Pre-trained language models have proven their unique powers in capturing implicit language features. However, most pre-training approaches focus on the word-level training objective, while sentence-level objectives are rarely studied. In this paper, we propose Contrastive LEArning for sentence Representation (CLEAR), which employs multiple sentence-level augmentation strategies in order to learn a noise-invariant sentence representation. These augmentations include word and span deletion, reordering, and substitution. Furthermore, we investigate the key reasons that make contrastive learning effective through numerous experiments. We observe that different sentence augmentations during pre-training lead to different performance improvements on various downstream tasks. Our approach is shown to outperform multiple existing methods on both SentEval and GLUE benchmarks.
@article{arxiv.2012.15466,
title = {CLEAR: Contrastive Learning for Sentence Representation},
author = {Zhuofeng Wu and Sinong Wang and Jiatao Gu and Madian Khabsa and Fei Sun and Hao Ma},
journal= {arXiv preprint arXiv:2012.15466},
year = {2021}
}