Factual accuracy
Compared with 0.87 for prompt-only generation.
← All publications
Findings of EMNLP 2025
Speaking at the Right Level: Literacy-Controlled Counterspeech Generation with RAG-RL
A RAG and reinforcement-learning framework that adapts health-misinformation corrections to a reader's health-literacy level.
The problem
Why this work matters
A single explanation cannot serve every reader equally well. Existing counterspeech systems often ignore health literacy, producing responses that may be accurate but too technical, too simple, or poorly aligned with the intended audience.
Methodology
How the approach works
Contributions
What this work contributes
- 01
Controlled-Literacy, a framework that combines level-aware retrieval with reinforcement learning.
- 02
A hybrid reward balancing readability alignment and simulated reader preference.
- 03
MisinfoLiteracy, a dataset of 440 health-misinformation posts paired with literacy-controlled counterspeech.
- 04
Cross-level evaluation showing that readers prefer responses matched to their intended literacy level.
Evaluation
Key results
Politeness
Compared with 0.41 for the prompt-only baseline.
Target distance ↓
Closer readability alignment than the 2.74 baseline.
User preference
Highest when the response matches the reader's level.