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

TL;DR

This work tackles health misinformation by introducing Controlled-Literacy, a literacy-aware counterspeech generator built on retrieval-augmented generation (RAG) and reinforcement learning (RL). It uses health-literacy level as a controllable factor, employing FKRE-based readability and simulated user preferences to shape outputs for low, medium, and high literacy audiences. A new MisinfoLiteracy dataset, along with cross-dataset evaluations on MisinfoCorrect and Check-COVID, demonstrates that literacy-aligned generation improves accessibility, politeness, and factual accuracy, with notable gains for smaller models. The framework advances equitable public health communication but acknowledges limitations in knowledge coverage and real-time information integration, outlining directions for finer-grained segmentation and human-centered evaluations.

Abstract

Health misinformation spreading online poses a significant threat to public health. Researchers have explored methods for automatically generating counterspeech to health misinformation as a mitigation strategy. Existing approaches often produce uniform responses, ignoring that the health literacy level of the audience could affect the accessibility and effectiveness of counterspeech. We propose a Controlled-Literacy framework using retrieval-augmented generation (RAG) with reinforcement learning (RL) to generate tailored counterspeech adapted to different health literacy levels. In particular, we retrieve knowledge aligned with specific health literacy levels, enabling accessible and factual information to support generation. We design a reward function incorporating subjective user preferences and objective readability-based rewards to optimize counterspeech to the target health literacy level. Experiment results show that Controlled-Literacy outperforms baselines by generating more accessible and user-preferred counterspeech. This research contributes to more equitable and impactful public health communication by improving the accessibility and comprehension of counterspeech to health misinformation

Figures (2)

• Figure 1: An example of health misinformation, paired with three counterspeech responses tailored to low, medium, and high health literacy levels.
• Figure 2: Overview of our Controlled-Literacy counterspeech generation framework tailored to users with different health literacy levels. (a) The GRPO training loop integrates evidence retrieval into the LLM policy and optimizes it using a hybrid reward function combining user preference (weight $\alpha$) and readability (weight $1{-}\alpha$). Rewards are aggregated through group computation to compute the advantage signal. (b) During inference, the model takes a health misinformation input and retrieves customized evidence to generate counterspeech adapted to low, medium, or high health literacy users. (c) The evidence retrieval module selects content from the knowledge base by filtering it according to the target readability range and user preference thresholds, ensuring personalized support for the generation process.