PaperlayerView exact revision on arXiv

Source-linked AI summary

Avoiding the Hypothesis-Only Bias in Natural Language Inference via Ensemble Adversarial Training

Joe Stacey, Pasquale Minervini, Haim Dubossarsky, Sebastian Riedel, Tim Rocktäschel

arXiv:2004.07790v5cs.LGcs.AIcs.CLstat.ML

TL;DR

NLI models can exploit annotation artefacts, and earlier adversarial training may leave bias recoverable from sentence representations. This paper trains an ensemble of adversaries alongside the NLI model to reduce that bias. The authors report better generalization than previous de-biasing approaches across 12 NLI datasets.

  • Problem

    NLI annotation artefacts create spurious correlations, and prior adversarial training may leave bias recoverable from frozen sentence representations.

    Source: §1 · p. 1

  • Method

    The model jointly fits the NLI task and trains multiple hypothesis-only adversaries, while the encoder is trained to reduce their predictive accuracy.

    Source: §3 · p. 4

  • Results

    The ensemble method produced less biased representations and more robust models that generalized better across 12 NLI datasets than previous de-biasing approaches.

    Source: §5 · p. 8

  • Takeaways & Limitations

    Higher-dimensional sentence representations are harder to de-bias and appear to require more adversarial classifiers.

    Source: §5 · p. 8

    A nonlinear classifier may still detect bias in representations where linear classifiers cannot, so preventing nonlinear bias recovery merits further experimentation.

    Source: §5 · p. 9

Abstract

from arXiv · show

Natural Language Inference (NLI) datasets contain annotation artefacts resulting in spurious correlations between the natural language utterances and their respective entailment classes. These artefacts are exploited by neural networks even when only considering the hypothesis and ignoring the premise, leading to unwanted biases. Belinkov et al. (2019b) proposed tackling this problem via adversarial training, but this can lead to learned sentence representations that still suffer from the same biases. We show that the bias can be reduced in the sentence representations by using an ensemble of adversaries, encouraging the model to jointly decrease the accuracy of these different adversaries while fitting the data. This approach produces more robust NLI models, outperforming previous de-biasing efforts when generalised to 12 other datasets (Belinkov et al., 2019a; Mahabadi et al., 2020). In addition, we find that the optimal number of adversarial classifiers depends on the dimensionality of the sentence representations, with larger sentence representations being more difficult to de-bias while benefiting from using a greater number of adversaries.

1 Introduction

NLI models can exploit annotation artefacts, including hypothesis-only bias, which can undermine generalization beyond the training dataset. This paper investigates whether ensemble adversarial training can reduce bias in sentence representations while preserving or improving generalization.

1 Introduction

  • Dataset-specific annotation artefacts let NLI models rely on spurious correlations, contributing to poor generalization across datasets.Hypothesis-only cues include negation and hypothesis length.Source: §1 · p. 1
  • Adversarial training asks a shared encoder to fit the NLI task while reducing an adversary’s ability to predict from bias-related information.The goal is sentence representations that incorporate less information about the artefacts.Source: §1 · p. 1
  • Prior adversarial training can improve cross-dataset generalization, but bias may remain recoverable from frozen representations, leaving the source of those gains unclear.The paper argues that effective de-biasing should correspond to an observed reduction of bias in the representations.Source: §1 · p. 1
  • An ensemble of adversarial classifiers reduces hypothesis-only bias and outperforms previous de-biasing efforts across 12 other NLI datasets.The authors also report that higher-dimensional representations are harder to de-bias and require more adversaries.Source: §1 · pp. 1–2

2 Related Work

Related work describes hypothesis-only bias and weak cross-dataset generalization in NLI, alongside adversarial and ensemble approaches to addressing dataset artefacts. The paper positions its approach as targeting bias in the representations themselves.

Hypothesis-only bias

  • On SNLI, models using only the hypothesis can reach 67% accuracy against a 34% majority baseline, using cues such as negation and sentence length.Poliak et al. (2018) identify SNLI and MultiNLI among human-elicited datasets with the largest hypothesis-only bias; the paper focuses on SNLI.Source: §2 · p. 2

Generalisation to Other Datasets

  • NLI models can perform poorly on other datasets and on hard examples, defined as examples a hypothesis-only model misclassifies.Prior work finds that only MultiNLI shares some of SNLI’s hypothesis-only bias (Belinkov et al., 2019a).Source: §2 · p. 2

Biases and Artefacts

  • The paper frames its approach against broader evidence of dataset artefacts, including models that predict from text while ignoring other input modalities.The cited examples span reading comprehension, visual question answering, and story completion.Source: §2 · pp. 2–3

Adversarial Training

  • Adversarial training can encourage task-relevant features that are invariant to domain changes by reducing a domain classifier’s accuracy.Ganin and Lempitsky (2015) use gradient reversal in joint training of task and domain classifiers.Source: §2 · p. 3
  • Earlier adversarial training improved cross-dataset generalization, but bias could remain recoverable from frozen representations and SNLI-hard performance degraded.Related work also reports that adversarial training can leave demographic information in representations (Elazar and Goldberg, 2018).Source: §2 · p. 3

Using Model Ensembles

  • Prior ensemble approaches include adversarial classifiers for demographic information and Product of Experts models for robust NLI, without targeting bias in underlying representations.The paper compares its results with Mahabadi et al. (2020) to quantify the benefits of representation-level de-biasing.Source: §2 · p. 3

3 Ensemble Adversarial Training

The section describes an ensemble-adversarial training procedure and the experiments used to test representation de-biasing, adversary choice, and cross-dataset performance. The setup varies representation dimensionality and adversary type, then evaluates frozen representations and model generalization.

