mlsanityData to 7 Oct 2026

Using Adversarial Attacks to Reveal the Statistical Bias in Machine Reading Comprehension Models

arXiv 2105.11136 ↗First posted 24 May 2021cs.CLcited by 29

Not on the map. Why?

Abstract (arXiv)

Pre-trained language models have achieved human-level performance on many Machine Reading Comprehension (MRC) tasks, but it remains unclear whether these models truly understand language or answer questions by exploiting statistical biases in datasets. Here, we demonstrate a simple yet effective method to attack MRC models and reveal the statistical biases in these models. We apply the method to the RACE dataset, for which the answer to each MRC question is selected from 4 options. It is found that several pre-trained language models, including BERT, ALBERT, and RoBERTa, show consistent preference to some options, even when these options are irrelevant to the question. When interfered by these irrelevant options, the performance of MRC models can be reduced from human-level performance to the chance-level performance. Human readers, however, are not clearly affected by these irrelevant options. Finally, we propose an augmented training method that can greatly reduce models' statistical biases.

Key facts. Cited by 29 papers in the map (data through 7 Oct 2026). Builds on 16 of its 27 references that are in the map. First posted on arXiv 24 May 2021.

Builds on

16 of 27 references.

References · 16

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  2. 2020Language Models are Few-Shot Learnerscited by 32,199
  3. 2019RoBERTa: A Robustly Optimized BERT Pretraining Approachcited by 13,266
  4. 2019ALBERT: A Lite BERT for Self-supervised Learning of Language Representationscited by 2,772
  5. 2017RACE: Large-scale ReAding Comprehension Dataset From Examinationscited by 851
  6. 2019Universal Adversarial Triggers for Attacking and Analyzing NLPcited by 709
  7. 2018Annotation Artifacts in Natural Language Inference Datacited by 656
  8. 2017Adversarial Examples for Evaluating Reading Comprehension Systemscited by 568
  9. 2017Generating Natural Adversarial Examplescited by 182
  10. 2018Seq2Sick: Evaluating the Robustness of Sequence-to-Sequence Models with Adversarial Examplescited by 93
  11. 2019What does BERT Learn from Multiple-Choice Reading Comprehension Datasets?cited by 30
  12. 2017Large-scale Cloze Test Dataset Created by Teacherscited by 27
4 more
  1. 2019Dual Co-Matching Network for Multi-choice Reading Comprehensioncited by 26
  2. 2020Benchmarking Robustness of Machine Reading Comprehension Modelscited by 18
  3. 2020STARC: Structured Annotations for Reading Comprehensioncited by 17
  4. 2019Option Comparison Network for Multiple-choice Reading Comprehensioncited by 14

Built on by

Cited by year

Citing papers per year: 2022: 6; 2023: 7; 2024: 7; 2025: 9.

29 papers.

Sort:

By area:

  1. 2024Advancing the Robustness of Large Language Models through Self-Denoised Smoothingcited by 15
  2. 2024PertEval: Unveiling Real Knowledge Capacity of LLMs with Knowledge-Invariant Perturbationscited by 13
  3. 2022A Prompting-based Approach for Adversarial Example Generation and Robustness Enhancementcited by 12
  4. 2022Divide and Conquer: Text Semantic Matching with Disentangled Keywords and Intentscited by 9
  5. 2023TextShield: Beyond Successfully Detecting Adversarial Sentences in Text Classificationcited by 8
  6. 2023DSRM: Boost Textual Adversarial Training with Distribution Shift Risk Minimizationcited by 7
  7. 2024Image-to-Text Logic Jailbreak: Your Imagination can Help You Do Anythingcited by 7
  8. 2022ROSE: Robust Selective Fine-tuning for Pre-trained Language Modelscited by 6
  9. 2023LEAP: Efficient and Automated Test Method for NLP Softwarecited by 6
  10. 2022Rethinking Textual Adversarial Defense for Pre-trained Language Modelscited by 5
  11. 2022Efficient Adversarial Training with Robust Early-Bird Ticketscited by 5
  12. 2025Are LLMs Stable Formal Logic Translators in Logical Reasoning Across Linguistically Diversified Texts?cited by 5
17 more
  1. 2022A Survey on Measuring and Mitigating Reasoning Shortcuts in Machine Reading Comprehensioncited by 4
  2. 2024Unveiling the Achilles' Heel of NLG Evaluators: A Unified Adversarial Framework Driven by Large Language Modelscited by 4
  3. 2024Towards Action Hijacking of Large Language Model-based Agentcited by 4
  4. 2025No Query, No Accesscited by 4
  5. 2025IM-BERT: Enhancing Robustness of BERT through the Implicit Euler Methodcited by 3
  6. 2025AutoEvoEval: An Automated Framework for Evolving Close-Ended LLM Evaluation Datacited by 3
  7. 2025The Strong, Weak and Benign Goodhart's law. An independence-free and paradigm-agnostic formalisationcited by 2
  8. 2025Text Adversarial Attacks with Dynamic Outputscited by 2
  9. 2025Style Attack Disguise: When Fonts Become a Camouflage for Adversarial Intentcited by 2
  10. 2024Subspace Defense: Discarding Adversarial Perturbations by Learning a Subspace for Clean Signalscited by 1
  11. 2024Simpler becomes Harder: Do LLMs Exhibit a Coherent Behavior on Simplified Corpora?cited by 1
  12. 2023EMBRACE: Evaluation and Modifications for Boosting RACEcited by 0
  13. 2023SCAT: Robust Self-supervised Contrastive Learning via Adversarial Training for Text Classificationcited by 0
  14. 2023SenTest: Evaluating Robustness of Sentence Encoderscited by 0
  15. 2023From text to multimodal: a survey of adversarial example generation in question answering systemscited by 0
  16. 2025Coordinated Robustness Evaluation Framework for Vision-Language Modelscited by 0
  17. 2025Innocence in the Crossfire: Roles of Skip Connections in Jailbreaking Visual Language Modelscited by 0