Optimal Transport for Unsupervised Hallucination Detection in Neural Machine Translation

The evolution of neural machine translation (Vaswani et al., 2017) has significantly improved translation quality, especially with transformer architectures. Despite these advances, hallucinations—translations that deviate from the source—remain a critical challenge (Lee et al., 2018; Müller et al., 2020). Prior work often relies on external quality estimation models (Rei et al., 2020; Feng et al., 2020) or heuristic attention heuristics (Berard et al., 2019), which are costly and limited in scope. Recent studies highlight the importance of internal attention patterns, but lack a unified, unsupervised detection framework. This paper situates itself at the intersection of attention analysis and optimal transport theory, proposing a novel approach to identify anomalous translations based on attention distribution divergence.

Detecting hallucinations in NMT remains difficult due to their rarity and diversity, especially in high-resource settings. Existing methods depend on external models trained on large annotated datasets, which are costly and not scalable. The core challenge is to develop a model-internal, unsupervised method capable of identifying translations that are disconnected from the source content. This requires quantifying the deviation of attention patterns without reference translations, which is complicated by the complex and variable nature of attention maps across different models and languages. Addressing this problem is crucial for deploying safer, more reliable translation systems in real-world scenarios.

The key innovations include: 1) leveraging the geometric properties of optimal transport (OT) to measure divergence in attention distributions; 2) designing a fully unsupervised detection framework that operates solely on internal attention features; 3) introducing a data-driven reference set (Wass-Data) to adapt to model-specific attention patterns; 4) combining uniform and data-driven distances into a robust detection score (Wass-Combo). These innovations enable precise, scalable, and model-agnostic hallucination detection, addressing the limitations of prior heuristic or supervised approaches and providing a solid theoretical basis grounded in OT theory.

• �� Extract the last-layer cross-attention matrix Ω(x) from the transformer model for each input.
• �� Compute the source attention mass distribution πM(x) by averaging across attention heads.
• �� In the single-reference scenario, define a uniform distribution u over source tokens.
• �� Calculate the Wasserstein-1 distance W(πM(x), u) using the L1 norm, which simplifies to total variation.
• �� For data-driven detection, construct a reference set Rheld from high-quality samples, based on COMET scores.
• �� For each test sample, compute pairwise Wasserstein distances to Rheld, then average the smallest k distances to get the anomaly score.
• �� Fuse the uniform and data-driven scores into a single detector (Wass-Combo) with a threshold for binary classification.
• �� Evaluate performance using AUROC and FPR@90TPR, tuning hyperparameters for optimal detection.
• �� Perform ablation studies to analyze the impact of reference set size, hyperparameters, and hallucination types.

Experiments utilize the WMT18 DE-EN dataset with human annotations for hallucinations, comparing against models like LaBSE and COMET. The detectors are tested under different scenarios: single uniform reference (Wass-to-Unif), data-driven reference (Wass-to-Data), and their combination (Wass-Combo). Hyperparameters such as δ, k, and thresholds are optimized via validation. The evaluation metrics include AUROC and FPR@90TPR, with multiple random seeds to ensure robustness. Additional experiments analyze detection across hallucination types, confirming the method’s versatility. Results demonstrate that Wass-Combo consistently outperforms baselines, with significant improvements in false positive rates and overall accuracy.

Wass-Combo achieves an AUROC of 87.17% and FPR of 0.07%, outperforming previous model-based and external detectors. Data-driven reference (Wass-Data) significantly improves detection over uniform reference, reducing FPR by over 30 points. The method effectively detects all hallucination types, with FPR as low as 0.15% for fully detached hallucinations. The fusion of both distances enhances robustness, confirming the complementarity of the approaches. The results validate the hypothesis that attention anomalies correlate with hallucinations and demonstrate the practical viability of OT-based internal feature analysis.

This detection framework can be integrated into real-time NMT systems to monitor translation quality, providing automatic alerts for potential hallucinations. It is suitable for deployment in production environments where reference translations are unavailable. The approach can also be extended to multilingual and low-resource settings by adjusting reference set construction. Additionally, it can serve as a safety layer in critical applications like medical or legal translation, ensuring content reliability. Long-term, the method could facilitate adaptive training, model fine-tuning, and explainability in neural translation systems, fostering greater user trust and system robustness.

The reliance on last-layer attention maps may limit detection in models with different architectures or attention mechanisms. Computational complexity of OT distance calculations poses scalability challenges, especially for large datasets. The method's effectiveness in extremely low-resource or multilingual contexts remains to be validated. Sensitivity to hyperparameters like thresholds and reference set size requires careful tuning. Future work should focus on multi-layer attention fusion, efficiency improvements, and broader language applicability to address these limitations.