PaperlayerView exact revision on arXiv

Source-linked AI summary

More Bang for Your Buck: Natural Perturbation for Robust Question Answering

Daniel Khashabi, Tushar Khot, Ashish Sabharwal

arXiv:2004.04849v2cs.CLcs.AIcs.LG

TL;DR

Models can be sensitive to small input changes, while constructing training sets from scratch is expensive. The paper expands seed examples with human-authored minimal perturbations and evaluates the approach on BoolQ. At a cost ratio of 0.6, perturbed-data models improve robustness and cross-dataset generalization over BOOLQ-trained models while retaining performance on the original task.

  • Problem

    Models can be sensitive to small input changes, and creating large training datasets by writing independent examples from scratch is expensive.

    Source: §1 · p. 1

  • Method

    The approach expands seed examples with human-authored minimal perturbations, often changing the answer, and evaluates training data under different perturbation cost ratios.

    Source: §1 · p. 1, §3 · p. 3

  • Results

    At cost ratio 0.6, perturbed-data models are 9% more robust to minor changes and generalize 4.5% better across datasets than BOOLQ-trained models, while retaining performance on the original task.

    Source: §1 · p. 2, §6 · p. 5

  • Takeaways & Limitations

    Perturbation clusters can yield cost-efficiently trained, robust models that generalize across datasets.

    Source: §1 · p. 2

    The study leaves the effects of dataset budget and large cost ratios for future work; over-generation can produce overly similar, less-informative variations.

    Source: §6 · p. 5

Abstract

from arXiv · show

While recent models have achieved human-level scores on many NLP datasets, we observe that they are considerably sensitive to small changes in input. As an alternative to the standard approach of addressing this issue by constructing training sets of completely new examples, we propose doing so via minimal perturbation of examples. Specifically, our approach involves first collecting a set of seed examples and then applying human-driven natural perturbations (as opposed to rule-based machine perturbations), which often change the gold label as well. Local perturbations have the advantage of being relatively easier (and hence cheaper) to create than writing out completely new examples. To evaluate the impact of this phenomenon, we consider a recent question-answering dataset (BoolQ) and study the benefit of our approach as a function of the perturbation cost ratio, the relative cost of perturbing an existing question vs. creating a new one from scratch. We find that when natural perturbations are moderately cheaper to create, it is more effective to train models using them: such models exhibit higher robustness and better generalization, while retaining performance on the original BoolQ dataset.

1 Introduction

The paper proposes expanding seed examples with human-authored minimal perturbations as a cheaper alternative to writing every example from scratch. On BoolQ, this approach improves robustness and cross-dataset generalization, with gains reported even at a moderate perturbation cost ratio.

1 Introduction

  • The method expands a small seed dataset with human-authored minimal perturbations, which are encouraged to change answers and form clusters with mixed labels.Figure 1 illustrates three seed instances expanding into ten instances, with two to four perturbations per seed.Source: §1 · p. 1, Figure 1 · p. 1
  • The motivation is that models can be sensitive to small input changes: a RoBERTa model’s BoolQ accuracy dropped 15% on locally perturbed questions that were no harder for humans.The paper asks whether a different training-set construction approach can alleviate this issue.Source: §1 · p. 2
  • At cost ratio 0.6, perturbed-data models are 9% more robust to minor changes and generalize 4.5% better across datasets than BOOLQ-trained models.The authors also report that performance on the original task is largely retained; on unseen, unperturbed questions from MultiRC, perturbed-data models outperform models trained on traditional data.Source: §1 · p. 2
  • Even at a cost ratio of 1.0, perturbed-example models remain competitive across evaluation sets, a potential use case when distinct training examples are scarce.The authors suggest minimal perturbations of limited real examples may be as effective as hypothetically having access to large amounts of real data.Source: §1 · p. 2
  • The proposed construction method combines independent example collection with minimal natural perturbations to produce cost-efficiently trained, robust models that generalize across datasets.The paper describes perturbing existing examples as often substantially cheaper than creating new examples from scratch.Source: §1 · pp. 1–2

2 Related Work

The paper distinguishes its use of human-authored perturbations for dataset construction from augmentation and adversarial work. It also connects its approach to minimal-pair datasets while noting that the value of larger clusters remains unclear.

