Topic and style: Authorship attribution with multi-vector meta-features and stacked classifier
2025, vol.17 , no.4, pp. 33-44
Article [2025-04-04]
Authorial rights are meaningful only when a document is promptly attributed to its legitimate creator. While an author’s unique topical preferences may pinpoint them within a group of thematically-diverse writers, attribution becomes considerably challenging in cases such as judicial opinions – e.g., attributing a criminal case judgement to one among hundreds of judges heading criminal courts – due to overlapping contextual jargon. In such situations, integrating both topical and stylistic cues can enhance the reliability and generalisability of authorship attribution. This work proposes a custom stacking ensemble model trained on diverse aspects of text, later giving rise to novel features: Multi-Vector Meta-Features. The proposed method demonstrates statistically significant boosted accuracy scores across three out of four datasets, outperforming baseline model and recent stacking ensemble models.
authorship attribution, bertopic, stacked classifier, multi-vector, natural language processing
https://doi.org/10.59035/BTRZ6248
B. Lavanya, R. Sowmiya. Topic and style: Authorship attribution with multi-vector meta-features and stacked classifier. International Journal on Information Technologies and Security, vol.17 , no.4, 2025, pp. 33-44. https://doi.org/10.59035/BTRZ6248