Legal text is hard to read for a reason. Statutes, regulations and privacy policies are long and deeply nested because each extra clause pins down an obligation, a condition or an exception. Automatic simplification shortens and plainens these texts, but it can also change what they mean legally. A rewrite that turns "shall" into "may", or drops "unless the claimant was discharged for misconduct", reads better and says something different. Current simplification systems measure readability and surface similarity, so they cannot see this failure.
We propose to treat legal simplification as a multi-attribute rewriting problem with two kinds of attributes. Some must change: reading level, length and how deeply clauses are nested. Others must not change: the clause's legal force (obligation, permission, prohibition, or mere recommendation), its conditions, its exceptions, and who it applies to. We represent the attributes that must not change with natural-language concept descriptions, borrowed from interpretable few-shot classification. These descriptions do two jobs. They act as a checker that flags rewrites whose legal profile has drifted, and they act as a signal the generator is conditioned on. The checker is meant to be useful regardless of which generator wins, including a prompted frontier model.
Core Claim
A small library of legal concepts is enough to detect legal drift that generic meaning-preservation measures miss. The library has two parts: descriptions of what a clause type typically looks like, and descriptions of what separates easily confused types. Generic measures here include semantic difference scores, embedding similarity, and sentence-level entailment. Using this checker to filter and balance training data should let a controllable rewriter simplify statutes to a target reading level with fewer substantive legal errors.
The decisive test is extrinsic. A simplified statute is legally faithful if questions about the law still get the same answers when they are asked of the simplified version instead of the original.
Where This Sits in Prior Work
Three prior lines each supply one piece. None combines them.
Controllable simplification supplies the "attributes to change" machinery. Controllable Sentence Simplification (ACCESS) prepends discrete ratio tokens for length, paraphrase amount, word rarity and parse-tree depth. It reports 41.87 SARI on WikiLarge, but only on Wikipedia text and with no notion of content that must survive. Unsupervised Simplification of Legal Texts (USLT) brings simplification to legalese. It substitutes rare words using Legal-BERT and splits nested clauses into core and context sentences. It is evaluated only by readability grade and a surface semantic-difference score, so it never checks whether obligations survive.
Semantic label embedding supplies the checker. Adaptive Concept Discovery for Interpretable Few-Shot Text Classification (StructCBM) describes each class two ways: prototypical concepts say what a class looks like, and discriminative concepts say how to tell two confusable classes apart. An LLM refines these descriptions when classification fails, and a small embedding model is then fine-tuned. StructCBM is purely a classifier and was never used to guide or audit generation. Its known weakness is that discriminative pairs grow quadratically with the number of classes. That weakness barely matters here, because the set of legal-invariant categories is small (on the order of ten).
Legal NLP supplies both the motivation and the test bed. AI for Statutory Simplification introduces StateCodes, with 1.8M statutes and 2.2M regulations. It also introduces LaborBench, 3,700 expert-derived unemployment-insurance questions. On LaborBench the best retrieval-plus-LLM pipeline reaches only 0.691 F1, with about 73% citation reliability. LeSeR reports that dense retrievers confuse voluntary guidance with binding obligations. That is exactly the distinction our discriminative concepts target. Regulatory Compliance through Doc2Doc Information Retrieval shows that legal-concept supervision lets models see past boilerplate, but only for retrieval, not generation.
One result constrains the design. Multiple-Attribute Text Rewriting shows that attribute information remains recoverable at over 90% accuracy from supposedly disentangled latents. StyleFlow argues the opposite, using invertible flows. We do not try to settle that dispute inside legal text. Legal terms such as "shall" are both style and substance, so we do not hope that a latent space will separate them. Instead we enforce the invariants outside the generator, where they can be checked.