Data augmentation

Adversarial perturbations

  • Unlike adversarial perturbation work that seeks examples to expose system weaknesses, this approach expands datasets cost-efficiently with human-authored changes that are harder to model.Prior rule-based approaches often use label-preserving, shallow perturbations, which can be reverse-engineered by models (Jia and Liang, 2017).Source: §2 · p. 2

Minimal-pairs in NLP

  • The value of minimal-pair datasets is not well understood, while findings about perturbation clusters could inform future dataset construction for such tasks.Minimal-pair datasets include Winograd schema datasets (Sakaguchi et al., 2020) and contrast sets (Gardner et al., 2020).Source: §2 · pp. 2–3

3 Expansion via Perturbation Clusters

The approach expands seed examples by asking crowdworkers to make minimal, natural perturbations, then independently verifies the resulting questions. It evaluates the value of perturbation clusters under a fixed budget as their size and relative cost vary.

3 Expansion via Perturbation Clusters

  • Crowdworkers make minimal changes to seed questions, often to flip their answers, while aiming to produce challenging, diverse questions.Workers are encouraged to change up to four terms; a trained RoBERTa model guides the challenge level, but questions need not fool it.Source: §3 · p. 3
  • An independent verification stage obtains gold labels and checks that each question is answerable from its paragraph in isolation.Questions lacking a consistent majority label or judged unanswerable are filtered out.Source: §3 · p. 3
  • The BOOLQ Expansion is a naturally perturbed dataset containing 17k questions, obtained by sampling from BOOLQ and derived from 4k seed questions.Source: §3 · p. 3
  • The perturbation cost ratio r is the cost of creating a perturbed question relative to writing a new one; at a fixed budget, lower r permits more questions.For example, when r = 0.5, one newly written question costs as much as two perturbed questions.Source: §3 · p. 3
  • At a fixed budget, the study compares subsets by maximum cluster size and cost ratio; singleton clusters resemble standard construction, while size-four clusters approximate the expanded BOOLQ data.Figure 2 also indicates that adding perturbations per cluster can increase accuracy, particularly when perturbations are cheap.Source: §3 · pp. 3–4, Figure 2 · p. 4
  • At a cost ratio of 0.6, models trained on perturbed datasets gain 3–5% over models trained on traditionally constructed data.Figure 3 compares singleton clusters with clusters of size four under a fixed total budget.Source: Figure 3 · p. 4

4 Experiments

The experiments compare models trained on BOOLQ subsets built under different cluster sizes and perturbation costs. Perturbation clusters can improve performance when they are sufficiently cheap, while larger clusters also add value when the number of clusters is fixed.

4 Experiments

  • RoBERTa-large is evaluated on BOOLQ-e for robustness, MULTIRC for generalization, and the original BOOLQ test set for retained performance.Each plotted point averages five random dataset subsamples, with standard-deviation error bars.Source: §4 · p. 4

4.1 Effect of Cluster Size (c)

  • With a fixed number of clusters, accuracy consistently rises as cluster size grows; with a fixed number of questions, cluster size does not affect quality on two of three datasets.The comparisons use r = 0 and r = 1, respectively.Source: §4.1 · pp. 4–5

4.2 Effect of Perturbations Cost Ratio (r)

  • At r = 0.33, perturbed training was about 10% more effective on BOOLQ-e and 5% on MULTIRC; results across datasets indicate value in clusters when r ≤ 0.6.The estimate assumes a $0.60 cost per newly written question and $0.20 per perturbed question, with total budget b = 1500.Source: §4.2 · p. 5
  • As perturbations become cheaper, larger clusters can be more cost-effective; even when they are not much cheaper, they retain performance comparable to independent examples.The authors identify such clusters as an alternative for expanding datasets when distinct examples are scarce.Source: §4.2 · p. 5

5 Discussion

The authors are cautiously optimistic that natural perturbation may generalize beyond yes/no questions, but note that its value depends on the relative cost of creating diverse, challenging perturbations. Their discussion assumes workers write questions from fixed contexts, an assumption that may not fit other dataset-generation pipelines.

