Score-based likelihood ratios for authorship verification with authorship embeddings
Journal of the Royal Statistical Society Series A: Statistics in Society · 04.11.2026
DOI: 10.1093/jrsssa/qnag115 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of the Royal Statistical Society Series A: Statistics in Society, online erschienen am 04.11.2026. Verfasst von Rachel Longjohn und 2 weiteren Autor:innen. Die Arbeit nennt 71 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Center for Statistics and Applications in Forensic Evidence, National Science Foundation, Hasso Plattner Institute (HPI) Research Center in Machine Learning and Data Science at the University of California, Irvine.
Abstract (Originalsprache)
We consider the problem of authorship verification (AV) in forensic science, with the goal of assessing whether two digital documents were written by the same author or two different authors, focusing primarily on short documents, such as social media messages, for which it may be challenging to extract traditional stylometric features. Recent work in deep learning has used data-driven neural network approaches for learning features that distinguish between authors, training a mapping from a document to a continuous-valued vector representation, which we refer to as an authorship embedding. In this paper, we investigate incorporating pre-trained authorship embeddings into a score-based likelihood ratio approach (SLR) for forensic AV for short documents. We investigate the feasibility of calculating SLRs based on these embeddings on several AV datasets, comparing them to a subselection of approaches that use manually-defined numeric features of the text. We find under our experimental settings that relative to the baselines we tested, the embedding-based SLRs are more accurate and have stronger discriminative performance. Based on these preliminary findings and the potential advantages of pre-trained methods in low data regimes, further exploration of the embedding-based approaches to forensic AV for short documents is warranted.
Förderung
- Center for Statistics and Applications in Forensic Evidence
- National Science Foundation
- Hasso Plattner Institute (HPI) Research Center in Machine Learning and Data Science at the University of California, Irvine
Zitieren
APA
Longjohn, R., Nelson, K., & Smyth, P. (2026). Score-based likelihood ratios for authorship verification with authorship embeddings. Journal of the Royal Statistical Society Series A: Statistics in Society. https://doi.org/10.1093/jrsssa/qnag115
Vancouver
Longjohn R, Nelson K, Smyth P. Score-based likelihood ratios for authorship verification with authorship embeddings. Journal of the Royal Statistical Society Series A: Statistics in Society. 2026. doi:10.1093/jrsssa/qnag115
BibTeX
@article{Longjohn2026,
author = {Longjohn, Rachel and Nelson, Kai and Smyth, Padhraic},
title = {Score-based likelihood ratios for authorship verification with authorship embeddings},
journal = {Journal of the Royal Statistical Society Series A: Statistics in Society},
year = {2026},
doi = {10.1093/jrsssa/qnag115}
}
EUVIMED