Research Program
Concept library. First we build a legal-invariant concept library at the clause level. The categories cover legal force (obligation, permission, prohibition, recommendation, definition), the presence and content of a condition, the presence and content of an exception, the responsible actor, and numeric or temporal thresholds such as deadlines, amounts and durations. For each category we write prototypical descriptions. For confusable pairs we write discriminative descriptions, such as "obligation vs. recommendation" and "prohibition vs. exception to a prohibition". The descriptions are seeded from statutory drafting conventions and refined with StructCBM's generate–predict–refine loop on a few hundred expert-labelled clauses from StateCodes unemployment-insurance titles.
Concept profiles. A small text–concept embedding model maps any clause $c$ to a concept profile $p(c)$: the vector of its similarity scores to each concept description. A rewrite can split one source clause $x$ into several sentences $y_1,\dots,y_m$. We first align these sentences to the source clause, then measure drift as
$$D(x,y)=\big\| p(x) - \textstyle\max_{j} p(y_j) \big\|_1 ,$$
where the element-wise maximum means a property counts as preserved if any output sentence carries it. Separately, threshold values are compared exactly after normalization. They are too brittle for embeddings, which echoes LeSeR's finding that legal precision needs exact matching.
Training data. Second, we build pseudo-parallel training data. A strong LLM rewrites StateCodes and privacy-policy clauses at several target reading levels. We keep only rewrites with low $D$ and matched thresholds. We then rebalance across cells defined by reading level and legal-force type, following Balancing the Effect of Training Dataset Distribution of Multiple Styles. That work found skewed style combinations cause failures on rare combinations. In legal text, "grade-6 reading level × conditional prohibition with exception" is exactly such a rare combination.
Rewriter. Third, we train a mid-sized sequence-to-sequence rewriter with ACCESS-style controls for length, word rarity and depth. It is also conditioned on the embedded concept descriptions of the source clause's profile, so the decoder is told in words what legal content it must carry over. This tests the open question flagged in the file: can semantic label embeddings steer generation, not just classification?
Baselines
Unedited original. The "do nothing" reference is the unedited original text. It has perfect legal fidelity by definition and the worst readability, so it anchors one end of the trade-off curve. Its LaborBench score is the reported 0.691 F1 for the Gemini retriever with DeepSeek V3 generator in AI for Statutory Simplification. We will re-run that exact pipeline on our state subset, so the comparison uses identical questions.
USLT. USLT is the strongest existing legal-specific simplifier. It is the only one in the file built for legalese, and it reports better readability scores than ACCESS, MUSS and LS-BERT. No numeric readability or semantic-difference values for it were found on Lacuna, so we will run it ourselves on the same clauses.
ACCESS on the same data. ACCESS is retrained on our pseudo-parallel legal data without concept filtering or concept conditioning. It isolates the value of control tokens alone.
Prompted frontier LLM. The strongest generator baseline is a frontier LLM prompted to simplify to the target grade while preserving every obligation, condition and exception. It is given a self-check pass. It receives the same few-shot examples and, for fairness, a comparable token budget to the one spent building our library and data.
Ablations of the key ingredient. Ablations replace our checker with three alternatives: a generic entailment filter, embedding-similarity and semantic-difference filters, and a random filter with the same acceptance rate. Further ablations drop the discriminative concepts, keeping only prototypical ones, and drop decoder conditioning.
What counts as winning. Beating the baselines means two things. First, at a matched target readability grade, we incur fewer expert-judged legal-invariant violations. Second, LaborBench F1 on the rewritten corpus falls less below the original-text F1 than it does for the prompted LLM and USLT. If the prompted LLM matches us on both, our generator offers no advantage, and the contribution reduces to the checker and the evaluation protocol. We will report that outcome as such.
Why not just…?
Why not just prompt a frontier LLM to preserve the obligations? This is the real threat, and we treat it as one. The evidence against it is indirect. BLESS finds that open models often copy their input, splitting or reordering sentences in only about 3.5% of edit operations versus about 12.5% for closed models. Its evidence that simplification and preservation trade off comes from Wikipedia, news and medicine, not law. The LaborBench error rate comes from question answering, not rewriting. So we do not assume the prompted LLM fails. We measure it with the same checker and the same LaborBench round-trip. Our checker is also how we would know it succeeds, and the checked LLM outputs are what our smaller rewriter is distilled from.
