Cracking double-blind review: Authorship attribution with deep learning.

Area of Science:

• Computational Linguistics
• Bibliometrics
• Scientific Publishing

Background:

• Double-blind peer review aims for unbiased evaluation but can be compromised by author recognition.
• Experienced researchers can often infer authorship, introducing bias into the peer-review process.

Purpose of the Study:

• To develop and evaluate a novel computational method for automated authorship attribution of scientific manuscripts.
• To assess the feasibility of using text content and bibliographic information for identifying authors of anonymous research papers.

Main Methods:

• A transformer-based neural network architecture was designed for authorship attribution.
• The largest authorship identification dataset to date was created using over 2 million arXiv papers.
• The model was trained and evaluated on subsets of arXiv data with up to 2,000 authors.

Main Results:

• The proposed method achieved up to 73% authorship attribution accuracy in specific arXiv subsets.
• A scaling analysis demonstrated the method's potential for larger datasets with increased computational resources.
• The study identified key textual and bibliographical features contributing to attribution accuracy.

Conclusions:

• Automated authorship attribution using neural networks is feasible and effective.
• This technology can potentially mitigate bias in scientific peer review by identifying potential author recognition.
• Open-sourced tools are provided to facilitate reproduction and further research in authorship attribution.