NeurIPS 2021

Timezone: »

Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models

Boxin Wang · Chejian Xu · Shuohang Wang · Zhe Gan · Yu Cheng · Jianfeng Gao · Ahmed Awadallah · Bo Li

@ None

in Dataset and Benchmark Poster Session 1 »

Large-scale pre-trained language models have achieved tremendous success across a wide range of natural language understanding (NLU) tasks, even surpassing human performance. However, recent studies reveal that the robustness of these models can be challenged by carefully crafted textual adversarial examples. While several individual datasets have been proposed to evaluate model robustness, a principled and comprehensive benchmark is still missing. In this paper, we present Adversarial GLUE (AdvGLUE), a new multi-task benchmark to quantitatively and thoroughly explore and evaluate the vulnerabilities of modern large-scale language models under various types of adversarial attacks. In particular, we systematically apply 14 textual adversarial attack methods to GLUE tasks to construct AdvGLUE, which is further validated by humans for reliable annotations. Our ﬁndings are summarized as follows. (i) Most existing adversarial attack algorithms are prone to generating invalid or ambiguous adversarial examples, with around 90% of them either changing the original semantic meanings or misleading human annotators as well. Therefore, we perform a careful ﬁltering process to curate a high-quality benchmark. (ii) All the language models and robust training methods we tested perform poorly on AdvGLUE, with scores lagging far behind the benign accuracy. We hope our work will motivate the development of new adversarial attacks that are more stealthy and semantic-preserving, as well as new robust language models against sophisticated adversarial attacks. AdvGLUE is available at https://adversarialglue.github.io.

[ Chat] [ Paper ]

Author Information

Boxin Wang (Department of Computer Science, University of Illinois, Urbana Champaign)

Chejian Xu (Zhejiang University)

Shuohang Wang (Microsoft)

Zhe Gan (Duke University)

Yu Cheng (Microsoft Research)

Jianfeng Gao (Microsoft Research, Redmond, WA)

Ahmed Awadallah (MICROSOFT RESEARCH)

I am passionate about using AI and Machine Learning to create intelligent user experiences that connect people to information. I lead a research and incubation team in Microsoft Research Technologies. Our work at the Language and Information Technologies team is focused on creating language understanding and user modeling technologies to enable intelligent experiences in multiple products. Our work has been shipped in several products such as Bing, Cortana, Office 365, and Dynamics 365. I have hands-on experience building and shipping state-of-the-art ML/AI algorithms. I also have experience building and managing world-class teams of scientists and engineers. My research interests are at the intersection of machine learning, language understanding, and information retrieval. A key part of my work involves using Machine Learning to model large-scale text and user behavior data with applications to intelligent assistants, search, user modeling, quality evaluation, recommendation and personalization. I received my Ph.D. from the department of Computer Science and Engineering at the University of Michigan Ann Arbor. I Invented, published, and patented new approaches in language understanding, information retrieval and machine learning. I published 60+ peer-reviewed papers in these areas and I am an inventor on 20+ (granted and pending) patents.