Matches in SemOpenAlex for { <https://semopenalex.org/work/W3208933101> ?p ?o ?g. }
- W3208933101 abstract "Large-scale pre-trained language models have achieved tremendous successacross a wide range of natural language understanding (NLU) tasks, evensurpassing human performance. However, recent studies reveal that therobustness of these models can be challenged by carefully crafted textualadversarial examples. While several individual datasets have been proposed toevaluate model robustness, a principled and comprehensive benchmark is stillmissing. In this paper, we present Adversarial GLUE (AdvGLUE), a new multi-taskbenchmark to quantitatively and thoroughly explore and evaluate thevulnerabilities of modern large-scale language models under various types ofadversarial attacks. In particular, we systematically apply 14 textualadversarial attack methods to GLUE tasks to construct AdvGLUE, which is furthervalidated by humans for reliable annotations. Our findings are summarized asfollows. (i) Most existing adversarial attack algorithms are prone togenerating invalid or ambiguous adversarial examples, with around 90% of themeither changing the original semantic meanings or misleading human annotatorsas well. Therefore, we perform a careful filtering process to curate ahigh-quality benchmark. (ii) All the language models and robust trainingmethods we tested perform poorly on AdvGLUE, with scores lagging far behind thebenign accuracy. We hope our work will motivate the development of newadversarial attacks that are more stealthy and semantic-preserving, as well asnew robust language models against sophisticated adversarial attacks. AdvGLUEis available at https://adversarialglue.github.io." @default.
- W3208933101 created "2021-11-08" @default.
- W3208933101 creator A5004007883 @default.
- W3208933101 creator A5021000040 @default.
- W3208933101 creator A5026746295 @default.
- W3208933101 creator A5034826937 @default.
- W3208933101 creator A5034995105 @default.
- W3208933101 creator A5041191241 @default.
- W3208933101 creator A5047233371 @default.
- W3208933101 creator A5066666034 @default.
- W3208933101 date "2021-11-04" @default.
- W3208933101 modified "2023-09-27" @default.
- W3208933101 title "Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models" @default.
- W3208933101 cites W2153579005 @default.
- W3208933101 cites W2243397390 @default.
- W3208933101 cites W2250539671 @default.
- W3208933101 cites W2251939518 @default.
- W3208933101 cites W2427527485 @default.
- W3208933101 cites W2607892599 @default.
- W3208933101 cites W2738015883 @default.
- W3208933101 cites W2759471388 @default.
- W3208933101 cites W2795038878 @default.
- W3208933101 cites W2798302089 @default.
- W3208933101 cites W2799194071 @default.
- W3208933101 cites W2892852825 @default.
- W3208933101 cites W2896457183 @default.
- W3208933101 cites W2911634294 @default.
- W3208933101 cites W2913470588 @default.
- W3208933101 cites W2922266396 @default.
- W3208933101 cites W2945067664 @default.
- W3208933101 cites W2951718443 @default.
- W3208933101 cites W2953084091 @default.
- W3208933101 cites W2953356739 @default.
- W3208933101 cites W2963126845 @default.
- W3208933101 cites W2963207607 @default.
- W3208933101 cites W2963310665 @default.
- W3208933101 cites W2963394326 @default.
- W3208933101 cites W2963859254 @default.
- W3208933101 cites W2963961878 @default.
- W3208933101 cites W2964082701 @default.
- W3208933101 cites W2964301649 @default.
- W3208933101 cites W2965373594 @default.
- W3208933101 cites W2970078867 @default.
- W3208933101 cites W2970449623 @default.
- W3208933101 cites W2970597249 @default.
- W3208933101 cites W2975059944 @default.
- W3208933101 cites W2990704537 @default.
- W3208933101 cites W2996403597 @default.
- W3208933101 cites W2996851481 @default.
- W3208933101 cites W3006647218 @default.
- W3208933101 cites W3013571468 @default.
- W3208933101 cites W3017003177 @default.
- W3208933101 cites W3024608270 @default.
- W3208933101 cites W3033187248 @default.
- W3208933101 cites W3034850762 @default.
- W3208933101 cites W3035164976 @default.
- W3208933101 cites W3035507081 @default.
- W3208933101 cites W3035688398 @default.
- W3208933101 cites W3035736465 @default.
- W3208933101 cites W3044324512 @default.
- W3208933101 cites W3101449015 @default.
- W3208933101 cites W3103934057 @default.
- W3208933101 cites W3104423855 @default.
- W3208933101 cites W3105662186 @default.
- W3208933101 cites W3120706522 @default.
- W3208933101 cites W3125455309 @default.
- W3208933101 cites W3128654100 @default.
- W3208933101 cites W3136077193 @default.
- W3208933101 cites W3168194750 @default.
- W3208933101 cites W3171654528 @default.
- W3208933101 hasPublicationYear "2021" @default.
- W3208933101 type Work @default.
- W3208933101 sameAs 3208933101 @default.
- W3208933101 citedByCount "0" @default.
- W3208933101 crossrefType "posted-content" @default.
- W3208933101 hasAuthorship W3208933101A5004007883 @default.
- W3208933101 hasAuthorship W3208933101A5021000040 @default.
- W3208933101 hasAuthorship W3208933101A5026746295 @default.
- W3208933101 hasAuthorship W3208933101A5034826937 @default.
- W3208933101 hasAuthorship W3208933101A5034995105 @default.
- W3208933101 hasAuthorship W3208933101A5041191241 @default.
- W3208933101 hasAuthorship W3208933101A5047233371 @default.
- W3208933101 hasAuthorship W3208933101A5066666034 @default.
- W3208933101 hasConcept C104317684 @default.
- W3208933101 hasConcept C119857082 @default.
- W3208933101 hasConcept C13280743 @default.
- W3208933101 hasConcept C137293760 @default.
- W3208933101 hasConcept C154945302 @default.
- W3208933101 hasConcept C162324750 @default.
- W3208933101 hasConcept C185592680 @default.
- W3208933101 hasConcept C185798385 @default.
- W3208933101 hasConcept C187736073 @default.
- W3208933101 hasConcept C204321447 @default.
- W3208933101 hasConcept C205649164 @default.
- W3208933101 hasConcept C2780451532 @default.
- W3208933101 hasConcept C37736160 @default.
- W3208933101 hasConcept C41008148 @default.
- W3208933101 hasConcept C55493867 @default.
- W3208933101 hasConcept C63479239 @default.
- W3208933101 hasConceptScore W3208933101C104317684 @default.