Source-linked AI summary
On Exposure Bias, Hallucination and Domain Shift in Neural Machine Translation
Chaojun Wang, Rico Sennrich
TL;DR
NMT exposure bias may contribute to hallucinations under domain shift, but its practical impact has remained debated. The paper studies this link using Minimum Risk Training across multiple datasets and domains, finding that MRT improves out-of-domain robustness, reduces hallucinations, and stabilizes large-beam search. These findings support sequence-level objectives when test domains are unknown or hallucinations are common, although the evaluation does not combine MRT with several orthogonal robustness techniques.
Problem
Exposure bias in standard NMT training has debated practical importance, particularly for hallucinations under domain shift.
Method
The paper evaluates Minimum Risk Training, a sequence-level objective that avoids exposure bias, across three datasets with multiple test domains.
Results
MRT improves quality more in out-of-domain settings, reduces hallucinations, and makes beam search with large beams more stable.
Takeaways & Limitations
Sequence-level objectives that reduce exposure bias may be especially useful when the test domain is unknown or hallucinations are common.
Takeaways & Limitations
The reported out-of-domain results lag slightly behind prior work, and combining MRT with orthogonal robustness techniques is left for future work.
Abstract
from arXiv · showhide
The standard training algorithm in neural machine translation (NMT) suffers from exposure bias, and alternative algorithms have been proposed to mitigate this. However, the practical impact of exposure bias is under debate. In this paper, we link exposure bias to another well-known problem in NMT, namely the tendency to generate hallucinations under domain shift. In experiments on three datasets with multiple test domains, we show that exposure bias is partially to blame for hallucinations, and that training with Minimum Risk Training, which avoids exposure bias, can mitigate this. Our analysis explains why exposure bias is more problematic under domain shift, and also links exposure bias to the beam search problem, i.e. performance deterioration with increasing beam size. Our results provide a new justification for methods that reduce exposure bias: even if they do not increase performance on in-domain test sets, they can increase model robustness to domain shift.
1 Introduction
NMT is vulnerable to hallucinations under domain shift, and the paper investigates whether exposure bias contributes to this problem and related beam-search instability. Experiments with MRT examine whether reducing exposure bias improves robustness beyond in-domain performance.
- Domain-shifted NMT translations can be fluent yet unrelated to the input, making hallucinations particularly misleading.
- Teacher forcing creates exposure bias by training on gold histories while inference uses potentially erroneous model predictions.This mismatch may increase reliance on previous predictions and exacerbate error propagation.
- The practical importance of error propagation remains debated, with prior work arguing that linguistic features explain part of the accuracy decline at higher time steps.
- The paper empirically links hallucination with exposure bias and examines the practical relevance of sequence-level objectives for predicting when they may help.
- Experiments span German→English and German→Romansh datasets across in-domain and out-of-domain settings, using Minimum Risk Training, which inherently avoids exposure bias.
- MRT improves quality more in out-of-domain settings, reduces hallucinations, and reduces the beam search problem.The analysis also reports that MLE overestimates random translations after the initial time steps, while MRT mitigates this problem.
2 Minimum Risk Training
The paper contrasts token-level maximum-likelihood training with Minimum Risk Training, a sequence-level objective designed to avoid exposure bias. MRT evaluates expected loss over a posterior approximated by sampled candidate translations.
- The standard NMT objective minimizes the negative log-likelihood of the training data.
- Teacher forcing uses groundtruth target histories during MLE training, unlike inference, where histories consist of model predictions.This training–inference mismatch is the exposure-bias problem addressed by MRT.
- Minimum Risk Training is a sequence-level objective that minimizes expected loss, or risk, under the posterior distribution.
- Because the search space is intractable, MRT approximates the posterior with a normalized subspace of sampled candidate translations.
- The candidate distribution is controlled by a sharpness hyperparameter, and the experiments generate candidates through random sampling without adding the reference translation.
3 Experiments
The experiments evaluate MRT across in-domain and domain-shifted translation settings, using multiple datasets, evaluation protocols, and comparisons with MLE and prior systems.
- 3.1 Data: The study evaluates German→English and German→Romansh under domain shift, including medical, IT, law, koran, subtitles, blogs, and a low-resource setting.DE→EN uses medical data for training and development, with four other domains for out-of-domain testing; DE→RM trains on law and tests on blogs.
