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

Xiaoying Song Anirban Saha Anik Dibakar Barua, Pengcheng Luo Junhua Ding Lingzi Hong · 2025

Read Speaking at the Right Level: Literacy-Controlled Counterspeech Generation with RAG-RL by Xiaoying Song Anirban Saha Anik Dibakar Barua, Pengcheng Luo Junhua Ding Lingzi Hong (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-01058.txt.

Introduction

Health falsehoods spread fast online, and replies that correct them often use words many readers cannot follow. Xiaoying Song, with five co-authors, builds a system that writes such replies at a reading level matched to its audience. Their method joins search-based generation with learning from rewards.

The system first pulls in facts written for a chosen literacy level, meaning how well a reader handles health text. A reward score then blends user likes with plain readability checks, guiding the output toward that level. Tests show the replies were easier to read and more liked than those from rival methods. Song and the team argue this can make public health messages fairer and clearer.

Catalog record

CREATORXiaoying Song Anirban Saha Anik Dibakar Barua, Pengcheng Luo Junhua Ding Lingzi Hong
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-01058

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Speaking at the Right Level: Literacy-Controlled Counterspeech Generation with RAG-RL Xiaoying Song Anirban Saha Anik Dibakar Barua, Pengcheng Luo Junhua Ding Lingzi Hong 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. 1 Introduction Health misinformation online poses a serious risk to public health as it can mislead people, cause bad health choices, and reduce trust in doctors and medical advice Do Nascimento et al. 2022 . A user on Reddit once claimed that “I won’t take a mammogram because it is the squishing that causes cancer.” Such health misinformation not only misleads individuals into avoiding appropriate medical actions but also increases public confusion Fridman et al. 2023 . Therefore, it is essential to curb the spread of health misinformation. Figure 1: An example of health misinformation, paired with three counterspeech responses tailored to low, medium, and high health literacy levels. Social media users actively participate in combating health misinformation by posting counterspeech Kuo and Chen 2025 ; Anik et al. 2025 ; Micallef et al. 2020 . However, user-generated responses often lack credible evidence, decreasing the effectiveness of counterspeech Yue et al. 2024a . While experts and fact-checkers can provide factual support in debunking health misinformation, it becomes less efficient when confronted with an increasing volume of health misinformation Yue et al. 2024b . Generating supportive counterspeech at scale remains a challenge. Recent studies have explored the use of Large Language Models (LLMs) to generate counterspeech in response to health misinformation Yue et al. 2024a ; Yue et al. 2024b ; Wang et al. 2024b ; He et al. 2023 . However, these methods often overlook a critical issue: Is the generated counterspeech accessible and persuasive to users with different health literacy levels? Health literacy refers to the capability to understand, judge, and use health information to make good choices and stay healthy World Health Organization 2024 . The health literacy level of the audience can significantly affect the effectiveness of counterspeech Liu et al. 2020 . Users with lower health literacy often struggle to comprehend counterspeech grounded in scientific research Shahid et al. 2022 ; Liu et al. 2020 and understand information that contains complex or technical language Centers for Disease Control and Prevention 2024 ; Chen et al. 2018 . In contrast, users with a higher literacy level may lose interest or perceive the content as oversimplified if it is tailored to a low comprehension level August et al. 2024 ; Martínez Silvagnoli et al. 2022 . For example, simplistic explanations such as “It is made from a tiny, weak part of the virus that cannot make you sick …” in Figure 1 (low-level counterspeech) may be inappropriate for users with advanced health literacy. Such mismatches in communication can critically undermine the effectiveness of counterspeech. In this work, we aim to generate counterspeech for users with diverse health literacy levels. To address this question, we propose a controllable RAG framework, the Controlled-Literacy . The approach uniquely retrieve knowledge adapting to diverse health literacy levels, considering that the complexity of retrieved knowledge directly influences the style and clarity of the generated responses Ke et al. 2025 . We then integrate Reinforcement Learning (RL) to further optimize the generation, ensuring that the outputs align with the target health literacy level. Our Controlled-Literacy framework is capable of producing counterspeech that is more accessible, user-preferred, polite, and also factually accurate when addressing health misinformation. We summarize our contribution as follows: (1) We introduce the novel insight that effective counterspeech should be aligned with the health literacy level of its target audience. (2) We propose Controlled-Literacy , a framework that integrates RAG and RL for generating accessible counterspeech that accounts for both objective readability and subjective user preferences. (3) A new dataset is curated, MisinfoLiteracy , consisting of health misinformation posts paired with counterspeech responses tailored to users with different health literacy levels. 2 Related Work 2.1 Counterspeech to Misinformation Generation Counterspeech has been proven effective in mitigating misinformation Peng and Grimmelmann 2024 ; Siebert and Siebert 2023 . Previous studies focus on generating counterspeech with desirable attributes. For example, Anik et al. 2025 and Yue et al. 2024a combined external scientific sources to generate evidence-based counterspeech in response to misinformation. Yue et al. 2024b generated factually accurate counterspeech by synthesizing contrastive arguments derived from fact-checking sources. Hong et al. 2024 adapted the generation of counterspeech according to preferred conversation outcomes, resulting in more positive conversation engagement. He et al. 2023 proposed a reinforcement learning-based framework to generate counterspeech with enhanced politeness, factuality, and refutational strength. Despite these efforts, existing studies have not yet investigated counterspeech generation tailored to users with different 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. 2.2 Accessible Language Generation Accessible language generation is widely explored in multiple domains, such as health care Yao et al. 2024 ; Rahman et al. 2024 ; Luo et al. 2022 ; Phatak et al. 2022 , education Wang et al. 2024a ; Malik et al. 2024 ; Rooein et al. 2023 , and finance Wang et al. ; Kosireddy et al. 2024 ; Perez-Rojas et al. 2023 . In health care, previous studies primarily focus on simplifying or summarizing complex medical terminology. Yao et al. 2024 introduced a dataset with medical terms paired with lay definitions, aiming to enhance the accessibility of medical information for non-experts. In education, studies have investigated tailoring learning materials to learners at varying proficiency levels. For instance, Malik et al. 2024 have experimented with methods to control the language proficiency level of LLM-generated content to suit different types of learners. In finance, the emphasis has been on democratizing access to financial information for a broader audience. Kosireddy et al. 2024 used small language models

Read the complete text