RobustSentEmbed: Robust Sentence Embeddings Using Adversarial Self-Supervised Contrastive Learning
Published in North American Chapter of the Association for Computational Linguistics, 2024
This paper introduces RobustSentEmbed, a method for obtaining robust sentence embeddings through an iterative collaboration between an adversarial perturbation generator and a PLM-based encoder. By generating high-risk perturbations in both token-level and sentence-level embedding spaces, RobustSentEmbed employs a contrastive learning objective combined with a token replacement detection objective to enhance similarity between original and adversarial embeddings.
Recommended citation: Javad Rafiei Asl, Prajwal Panzade, Eduardo Blanco, Daniel Takabi, Zhipeng Cai. "RobustSentEmbed: Robust Sentence Embeddings Using Adversarial Self-Supervised Contrastive Learning" North American Chapter of the Association for Computational Linguistics (NAACL 2024). https://arxiv.org/abs/2403.11082