CSE PhD student Animesh Nighojkar and CSE Assistant Professor John Licato’s paper, “Improving Paraphrase Detection with the Adversarial Paraphrasing Task,” was accepted as a full paper in the main track for the 2021 Annual Meeting of the Association for Computational Linguistics (ACL). In this paper, Nighojkar and Licato showed that existing benchmarks and measures for detecting paraphrases (whether two sentences mean the same thing) are severely limited, as they rely too much on word-level meanings rather than sentence-level meaning. To address this, they used an adversarial paradigm design where human participants intentionally tried to trick the AI models in order to help them improve. The resulting data sets yielded immediate improvement in the performance of the models. This has implications for future work involving plagiarism detection, argument analysis, misleading news headline detection, and more.
The Annual Meeting of the ACL is considered by some to be the most prestigious publication venue in the subfield of NLP / Computational Linguistics. The target acceptance rate for the 2021 ACL meeting is estimated to be between 21 and 24%. The conference will take place in August.