Abstract
Word sense disambiguation is a well-known source of translation errors in NMT. We posit that some of the incorrect disambiguation choices are due to models' over-reliance on dataset artifacts found in training data, specifically superficial word co-occurrences, rather than a deeper understanding of the source text. We introduce a method for the prediction of disambiguation errors based on statistical data properties, demonstrating its effectiveness across several domains and model types. Moreover, we develop a simple adversarial attack strategy that minimally perturbs sentences in order to elicit disambiguation errors to further probe the robustness of translation models. Our findings indicate that disambiguation robustness varies substantially between domains and that different models trained on the same data are vulnerable to different attacks.
Introduction
Word Sense Disambiguation (WSD) remains a fundamental challenge for Neural Machine Translation (NMT). While humans leverage global context to resolve ambiguities—such as determining whether "spring" refers to a season or a water source—NMT models frequently rely on "dataset artifacts." These are statistical patterns in training data that may not reflect true linguistic logic. This research investigates whether NMT models over-rely on these superficial co-occurrences and demonstrates how this reliance can be exploited through model-agnostic adversarial attacks.
Methodology
The study introduces a framework to identify and utilize statistical biases through "attractors"—source words that strongly co-occur with specific senses of a homograph in the training corpus.
Quantifying Bias
Two primary metrics are used to measure context-driven bias:
- FREQ: Raw co-occurrence frequency between words.
- PPMI (Positive Point-wise Mutual Information): A measure of how much more likely words are to appear together than by chance.
The researchers hypothesized that the difference in frequency between attractors for the incorrect sense versus the correct sense (FREQ_DIFF) would be a strong predictor of WSD errors.
Adversarial Attack Strategies
To test model robustness, the authors designed four adversarial strategies involving minimal perturbations with adjectives:
- IH/RH: Inserting or replacing an adjective that directly modifies the homograph.
- InH/RnH: Inserting or replacing an adjective that modifies a different noun in the sentence.
To maintain naturalness, a GPT-2 language model was used to filter out perturbations that significantly increased sentence perplexity.
Experimental Setup
The experiments focused on English → German translation across two domains: OpenSubtitles2018 (OS18) for casual conversation and WMT19 (News Task) for formal language. Three architectures were evaluated: Transformer, LSTM, and ConvS2S. English homographs were mapped to German sense clusters using BabelNet to objectively identify disambiguation failures.
Results and Analysis
The Impact of Statistical Bias
The results confirm that NMT models are highly susceptible to biased contexts. When evaluated on "challenge sets"—sentences specifically sampled to contain misleading attractors—error rates spiked dramatically. As shown in Figure 1, error rates for the Transformer model on OS18 increased from 3.83% in random sets to 23.4% in challenge sets, a factor of over 6x.
The study found a strong correlation (up to 0.73 Rank Biserial Correlation) between the proposed bias measures (specifically FREQ_DIFF) and actual translation errors, supporting the "shortcut hypothesis" that models prioritize statistical co-occurrences over deep context.
Adversarial Attack Success
The targeted adversarial attacks proved highly effective, particularly when the attractor was placed in close proximity to the ambiguous word. Figure 2 illustrates the success rates of different attack strategies. The IH (Insert-Homograph) and RH (Replace-Homograph) strategies, which modify the homograph's direct context, consistently outperformed the InH and RnH strategies.
Notably, the OS18 Transformer model was the most vulnerable, with a success rate of 24.39% for targeted attacks. Human evaluation confirmed the subtlety of these attacks; bilingual judges rated adversarial sentences nearly as natural as the originals (3.94 vs 4.18 on a 5-point scale).
Conclusion
This research demonstrates that current NMT models are significantly biased by training data artifacts, leading to predictable WSD failures. While Transformers generally perform better than older architectures, they remain highly vulnerable to simple, attractor-informed adversarial perturbations. The low transferability of successful attacks between different models suggests that specific vulnerabilities are often tied to the random seeds and specific training dynamics of individual model instances.
Summary updated