Generalizing Authorship Attribution Across Genres

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 222 papers26 conceptsfrozen 2026-10-04

Authorship attribution—the task of identifying the writer of an anonymous text—is a cornerstone of forensic linguistics and digital security. However, traditional models suffer from a fundamental fragility: they are often "topic-bound." If a model learns to identify an author based on their political blog posts, it frequently fails when asked to identify that same author writing a movie review.

This failure occurs because the model confuses what is being said (the topic) with how it is being said (the style). Research in this field is shifting toward identifying "invariant" stylistic markers that persist even when the subject matter, vocabulary, and genre change.

The Topic Trap and Domain Shift

The core challenge in modern authorship attribution is domain shift. As Damon L. Woodard and Kalaivani Sundararajan argue in What represents "style" in authorship attribution?, models often overfit to domain-specific vocabulary. For example, a model might associate the word "policy" with a specific author simply because that author wrote the only political texts in the training set.

When the domain changes—moving from formal news to informal social media—these lexical cues vanish. Fredrik Johansson highlights that real-world forensic evidence is rarely clean; it spans diverse platforms like private messages and public articles, where the "authorial voice" adapts to fit the medium. To solve this, researchers are looking for features that are "blind" to topic but "sensitive" to habit.

Extracting the Multi-Layered Fingerprint

If style is not just in the words we choose, where is it? Recent work suggests that authorial identity is encoded across multiple levels of linguistic abstraction, from punctuation habits to complex syntactic structures.

Martijn Bartelds and Wietse de Vries demonstrated in Improving Cross-domain Authorship Attribution by Combining Lexical and Syntactic Features that combining character n-grams with deep syntactic information, such as dependency relation chains, allows models to generalize much better across different fictional universes in fanfiction.

Taking this further, the LIGHT framework, developed by Milad Alshomary and Kathleen McKeown, posits that style is a multi-level phenomenon hidden within the layers of a Transformer. While standard models only look at the final output layer (which is heavily influenced by semantics and topic), Layered Insights: Generalizable Analysis of Authorial Style by Leveraging All Transformer Layers shows that middle layers often capture the structural and stylistic nuances that remain stable across domains. By aggregating representations from every layer, they achieved performance gains of up to 50% in out-of-domain scenarios.

Measuring Style as Divergence

Another intuitive approach to cross-domain attribution is to stop looking for who an author is and start looking at how they deviate from the norm.

In Improving Authorship Verification using Linguistic Divergence, Yifan Zhang and colleagues introduced DV-Distance. This method uses a pre-trained language model to establish a "Normal Writing Style" (NWS). The model then calculates a "deviation vector" for a specific text. The intuition is that while an author's vocabulary changes across genres, the specific ways they "break" or "bend" standard linguistic patterns remain consistent. This unsupervised approach effectively bypasses the need for domain-specific training data.

The LLM Challenge and Uncertainty

The rise of Large Language Models (LLMs) introduces two new complications: LLMs as tools for attribution and LLMs as "erasers" of style.

  1. LLMs as Detectors: Baixiang Huang and Kai Shu show in Can Large Language Models Identify Authorship? that GPT-4 can outperform specialized models when using Linguistically Informed Prompting (LIP). By explicitly telling the model to look for phrasal verbs and punctuation, they leverage the LLM's vast internal knowledge of linguistic variation.
  2. The Erasure Problem: Conversely, Nafis Irtiza Tripto and Adaku Uchendu explore the "Ship of Theseus" paradox in A Ship of Theseus: Curious Cases of Paraphrasing in LLM-Generated Texts. They found that even a single round of AI paraphrasing can overwrite a human's stylistic signature with the LLM's own "average" style, making attribution nearly impossible.

To handle these increasingly complex and "noisy" environments, researchers like Iqra Zahid are integrating Bayesian uncertainty into attribution models. Their framework, BEDAA, described in BEDAA: Bayesian Enhanced DeBERTa for Uncertainty-Aware Authorship Attribution, allows a model to signal when it is "unsure" due to a domain shift, providing a layer of transparency essential for forensic applications.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 17 papers
  2. Version 1 · 2026-10-04 · 22 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_cross_domain_authorship_attribution_6faf4f9e,
  title = {cross-domain authorship attribution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-domain-authorship-attribution/rdr_6faf4f9e862941128a71fd3d66bc7e5f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_generalizing_authorship_attribution_acro_6faf4f9e_v1,
  title = {Generalizing Authorship Attribution Across Genres},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generalizing-authorship-attribution-across-genres/rdr_6faf4f9e862941128a71fd3d66bc7e5f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}