Prevent the Language Model from being Overconfident in Neural Machine Translation

• NMT hallucination produces fluent but inadequate translations when excessive attention to partial translations causes source segments to be neglected.Source: §1 · p. 1
• The paper defines the Margin as the NMT token probability minus the LM token probability to indicate LM overconfidence.Source: §1 · pp. 1–2
• MTO and MSO maximize the Margin, with MSO adding a dynamic weighting function to reduce the effect of training sentences containing many negative-Margin tokens.Source: §1 · p. 2
• The cited passage reports significant improvements of +1.36, +1.50, and +0.63 BLEU points for English-to-French results, respectively.Source: §1 · p. 2
• Human evaluation reports improvements in both translation adequacy and fluency.Source: §1 · p. 2

• NMT predicts each target word using the source sentence together with the preceding partial translation.Source: §2 · p. 2
• The LM component predicts from the partial translation alone, while NMT additionally conditions prediction on the source sentence.Source: §2 · p. 2
• Overemphasizing translation fluency can cause NMT to neglect source segments, producing inadequate translations.Source: §2 · p. 2

3.1 Margin between the NMT and the LM

• The Margin is the difference between NMT and LM probabilities for each token, with smaller values indicating greater LM overconfidence and possible source neglect.The NMT probability conditions on the source and partial translation, whereas the LM probability conditions only on the partial translation.Source: §3.1 · pp. 2–3

3.2 Margin-based Token-level Objective

• The Margin-based Token-level Objective (MTO) adds a Margin loss to cross-entropy to maximize the Margin, weighting poorly learned tokens more strongly.The weighting factor 1 − pNMT(t) reduces attention to tokens already predicted with high probability.Source: §3.2 · p. 3
• MTO uses monotonically decreasing Margin functions, including Linear, Cube, Quintic, and Log; nonlinear functions are more stable near zero Margin than Linear.Figure 1 compares their slopes, with α in Log set to 10.Source: §3.2 · p. 3, Figure 1 · p. 3

3.3 Margin-based Sentence-level Objective

• MSO measures the sentence-level ratio of negative-Margin tokens and ignores losses for sentences exceeding threshold k, which may be inadequate or unrelated to their sources.The sentence-level weight is dynamic and equals one only when R(x,y) is below k; otherwise it is zero.Source: §3.3 · p. 4

3.4 Two-stage Training

• The method jointly pretrains NMT with an auxiliary LM, then finetunes NMT using MTO and MSO; the LM is not used during inference.Joint training aligns the auxiliary LM with NMT’s language-model mechanism while avoiding abundant extra parameters.Source: §3.4 · p. 4

4 Experimental Settings

• Experiments use WMT14 En→De, WMT14 En→Fr, and WMT19 Zh→En datasets with task-specific training and validation/test splits.Training data sizes are 4.5M, about 36M, and 20.5M sentence pairs, respectively.Source: §4 · p. 5
• The study measures case-sensitive BLEU and selects the best validation model for test-set evaluation, with paired bootstrap resampling for significance testing.Source: §4 · p. 5
• The implementation uses Transformer-Base settings with six encoder and decoder layers, 512-dimensional hidden states, eight V100 GPUs, and beam size 5.Source: §4 · p. 5
• Training uses 150K jointly pretrained steps followed by 150K finetuning steps, while the vanilla Transformer and NMT+LM baselines use 300K steps.Source: §4 · p. 5

5 Results and Analysis

• +1.36, +0.63, and +1.50 BLEU improve Transformer baselines on En→De, En→Fr, and Zh→En, respectively.The improvements are reported for the proposed approaches across the three benchmark tasks.Source: §5.1 · p. 6, §5.2 · p. 6, Table 1 · p. 6

5.1 Results on En→De

• +0.50 BLEU from MTO over NMT+LM and +0.11 BLEU from MSO over MTO are reported on En→De.The authors attribute the additional MSO gain to reducing the impact of harmful training data.Source: §5.1 · p. 6

5.2 Results on En→Fr and Zh→En

• +0.42 and +1.04 BLEU from MTO over NMT+LM, followed by +0.14 and +0.31 BLEU from MSO over MTO, are reported on En→Fr and Zh→En.Source: §5.2 · p. 6

