Authorship Attribution (AA) seeks to determine the authorship of texts by examining distinctive writing styles. Although current AA methods have shown promising results, they often underperform in scenarios with significant topic shifts. This limitation arises from their inability to effectively separate topical content from the author’s stylistic elements. Furthermore, most studies have focused on individual-level AA, overlooking the potential of regional-level AA to uncover linguistic patterns influenced by cultural and geographical factors. To bridge these gaps, this paper introduces ContrastDistAA, a novel framework that leverages contrastive learning and mutual information maximization to disentangle content and stylistic features in latent representations for AA. Our extensive experiments demonstrate that ContrastDistAA surpasses existing state-of-the-art models in both individual and regional-level AA tasks. This breakthrough not only improves the accuracy of authorship attribution but also broadens its applicability to include regional linguistic analysis, making a substantial contribution to the field of computational linguistics.
