Most representation learning on text-attributed graphs treats semantics as a flat space, even though human knowledge is organized hierarchically. A new arXiv preprint argues that this flat assumption overlooks the coarse-to-fine granularity inherent in concepts, and proposes a framework that explicitly recovers such hierarchies.
The method, described as a variational EM-based LLM-GNN synergy, alternates between a large language model and a graph neural network to jointly infer latent hierarchical structure. The LLM contributes textual understanding while the GNN captures graph topology, and the variational EM procedure coordinates the two.
Because the abstract is truncated, specific experimental results are not available in the source. The contribution, as described, is conceptual: a way to move beyond flat semantic spaces in graph representation learning, with potential implications for tasks that depend on taxonomic or part-whole relationships.