Abstract
Is bias amplified when neural machine translation (NMT) models are optimized for speed and evaluated on generic test sets using BLEU? We investigate architectures and techniques commonly used to speed up decoding in Transformer-based models, such as greedy search, quantization, average attention networks (AANs) and shallow decoder models and show their effect on gendered noun translation. We construct a new gender bias test set, SimpleGEN, based on gendered noun phrases in which there is a single, unambiguous, correct answer. While we find minimal overall BLEU degradation as we apply speed optimizations, we observe that gendered noun translation performance degrades at a much faster rate.
Introduction
In production environments, the speed of Neural Machine Translation (NMT) is a primary concern. Engineers often employ various "speed-accuracy trade-offs" to reduce latency. Traditionally, the "accuracy" component of this trade-off is measured using BLEU, which compares machine output to human references. However, BLEU is a coarse metric that may fail to capture specific linguistic failures, such as the incorrect assignment of gender based on societal stereotypes. This study investigates whether optimizing for speed inadvertently amplifies gender bias, specifically in cases where the source text provides unambiguous gender signals.
Methodology
The researchers examined several standard optimization techniques:
- Search Optimization: Reducing beam size to 1 (greedy search).
- Architecture Modifications: Utilizing "Shallow Decoders" with fewer layers.
- AANs: Implementing Average Attention Networks to replace standard self-attention.
- Quantization: Converting 32-bit floats to 8-bit integers (INT8).
These techniques were evaluated on English-Spanish (WMT13) and English-German (WMT14) models, both individually and in "stacked" configurations.
The SimpleGEN Dataset
To measure bias precisely, the authors developed SimpleGEN. Unlike datasets that probe ambiguous gender, SimpleGEN uses templates to create sentences where the correct gender is explicitly signaled by context words like "lady," "man," "mother," or "father."
The dataset distinguishes between "pro-stereotypical" (e.g., a female nurse) and "anti-stereotypical" (e.g., a female mechanic) contexts. A model failure occurs when it ignores the explicit signal (e.g., "lady") and reverts to a stereotypical translation.
Key Insights: The BLEU-Bias Decoupling
The most significant finding is that BLEU scores are "bias-blind." As models are optimized for speed, BLEU scores remain relatively stable, while accuracy on anti-stereotypical gender translation drops precipitously.
The "Anti-Stereotypical" Tax
While models handle pro-stereotypical sentences well, speed optimizations cause a "crater" in performance for anti-stereotypical cases. Average Attention Networks (AANs) were particularly problematic, as their simplified mechanism often fails to capture the specific lexical cues (like "woman" or "guy") necessary for correct gender agreement.
Results and Discussion
When optimizations were stacked to achieve an 80% reduction in decoding time, BLEU scores only decreased by 6-9%. In contrast, gender translation accuracy in anti-stereotypical contexts plummeted by 22% to 41%. This suggests that the algorithmic "shortcuts" taken for speed disproportionately rely on the most frequent patterns in training data—often stereotypes—at the expense of nuanced contextual signals.
Limitations
The study is limited by its use of synthetic template-based data and its focus on the linguistic gender binary. Future research should explore mitigation strategies that preserve gender signals during the optimization process.
Conclusion
Efficiency in NMT comes with a hidden cost. Relying solely on aggregate metrics like BLEU can mask significant degradations in model fairness. As AI systems are optimized for real-world deployment, explicit monitoring for bias amplification is essential to ensure that speed does not come at the expense of equity.
Summary updated