Abstract
Neural Machine Translation (NMT) models often take "statistical shortcuts" by over-relying on superficial word co-occurrences rather than truly understanding context. This paper introduces a method to quantify these lexical biases and demonstrates a model-agnostic adversarial attack that triggers translation errors by simply adding or replacing a single "attractor" word.
Introduction
Word Sense Disambiguation (WSD) remains a persistent challenge in NMT. For instance, the English word spring can refer to a season, a mechanical object, or a water source. While humans use global context to disambiguate, NMT models frequently fail. The authors argue these failures are driven by dataset artifacts: if a training corpus frequently pairs "hot" with "water source" (e.g., hot spring), a model might incorrectly translate "the hot spring of 1988" (referring to a warm season) as a geothermal pool. This study probes these "lazy" behaviors through statistical quantification and targeted adversarial attacks.
Methodology
The researchers developed a two-pronged approach to investigate model bias:
1. Quantifying Bias
They identified "attractors"—context words strongly associated with specific senses of a homograph. Association was measured using:
- FREQ: Raw co-occurrence frequency in training data.
- PPMI: Positive Point-wise Mutual Information, measuring associations beyond chance.
2. Model-Agnostic Adversarial Attacks
To test if these biases drive errors, they developed a model-agnostic attack that perturbs input text without requiring gradient access. Strategies included:
- Insertion/Replacement: Adding or swapping adjectives near the homograph to "pull" the model toward an incorrect sense.
- Filtering: Using BERT and GPT-2 to ensure perturbed sentences remained grammatically fluent and semantically plausible.
Experiments were conducted on English-to-German translation using OpenSubtitles2018 (OS18) and WMT19 News datasets across Transformer, LSTM, and Convolutional (ConvS2S) architectures.
Key Results
Error Amplification
The study found that statistical bias is highly predictive of WSD errors. On a "challenge set" of sentences characterized by high statistical bias (where attractors contradict the correct sense), WSD error rates were significantly higher—up to 6.1 times higher than on randomly selected sentences.
Attack Success
The adversarial attacks achieved success rates between 4.6% and 24.4%. These represent "targeted" flips where a previously correct model was led astray by a single word change. The success of these attacks confirms that models are indeed relying on the identified lexical shortcuts.
Insights on Model Robustness
- Proximity Matters: Perturbations placed directly in front of the ambiguous word were the most effective at triggering errors.
- Low Transferability: Adversarial examples that tricked a Transformer often failed against an LSTM trained on the same data, suggesting different architectures develop unique "lazy" patterns.
- Human Validation: Bilingual judges confirmed that while adversarial sentences were slightly less natural, they remained as clear to humans as the originals, highlighting that models fail where humans do not.
Conclusion
NMT models are susceptible to lexical traps regardless of architecture. While the success rates of these simple attacks are modest, they reveal a fundamental reliance on shallow correlations. Future work will explore if multilingual models offer better resilience and extend these attacks to other NLP tasks.
Summary updated