Legal Semantic Engineering: Reconciling Probabilistic Generation with Rigid Normative Constraints

ACL ARR 2026 January Submission2744 Authors

Keywords: Legal Machine Translation, Deontic Drift, Multi-Agent Systems, Zipf-Mandelbrot Law, Hierarchical Control.

Abstract: Legislative translation demands the precise preservation of normative intent. However, Large Language Models (LLMs) frequently suffer from **Deontic Drift**-a systemic failure where models prioritize probabilistic fluency over rigid normative mandates. By analyzing a five-jurisdiction benchmark via Zipf-Mandelbrot modeling, we characterize this failure as a structural distributional mismatch: the high-concentration mandatory monopoly of source legal terms diverges significantly from the granular, dispersed probability distributions of target languages. To bridge this gap, we propose **Legal Semantic Engineering (LSE)**, a framework that introduces vertical hierarchical control as a robust alternative to horizontal multi-agent collaboration. Through an Anchoring-Shaping-Polishing (ASP) pipeline, LSE explicitly decouples normative logic validation from stochastic text generation. Experiments on a trilingual legislative benchmark demonstrate that LSE is highly robust to backbone variations; implementations using DeepSeek, GPT, and Gemini all significantly surpass strong horizontal agent baselines. Furthermore, our analysis unveils the gain-interference-rescue dynamics, quantitatively illustrating the necessary trade-offs between linguistic fluency and legal fidelity.

Paper Type: Long

Research Area: Machine Translation

Research Area Keywords: Legal NLP, Constrained Decoding, Evaluation Metrics

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models, Data resources, Data analysis

Languages Studied: Chinese, English, Japanese, German

Submission Number: 2744