Contrastive Learning for Authorship Verification

Peter Kirby1 min readarXivSep 23, 2026

Key Takeaway for Students

Our results show that contrastive learning outperforms a classification-based approach to authorship verification under the tested settings. We identify loss function, batch size, training duration, pre-trained model, input context length, and random text span data augmentation as important factors of model performance. Based on these considerations, we develop a ModernBERT Bi-

98.4% accuracyoutperforms a classification-based approach to authorship verification

Reading in Plain English mode — toggle below to switch

This paper demonstrates a highly effective approach to authorship verification, a critical task in many NLP applications. Understanding contrastive learning and its optimization factors can significantly improve model performance for similar classification and verification problems.

Read original on arXiv

Career Insight

high demand

Why it matters

Skills to learn

Implement contrastive learning loss functions (e.g., triplet loss, InfoNCE)Optimize batch size and training duration for contrastive learning setupsApply random text span data augmentation techniquesFine-tune pre-trained BERT-like models for specific downstream tasksEvaluate model performance on authorship verification benchmarks like PAN

Interview questions

Explain the core difference between a classification-based approach and a contrastive learning approach for authorship verification. What are the advantages of contrastive learning in this context?
How would you select an appropriate batch size and input context length when fine-tuning a pre-trained language model for a contrastive learning task, considering both performance and computational constraints?
Design a system for real-time authorship verification for a large-scale content platform. Discuss how you would integrate a ModernBERT Bi-Encoder, handle data augmentation, and ensure low-latency inference.

Relevant for ~24 months · generated by the enrichment pipeline