Graph vs. Transformer Embeddings Show Similar Local but Divergent Global Topology
June 8, 2026
Comparing CamemBERT-style transformer embeddings against lexical co-occurrence graphs on a French civic corpus, transformer geometries show poor global distributional structure while graphs yield more interpretable semantic organization. Local topology is similar between the two approaches, but global structure diverges significantly, suggesting the two are complementary rather than interchangeable.
HOW THIS AFFECTS YOU
●
researcherWorth watching because it offers a methodology for comparing embedding topology vs. graph structure, with implications for interpretability and hybrid semantic modeling.