ModernBERT Bi-Encoder Achieves 98.4% Accuracy in Authorship Verification
September 24, 2026
A ModernBERT Bi-Encoder using contrastive learning reaches 98.4% accuracy on the PAN21 authorship verification task. The study identifies loss function, batch size, and random text span data augmentation as critical drivers of performance over traditional classification methods.
HOW THIS AFFECTS YOU
●
researcherYou can leverage contrastive learning and specific augmentation techniques to outperform classification baselines in forensic linguistics tasks.