Contrastive learning boosts authorship verification accuracy to 98 percent
Contrastive Learning for Authorship Verification
Computation and LanguageMachine Learning
Summary
Determining if two texts come from the same author can be tricky. The authors show that a method called contrastive learning works better than traditional classification approaches for this task. They found that several factors like the training setup and data preparation influence performance. Their final model, based on a version of BERT, reached very high accuracy on a standard authorship test.
What this means in practice
- •For forensic analysts: Quickly determine if two sets of writing were produced by the same author in investigations using the new contrastive learning model.
- •For publishing editors: Verify authorship consistency across submissions to detect potential plagiarism or ghostwriting with improved accuracy.
Authors
Peter Kirby
Abstract
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-Encoder model that achieves 98.4% accuracy on the PAN21 authorship verification task.