A generalized solution to verify authorship and detect style change in multi-authored documents

Abstract

Identifying changes in style can be used to detect multi-authored social media accounts, plagiarism, compromised accounts, and author contributions in long documents. We propose an approach to recognize changes in authorship using large language models. Our approach leverages sentence-level contextual embeddings and semantic relationships. First we expand the training set by adding adversarial examples to the minority class [5], [13], [17]. Then we fine-tune a sequence classification transformer model to detect style change. Our approach outperforms all baselines of PAN21 with macro F1-scores of 0.80, 0.74, and 0.70 for detecting style changepoint between paragraphs, closed-set author ID per paragraph, and style changepoint between sentences, respectively. Our approach also performs better than the leading competitors in PAN22. Also, we achieved a five percent improvement in macro F1-score (0.78) on the newly introduced DarkReddit+ dataset for authorship verification.