Hierarchical Graph Convolutional Networks for Structured Long Document Classification.

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning
• Artificial Intelligence

Background:

• Long document classification (LDC) is crucial due to the increasing volume of text data.
• Current LDC methods often overlook document structure, limiting performance on structured texts.
• Pretrained language models have advanced LDC but struggle with inherent document organization.

Purpose of the Study:

• To develop a novel method for structured long document classification.
• To effectively incorporate document structure into LDC models.
• To improve the representation of long documents with inherent organizational information.

Main Methods:

• Proposed a hierarchical graph convolutional network (HGCN) for structured LDC.
• Developed a section graph network to model document macrostructure.
• Designed a word graph network with a decoupled graph convolutional block for fine-grained feature extraction.
• Introduced an interaction strategy to integrate macrostructure and fine-grained features.

Main Results:

• The proposed HGCN model demonstrated superior performance on structured LDC tasks.
• Experiments on four structured and one unstructured dataset confirmed the model's effectiveness.
• The method outperformed existing state-of-the-art classification techniques.

Conclusions:

• The hierarchical graph convolutional network (HGCN) effectively leverages document structure for improved LDC.
• Integrating macrostructural and fine-grained features enhances text representation.
• The proposed approach offers a significant advancement for classifying structured long documents.