Rethinking the Authorship Verification Experimental Setups
Methods & evidenceAuthorship & stylometry
The short version
A closer look at what authorship-verification benchmarks actually measure. The authors propose new PAN dataset splits to separate topic and author effects, evaluate neural baselines, and find that named entities can bias model decisions. Removing them improves generalization in their experiments.
Notes for the reading desk
- Evaluation splits influence which signals a model can exploit.
- Strong benchmark performance may partly reflect topic or named-entity cues.
- Testing on a separate corpus helps assess generalization beyond a benchmark.
Citation
Brad, F., Manolache, A., Burceanu, E., Barbalau, A., Ionescu, R. T., & Popescu, M. (2022). Rethinking the Authorship Verification Experimental Setups. Proceedings of EMNLP, 5634–5643. https://doi.org/10.18653/v1/2022.emnlp-main.380
Summary and reading notes are editorial guides. The linked paper is the original source.