XAI-CLASS: Explanation-Enhanced Text Classification with Extremely Weak Supervision

Minimal annotation requirements reduce costs: XAI-CLASS requires extremely weak supervision (minimal or no human annotation) for text classification, addressing a major practical barrier where traditional methods demand large amounts of manually labeled data that are "time-consuming and labor-intensive."

Explanations improve AI reliability: The method incorporates word saliency prediction—essentially identifying which words are most important for classification decisions—as a core component. This dual focus on both predictions and explanations helps ensure the AI system's reasoning is transparent and verifiable, not just accurate.

AI reasoning becomes auditable: By generating explanations alongside classifications through a multi-round question-answering process, XAI-CLASS creates a system where stakeholders can understand why documents are categorized a certain way, supporting accountability and trust in automated decision-making systems.

Mutual reinforcement improves performance: The method uses a feedback loop where class labels and explanation words enhance each other during training, suggesting that transparency mechanisms (explanations) can actually improve overall system performance rather than being a separate compliance requirement.

Practical scalability for real-world deployment: By reducing annotation burden while maintaining or improving accuracy through explanation-enhanced training, this approach makes responsible AI more feasible for organizations with limited resources to label training data.