Transductive Text Classification With Concept Bipartite Graphs and LLM-Assisted Labeling

Graph-based text classification has become an effective paradigm for modeling corpus-level structure, particularly under limited supervision. However, most existing approaches rely on document&#x2013;word representations and do not jointly address semantic abstraction, connectivity robustness, and supervision scarcity in transductive settings. We investigate transductive text classification in low-label regimes using corpus-level document&#x2013;concept bipartite graphs, where documents are connected to automatically extracted keyphrases represented by contextual embeddings. This concept-level representation replaces traditional document&#x2013;word graphs with a more compact and semantically expressive structure. To ensure global information propagation, we introduce a <inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula>-nearest concept linking mechanism that enforces graph connectivity and prevents isolated subgraphs. Graph neural networks (GCN and GAT) are applied to this structure, and LLM-assisted labeling is incorporated as weak supervision by generating pseudo-labels with a resource-efficient language model and combining them with limited human annotations. Multilevel bipartite coarsening is used as a complementary strategy to reduce graph size and examine the accuracy&#x2013;efficiency trade-off. Experiments on 12 datasets show that attention-based message passing is generally more effective than standard convolutions on document&#x2013;concept graphs, that <inline-formula> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula>-NN concept linking enhances structural robustness, and that LLM-generated labels provide the greatest benefit in extremely low-label settings, with diminishing returns as human supervision increases.