AI 研究arXiv cs.LG21 Oct, 12:00 pm SGTSIGNAL 7BB5C0

Topic-Graph Spherical Fusion Improves Unsupervised Legal Document Clustering

ORIGINAL / Hybrid Topic-Semantic Labeling and Graph Embeddings for Unsupervised Legal Document Clustering

Introduces TGSF, a framework that fuses dense topic-semantic discovery with document relational graph embeddings on a joint unit hypersphere for unsupervised legal document clustering. It outperforms semantic-only, structural-only, LDA, NMF, and TF-IDF baselines on three legal benchmarks, improving NMI and ARI, and provides hyperparameter sensitivity diagnostics.

01 ABSTRACT

The paper proposes TGSF, a dual-manifold representation learning framework that obtains topic-semantic and graph-structural representations via Top2Vec and Node2Vec, L2-normalizes them, fuses them on a unit hypersphere, and clusters with KMeans. On ACORD, CUAD, and LEDGAR, TGSF improves NMI and ARI over semantic-only, structural-only, LDA, NMF, and TF-IDF baselines. The authors also characterize hyperparameter sensitivity using functional ANOVA and Pareto frontier diagnostics.

02 KEY FINDINGS

TGSF combines Top2Vec topic semantics with Node2Vec graph embeddings via modality-weighted concatenation on a joint unit hypersphere.
On legal document benchmarks, TGSF improves NMI 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.
Hyperparameter sensitivity is evaluated via 216 factorial grid configurations and 240 Bayesian/multi-objective optimization trials per dataset (720 trials total).

03 RELEVANCE

Legal documents are annotation-scarce and terminology-heavy, making unsupervised clustering useful for knowledge discovery. TGSF offers a dual-manifold fusion approach that does not rely on expert labels, validates the joint use of semantic and structural information on multiple legal benchmarks, and provides hyperparameter analysis references for reproducible legal text mining.

04 EVIDENCE BOUNDARY

Evaluation covers only three English legal benchmarks (ACORD, CUAD, LEDGAR); generalization to other languages, jurisdictions, or document types is unknown. The method still depends on hyperparameter settings such as the number of KMeans clusters, and comparisons are limited to the listed baselines without contrast to more recent deep clustering methods; the ANOVA and Pareto diagnostics characterize sensitivity only and do not provide causal explanations.

Return to the primary source

AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.CL · ARXIV CS.LG

Read original ↗