.issn.1000-0801.2025136 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Addressing the issue of low classification accuracy raised by the poor performance of the model, which is caused by the difficulty in learning from dynamic aggregation unknown neighboring nodes of graph data and insufficient fusion of semantic features, a model named graph attention text classification(GATC) based on graph neural network (GNN) and attention mechanism was proposed. Firstly, an inductive learning of graph neural network model was constructed, and dynamic embedding the unknown neighboring node was implemented by using an aggregation function to enhance the model’s generalization ability. Secondly, the reasoning cache size of key-value was reduced by the introduction of multi-head latent attention mechanism that utilized the low-rank key-value joint compression technology, which significantly diminished memory usage and improved the performance of the model. Finally, the integration of GNN and gated recurrent unit (GRU) network models further captured the semantic feature information of structural and temporal attributes for graph data, resulting in achieving efficient feature fusion and improving the classification accuracy of the model. The experimental results show that the proposed method not only is effective, but also improves the accuracy of classification that is increased at least 4.0%, 2.4% and 3.1% on the CSI 100，CSI 300 and Rus 1K datasets, respectively, compared with the algorithm ADGL+MLA (adaptive dynamic graph learning+multi-head latent attention). |
|---|---|
| ISSN: | 1000-0801 |