SIDU-TXT: An XAI Algorithm for NLP with a Holistic Assessment Approach

XAI methods for text need better evaluation standards: The paper addresses a significant gap in responsible AI by proposing a comprehensive three-tiered evaluation framework (Functionally-Grounded, Human-Grounded, and Application-Grounded) for assessing explainability methods in natural language processing, since no unified standard currently exists.

SIDU-TXT shows promise for understanding AI text decisions: The new method generates word-level explanations by highlighting contextually significant words that influence model predictions, offering a more granular way to understand why AI systems make specific decisions about text compared to existing methods like LIME and Grad-CAM.

Explainability performance varies by task complexity: While SIDU-TXT performed well on sentiment analysis of movie reviews, its performance was comparable to other methods in the more complex legal domain of asylum decision-making, suggesting that explainability methods may need task-specific refinement for high-stakes applications.

Human evaluation matters for responsible AI: The paper emphasizes that technical performance alone is insufficient—human-grounded evaluation is essential to ensure that AI explanations are actually understandable and useful to real people who need to trust and verify AI decisions.

High-stakes domains require careful explainability assessment: The inclusion of asylum decision-making as a test case highlights that explainability methods must be rigorously validated in sensitive applications where AI decisions significantly impact people's lives.