Text simplification systems are told who they are writing for through a label: a reading-level tag such as "A2", a compression ratio, or an operation code such as <DEL>. A label is an arbitrary symbol. The model learns what it means only from many paired examples of hard and easy sentences, and for most audiences those pairs do not exist. This proposal treats the audience differently. We would describe the target reader in plain language and train one model to simplify for whatever reader the description names. A teacher, clinician, or accessibility editor could then define a new audience by writing a paragraph instead of collecting a parallel corpus. Examples of such a paragraph are the official CEFR descriptor of what an A2 learner can understand, or a professional guideline for readers with cognitive disabilities.
The idea comes from classification work, where swapping class IDs for natural-language descriptions lets a pretrained model use what it already knows about the categories. We bring that idea to generation under the condition that defines low-resource simplification: no aligned complex–simple pairs. Training data is manufactured by over-generating candidate rewrites and keeping only those that a description-matching verifier judges to fit the stated reader while preserving meaning. The claim we test is narrow and can fail. On audiences seen in training, descriptions will probably do no better than tags. The test that matters is audiences never seen in training: a new rubric or a held-out operation. There, describing the reader should work and a tag cannot.
Core Claim
A single simplifier, conditioned on the text of an audience description and trained only on verifier-filtered pseudo-parallel data, can hit targets that were never paired with training examples. The proof is a held-out simplification operation and a cross-rubric transfer to cognitive-accessibility simplification. In both cases discrete control tokens, ordinal level numbers, and zero-shot prompting of instruction-tuned models should fall short.
This differs from the closest work in three specific ways. Analysing Zero-Shot Readability-Controlled Sentence Simplification puts CEFR descriptors into prompts of frozen instruction-tuned models. It finds that the gains vary from model to model and that large level drops remain unsolved. It never trains a model to follow descriptors, which is what we do. Aligning Sentence Simplification with ESL Learner's Proficiency for Language Acquisition avoids parallel data by using reinforcement learning on a small model, with a separate low-rank adapter for each proficiency band. It names merging those adapters into one proficiency-aware model as future work. A description-conditioned model is one concrete way to do that merge, and it also extends past the three bands that adapter set covers. Cognitive Simplification Operations Improve Text Simplification distils professional accessibility manuals into seven operations but encodes them as opaque tokens. It assigns those tokens with noisy heuristics; precision on additions was low. We keep the manuals' definitions as the conditioning signal itself.
Research Program
The work has three parts.
The first part is a description-matching verifier, a model that reads a sentence and a description of a reader and says whether the reader could handle the sentence. Its design follows Sren Kierkegaard at SemEval-2023 Task 4: Label-aware text classification using Natural Language Inference. That system recasts multi-label classification as entailment between a premise and a handwritten hypothesis per label, and it improved substantially over a plain BERT classifier. We train the same kind of entailment model on CEFR-SP, the roughly 17,000 sentences that experts labelled A1 to C2. Each sentence is paired with the descriptor of its own level as a positive and with other levels' descriptors as negatives. That paper also supplies a warning. Entailment reformulation does not cure label skew: their rare label "Humility" reached only 0.18 F1. The rare extreme CEFR levels may behave the same way, so we balance negatives by level and report per-level calibration. To stop the model from memorising descriptor strings, every descriptor appears in several paraphrases written by an LLM and checked by hand.
The second part is pseudo-parallel data. We start from unaligned sentences: the complex side of WikiAuto and CEFR-SP sentences at B2 and above. An instruction-tuned model over-generates many rewrites per sentence under varied target descriptions. Each candidate is kept for a given description only if two checks pass. The verifier must judge that the candidate fits that description. Separately, semantic similarity to the source must clear a threshold, and the candidate must not be a near-copy; the zero-shot study found exact copying to be a common failure. This replaces MUSS's mining of one-to-one paraphrases (MUSS: Multilingual Unsupervised Sentence Simplification by Mining Paraphrases). Because the generator is free to split sentences, the pipeline can produce the one-to-many rewrites that one-to-one embedding alignment cannot recover.
For operations, we reuse the silver operation labels on WikiAuto from the cognitive-simplification work. We render each label as its manual definition, not as a token. Pairs whose labels the verifier contradicts are dropped, which should filter some of the heuristic noise.
The third part is the generator itself. We use one Phi-3-mini model, the same backbone as the ESL paper so the comparison is clean. It is fine-tuned on (description, source, filtered rewrite) triples. Optionally, a short reinforcement stage follows, rewarding the verifier score. In that stage the verifier is frozen and a separately trained level assessor is held out for evaluation, so the generator cannot game its own judge.
The generalisation tests are built into training. We withhold one of the seven cognitive operations, Explicitation, along with all of its definitional text. We also withhold the cognitive-accessibility register entirely. That register is evaluated only on FestAbility: 321 sentences simplified live by professional simplifiers, rich in reordering and explicitation. There the model receives a description written from the cognitive-simplification manuals and nothing else.
Baselines
The naive reference is copying the source sentence. Its SARI is low, but it shows how much of any CEFR-accuracy gain simply comes from inputs already near the target level. The second naive reference is the instruction-tuned model used to over-generate, applied directly with the description in its prompt. This tells us whether filtering and fine-tuning add anything beyond the generator that produced the data.
