Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cross-Genre Authorship Attribution via LLM-Based Retrieve-and-Rerank

Shantanu Agarwal, Joel Barry, Steven Fincke, Scott Miller

TL;DR

This paper tackles cross-genre authorship attribution by framing it as a ranking task and applying a two-stage retrieve-and-rerank pipeline powered by fine-tuned large language models. The retriever is a bi-encoder, trained with contrastive loss on carefully curated, semantically diverse author pairs to emphasize stylistic signals over content. The reranker, a cross-encoder, is trained with diversified negative sampling and applied to the top retrieved candidates to boost end-to-end accuracy. Empirical results on HIATUS HRS1/HRS2 show large gains over previous SOTA, with the reranker delivering meaningful improvements when negatives are sampled from multiple similarity regimes, illustrating both the method’s potential and its computational trade-offs for scalable cross-genre attribution.

Abstract

Authorship attribution (AA) is the task of identifying the most likely author of a query document from a predefined set of candidate authors. We introduce a two-stage retrieve-and-rerank framework that finetunes LLMs for cross-genre AA. Unlike the field of information retrieval (IR), where retrieve-and-rerank is a de facto strategy, cross-genre AA systems must avoid relying on topical cues and instead learn to identify author-specific linguistic patterns that are independent of the text's subject matter (genre/domain/topic). Consequently, for the reranker, we demonstrate that training strategies commonly used in IR are fundamentally misaligned with cross-genre AA, leading to suboptimal behavior. To address this, we introduce a targeted data curation strategy that enables the reranker to effectively learn author-discriminative signals. Using our LLM-based retrieve-and-rerank pipeline, we achieve substantial gains of 22.3 and 34.4 absolute Success@8 points over the previous state-of-the-art on HIATUS's challenging HRS1 and HRS2 cross-genre AA benchmarks.

Cross-Genre Authorship Attribution via LLM-Based Retrieve-and-Rerank

TL;DR

This paper tackles cross-genre authorship attribution by framing it as a ranking task and applying a two-stage retrieve-and-rerank pipeline powered by fine-tuned large language models. The retriever is a bi-encoder, trained with contrastive loss on carefully curated, semantically diverse author pairs to emphasize stylistic signals over content. The reranker, a cross-encoder, is trained with diversified negative sampling and applied to the top retrieved candidates to boost end-to-end accuracy. Empirical results on HIATUS HRS1/HRS2 show large gains over previous SOTA, with the reranker delivering meaningful improvements when negatives are sampled from multiple similarity regimes, illustrating both the method’s potential and its computational trade-offs for scalable cross-genre attribution.

Abstract

Authorship attribution (AA) is the task of identifying the most likely author of a query document from a predefined set of candidate authors. We introduce a two-stage retrieve-and-rerank framework that finetunes LLMs for cross-genre AA. Unlike the field of information retrieval (IR), where retrieve-and-rerank is a de facto strategy, cross-genre AA systems must avoid relying on topical cues and instead learn to identify author-specific linguistic patterns that are independent of the text's subject matter (genre/domain/topic). Consequently, for the reranker, we demonstrate that training strategies commonly used in IR are fundamentally misaligned with cross-genre AA, leading to suboptimal behavior. To address this, we introduce a targeted data curation strategy that enables the reranker to effectively learn author-discriminative signals. Using our LLM-based retrieve-and-rerank pipeline, we achieve substantial gains of 22.3 and 34.4 absolute Success@8 points over the previous state-of-the-art on HIATUS's challenging HRS1 and HRS2 cross-genre AA benchmarks.
Paper Structure (17 sections, 4 equations, 1 figure, 9 tables, 1 algorithm)

This paper contains 17 sections, 4 equations, 1 figure, 9 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methodology
  3. Task definition
  4. Retriever
  5. Reranker
  6. Experiments
  7. Curating training data for cross-genre AA
  8. Test Dataset
  9. Metrics
  10. Baseline Models
  11. Implementation Details
  12. Results
  13. Related Work
  14. Conclusion
  15. Training data
  16. ...and 2 more sections

Figures (1)

  • Figure 1: Negative documents, i.e. documents not written by the author of the query document, can be categorized into three distinct groups: $n(d_q)$ are documents close to the query document, $n(d_q^+)$ are close to the positive document and $n(r)$ are random documents with no particular bearing to either $d_q$ or $d_q^+$.