Argumentative Explanations for Pattern-Based Text Classifiers

Even "interpretable" AI models need better explanations: While pattern-based logistic regression is designed to be transparent, the paper found that standard explanation methods fail to show how different features interact with each other, resulting in explanations that don't feel realistic to humans—highlighting that interpretability alone doesn't guarantee understandable explanations.

Computational argumentation can reveal model reasoning: The researchers developed AXPLR, which treats the model's decision-making features as arguments that either support or attack each other, similar to how humans construct logical arguments. This approach uncovers where the model agrees and disagrees with itself, making its reasoning more transparent.

Feature relationships matter for trust: The paper demonstrates that explanations accounting for how features interact (rather than treating them independently) are more plausible to humans, suggesting that responsible AI explanations should capture the nuanced relationships between different factors influencing a decision, not just list individual reasons.

Formal methods can improve explanation quality: By grounding their explanation method in established computational argumentation theory and testing it against formal properties, the authors show that rigorous mathematical frameworks can help ensure explanations are not just intuitive but also theoretically sound—an important consideration for high-stakes applications.