CDAE Enhances BERT Embedding Robustness to Textual Perturbations
July 31, 2026
The Contrastive Denoising Autoencoder (CDAE) improves pretrained BERT embedding stability against synonym substitution and masking. It achieves this by jointly optimizing contrastive and reconstruction objectives to learn perturbation-invariant representations.
HOW THIS AFFECTS YOU
●
researcherThis offers a lightweight way to improve representation stability without full model retraining.