Posterior Differential Regularization with f-divergence for Improving Model Robustness

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1078–1089 June 6–11, 2021. ©2021 Association for Computational Linguistics 1078 Posterior Differential Regularization with f-divergence for Improving Model Robustness Hao Cheng1, Xiaodong Liu1, Lis Pereira2, Yaoliang Yu3, Jianfeng Gao1 1Microsoft Research 2Ochanomizu University 3University of Waterloo & Vector Institute {chehao,xiaodl,jfgao}@microsoft.com {kanashiro.pereira}@ocha.ac.jp {yaoliang.yu}@uwaterloo.ca Abstract We address the problem of enhancing model robustness through regularization. Specifi- cally, we focus on methods that regularize the model posterior difference between clean and noisy inputs. Theoretically, we provide a con- nection of two recent methods, Jacobian Regu- larization and Virtual Adversarial Training, un- der this framework. Additionally, we general- ize the posterior differential regularization to the family of f-divergences and characterize the overall framework in terms of Jacobian ma- trix. Empirically, we compare those regular- izations and standard BERT training on a di- verse set of tasks to provide a comprehensive profile of their effect on model generalization. For both fully supervised and semi-supervised settings, we show that regularizing the poste- rior difference with f-divergence can result in well-improved model robustness. In particu- lar, with a proper f-divergence, a BERT-base model can achieve comparable generalization as its BERT-large counterpart for in-domain, adversarial and domain shift scenarios, indicat- ing the great potential of the proposed frame- work for enhancing NLP model robustness.1 1 Introduction Although recent neural network based models have achieved great success in a wide range of natu- ral language processing (NLP) tasks, these models may still suffer catastrophic degradation in out-of- domain generalization to datasets with domain shift or adversarial scenarios (Nie et al., 2019; Hsieh et al., 2019). For example, large-scale pretrained neural language models (Devlin et al., 2019; Liu et al., 2019) have better generalization, but experi- ence performance reduction caused by domain shift (Hendrycks et al., 2020). Textual entailment mod- els trained on MNLI (Williams et al., 2018) have picked up superficial cues focusing on either the 1Code is available at https://github.com/ hao-cheng/f-divergence. 50 80 in-domain out-of-domain 39% 25% 31% Adversarial QA Domain Shift QA Adversarial NLI Figure 1: BERT-based question answering (QA) and natural language inference (NLI) model robustness to- wards domain shift and adversarial attack. The num- bers in percentage indicate the corresponding rela- tive performance drop between in-domain and out-of- domain test sets. presence of certain keywords (Gururangan et al., 2018) or whether similar words are mentioned in the sentence pairs (McCoy et al., 2019). Jia and Liang (2017) have also shown that SQuAD models are very easily distracted by irrelevant sentences that contain many question words regardless of context, despite of their human-performance on in-domain data. As shown in Figure 1, three BERT- based (Devlin et al., 2019) models perform well for in-domain evaluation data, but transfer poorly to out-of-domain datasets with domain shift or ad- versarial attack, i.e. more than 25% relative perfor- mance reduction. Achieving good generalizations towards datasets with domain shift has been a long-standing goal of domain adaptation. Various methods (Blitzer et …

Content truncated to 500 words