Evaluating Explainable AI Attribution Methods in Neural Machine Translation via Attention-Guided Knowledge Distillation

Attention-based methods outperform gradient-based approaches: The study found that Attention, Value Zeroing, and Layer Gradient × Activation consistently produced the largest improvements in translation quality (measured by BLEU scores), while popular gradient-based methods like Saliency, Integrated Gradients, and DeepLIFT showed smaller and less consistent results across different language pairs.

Different XAI methods reveal different types of information: The research demonstrates that various explainability techniques capture distinct signals about how neural translation models work, suggesting that no single attribution method provides a complete picture—practitioners may need multiple approaches to fully understand model behavior.

Systematic evaluation of explainability is now possible for translation models: By using a "student model" approach where attribution maps guide learning, the researchers created an automated way to measure which explanation methods are actually useful, moving beyond subjective assessments to quantifiable performance metrics across German-English, French-English, and Arabic-English translation tasks.

Explainability methods should be validated before deployment: The significant variation in how well different attribution methods performed highlights the importance of testing explanation techniques before relying on them to understand AI systems in real-world applications, rather than assuming all popular XAI methods are equally reliable.