Probing Neural Network Comprehension of Natural Language Arguments

Abstract

We are surprised to find that BERT's peak performance of 77% on the Argument Reasoning Comprehension Task reaches just three points below the average untrained human baseline. However, we show that this result is entirely accounted for by exploitation of spurious statistical cues in the dataset. We analyze the nature of these cues and demonstrate that a range of models all exploit them. This analysis informs the construction of an adversarial dataset on which all models achieve random accuracy. Our adversarial dataset provides a more robust assessment of argument comprehension and should be adopted as the standard in future work.

arXiv commentACL 2019 (Updated Version)

References in corpus (2)

• Measuring the tendency of CNNs to Learn Surface Statistical Regularities
• The Argument Reasoning Comprehension Task: Identification and Reconstruction of Implicit Warrants

Cited by in corpus (22)

• Language Models are Few-Shot Learners
• Is BERT Really Robust? A Strong Baseline for Natural Language Attack on Text Classification and Entailment
• Do Explanations Reflect Decisions? A Machine-centric Strategy to Quantify the Performance of Explainability Algorithms
• Data and its (dis)contents: A survey of dataset development and use in machine learning research
• Gradient Starvation: A Learning Proclivity in Neural Networks
• Discrete and continuous representations and processing in deep learning: Looking forward
• Fairness via Representation Neutralization
• Beyond Leaderboards: A survey of methods for revealing weaknesses in Natural Language Inference data and models
• Pre-Trained Models: Past, Present and Future
• Behind the Scene: Revealing the Secrets of Pre-trained Vision-and-Language Models
• Connecting the Dots: A Knowledgeable Path Generator for Commonsense Question Answering
• A Hybrid Neural Network Model for Commonsense Reasoning
• Exploring Neural Models for Parsing Natural Language into First-Order Logic
• Stress Test Evaluation of Transformer-based Models in Natural Language Understanding Tasks
• Counterfactual Story Reasoning and Generation
• An Algorithm for Routing Capsules in All Domains
• On the Interplay Between Fine-tuning and Composition in Transformers
• Geometry matters: Exploring language examples at the decision boundary
• WaLDORf: Wasteless Language-model Distillation On Reading-comprehension
• The distance between the weights of the neural network is meaningful
• Deconfounding age effects with fair representation learning when assessing dementia
• Dependency-Based Neural Representations for Classifying Lines of Programs