CAF-Gen: Multi-Agent System Enriches Argument Structures with Creator-Reviewer Pipeline
Iterative feedback loop fixes single-pass AI failures in formal argumentation.
Current Argument Mining (AM) techniques can identify basic claims and premises, but they fall short when capturing richer structural features required by advanced schemas like the Carneades Argumentation Framework (CAF), which includes premise types, proof standards, and argument schemes. This limitation hinders the formalization of complex reasoning from natural text, a key challenge in computational linguistics.
CAF-Gen addresses this by employing an iterative Creator-Reviewer multi-agent pipeline. The creator agent generates candidate structures, while the reviewer agent validates them for CAF compliance, feeding back corrections. This cycle mitigates structural instability common in single-pass models. Experimental results demonstrate strong alignment with original annotations and richer, more robust argument models. The work is accepted at ICCCI 2026.
- Uses a Creator-Reviewer pipeline to iteratively validate and refine argument structures.
- Targets the Carneades Argumentation Framework (CAF), capturing premise types and proof standards.
- Outperforms single-pass generative models in structural stability and annotation alignment.
Why It Matters
Enables AI systems to logically parse complex arguments, improving transparency and reasoning in automated decision-making.