Abstract
Natural Language Inference (NLI) models often rely on spurious correlations in human-annotated datasets, such as "hypothesis-only" biases, rather than learning true logical reasoning. This work introduces Ensemble Adversarial Training, a method that employs an ensemble of adversarial classifiers to de-bias sentence representations. By forcing the encoder to hide information from multiple adversaries simultaneously, the resulting models rely less on dataset artifacts and demonstrate superior generalization. We evaluate our approach across 12 NLI benchmarks, showing significant improvements in both bias reduction and cross-dataset performance.
Introduction
Natural Language Inference (NLI) is the task of determining if a "premise" sentence entails, contradicts, or is neutral toward a "hypothesis" sentence. Ideally, a model must understand the relationship between both sentences. However, human-annotated datasets like SNLI contain "annotation artifacts"—spurious correlations that allow a model to guess the correct answer by looking only at the hypothesis. For example, annotators often use negation words like "no" or "never" specifically for contradiction examples.
Neural networks are "lazy" learners; they often exploit these "hypothesis-only" biases rather than learning true logical reasoning. While previous researchers tried using adversarial training (where a secondary model tries to guess the label from the hypothesis-only representation), the bias often remained "hidden" in the representations. If you froze the model and trained a new classifier on its outputs, that new classifier could often "re-learn" the bias, proving it was never truly removed.
Methodology
The authors propose Ensemble Adversarial Training. Instead of one adversary, they use an ensemble of $n$ hypothesis-only adversarial classifiers.
- The Minimax Game: The training objective is a tug-of-war. The sentence encoder and task classifier try to minimize standard cross-entropy loss (accuracy). Simultaneously, the encoder tries to maximize the loss of the $n$ adversaries.
- Gradient Reversal: They use a Gradient Reversal Layer (GRL). During the backpropagation step, the gradients from the adversaries are multiplied by a negative constant before hitting the encoder, effectively telling the encoder to change its representations so these adversaries fail.
- The "Re-learning" Test: To prove the bias is actually gone, the authors freeze the trained encoder and attempt to train 20 independent classifiers to see if they can still extract the hypothesis-only bias. If these new classifiers fail, the de-biasing was successful.
Results and Discussion
Experiments were conducted varying the dimensionality of the sentence representations (from 256 to 2048) and the number of adversaries (up to 20). The primary training and "bias-source" dataset was SNLI, with generalization tested on 12 different datasets including SNLI-hard, MultiNLI, and various GLUE tasks.
Key Insights
- Bias is "Sticky": A single adversary is often insufficient. The encoder can find a way to satisfy one adversary while still keeping the biased information tucked away in a different "corner" of the high-dimensional vector space.
- Dimensionality Matters: Larger sentence representations (e.g., 2048 dimensions) are significantly harder to de-bias than smaller ones. Bigger vectors provide more "nooks and crannies" for spurious correlations to hide.
- The Ensemble Advantage: Using multiple adversaries forces the encoder to create a representation that is more robustly invariant to the bias. As the number of adversaries increases, the "re-learnable" bias consistently drops.
Performance Metrics
The ensemble approach (using 20 adversaries) outperformed previous state-of-the-art de-biasing methods like "Product of Experts" (PoE). Across 12 datasets, the ensemble model improved performance by +2.5 points over the baseline. For a 2048-dimensional representation, a single adversary only reduced the "re-learnable" bias accuracy to 62%, whereas 20 adversaries dropped it to 53% (approaching the 33% random baseline).
Conclusion
Ensemble Adversarial Training provides a robust framework for de-biasing NLI models. By addressing the "nooks and crannies" of high-dimensional representations where bias can hide, this method ensures that models learn more generalizable features. While training multiple adversaries increases computational overhead, the resulting models achieve statistically significant improvements on hard datasets like SNLI-hard, suggesting they truly learn to ignore "easy" shortcuts in favor of logical reasoning.
Summary updated