5 Discussion

  • Whether perturbation construction generalizes to other tasks remains an open question, and its value may depend on the cost of producing diverse, challenging changes.The authors chose yes/no questions as a comparatively underexplored QA area and express cautious optimism about broader applicability.Source: §5 · p. 5
  • The proposed workflow assumes workers write questions from fixed contexts, which may not hold when a pipeline starts from an existing question set.The authors suggest applying perturbations at another annotation stage in such cases.Source: §5 · p. 5

6 Conclusion

The paper concludes that expanding seed examples with natural perturbations can yield robust models that generalize while preserving BOOLQ performance when perturbations are moderately cheap. It also releases the BOOLQ perturbation resource and flags limits to the value of ever-larger clusters.

6 Conclusion

  • Models trained on BOOLQ perturbations are more robust to minor variations and generalize better while preserving original BOOLQ test performance when perturbations are moderately cheap.The authors report notable gains even at a cost ratio of 0.6.Source: §6 · p. 5
  • The study provides the BOOLQ natural-perturbation resource, while its central focus is the value of perturbations for robust model design.The resource was constructed during the study.Source: §6 · p. 5
  • Over-generating perturbations can produce overly similar, less informative variants, so larger clusters are valuable only up to a point.The effects of total dataset budget and large cluster sizes are left for future study.Source: §6 · p. 5

A Question Perturbations: Further Details

The annotation process separates yes- and no-question perturbations into phases after a pilot found that workers used different strategies for each. Annotators qualify through instructions and a test, with additional worker eligibility requirements.

A Question Perturbations: Further Details

B Details of ROBERTA Training

The authors fine-tune ROBERTA for two-way questions using a passage, question, and candidate answer as input. They select models using validation accuracy after a small hyperparameter sweep.

B Details of ROBERTA Training

C Performances Across Datasets

Across the reported datasets, systems fine-tuned on BOOLQ++ generalize better than systems trained on BOOLQ alone. ROBERTA also shows substantial performance degradation outside within-BOOLQ evaluation.

C Performances Across Datasets

D Cluster-Level Evaluation

The cluster-level evaluation measures whether models answer related questions consistently, using consensus scores over subsets of each cluster. Machine solvers lose consensus more steeply than humans as subset size grows.

D Cluster-Level Evaluation

E Rule-Based Perturbations

The paper compares human-authored perturbations with cheaper rule-based alternatives, which it characterizes as limited in sophistication and usefulness for model training.

E Rule-Based Perturbations

  • Rule-based paraphrase systems are presented as a potentially cheaper alternative to human-annotated perturbations.The paper reports putting effort into machine-generated baselines but excluding them from the main text because they were not reasonably sophisticated.Source: E Rule-Based Perturbations · p. 8
  • Automatic question rephrasing generally preserves answers and omits the question’s provided context, limiting perturbation richness.The authors note that more interesting and diverse rule-based perturbations can be difficult to develop and may be reverse-engineered by QA models.Source: E Rule-Based Perturbations · p. 8
  • The examples contrast a machine paraphrase that retains the original FALSE answer with human perturbations that take the provided context into account.The original question asks whether there will be a season four of Da Vinci’s Demons; the listed machine paraphrases retain FALSE.Source: E Rule-Based Perturbations · p. 8
  • Machine-generated perturbations were generally minor and did not provide a useful enough signal to improve model accuracy.The authors therefore did not include their rule-based/machine-generated baselines in the main text because they were not reasonably sophisticated.Source: E Rule-Based Perturbations · p. 8

Related papers

Loading 2004.04849v2…