The strongest existing zero-shot method is GPT-4-Turbo with target level, descriptor, and examples, as in Analysing Zero-Shot Readability-Controlled Sentence Simplification. That paper reports 52.45% adjacency CEFR accuracy for this setting on CEFR-SP. We rerun the same prompt on our CEFR-SP test split for a matched comparison. Beating it means higher adjacency accuracy and lower RMSE, with no drop in semantic similarity or rise in exact copies, and especially on drops of four or more levels.
The strongest trained, parallel-free system for proficiency control is the multi-adapter reinforcement approach of Aligning Sentence Simplification with ESL Learner's Proficiency for Language Acquisition. It reports a LENS score of 67.29 for its A-level adapter against 41.37 for a T5 baseline. We train its three adapters on the same backbone and with the same unaligned source text. Our verifier stands in for its sentence-level reward model, so that only the conditioning mechanism differs. The two systems get matched GPU hours and the same hyperparameter search budget. Beating it means matching its LENS on seen bands with one model instead of three, and exceeding it where its design cannot reach.
On ASSET, the reference is MUSS, which reports 42.65 SARI. On FestAbility, the reference is BART-large with the operation classifier from Cognitive Simplification Operations Improve Text Simplification, which reports 27.13 SARI. On FestAbility our model sees no in-register training data. Matching 27.13 would therefore already be a meaningful result; exceeding it would be strong evidence.
The ablation that isolates the key ingredient is an identical model trained on identical filtered data, with the description replaced by a discrete tag. We run three variants. The first uses <A2>-style level tags and <EXPL>-style operation tags, the same scheme as CUT's control tokens (CUT: Controllable Unsupervised Text Simplification). The second uses an ordinal number from 1 to 6 for level. The third gives each label a mismatched description. If the mismatched-description model matches the real one, the descriptions are functioning only as IDs, and the core claim fails.
Why not just…?
Why not just use discrete control tokens, which already work? They do work on seen targets: CUT reports 46.88 SARI on Newsela, and operation tags lifted T5 from 32.01 to 38.13 SARI on ASSET. We expect our tag ablation to tie with description conditioning on seen CEFR levels, and we will say so if it does. A tag that never occurred in training carries no meaning, though, so the held-out Explicitation and FestAbility conditions are exactly where tags should fail. That contrast is the experiment.
Why not just prompt a strong instruction-tuned LLM with the descriptor? This has been tried. Gains from descriptors were inconsistent across models, and the best configuration reached 52.45% adjacency accuracy. The open question is whether training a model to follow descriptions fixes what prompting cannot. Our generator is built from that kind of prompting, so the direct-prompt baseline measures this exactly.
Why not just filter with an ordinary six-way level classifier and condition on level numbers? This is the most serious threat. An ordinal model can interpolate within the CEFR scale and may equal ours on CEFR-SP. It has no input for "a reader with a cognitive disability who needs implicit references made explicit." The only option would be to map that reader onto some CEFR number by hand, and we will try the best such mapping on FestAbility. If the hand-mapped ordinal model matches ours there, descriptions add nothing beyond a scale, and we would narrow the claim to robustness under paraphrased descriptors.
Why not just mine paraphrases as MUSS does? One-to-one embedding mining cannot produce sentence splits. ASSET reports that 20.2% of its references involve splitting. Comparing split rate and SARI on the splitting subset of ASSET against MUSS will show whether generate-and-verify recovers those transformations.
Proposed Experiments
The first decisive check is the held-out operation. We evaluate on WikiAuto–ASSET sentences whose references involve explicitation. The question is whether conditioning on the unseen Explicitation definition increases explicitation, measured by human annotation and the cognitive-simplification operation detector, while SARI holds. Both the tag model and the mismatched-description model should show no such increase.
The second check is FestAbility zero-shot: SARI against the reported 27.13, plus human ratings of simplicity and meaning preservation from annotators blind to the system.
The third is CEFR-SP control. We report adjacency accuracy, RMSE, and Spearman correlation, scored by an assessor trained separately from the verifier, broken down by the size of the level drop, and compared with GPT-4-Turbo and the three-adapter system. Each condition is run under original and paraphrased descriptors; a large drop under paraphrase would show the model is reading strings, not meaning.
Throughout, ASSET is the primary transformation-quality test, because its authors show that n-gram metrics penalise valid rewrites (ASSET: A Dataset for Tuning and Evaluation of Sentence Simplification Models with Multiple Rewriting Transformations). For that reason, SARI is paired with LENS and human judgement in every headline comparison.
References
- Analysing Zero-Shot Readability-Controlled Sentence Simplification
- Aligning Sentence Simplification with ESL Learner's Proficiency for Language Acquisition
- Cognitive Simplification Operations Improve Text Simplification
- Sren Kierkegaard at SemEval-2023 Task 4: Label-aware text classification using Natural Language Inference
- MUSS: Multilingual Unsupervised Sentence Simplification by Mining Paraphrases
- CUT: Controllable Unsupervised Text Simplification
- ASSET: A Dataset for Tuning and Evaluation of Sentence Simplification Models with Multiple Rewriting Transformations