Benchmarking Robustness of Machine Reading Comprehension Models

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 634–644 August 1–6, 2021. ©2021 Association for Computational Linguistics 634 Benchmarking Robustness of Machine Reading Comprehension Models Chenglei Si1∗ , Ziqing Yang2, Yiming Cui2,3, Wentao Ma2, Ting Liu3, Shijin Wang2,4 1 University of Maryland, College Park, MD, USA 2 Joint Laboratory of HIT and iFLYTEK Research (HFL), Beijing, China 3 Research Center for SCIR, Harbin Institute of Technology, Harbin, China 4 State Key Laboratory of Cognitive Intelligence, Beijing, China 1clsi@terpmail.umd.edu 2{zqyang5,ymcui,wtma,sjwang3}@iflytek.com 3{ymcui,tliu}@ir.hit.edu.cn Abstract Machine Reading Comprehension (MRC) is an important testbed for evaluating models’ natural language understanding (NLU) ability. There has been rapid progress in this area, with new models achieving impressive perfor- mance on various benchmarks. However, ex- isting benchmarks only evaluate models on in- domain test sets without considering their ro- bustness under test-time perturbations or ad- versarial attacks. To fill this important gap, we construct AdvRACE (Adversarial RACE), a new model-agnostic benchmark for evaluat- ing the robustness of MRC models under four different types of adversarial attacks, includ- ing our novel distractor extraction and gener- ation attacks. We show that state-of-the-art (SOTA) models are vulnerable to all of these attacks. We conclude that there is substan- tial room for building more robust MRC mod- els and our benchmark can help motivate and measure progress in this area. We release our data and code at https://github.com/ NoviScl/AdvRACE. 1 Introduction The goal of Machine Reading Comprehension (MRC) is to examine whether the model can under- stand the text and perform certain types of reason- ing. To this end, many MRC benchmarks have been constructed in different domains, styles, and lan- guages (Rajpurkar et al., 2016; Lai et al., 2017; Dua et al., 2019b; Cui et al., 2019, inter alia). While most of these benchmarks use leaderboards to com- pare different models’ performance on in-domain test sets, they have ignored the important aspect of evaluating models’ robustness. The research on robustness of MRC models can be generally categorised into two directions: gen- eralization to out-of-domain distributions and ro- ∗ Work done during an internship at HFL. bustness under test-time perturbations. Regard- ing generalization to out-of-domain distributions, Talmor and Berant (2019) and the MRQA shared task (Fisch et al., 2019) investigated how well do MRC models trained on source datasets generalize on unseen datasets in different domains. On evaluating MRC models under test time per- turbations, previous works (Jia and Liang, 2017; Gan and Ng, 2019) evaluate models under only one specific attack. However, truly robust models should perform well on different types of perturbed test inputs instead of just one. In this work, we cover diverse types of adversarial attacks to allow for a more comprehensive evaluation of model ro- bustness. Dua et al. (2019a) performed a set of synthetic perturbations to test models’ robustness. However, their perturbations are only applicable to specific types of questions, yielding limited num- ber of valid examples, limiting their usefulness in revealing model weakness. This may be because that they only performed perturbations on …

Content truncated to 500 words