Why not just filter with a generic entailment or similarity model? Generic similarity may score "the employer shall notify" and "the employer may notify" as near-identical, because they share almost every token. We do not yet know how often generic sentence-level entailment catches such flips; Experiment 1 measures this. The guidance-versus-obligation confusion that LeSeR reports for dense retrievers is a warning sign. The checker-comparison experiment settles this directly. If a generic entailment model matches our concept checker's agreement with experts, we drop the library and use entailment.
Why not just use ACCESS control tokens on legal data? ACCESS controls how much the text changes but has no slot for what must not change. Its authors also note that its attributes stay correlated rather than independent. The ACCESS-on-same-data baseline and the no-conditioning ablation separate two questions: whether gains come from the controls, from the data filtering, or from concept conditioning.
Why not just use USLT's word-substitution pipeline? Substituting rare words rarely touches modal verbs. Its clause splitting can still orphan an exception from its rule, and it offers no target reading level. Running it under our invariant evaluation shows whether its apparent safety is real.
Proposed Experiments
Experiment 1: does the checker catch what matters? Legal annotators label about 800 source–rewrite pairs drawn from all systems for invariant violations: changed legal force, lost condition, lost exception, changed actor, changed threshold. We compare our concept checker, generic entailment, embedding similarity and semantic difference on how well each detects these violations, broken down by violation type. StructCBM reports 96.43% agreement between its automatic concept annotations and humans, but that was in news and other domains. We treat legal agreement as unknown until measured.
Experiment 2: the readability–fidelity curve. We generate rewrites of held-out StateCodes unemployment-insurance clauses at three target grades. We also generate rewrites of privacy-policy clauses, which have doubled in length over two decades per Privacy Policies over Time. Readability is measured by Flesch–Kincaid grade and Dale–Chall score. Fidelity is measured by expert violation rate. We also count edit operations, BLESS-style, to expose any copying.
Experiment 3: the extrinsic round-trip, which is the decisive check. We replace the statute corpus behind the LaborBench pipeline with each system's rewrites and re-ask all questions in our state subset. We report the F1 drop relative to originals and the change in citation reliability, broken down by state, given the reported spread from 0.52 to 0.81. As a secondary invariance check, we run LexGLUE's UNFAIR-TOS and LEDGAR classifiers (LexGLUE, code at coastalcph/lex-glue) on original and rewritten clauses. We report how often the predicted label stays the same.
How each outcome would be read. If our rewriter loses less LaborBench F1 than the prompted LLM at equal readability, the full proposal holds. If the two tie but our checker beats generic filters in Experiment 1, the contribution is the checker and the evaluation protocol. If generic entailment ties our checker, the concept library is not needed, and we will say so.
References
- Controllable Sentence Simplification
- Unsupervised Simplification of Legal Texts
- Adaptive Concept Discovery for Interpretable Few-Shot Text Classification
- AI for Statutory Simplification: A Comprehensive State Legal Corpus and Labor Benchmark
- 1-800-SHARED-TASKS at RegNLP: Lexical Reranking of Semantic Retrieval (LeSeR) for Regulatory Question Answering
- Regulatory Compliance through Doc2Doc Information Retrieval: A case study in EU/UK legislation where text similarity has limitations
- Multiple-Attribute Text Rewriting
- StyleFlow: Disentangle Latent Representations via Normalizing Flow for Unsupervised Text Style Transfer
- Balancing the Effect of Training Dataset Distribution of Multiple Styles for Multi-Style Text Transfer
- BLESS: Benchmarking Large Language Models on Sentence Simplification
- Privacy Policies over Time: Curation and Analysis of a Million-Document Dataset
- LexGLUE: A Benchmark Dataset for Legal Language Understanding in English