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

Daniel Khashabi, Tushar Khot, Ashish SabharwalEMNLP 2020 · OpenReview

Abstract: Daniel Khashabi, Tushar Khot, Ashish Sabharwal. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP).

2020.

Introduction

Training state-of-the-art Natural Language Processing (NLP) models is an expensive endeavor that often yields "brittle" systems. These models frequently achieve high scores on standard benchmarks but fail when faced with minor, non-essential input changes. For instance, a model might correctly answer a question about a historical fact but fail if the date in the question is slightly altered.

Figure 1: Conceptual overview of natural perturbations. A seed question (left) is minimally edited by humans to create a cluster of related questions (right) with flipped answers, forcing models to learn precise decision boundaries.

This paper proposes a more efficient and effective data collection strategy: instead of writing every example from scratch, annotators create "natural perturbations" of existing seed examples. This approach produces "perturbation clusters" that are significantly cheaper to build and result in models with higher robustness and better generalization across datasets.

The Natural Perturbation Framework

The authors introduce a two-step framework for expanding a small "seed" dataset into a larger, more robust training set:

  1. Natural Minimal Perturbations: Crowdworkers are provided with a seed question and its corresponding answer. They are tasked with making the smallest possible change (ideally under four words) that "flips" the answer (e.g., changing a "Yes" to a "No"). Unlike rule-based machine perturbations, these are "natural"—authored by humans who understand linguistic context and nuance.
  2. Quality Verification: To ensure validity, a second set of annotators verifies the perturbed questions in isolation, confirming that the new labels are correct.

A central concept in this study is the perturbation cost ratio ($r$), defined as the relative cost of perturbing an existing question versus creating a new one from scratch. The authors investigate how model performance scales as this ratio and the size of the perturbation clusters vary.

Experimental Setup and Datasets

The research primarily utilizes BoolQ, a dataset of naturally occurring yes/no questions, as the source for "seed" examples. From this, the authors developed:

  • $BOOLQ_\omega$: An expanded dataset containing approximately 17,000 questions derived from 4,000 seeds, with each seed having a cluster of 2–4 minimal perturbations.
  • $BOOLQ-\epsilon_\omega$: An expert-generated test set designed to evaluate model robustness against subtle, local changes.
  • MultiRC: An "out-of-domain" evaluation set used to test how well models generalize to different reading comprehension tasks.

Key Insights

The core philosophy of this approach is that clusters matter more than individual points. Traditional datasets are collections of independent data points. In contrast, training on perturbation clusters forces the model to learn the precise decision boundaries between answers rather than relying on keyword memorization.

Furthermore, the authors argue that label flipping is crucial. While many automated augmentation techniques (like back-translation) preserve the original label, changing the label through minimal edits provides a much stronger learning signal. Human-authored perturbations are superior to rule-based ones because they capture nuances that models cannot easily "reverse-engineer."

Results and Analysis

The experimental results demonstrate significant gains across multiple dimensions:

  • Robustness Boost: Models trained on $BOOLQ_\omega$ showed a 9% improvement in robustness on the $BOOLQ-\epsilon_\omega$ challenge set.
  • Enhanced Generalization: These models performed 4.5% better on the MultiRC dataset, suggesting a shift from simple pattern matching to a deeper logical understanding.
  • Efficiency and the "60% Rule": The authors found that as long as the cost of perturbing a question is less than 60% of the cost of writing a new one, using perturbations is more effective for training.
Figure 6: Model performance across BoolQ-e, MultiRC, and BoolQ as a function of max cluster size (c) and cost ratio (r). The data shows that increasing cluster size generally aids robustness and generalization, remaining competitive even when perturbation costs are high.

Importantly, these gains do not come at the expense of standard performance; the models retained their high accuracy on the original BoolQ test set. Even when the cost of perturbation was equal to the cost of a new example ($r=1.0$), cluster-based training remained competitive, proving that the structure of the data is a vital component of model performance.

Conclusion and Future Work

This work highlights a path toward more robust NLP systems through smarter data annotation. By focusing on natural, minimal perturbations, researchers can build datasets that challenge models to understand the underlying logic of a task.

While successful for boolean question answering, future work should explore how this approach scales to more complex tasks like extractive QA or summarization. Additionally, further research is needed to determine the point of diminishing returns when over-generating perturbations for a single seed and to refine economic models for crowdsourcing these tasks.

Summary updated

BibTeX
@inproceedings{khashabi2020more,
  author = {Daniel Khashabi and Tushar Khot and Ashish Sabharwal},
  title = {{More Bang for Your Buck: Natural Perturbation for Robust Question Answering}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  url = {https://openreview.net/forum?id=unpHog5w3yH},
}

Loading related content...

Figure