Enhancing the Comprehensibility of Text Explanations via Unsupervised Concept Discovery

Concept-based explanations can bridge AI and human understanding: The paper demonstrates that explaining AI decisions through human-like concepts is more effective than traditional methods, addressing a key challenge in making AI systems trustworthy and interpretable to non-experts.

Current text explanation methods have a critical gap: Existing approaches either require manual labeling of concepts (limiting their scalability) or automatically discover concepts that people find confusing—both problems undermine user trust in AI systems, according to the research.

Large language models can evaluate explanation quality: ECO-Concept uses LLMs to assess whether automatically discovered concepts are actually comprehensible to humans, then uses this feedback to improve the AI model—suggesting a practical role for LLMs in the responsible AI pipeline.

Automatic concept discovery without human annotation is now feasible: The framework successfully discovers meaningful, understandable concepts without requiring experts to pre-define them, making concept-based explanations more practical for real-world deployment across different applications.

Comprehensibility matters as much as accuracy: The paper's validation shows that explanations rated as more comprehensible by human evaluation standards outperformed previous methods, emphasizing that responsible AI requires explanations people can actually understand, not just technically sophisticated ones.