Contextual Breach: Assessing the Robustness of Transformer-based QA Models

2024-09-17

arXiv.org

AI Summary

Existing studies on Transformer-based question answering (QA) models lack fine-grained adversarial noise benchmarks and unified quantitative metrics for evaluating robustness to context perturbations. Method: We construct the first comprehensive adversarial context perturbation benchmark—built upon SQuAD—with seven distinct noise types and five intensity levels, and propose a standardized robustness measurement framework enabling cross-noise-type and cross-intensity comparisons. Contribution/Results: Empirical evaluation reveals that mainstream QA models (e.g., BERT, RoBERTa) exhibit high sensitivity to context perturbations, with performance degradation following a pronounced nonlinear pattern; moreover, different noise types induce markedly heterogeneous impacts. This work establishes a reproducible benchmark, comparable evaluation metrics, and critical empirical evidence to advance the design and assessment of robust QA systems.

Technology Category

• Natural Language Processing: Question Answering
• Machine Learning: Adversarial Learning & Robustness
• Computer Vision: Adversarial Attacks & Robustness

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

Contextual question-answering models are susceptible to adversarial perturbations to input context, commonly observed in real-world scenarios. These adversarial noises are designed to degrade the performance of the model by distorting the textual input. We introduce a unique dataset that incorporates seven distinct types of adversarial noise into the context, each applied at five different intensity levels on the SQuAD dataset. To quantify the robustness, we utilize robustness metrics providing a standardized measure for assessing model performance across varying noise types and levels. Experiments on transformer-based question-answering models reveal robustness vulnerabilities and important insights into the model's performance in realistic textual input.

Problem

• Assessing transformer QA models' vulnerability to adversarial context perturbations
• Evaluating model robustness across seven noise types and five intensity levels
• Quantifying performance degradation in realistic noisy text scenarios

Innovation

• Created dataset with seven adversarial noise types
• Applied five intensity levels on SQuAD dataset
• Used robustness metrics to evaluate transformer models