Evaluating Explainable AI Attribution Methods in Neural Machine Translation via Attention-Guided Knowledge Distillation
A new study uses 'attention-guided knowledge distillation' to rank which XAI methods actually help AI models learn better.
A team of researchers has developed a new, automated framework to evaluate which Explainable AI (XAI) attribution methods are most useful for improving neural machine translation models. The core innovation is an 'attention-guided knowledge distillation' approach: they extract attribution maps from a large 'teacher' model (like Marian-MT or mBART) and inject them into the attention mechanism of a smaller 'student' model to guide its learning. By measuring the student's performance gains, they can quantify the practical utility of different XAI techniques. This moves beyond simply visualizing model decisions to actively testing if those explanations contain meaningful, transferable knowledge.
The study tested numerous popular XAI methods across three translation tasks (German-English, French-English, Arabic-English). The results were clear: attribution methods derived from the model's internal attention mechanism (Attention, Value Zeroing, Layer Gradient × Activation) consistently provided the largest boosts in translation quality, measured by BLEU and chrF scores. In contrast, several gradient-based methods (Saliency, Integrated Gradients, DeepLIFT) led to smaller, less reliable improvements. This suggests that attention-based attributions better capture the crucial alignment between source and target words in sequence-to-sequence tasks.
To further validate their findings, the researchers built an 'Attributor' transformer—a model trained to predict the teacher's attribution map for a given sentence pair. They found a direct correlation: the more accurately the Attributor could reproduce a high-quality map, the more beneficial injecting that map was for the student model's performance. This work, shared on arXiv, provides a concrete methodology for benchmarking XAI tools and reveals that the choice of explanation method has a real, measurable impact on downstream AI system performance. The code is publicly available on GitHub for replication and further research.
- The study introduced a novel evaluation framework using 'attention-guided knowledge distillation' to test the utility of XAI attribution methods by measuring student model performance gains.
- Attention, Value Zeroing, and Layer Gradient × Activation methods led to the most consistent BLEU score improvements (up to +1.5) across three language pairs, outperforming gradient-based methods like Saliency and Integrated Gradients.
- The team created an 'Attributor' transformer that learns to predict useful attribution maps, establishing a direct link between attribution map accuracy and downstream task performance.
Why It Matters
This provides a concrete benchmark for choosing XAI tools that actually improve model performance, not just generate visualizations.