Researchers at MIT have created a framework—TextFooler—that brought down the prediction accuracy of certain NLP models from 90% down to under 20% by simply using synonyms in place of certain words.

In the scope of this paper, a paradigm is a general modeling framework or a distinct set of methodologies to solve a class of tasks. For instance, sequence labeling is a mainstream paradigm for named ...