5.3 Human Evaluation

• MTO and MSO improve adequacy by 0.22 and 0.37 scores over the NMT baseline, while their further fluency gains over NMT+LM are 0.01 and 0.05.NMT+LM itself improves adequacy by 0.08 and fluency by 0.2 over NMT.Source: §5.3 · p. 7, Table 3 · p. 7

5.4 Analysis

• MSO shifts token Margins away from 0 toward [0.75, 1.0], while MTO and MSO reduce negative-Margin tokens and increase average Margin from 0.33 to 0.44.The average Margin increase is 33% relative to NMT+LM.Source: §5.4 · p. 7, Figure 3 · p. 7, Table 4 · p. 7
• Among the tested Margin functions, all improve over NMT and NMT+LM, with Quintic performing best; MSO also outperforms MTO at most finetuning steps.Source: §5.4 · pp. 7–8, Table 5 · p. 8, Figure 4 · p. 8
• Removing MTO’s weight decreases test BLEU by 1.09 relative to weighted MTO, showing that the weight 1 −pNMT (t) is important.Source: §5.4 · p. 8, Table 6 · p. 8
• Translation examples show that the proposed approach captures more source segments that Transformer mistranslates or neglects.Source: §5.4 · p. 8

6 Related Work

• Prior approaches improve adequacy through coverage tracking, enhanced cross-attention, source-side representations, source partitioning, and multi-task learning.Source: §6 · p. 8
• Other methods optimize reconstruction, coverage, or evaluation-based adequacy objectives, whereas this work directly maximizes the Margin between NMT and LM probabilities.Source: §6 · pp. 8–9
• Language models have been used for fusion, representation enhancement, prior knowledge, and data augmentation, but this work requires no additional data and avoids LM use during inference.Source: §6 · p. 9
• Unlike prior data-selection methods, MSO dynamically combines language and translation models for data selection during training.Source: §6 · p. 9

• Margin-based token-level and sentence-level objectives maximize the Margin to prevent LM overconfidence.Source: §7 · p. 9
• Experiments on three large-scale translation tasks demonstrate the effectiveness and superiority of the proposed approaches.Source: §7 · p. 9
• Human evaluation verifies improvements in both translation adequacy and fluency.Source: §7 · p. 9

• The LM loss remains stable after about 80K training steps for En→De, Zh→En, and En→Fr.This stability indicates that the LM converges during the pretraining stage.Source: A Loss of the Language Model · p. 12, Figure 5 · p. 12
• LM convergence provides the basis for evaluating the pretrained language model in the subsequent approach.Source: A Loss of the Language Model · p. 12
• Figure 5 tracks LM validation loss during pretraining across the three translation tasks.Source: Figure 5 · p. 12

• The best k is 30% for En→De and Zh→En, and 40% for En→Fr.These thresholds correspond to the proportion of negative-Margin tokens in a target sentence.Source: B Hyperparameters Selection · p. 12
• All three datasets perform better for λM ∈[5, 10], with peaks at λM =5, 8, and 8 for the three tasks, respectively.Source: B Hyperparameters Selection · p. 12
• Figure 6 reports validation BLEU scores for different λM values and MSO thresholds k across WMT14 En→De, WMT14 En→Fr, and WMT19 Zh→En.Source: Figure 6 · p. 12
• Sentences with negative-Margin token proportions above 30% or 40% are most likely to be hallucinations.Source: B Hyperparameters Selection · p. 12

• MSO outperforms the Transformer baseline in translation adequacy across the illustrated examples.Figure 7 compares source, reference, baseline, and MSO translations, highlighting baseline errors and corrections.Source: C Case Study · p. 12, Figure 7 · p. 13
• MSO translates source segments that the baseline hallucinates over, under-translates, or neglects.Source: C Case Study · p. 12
• MSO corrects a baseline error that reverses the chronological order of the source sentence.Source: C Case Study · p. 12
• Figure 7 marks baseline mistranslations in red and baseline errors corrected by MSO in blue.Source: Figure 7 · p. 13