metadata

annotations_creators:
  - found
language:
  - en
language_creators:
  - found
license:
  - afl-3.0
multilinguality:
  - monolingual
pretty_name: AuthorMix
size_categories:
  - 10K<n<100K
source_datasets:
  - original
  - extended|blog_authorship_corpus
tags:
  - authorship obfuscation
  - authorship attribution
  - authorship classification
  - authorship style
  - style
  - author
task_categories:
  - text-classification
  - other
task_ids:
  - multi-class-classification
  - multi-label-classification

Dataset Card for AuthorMix

Table of Contents
Dataset Description
- Dataset Summary
- Languages
Dataset Structure
- Data Instances
- Data Splits
Dataset Creation
- Curation Rationale
- Source Data
Additional Information
- Citation Information
- Contributions

Dataset Description

AUTHORMIX, was originally created for authorship obfuscation task and had data from four distinct domains: presidential speeches, early-1900s fiction novels, scholarly articles, and diary-style blogs. Altogether, AUTHORMIX contains over 30k high-quality paragraphs from 14 authors.

This work was created in the paper: StyleRemix: Interpretable Authorship Obfuscation via Distillation and Perturbation of Style Elements, which should be cited if used in future work.

Homepage:
Repository:
Paper:
Point of Contact: [email protected], [email protected]

Dataset Summary

For the presidential domain, we curate and clean speeches from George W. Bush, Barack Obama, and Donald Trump. For novel domain, we choose a collection of early 1900s fiction writers with strong writing styles: Ernest Hemingway, F. Scott Fitzgerald, and Virginia Woolf. We select these presidents/authors due to their diverse styles but similar eras to minimize content discrepancies. Lastly, we alter two existing datasets to match the formality of our new domains: the Extended-Brennan Greenstad, a collection of "scholarly" short (500-word) paragraphs gathered from Amazon Mechanical Turk (AMT), and the Blog Authorship corpus, a collection of blogs (diary-style entries) that were posted to blog.com.

Languages

English

Dataset Structure

All four domains can be found in the AuthorMix.json file, which contains a dictionary with a key for each domain.

Data Instances

Each domain contain the paragraphs (up to 5 sentences each) for each author in the domain.

Data Splits

No preset data splits available.

Dataset Creation

For the presidential domain, we curate and clean a novel collection of high-quality presidential speeches from George W. Bush (n=38), Barack Obama (n=29), and Donald Trump (n=26), transcribed by the Miller Center (https://data.millercenter.org) at the University of Virginia. We broke the speeches naturally into paragraphs and then selected all paragraphs between 2-5 sentences. This resulted in a total of n= 13K paragraphs.

Similarly, we also decided to develop a new collection of early 1900s fiction writers from the with strong writing styles, therefore we choose text from books by Ernest Hemingway, F. Scott Fitzgerald, and Virginia Woolf which were collected from Project Gutenberg (https://www.gutenberg.org). We selected the top 4 most popular books on Project Gutenberg for each author and then again, used the natural paragraphs from each author. We selected all paragraphs between 2-5 sentences. This resulted in a total of n= 9K paragraphs.

Lastly, we altered the existing data from two current datasets, the Extended-Brennan Greenstad which is a collection of "scholarly" short (500-word) paragraphs gathered from Amazon Mechanical Turk (AMT) and the Blog Authorship corpus, a collection of blogs (diary-style entries) that were posted to blog.com. For the AMT dataset, we used authors "h", "pp", and "qq" and we artificially created paragraphs by chunking the text into a random collection of 2-5 sentences (as the text is not naturally broken into paragraphs). For the Blog dataset, we used authors "5546", "11518", "25872", "30102", "30407", we used the natural paragraphs. Then, to match the speech and novel domains, we edited to include all paragraphs between 2-5 sentences and at least 3 words. This resulted in n= 500 and n= 8K paragraphs for the AMT and Blog accordingly.

Curation Rationale

This dataset was created specifically for authorship obfuscation. We wanted to create a diverse high-quality dataset, which had domains of writings with similar topics but strong stylstic writing.