3 Ensemble Adversarial Training

  • The encoder and premise–hypothesis model fit NLI data while multiple hypothesis-only adversaries predict the label from the hypothesis representation and the encoder reduces their accuracy.The objective uses gradient reversal; the experiments vary representation dimensions and adversary counts.Source: §3 · pp. 3–4

3.1 Significance Testing

  • Significance testing repeats one- and five-adversary experiments across ten random seeds and uses bootstrap testing with a Bonferroni correction factor of four.P-values below 0.05 are treated as significant; Mann–Whitney U-test p-values are also reported.Source: §3.1 · p. 4

3.2 Using Deeper Adversaries

  • The deeper-adversary experiments compare linear and nonlinear classifiers during training and bias re-learning, including 512- and 2,048-dimensional representations.Figure 1 displays re-learned bias accuracy, independent bias-classifier accuracy at training’s end, and main NLI accuracy.Source: §3.2 · pp. 4–5, Figure 1 · p. 5

3.3 Evaluating De-biased Sentence Encoders

  • The evaluation compares SNLI-trained models with baseline and adversarial variants on SNLI-hard and 12 other NLI datasets, including a Product of Experts comparison.The cross-dataset setup follows Belinkov et al. (2019a), using a bidirectional LSTM for like-for-like comparison.Source: §3.3 · pp. 5–6
  • Table 1 reports maximum accuracy from 20 bias classifiers re-learning bias from frozen representations, across tested adversary counts and representation dimensions.The caption identifies tested counts of 0, 1, 5, 10, and 20 adversaries.Source: §3 · p. 4, Table 1 · p. 6

4 Results

An ensemble of adversaries reduces recoverable hypothesis-only bias, and the resulting models generalize better across other NLI datasets. The results also show that effective de-biasing depends on representation dimensionality and adversary strength.

4 Results

  • For 2,048-dimensional representations, increasing adversaries from 1 to 20 reduces hypothesis-only classifier accuracy from 62% to 53%.The paper reports that ensembles reduce bias across each tested representation dimensionality.Source: §4 · p. 6

4.4 Discussion

  • The optimal adversary count rises with representation dimension: 20 for 2,048 dimensions and 5 for 256, while gains plateau for 256, 512, and 1,028 dimensions.The discussion also links weaker de-biasing by a single adversary at higher dimensions to adversary strength.Source: §4 · p. 6, §4.4 · p. 8

4.1 Evaluating Multiple Adversaries

  • Five adversaries significantly reduce mean re-learnt bias versus one for 2,048-, 1,024-, and 512-dimensional representations, but not for 256 dimensions.The reported p-values are below 0.05 for the three larger dimensions.Source: §4.1 · p. 7, Table 2 · p. 8

4.2 Using Deeper Adversaries

  • When bias is relearned with a nonlinear classifier, a nonlinear adversary leaves less recoverable bias than a linear adversary; matching classifier types may help, but needs further testing.The best adversary depends on the classifier used to relearn bias.Source: §4.2 · p. 7

4.3 Evaluating De-biased Encoders

  • 0.9 points higher on average than one-adversary models, ensembles also outperform the Product of Experts approach; one-adversary models exceed the baseline by 1.6 points.The comparison covers 12 NLI datasets.Source: §4.3 · p. 8, Table 4 · p. 8

5 Conclusions

An ensemble of adversarial classifiers reduces hypothesis-only bias in NLI sentence representations and improves generalization across 12 datasets over previous approaches. Higher-dimensional representations are harder to de-bias and appear to require more adversarial classifiers.

  • Ensemble adversarial training produced less biased sentence representations that generalised better across 12 NLI datasets, outperforming previous approaches (Belinkov et al., 2019b; Mahabadi et al., 2020).Source: §5 · p. 8
  • Higher-dimensional sentence representations are harder to de-bias, and their optimal number of adversarial classifiers appears higher.Source: §5 · p. 8
  • On SNLI-hard, ensemble-trained models performed better than models using only one adversarial classifier.Source: §5 · p. 8
  • The authors report that these results match the expected behavior of models that no longer use hypothesis-only bias to inform predictions.Source: §5 · p. 9
  • The evidence establishes bias removal for linear classifiers, but nonlinear classifiers may still detect bias in the representations and require further study.Source: §5 · p. 9
Loading 2004.07790v5…