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

2026-03-11

AI Summary

This study addresses the lack of a systematic, automated evaluation framework for existing explainable AI attribution methods in sequence-to-sequence models. The authors propose an attention-guided knowledge distillation approach that leverages attribution maps generated by a teacher model as structured supervision signals. These maps are reconstructed by an Attributor Transformer, and attribution information—extracted using the Inseq library—is injected into the student model through four fusion operators (addition, multiplication, averaging, and replacement) that integrate attribution scores with attention mechanisms. Experiments on de-en, fr-en, and ar-en translation tasks demonstrate that Attention, Value Zeroing, and Layer Gradient×Activation significantly improve BLEU and chrF scores, confirming that the proposed framework effectively evaluates how well attribution methods capture source–target alignment signals.

Technology Category

Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP ModelsMachine Learning: Transparent, Interpretable, Explainable MLComputer Vision: Interpretability, Explainability, and Transparency

Application Category

Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for rankingSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMsUser Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

The study of the attribution of input features to the output of neural network models is an active area of research. While numerous Explainable AI (XAI) techniques have been proposed to interpret these models, the systematic and automated evaluation of these methods in sequence-to-sequence (seq2seq) models is less explored. This paper introduces a new approach for evaluating explainability methods in transformer-based seq2seq models. We use teacher-derived attribution maps as a structured side signal to guide a student model, and quantify the utility of different attribution methods through the student's ability to simulate targets. Using the Inseq library, we extract attribution scores over source-target sequence pairs and inject these scores into the attention mechanism of a student transformer model under four composition operators (addition, multiplication, averaging, and replacement). Across three language pairs (de-en, fr-en, ar-en) and attributions from Marian-MT and mBART models, Attention, Value Zeroing, and Layer Gradient $\times$ Activation consistently yield the largest gains in BLEU (and corresponding improvements in chrF) relative to baselines. In contrast, other gradient-based methods (Saliency, Integrated Gradients, DeepLIFT, Input $\times$ Gradient, GradientShap) lead to smaller and less consistent improvements. These results suggest that different attribution methods capture distinct signals and that attention-derived attributions better capture alignment between source and target representations in seq2seq models. Finally, we introduce an Attributor transformer that, given a source-target pair, learns to reconstruct the teacher's attribution map. Our findings demonstrate that the more accurately the Attributor can reproduce attribution maps, the more useful an injection of those maps is for the downstream task. The source code can be found on GitHub.

Problem

• Explainable AI
• Attribution Methods
• Neural Machine Translation
• Sequence-to-Sequence Models
• Evaluation

Innovation

• Attention-Guided Knowledge Distillation
• Explainable AI Evaluation
• Attribution Methods
• Neural Machine Translation
• Transformer Attribution