[Paper Review] Reducing Hallucinations in Neural Machine Translation with Feature Attribution

Joël Tang, Marina Fomicheva+1 more|arXiv (Cornell University)|Nov 17, 2022

Topic Modeling4 citations

TL;DR

This paper proposes a novel auxiliary loss function that reduces hallucinations in neural machine translation by leveraging feature attribution to identify and penalize models that overly rely on generated target tokens while ignoring source input. By minimizing the entropy of source token attributions, the method improves robustness on low-quality data, achieving a 13.75 BLEU point gain on hallucinated examples without retraining from scratch.

ABSTRACT

Neural conditional language generation models achieve the state-of-the-art in Neural Machine Translation (NMT) but are highly dependent on the quality of parallel training dataset. When trained on low-quality datasets, these models are prone to various error types, including hallucinations, i.e. outputs that are fluent, but unrelated to the source sentences. These errors are particularly dangerous, because on the surface the translation can be perceived as a correct output, especially if the reader does not understand the source language. We present a case study focusing on model understanding and regularisation to reduce hallucinations in NMT. We first use feature attribution methods to study the behaviour of an NMT model that produces hallucinations. We then leverage these methods to propose a novel loss function that substantially helps reduce hallucinations and does not require retraining the model from scratch.

Motivation & Objective

• To investigate the root causes of hallucinations in neural machine translation, particularly in models trained on low-quality parallel data.
• To analyze model behavior on hallucinated outputs using feature attribution methods to identify patterns of attention failure.
• To develop a differentiable, post-hoc loss function that reduces hallucinations without retraining the model from scratch.
• To evaluate the effectiveness of the proposed loss on both hallucinated and non-hallucinated examples, ensuring no degradation on high-quality translations.

Proposed method

• Feature attribution via Integrated Gradients is applied to compute relevance scores of source and target tokens to the model's output, quantifying their contribution to translation decisions.
• The entropy of attribution scores across source tokens is computed as a proxy for model reliance on source input; high entropy indicates poor alignment and potential hallucination.
• A novel auxiliary loss is introduced that penalizes high entropy in source token attributions, encouraging the model to attend more consistently to source content.
• The loss is combined with standard cross-entropy and label-smoothing during fine-tuning, with a weight of 5 to balance magnitude with the main loss.
• The method is evaluated on the MLQE-PE dataset for Romanian-to-English translation, using BLEU scores between back-translated outputs and original sources as the primary metric.
• Fine-tuning is performed on a strong Transformer base model using the fairseq library, with hyperparameters kept consistent across baseline and proposed models.

Experimental results

Research questions

• RQ1 What patterns in the source input trigger hallucinations in NMT models, especially under low-resource or noisy data conditions?
• RQ2 Can feature attribution methods quantitatively identify hallucination-prone behavior in neural translation models?
• RQ3 Is the entropy of source token attribution scores a reliable indicator of hallucination occurrence?
• RQ4 Can a differentiable, post-hoc loss based on attribution entropy effectively reduce hallucinations without retraining from scratch?
• RQ5 Does the proposed method preserve translation quality on non-hallucinated examples, ensuring no negative impact on overall performance?

Key findings

• The proposed auxiliary loss reduced hallucinations significantly, achieving a 13.75 BLEU point improvement on the hallucinated subset (from 10.71 to 24.46 BLEU).
• The method achieved a 9.35 BLEU point gain on the overall test set (from 28.90 to 38.25 BLEU), demonstrating broad effectiveness.
• High entropy in source token attribution scores was consistently linked to hallucinated outputs, validating its use as a diagnostic signal.
• The model maintained high performance on non-hallucinated examples, with a 9.61 BLEU point gain (from 35.29 to 44.90), showing no degradation on correct translations.
• The method successfully mitigated various hallucination types, including detached, oscillatory, and mixed cases, as shown in qualitative examples.
• Even with OOV tokens present, the fine-tuned model preserved the core meaning of the source sentence, outperforming the baseline in robustness.