Legal documents are hard to cluster automatically: they use specialized jargon, vary widely in length, and rarely come with expert annotations. A new arXiv preprint proposes Topic-Graph Spherical Fusion (TGSF), which treats each document as both a bag of topics and a node in a relational graph. The method builds a connected multi-relational document-topic graph using topic centroid bridges and temperature-sharpened k-nearest-neighbor edges, then fuses the two L2-normalized representations before KMeans partitioning.

On three legal benchmarks, TGSF improved normalized mutual information over semantic-only, structural-only, and classical baselines: from 0.606 to 0.653 on LEDGAR, from 0.228 to 0.284 on ACORD, and from 0.535 to 0.570 on CUAD. Adjusted Rand index scores were 0.425, 0.260, and 0.282 respectively. The authors attribute the gains to jointly modeling topical content and document relationships rather than relying on either signal alone.

The study also includes a large-scale sensitivity analysis—648 factorial grid runs plus 720 Bayesian and Pareto optimization trials—to identify which hyperparameters matter most. Because this is a single preprint, the results have not yet been independently replicated, but the consistent gains across three datasets suggest the hybrid approach is worth attention for legal text mining and other low-annotation domains.