Paper 2 — Series 1: Embedding & Legal NLP

The Geometry of Legal Language

Embedding Structure Across Authoritative Dictionaries

Joel Thorarinson May 2026 16 pages 33 references arXiv-ready

Abstract

Legal terms carry different meanings depending on which dictionary defines them and which jurisdiction applies them. We embed 6,200 terms from Bouvier's Law Dictionary (1856), 46 from Webster's 1913 (general-language control), 19 cross-jurisdictional terms absent from English dictionaries, and 233 jurisdiction-scored leverage words spanning 27 jurisdictions. We report three findings. First, definitional drift: the same term defined by a legal dictionary and a general dictionary maps to measurably different embedding locations (mean drift 23.0% across 44 terms, up to 36.4% for “attainder”), and the same term across jurisdictions drifts up to 33.7% (“material change,” Nebraska vs. Alaska). Second, leverage–vulnerability correlation: legal leverage score correlates positively with PCA compression damage (rs = +0.22, p < 0.01, 95% CI [0.09, 0.34]), so the terms that determine custody, imprisonment, and constitutional rights are disproportionately destroyed by dimensionality reduction. Third, a model-level variance floor: PCA at 16 dimensions captures only 33% of variance for legal terms and 30.5% for Wikipedia articles under nomic-embed-text, a gap of 2.8 percentage points that is too small to attribute to domain effects; the 50-point gap with E5-Mistral benchmarks is architectural. Together, these results show that dictionary provenance and jurisdictional context — not just dimensionality — determine what a retrieval system can and cannot distinguish.
PDF Download Dictionary Collapse Explorer

Key Findings

Top 5 Definitional Drift Terms

TermSimilarityDrift
attainder0.63636.4%
consideration0.68431.6%
covenant0.69830.2%
chattel0.71029.0%
escrow0.71728.3%

Figures

Definitional drift heatmap across Bouvier's and Webster's legal terms
Figure 3: Drift heatmap. Cosine similarity between Bouvier's (legal) and Webster's (general) definitions for shared terms. Darker cells indicate greater definitional drift.
t-SNE visualization of Bouvier's vs Webster's embeddings
Figure 4: t-SNE visualization. Bouvier's legal terms (blue) and Webster's general terms (orange) in 2D projection. Overlapping terms highlight where legal and general meanings diverge.
Leverage score vs PCA compression damage correlation
Figure 9: Leverage–vulnerability correlation. Legal leverage score vs. PCA compression damage for 233 terms. Higher-stakes terms (custody, contempt) suffer more compression damage (rs = +0.22).

Citation

BibTeX
@article{thorarinson2026geometry,
  title={The Geometry of Legal Language: Embedding Structure Across Authoritative Dictionaries},
  author={Thorarinson, Joel},
  year={2026},
  month={May},
  pages={1--16},
  note={arXiv preprint (forthcoming)},
  keywords={legal NLP, dictionary embeddings, Bouvier's Law Dictionary, semantic drift, cross-jurisdictional}
}
APA
Thorarinson, J. (2026). The Geometry of Legal Language: Embedding Structure Across Authoritative Dictionaries. arXiv preprint (forthcoming).

Authors

JT
Joel Thorarinson
Coherence Research Group · ORCID 0000-0002-0553-842X

Keywords

legal NLP dictionary embeddings Bouvier's Law Dictionary definitional drift cross-jurisdictional PCA compression leverage words semantic drift embedding geometry

Related Papers