Evaluating the Effectiveness of XAI Techniques for Encoder-Based Language Models

LIME is the most reliable explainability method overall: The study found that LIME (a model simplification technique) consistently outperformed five other explainability approaches across multiple evaluation metrics, particularly excelling at "Human-reasoning Agreement" with a score of 0.9685 on the largest model tested (DeBERTa-xlarge), meaning its explanations aligned best with how humans understand AI decisions.

Different XAI techniques have different strengths: Rather than one perfect solution, the research found that different explainability methods excel at different tasks—for example, Attention Mechanism Visualization was most robust to small changes in input, while Layer-wise Relevance Propagation performed best at contrastive explanations (explaining why the model chose one answer over another).

Explainability becomes more important as AI models grow larger and more complex: The study tested six different language models ranging from small (TinyBERT) to very large (DeBERTa-xlarge), finding that LIME's advantages in reliability and consistency became more pronounced with larger, more complex models—suggesting that transparency tools become increasingly critical as AI systems scale up.

A standardized evaluation framework is needed for responsible AI deployment: The paper proposes four key metrics (Human-reasoning Agreement, Robustness, Consistency, and Contrastivity) for evaluating explainability techniques, addressing a gap in how organizations can systematically assess whether their transparency tools actually work before deploying AI systems.