Geometry of Semantic Space: Comparative Study of Discrete and Continuous Models
Researchers have published a comparative study on the semantic geometry of natural language processing models, contrasting supervised vector embeddings like CamemBERT with graph-based models. The study found that while transformer embeddings perform well, their semantic organization can be less clear than that of graph-based models. Applying the analysis to a French debate corpus revealed similar local structures but distinct overall topologies, suggesting that combining deep learning with graph structures could lead to more stable and interpretable AI. AI
IMPACT Suggests graph-based models may offer more interpretable semantic structures than current transformer embeddings.