- 3.2 Models and training: The experiments use Transformer models, pre-trained with MLE and then fine-tuned with MRT using a smoothed 1-BLEU loss.MRT is implemented in Nematus, while preprocessing uses Moses tokenization and truecasing with shared 32,000-unit BPE.
- 3.3 Evaluation: Domain-shift results average three independent training runs and report standard deviations to account for optimizer instability.BLEU reporting differs by dataset: lowercased tokenized BLEU is used for IWSLT’14, while cased detokenized SacreBLEU is used otherwise.
- 3.3 Evaluation: Manual evaluation uses two native German speakers who rate fluency and adequacy across approximately 3,600 annotations.Agreement is reported with kappa coefficients, and statistical significance is assessed using two-tailed Fisher’s exact tests.
- 3.4 Results: On IWSLT’14 German→English, the MLE baseline reaches 34.7 BLEU, while MRT improves it by 0.5 BLEU.The baseline is described as competitive and is compared with convolutional and dynamic-convolution systems from prior work.
- 3.4 Results: For domain-shifted data, MRT raises average OOD BLEU by 0.8 for DE→EN and 0.7 for DE→RM, but does not consistently improve in-domain performance.The OOD results remain slightly behind Müller et al. (2019), whose reconstruction, subword-regularization, and noisy-channel techniques are orthogonal to MRT.
4 Analysis
The analysis connects exposure bias to hallucinations and beam-search deterioration under domain shift. MRT increases discrimination between references and distractors, reduces hallucinations, and makes performance less sensitive to larger beams.
- 4.1 Hallucination: Hallucinations are evaluated as fluent but source-unrelated translations, with MRT reducing their out-of-domain frequency and sometimes improving adequacy.An example shows MRT attempting to cover the source even when judged disfluent and inadequate, which the authors prefer to a complete hallucination.
- 4.1 Hallucination: Out-of-domain hallucinations occur at 33–35%, compared with 1–2% in-domain; MRT reduces them by up to 21% relative.The reduction is statistically significant at p < 0.05, although BLEU and hallucination rates do not correlate perfectly.
- 4.2 Uncertainty Analysis: MLE assigns similar probabilities to out-of-domain references and random in-domain distractors after the initial steps, allowing early errors to propagate.In-domain references receive much higher probabilities initially, reducing the chance that hallucinations remain in the beam or overtake good translations.
- 4.2 Uncertainty Analysis: MRT increases certainty more sharply for references than distractors, widening their probability gap and imposing a small penalty on hallucinated sequences at each time step.The authors associate this with greater reliance on the source signal and a lower risk of error propagation and hallucination.
- 4.3 Beam Size Analysis: With larger beams, hallucinations are more likely to survive early pruning and become top-scoring because MLE assigns them high later-step probabilities.MRT reduces this beam-search problem, though it does not eliminate it.
- 4.3 Beam Size Analysis: MRT outperforms MLE by 0.5–0.8 BLEU at beam size 4 and 0.6–1.5 BLEU at beam size 50 on out-of-domain evaluation.Manual evaluation with N=200 per system shows hallucinations increase with beam size, while MRT consistently reduces their proportion.
5 Conclusions
The conclusions connect teacher-forcing exposure bias with domain-shift failures, hallucinations, and beam-size deterioration. They argue that MRT can improve robustness and reduce hallucinations even without improving in-domain performance, while the practical importance of exposure bias varies by dataset.
- MLE-induced exposure bias is connected to poor domain-shift performance, hallucinated translations, and deteriorating performance as beam size increases.
- MRT can increase performance under domain shift, substantially reduce hallucinations, and make large-beam search more effective even without improving an in-domain test set.
- The practical impact of exposure bias varies substantially across datasets, supporting sequence-level objectives that reduce or eliminate it.
A Appendix
The appendix documents NMT system configurations across three datasets and illustrates how MRT training changes token probabilities for references versus distractors.
- Table 7 lists NMT configurations used for pre-training and fine-tuning across three datasets.
- Bracketed hyperparameter options identify the settings used during MRT fine-tuning.
- Figure 2 shows a widening gap between out-of-domain reference and in-domain distractor per-token probabilities across MRT checkpoints.
- Figure 3 compares per-token probabilities for in-domain references and distractors, including a direct comparison between the MLE baseline and final MRT model.