A hybrid re-fusion model for text classification

Area of Science:

• Natural Language Processing
• Machine Learning
• Deep Learning

Background:

• Text classification is crucial for organizing text data.
• Existing models like BertGCN have limitations in handling long sequences and deep networks.
• BERT struggles with long-distance dependencies, while GCN faces over-smoothing issues.

Purpose of the Study:

• To propose the XLG-Net model for enhanced text classification performance.
• To overcome limitations of BERT and GCN in complex text classification tasks.
• To improve accuracy and robustness for both long and short texts.

Main Methods:

• Integrating XLNet for improved long-distance dependency capture and complex structure understanding.
• Utilizing GCNII to mitigate the over-smoothing problem in Graph Convolutional Networks.
• Applying the DoubleMix approach to XLNet for hybrid hidden state mixing.

Main Results:

• XLG-Net demonstrates significant performance improvements on four benchmark text classification datasets.
• The model effectively handles long-distance dependencies and complex language structures.
• Improved accuracy and robustness were observed for both long and short text classification.

Conclusions:

• XLG-Net offers a superior approach to complex text classification tasks.
• The integration of XLNet and GCNII effectively addresses previous model limitations.
• XLG-Net shows strong potential for advancing natural language processing applications.