FreeLB: Enhanced Adversarial Training for Natural Language Understanding

Abstract

Adversarial training, which minimizes the maximal risk for label-preserving input perturbations, has proved to be effective for improving the generalization of language models. In this work, we propose a novel adversarial training algorithm, FreeLB, that promotes higher invariance in the embedding space, by adding adversarial perturbations to word embeddings and minimizing the resultant adversarial risk inside different regions around input samples. To validate the effectiveness of the proposed approach, we apply it to Transformer-based models for natural language understanding and commonsense reasoning tasks. Experiments on the GLUE benchmark show that when applied only to the finetuning stage, it is able to improve the overall test scores of BERT-base model from 78.3 to 79.4, and RoBERTa-large model from 88.5 to 88.8. In addition, the proposed approach achieves state-of-the-art single-model test accuracies of 85.44\% and 67.75\% on ARC-Easy and ARC-Challenge. Experiments on CommonsenseQA benchmark further demonstrate that FreeLB can be generalized and boost the performance of RoBERTa-large model on other tasks as well. Code is available at \url{https://github.com/zhuchen03/FreeLB .

arXiv commentAdding results with ALBERT

Cited by in corpus (70)

• Pre-trained Models for Natural Language Processing: A Survey
• A Survey on Data Augmentation for Text Classification
• Fine-Tuning Pretrained Language Models: Weight Initializations, Data Orders, and Early Stopping
• On the Stability of Fine-tuning BERT: Misconceptions, Explanations, and Strong Baselines
• Do Transformers Really Perform Bad for Graph Representation?
• RobustBench: a standardized adversarial robustness benchmark
• Generative Data Augmentation for Commonsense Reasoning
• A Simple but Tough-to-Beat Data Augmentation Approach for Natural Language Understanding and Generation
• Adversarial Training for Large Neural Language Models
• InfoBERT: Improving Robustness of Language Models from An Information Theoretic Perspective
• Understanding and Improving Fast Adversarial Training
• Adversarial Robustness: From Self-Supervised Pre-Training to Fine-Tuning
• A Closer Look at the Robustness of Vision-and-Language Pre-trained Models
• Scalable Multi-Hop Relational Reasoning for Knowledge-Aware Question Answering
• An Empirical Survey of Data Augmentation for Limited Data Learning in NLP
• Better Robustness by More Coverage: Adversarial Training with Mixup Augmentation for Robust Fine-tuning
• Better Fine-Tuning by Reducing Representational Collapse
• TAVAT: Token-Aware Virtual Adversarial Training for Language Understanding
• Adversarial Reinforced Instruction Attacker for Robust Vision-Language Navigation
• On the Effectiveness of Adapter-based Tuning for Pretrained Language Model Adaptation
• GNN is a Counter? Revisiting GNN for Question Answering
• Defending Against Backdoor Attacks in Natural Language Generation
• Self-Guided Contrastive Learning for BERT Sentence Representations
• Pair the Dots: Jointly Examining Training History and Test Stimuli for Model Interpretability
• Adversarial Feature Augmentation and Normalization for Visual Recognition
• Contrastive Learning with Adversarial Perturbations for Conditional Text Generation
• ASR-GLUE: A New Multi-task Benchmark for ASR-Robust Natural Language Understanding
• Improved OOD Generalization via Adversarial Training and Pre-training
• Fusing Context Into Knowledge Graph for Commonsense Question Answering
• Adversarial Training for Machine Reading Comprehension with Virtual Embeddings
• ExplanationLP: Abductive Reasoning for Explainable Science Question Answering
• Data-Efficient GAN Training Beyond (Just) Augmentations: A Lottery Ticket Perspective
• DropAttack: A Masked Weight Adversarial Training Method to Improve Generalization of Neural Networks
• Improving the Robustness of Transformer-based Large Language Models with Dynamic Attention
• Dynamic Semantic Graph Construction and Reasoning for Explainable Multi-hop Science Question Answering
• Consistency Regularization for Cross-Lingual Fine-Tuning
• TREATED:Towards Universal Defense against Textual Adversarial Attacks
• HiddenCut: Simple Data Augmentation for Natural Language Understanding with Better Generalization
• Adversarial Parameter Defense by Multi-Step Risk Minimization
• Transformer-based Language Model Fine-tuning Methods for COVID-19 Fake News Detection
• Fine-Tuning Pre-trained Language Model with Weak Supervision: A Contrastive-Regularized Self-Training Approach
• Contextual Text Style Transfer
• Improving Robustness and Generality of NLP Models Using Disentangled Representations
• Which *BERT? A Survey Organizing Contextualized Encoders
• Deepening Hidden Representations from Pre-trained Language Models
• How to Select One Among All? An Extensive Empirical Study Towards the Robustness of Knowledge Distillation in Natural Language Understanding
• How Does Adversarial Fine-Tuning Benefit BERT?
• CAP: Co-Adversarial Perturbation on Weights and Features for Improving Generalization of Graph Neural Networks
• Improving Commonsense Question Answering by Graph-based Iterative Retrieval over Multiple Knowledge Sources
• Are Adversarial Examples Created Equal? A Learnable Weighted Minimax Risk for Robustness under Non-uniform Attacks
• Enhancing Transformers with Gradient Boosted Decision Trees for NLI Fine-Tuning
• Reweighting Augmented Samples by Minimizing the Maximal Expected Loss
• MUSE: Multi-Scale Temporal Features Evolution for Knowledge Tracing
• AT-BERT: Adversarial Training BERT for Acronym Identification Winning Solution for SDU@AAAI-21
• CLINE: Contrastive Learning with Semantic Negative Examples for Natural Language Understanding
• Counterfactual Adversarial Learning with Representation Interpolation
• ARCH: Efficient Adversarial Regularized Training with Caching
• Adversarial Skill Learning for Robust Manipulation
• Noise Stability Regularization for Improving BERT Fine-tuning
• Searching for an Effective Defender: Benchmarking Defense against Adversarial Word Substitution
• Deep Neural Networks with Short Circuits for Improved Gradient Learning
• IIE-NLP-Eyas at SemEval-2021 Task 4: Enhancing PLM for ReCAM with Special Tokens, Re-Ranking, Siamese Encoders and Back Translation
• MaxVA: Fast Adaptation of Step Sizes by Maximizing Observed Variance of Gradients
• A Variational Inequality Approach to Bayesian Regression Games
• MvSR-NAT: Multi-view Subset Regularization for Non-Autoregressive Machine Translation
• Improving Gradient-based Adversarial Training for Text Classification by Contrastive Learning and Auto-Encoder
• Virtual Data Augmentation: A Robust and General Framework for Fine-tuning Pre-trained Models
• Multilingual and Multilabel Emotion Recognition using Virtual Adversarial Training
• Initializing Perturbations in Multiple Directions for Fast Adversarial Training
• Achieving Model Robustness through Discrete Adversarial Training