Speaking at the Right Level: Literacy-Controlled Counterspeech Generation with RAG-RL

Abstract
Health misinformation poses a critical risk to public well-being. Existing counterspeech approaches often generate uniform responses that fail to account for the audience’s health literacy level, limiting accessibility and effectiveness.

We propose Controlled-Literacy, a framework that combines Retrieval-Augmented Generation (RAG) with Reinforcement Learning (RL) to adapt counterspeech to different health literacy levels (low, medium, high). The framework retrieves evidence aligned with the target literacy level and optimizes generation using a hybrid reward function incorporating both readability (Flesch-Kincaid Reading Ease) and user preference ratings.

Experiments demonstrate that Controlled-Literacy outperforms baselines in Politeness, Readability Alignment (Target Distance), User Preference, and Factual Accuracy, producing counterspeech that is accessible, respectful, and factually grounded. Cross-evaluation further shows that user preference peaks when counterspeech is tailored to the correct literacy level.

Keywords: Health Literacy, Counterspeech, Retrieval-Augmented Generation, Reinforcement Learning, Trustworthy AI

Contributions

1. Health Literacy-Aware Counterspeech Framework
   • Introduces Controlled-Literacy, integrating RAG + RL to tailor responses to users’ health literacy.
2. Hybrid Reward Design
   • Combines readability-based scoring with simulated user preference ratings for effective optimization.
3. Dataset Creation – MisinfoLiteracy
   • A curated dataset of 440 Reddit posts with health misinformation, annotated and paired with counterspeech at different literacy levels.
4. Empirical Validation
   • Achieves significant gains across politeness, factual accuracy, and target readability alignment compared to RAG and prompting-only baselines.

Method Overview

Figure: The Controlled-Literacy framework pipeline.

• Evidence Retrieval Module filters documents based on readability (FKRE) and user preference scores.
• Counterspeech Generator produces responses optimized with Group Relative Policy Optimization (GRPO).
• Reward Function balances readability alignment with subjective accessibility.

Experimental Results

Controlled-Literacy demonstrates superior performance:

• Politeness: 0.84 (LLaMA-8B) vs. 0.41 in prompt-only baseline
• Target Distance (↓): 0.90 vs. 2.74 baseline (closer to intended readability)
• User Preference: 0.74 vs. 0.71 (RAG baseline)
• Factual Accuracy: 0.91 vs. 0.87 (prompt-only baseline):contentReference[oaicite:1]{index=1}

Cross-evaluation confirms that audiences rate responses highest when counterspeech matches their literacy level, reinforcing the need for tailored generation.

Resources

• 📄 Paper: arXiv:2509.01058v3

BibTeX

@inproceedings{song2025literacy,
  title={Speaking at the Right Level: Literacy-Controlled Counterspeech Generation with RAG-RL},
  author={Song, Xiaoying and Anik, Anirban Saha and Barua, Dibakar and Luo, Pengcheng and Ding, Junhua and Hong, Lingzi},
  booktitle={Findings of the Association for Computational Linguistics: EMNLP 2025},
  year={2025}
}

Recommended citation: Xiaoying Song, Anirban Saha Anik, Dibakar Barua, Pengcheng Luo, Junhua Ding, Lingzi Hong. (2025). "Speaking at the Right Level: Literacy-Controlled Counterspeech Generation with RAG-RL." Findings of the Association for Computational Linguistics: EMNLP 2025.
Download Paper