Scholarlink

<p>Text classification methods based on Graph Convolutional Networks (GCNs) have excelled in capturing structural relationships in texts. However, they face limitations in representing contextual semantics. In this study, we propose a text classification method that enhances GCNs with a Large Language Model (LLaMA2) for broad learning. The method integrates the global structural features extracted by the GCNs with the deep semantic features obtained after fine-tuning LLaMA2, overcoming the limitations of semantic understanding of GCNs. The fused features are subsequently reconstructed using a sparse BLS-based autoencoder, enhancing the model’s ability to capture high-dimensional semantic features. Finally, the reconstructed features are processed through the feature mapping and enhancement layers of the BLS, which improves generalization and completes the classification task, leading to consistent improvements in accuracy and robustness across multiple datasets. The experimental results demonstrate its advantages over existing models in multiple datasets, validating the effectiveness and adaptability of the proposed approach.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLMs enhanced graph convolutional broad sparse fusion networks for text classification

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Bolin Wang,
  • Xueyu Tang,
  • Jiana Meng,
  • Di Zhao,
  • Hongfei Lin

摘要

Text classification methods based on Graph Convolutional Networks (GCNs) have excelled in capturing structural relationships in texts. However, they face limitations in representing contextual semantics. In this study, we propose a text classification method that enhances GCNs with a Large Language Model (LLaMA2) for broad learning. The method integrates the global structural features extracted by the GCNs with the deep semantic features obtained after fine-tuning LLaMA2, overcoming the limitations of semantic understanding of GCNs. The fused features are subsequently reconstructed using a sparse BLS-based autoencoder, enhancing the model’s ability to capture high-dimensional semantic features. Finally, the reconstructed features are processed through the feature mapping and enhancement layers of the BLS, which improves generalization and completes the classification task, leading to consistent improvements in accuracy and robustness across multiple datasets. The experimental results demonstrate its advantages over existing models in multiple datasets, validating the effectiveness and adaptability of the proposed approach.