UCert: A Formal Certificate of Unexplainability for TAG Neural Networks

ACL ARR 2026 January Submission1984 Authors

Keywords: text-attributed graphs, graph neural networks, explanation, certifiable guarantees, natural language explanations

Abstract: Natural language explanations for classifiers operating on text-attributed graphs present an inherent tension: explanations should be both readable to humans and verifiable by machines. We introduce UCert, a unified framework that places multiple explanation paradigms within a single evaluation and generation interface and that supports a certificate-based reverse-explanation mode. The framework combines saliency-driven pseudo-labeling, iterative refinement for fluent explanations, adversarial probing to compress evidence into a concise bottleneck, and an executable domain-specific language for machine-checkable rationales. UCert additionally issues formally checkable Unexplainability Certificates which assert that no short symbolic explanation can reproduce a classifier's behavior within a defined perturbation model. We demonstrate that this hybrid design clarifies practical trade-offs between readability, automated faithfulness proxies, and machine-checkable guarantees for text-attributed graph models. Our code is available at: https://anonymous.4open.science/r/tacer-76B5/README.md.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: text-attributed graphs,graph neural networks,explanation,certifiable guarantees,natural language explanations

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 1984