Contrastive Disentanglement for Authorship Attribution

Anonymous

Abstract: Authorship Attribution (AA) aims to identify the authorship of texts by analyzing distinctive writing styles. While current AA methods have yielded promising performance, these approaches commonly exhibit suboptimal performance in contexts where the subject matter varies significantly (i.e., topic-shift scenarios). This limitation stems from their inadequacy in differentiating between the topical content and the author's stylistic signature. Additionally, existing studies predominantly focus on AA at an individual level, thereby neglecting the exploration of regional-level AA, which could reveal common linguistic patterns influenced by cultural and geographical factors. Addressing these gaps, this paper introduces ContratDistAA, a novel framework employing contrastive learning coupled with mutual information maximization to segregate content from stylistic features in latent representations for AA tasks. Our comprehensive experimental evaluations reveal that ContratDistAA outperforms existing state-of-the-art models in both individual and regional-level AA scenarios. This advancement not only enhances the accuracy of authorship attribution but also expands its applicability to encompass regional linguistic analysis, thus contributing significantly to the broader field of computational linguistics.

Paper Type: long

Research Area: Computational Social Science and Cultural Analytics

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data resources, Data analysis

Languages Studied: English