Reproduce “Graph Convolutional Networks for Text Classification” on OpenResearch

About this paper

Liang Yao, Chengsheng Mao, Yuan Luo

A novel Graph Convolutional Network (GCN) architecture, Text GCN, is introduced for text classification, modeling an entire corpus as a unified graph of words and documents. This approach achieves state-of-the-art accuracy on several benchmark datasets and demonstrates robust performance in low-resource, semi-supervised settings by learning predictive embeddings directly from the corpus.

Key insights

• Representing the entire text corpus as a heterogeneous graph allows GCNs to inherently capture global word co-occurrence patterns that traditional sequential models often miss.
• By starting with one-hot initial node features, the model learns all necessary word and document embeddings directly from the target corpus under supervised guidance, reducing reliance on external knowledge.
• The semi-supervised training objective effectively propagates label information through the graph, enabling strong classification performance even with a small subset of labeled documents.

Summary by alphaXiv.