gregorgabrovsek/authorship-attribution-on-short-texts-in-slovenian-language

Authorship Attribution on Short Texts in Slovenian Language

Link to the paper

This repository contains the code used in the project titled "Authorship Attribution on Short Texts in Slovenian Language". The project investigates the task of authorship attribution on short texts in Slovenian using two BERT language models.

The code can be used to fine-tune BERT models on short texts in Slovenian for authorship attribution. The datasets used in this project are publicly available on Huggingface Datasets and the trained Slovenian BERT models are available on Huggingface Models.

The project demonstrates the feasibility of using BERT models for authorship attribution in short texts and provides a solid starting point for systems tackling the growing problem of misinformation, hate speech, and the manipulation of public opinion.

Installation

To install the necessary packages, you have two options: either use the requirements.txt file or use Poetry.

If you use Poetry, you can install the necessary packages by running the following command:

poetry install

If you use the requirements.txt file, you can install the necessary packages by running the following command:

pip install -r requirements.txt

Usage

Generating the IMDb1m datasets

To generate the IMDb1m dataset to use with our trainer code, perform the following steps:

1. Download the IMDb1m dataset from here (make sure to download from the link with the title IMDb 1 million), extract the files and set the review_texts_file_path and posts_texts_file_path config variables in the config.yaml file
2. Enter your HuggingFace credentials into the config file.
3. Run the main.py file to generate the dataset.

Generating the RTV SLO datasets

1. Run main.py to retrieve the comment data from the RTV SLO website.
2. Run prepare_dataset.py to prepare the dataset for training.
3. Upload the dataset to HuggingFace Datasets manually.

Training the models

To train the models, run trainer.py. When running the file, you have to specify the following arguments:

• dataset_size: The number of the authors in the dataset (5, 10, 20, 50 or 100 for RTV SLO, and 5, 10, 25, 50 or 100 for IMDb1m)
• has_ooc: Whether the dataset has an out-of-corpus author (0 for false, 1 for true)
• use_multilingual_bert: Whether to use the multilingual BERT model (0 for false/SloBERTa, 1 for true/mBERT)
• use_imdb: Whether to use the IMDb1m dataset (0 for false/RTV SLO, 1 for true/IMDb1m)

Before starting the script, you have to edit the huggingface_username variable in line 11.

Citation

You can cite the paper using the following BibTeX snippet:

@article{gabrovvsek2023authorship,
  title={Authorship Attribution on Short Texts in the Slovenian Language},
  author={Gabrov{\v{s}}ek, Gregor and Peer, Peter and Emer{\v{s}}i{\v{c}}, {\v{Z}}iga and Batagelj, Borut},
  journal={Applied Sciences},
  volume={13},
  number={19},
  pages={10965},
  year={2023},
  publisher={MDPI}
}