[ACL 2026] Right at My Level: A Unified Multilingual Framework for Proficiency-Aware Text Simplification

Code and data for “Right at My Level: A Unified Multilingual Framework for Proficiency-Aware Text Simplification”. [paper]

• Venue: ACL 2026 Main Conference
• Authors: Jinhong Jeong, Junghun Park, Youngjae Yu

About

This repository provides the vocabulary data and the training code used in the paper “Right at My Level: A Unified Multilingual Framework for Proficiency-Aware Text Simplification”.

Overview

We propose Re-RIGHT, a unified reinforcement learning framework for adaptive multilingual text simplification without parallel corpora to help second language (L2) learners. The Re-RIGHT framework overcomes the limitations of state-of-the-art LLMs at easier proficiency levels and in non-English languages.

Highlights

• We show that even state-of-the-art LLMs (GPT-5.2, Gemini 2.5) struggle with lexical simplification at easier proficiency levels and in non-English languages.
• Trained with our Re-RIGHT framework, a compact 4B policy model outperforms much larger LLM baselines, aligned with the proficiency standards across four languages (English, Japanese, Korean, and Chinese).

Usage

Run all commands from the repository root.

Setup

pip install -r requirements.txt
python -m nltk.downloader stopwords

Data preprocessing

The cleaned vocabulary lists are included in data/wordlist_{en,ja,ko,zh}.csv. To build the Wikipedia dataset used for training and evaluation:

python src/preprocessing/download_articles.py
python src/preprocessing/parse_wikitext.py
python src/preprocessing/make_dataset.py

This creates data/wikipedia/dataset/all/.

Zero-shot inference

python src/experiments/zero-shot_test.py \
  --model Qwen/Qwen3-4B-Instruct-2507 \
  --gpu 1

Outputs are saved to results/llm_test/.

For API-based models, set API_BASE_URL and API_KEY in .env.local, then run:

python src/experiments/zero-shot_test_api.py --model gpt-5.2

GRPO training

The training script starts a vLLM evaluator server and launches GRPO training:

bash scripts/train_grpo.sh

By default, it trains Qwen/Qwen3-4B-Instruct-2507 and writes checkpoints to results/grpo/Qwen3-4B-Instruct-2507/. The script assumes multiple GPUs; adjust CUDA_VISIBLE_DEVICES, MODEL_ID, EVALUATOR_MODEL_ID, and OUTPUT_DIR as needed.

Evaluation

Run evaluation on any available result files:

bash scripts/evaluate_all.sh

Evaluation summaries and sample-level scores are saved to results/evaluation/.

Citation

If this repository supports your research, please cite:

@inproceedings{jeong-etal-2026-right,
    title = "Right at My Level: A Unified Multilingual Framework for Proficiency-Aware Text Simplification",
    author = "Jeong, Jinhong  and
      Park, Junghun  and
      Yu, Youngjae",
    editor = "Liakata, Maria  and
      Moreira, Viviane P.  and
      Zhang, Jiajun  and
      Jurgens, David",
    booktitle = "Proceedings of the 64th Annual Meeting of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.1086/",
    pages = "23680--23706",
    ISBN = "979-8-89176-390-6",
    abstract = "Text simplification supports second language (L2) learning by providing comprehensible input, consistent with the Input Hypothesis. However, constructing personalized parallel corpora is costly, while existing large language model (LLM)-based readability control methods rely on pre-labeled sentence corpora and primarily target English. We propose Re-RIGHT, a unified reinforcement learning framework for adaptive multilingual text simplification without parallel corpus supervision. We first show that prompting-based lexical simplification at target proficiency levels (CEFR, JLPT, TOPIK, and HSK) performs poorly at easier levels and for non-English languages, even with state-of-the-art LLMs such as GPT-5.2 and Gemini 2.5. To address this, we collect 43K vocabulary-level data across four languages (English, Japanese, Korean, and Chinese) and train a compact 4B policy model using Re-RIGHT, which integrates three reward modules: vocabulary coverage, semantic preservation, and coherence. Compared to the stronger LLM baselines, Re-RIGHT achieves higher lexical coverage at target proficiency levels while maintaining original meaning and fluency."
}

For any questions or issues, feel free to contact the authors!