Enron AV Corpus

Summary:------------------- The "Enron AV Corpus" represents an authorship verification (AV) corpus derived from the well-known "Enron Email Dataset." It was transformed into the same standardized format used by the PAN Authorship Verification corpora from 2013–2015 (http://pan.webis.de). With only minor modifications, it is also compatible with the PAV AV corpora released between 2020 and 2022. This corpus aims to support the research community in Digital Text Forensics by providing a shared resource for benchmarking and comparing AV methods.

Structure: ------------------- The corpus consists of 160 AV cases in total, divided into training and test splits containing 64 and 96 AV cases, respectively. Both splits are strictly balanced with respect to same-authorship and different-authorship cases. Each AV case comprises up to five documents (plain-text files). Four of these documents represent writing samples from the known (true) author, while the remaining document corresponds to the text of the unknown author whose authorship is to be verified. Each document is constructed from multiple short emails written by the same author in order to provide sufficient textual material to capture the author's writing style. The length of each text ranges from approximately 1 to 4 kilobytes.

Preprocessing: ------------------- All texts in the corpus underwent the same preprocessing procedure: de-duplication, removal of URLs, newlines/tabs, normalization of UTF-8 symbols and replacement of multiple consecutive whitespace characters with a single blank space. In addition, all email headers and other metadata (including signatures) were removed from each document so that only pure natural-language text written by a single author remained. Furthermore, all texts were topic-masked such that only stylistically relevant text units remained (e.g., function words, punctuation marks, POS tags, etc.). The topic-masking procedure applied in this work was POSNoise. Finally, the filenames of the underlying texts were anonymized to ensure GDPR compliance.

Paper: ------------------- A more detailed description of the "Enron AV Corpus" is provided in the supplementary materials (see Section 1.1.) accompanying the paper: Grammar as a Behavioral Biometric: Using Cognitively Motivated Grammar Models for Authorship Verification

Citing the Corpus:------------------- If you use this corpus in your research, please cite the following paper:

Andrea Nini, Oren Halvani, Lukas Graner, Sophie Titze, Valerio Gherardi and Shunichi Ishihara. Grammar as a Behavioral Biometric: Using Cognitively Motivated Grammar Models for Authorship Verification. Humanities and Social Sciences Communications (Nature) 13, 455 (2026).

Bibtex:

@Article{NiniLambdaG:2026,    author = {Nini, Andrea and Halvani, Oren and Graner, Lukas and Titze, Sophie and Gherardi, Valerio and Ishihara, Shunichi},    title = {{Grammar as a Behavioral Biometric: Using Cognitively Motivated Grammar Models for Authorship Verification}},    journal = {Humanities and Social Sciences Communications},    year = {2026},    month = {Mar},    day = {03},    volume = {13},    number = {1},    pages = {455},    abstract = {Authorship Verification (AV) is a key area of research in digital text forensics, which addresses the fundamental question of whether two texts were written by the same person. Numerous computational approaches have been proposed over the last two decades in an attempt to address this challenge. However, existing AV methods often suffer from high complexity, low explainability, and especially from a lack of clear scientific justification. We propose a simpler method based on modeling the grammar of an author following Cognitive Linguistics principles. These models are used to calculate $\lambda$G (LambdaG): the ratio of the likelihoods of a document given the candidate's grammar versus given a reference population's grammar. Our empirical evaluation, conducted on 12 datasets and compared against seven baseline methods, demonstrates that LambdaG achieves superior performance, including against several neural network-based AV methods. LambdaG is also robust to small variations in the composition of the reference population and provides interpretable visualizations, enhancing its explainability. We argue that its effectiveness is due to the method's compatibility with Cognitive Linguistics theories, predicting that a person's grammar is a behavioral biometric.},    issn = {2662-9992},    doi = {10.1057/s41599-025-06340-3},    url = {https://doi.org/10.1057/s41599-025-06340-3}}