Contrastive Learning for Authorship Verification

TL;DR

This paper presents a contrastive learning approach for authorship verification, outperforming classification-based methods. The proposed ModernBERT Bi-Encoder model achieves 98.4% accuracy on the PAN21 task. It identifies key factors influencing model performance.

Why It Matters

Authorship verification has significant implications for plagiarism detection, copyright protection, and digital forensics. Accurate verification can prevent intellectual property theft and ensure authenticity. This research contributes to the development of reliable authorship verification systems.

Key Findings

• contrastive learning outperforms classification-based approach
• importance of loss function and batch size
• pre-trained model and input context length impact performance