## 論文リンク https://arxiv.org/pdf/2107.00440.pdf ## 概要 意味は似ているが推論を間違えさせる敵対的サンプルを正例、意味が正反対なサンプル(contrastive examples)を負例として contrastive learning。 sentiment analysis、reading comprehension、NLIなどいくつかのデータセットで性能向上かつよりロバストに また、contrastive examplesにもロバストなモデルを実現 <img width="683" alt="Screen Shot 2021-10-05 at 16 05 25" src="https://user-images.githubusercontent.com/44390274/135975952-ad7e64c1-fc47-413b-bdeb-b3d2bccc94e7.png">