Code from the paper "Cross-Domain Authorship Attribution Using Pre-trained Language Models".
This repository contains sub-modules, create a clone with --recurse-submodules option.
git clone --recurse-submodules https://github.com/GBarlas/mhc
Enter to mhc folder.
cd mhc
Install a conda environment for each pre-trained language model and MHC. All the environments require gpu support, for cpu support only, you may refer to: BERT, ELMO, GPT-2, ULMFiT and MHC.
conda env create -f "envs/mhc-bert.yml"
conda env create -f "envs/mhc-elmo.yml"
conda env create -f "envs/mhc-gpt-2.yml"
conda env create -f "envs/mhc-ulmfit.yml"
conda env create -f "envs/mhc.yml"
Download the weights of the pre-trained language models.
sh download_weights.sh
You may use CMCC, PAN18 or your own corpus. The corpus_to_csv.py script will help you to create a compatible csv file. To create the csv of CMCC or PAN18 just run the script with the required arguments. If you like to create a csv for another corpus you have to add a function into the corpus_to_csv.py, further information you may find into the script.
usage: corpus_to_csv.py [-h] {CMCC,PAN18} CORPUS-PATH CSV-FILENAME
positional arguments:
{CMCC,PAN18} Choose one.
CORPUS-PATH Where are the data?
CSV-FILENAME The filename of the output csv.
optional arguments:
-h, --help show this help message and exit
To create an AA problem, run the script create_problem.py with a selected corpus csv file and the UI will guide you step by step on how to define the training, evaluation and test set by selecting the approaprete instances from the corpus.
usage: create_problem.py [-h]
[--columns_and_desirables COLUMNS_AND_DESIRABLES]
corpus csv_filename
positional arguments:
corpus Choose a csv file created from "corpus_to_csv.py".
csv_filename The filename of the output csv.
optional arguments:
-h, --help show this help message and exit
--columns_and_desirables COLUMNS_AND_DESIRABLES
This option must have a specific format. Run the
script once without this option and you will get it.
Extract and save the representations that will be used as inputs to MHC. The representation from every layer of the selected language models will be saved along with the tokens, they have been created from.
!!! Don't forget to activate the coresponding conda environment conda activate (mhc-bert|mhc-elmo|mhc-gpt-2|mhc-ulmfit)
usage: extract_representations.py [-h] [--config_file CONFIG_FILE]
[-e ELMO_TOKENS_PATH]
{bert,elmo,ulmfit,gpt-2} dataset output_path
positional arguments:
{bert,elmo,ulmfit,gpt-2}
Choose one.
dataset Select a dataset or a problem.
output_path Where would you like to save the representations?
optional arguments:
-h, --help show this help message and exit
--config_file CONFIG_FILE
Choose a configuration file other than "config.ini"
-e ELMO_TOKENS_PATH, --elmo_tokens_path ELMO_TOKENS_PATH
If "elmo" is selected, you need to define the tokens
folder path created from "ulmfit" from the same csv
file.
mhc.py script will train and evaluate the MHC.
!!! Don't forget to activate the mhc conda environment conda activate mhc
usage: mhc.py [-h] -r REPRESENTATIONS_PATH -p PROBLEM -o OUTPUT_FILENAME
[--epochs EPOCHS] [--load-model LOAD_MODEL]
[--save-model-path SAVE_MODEL_PATH]
[--vocabulary-size VOCABULARY_SIZE] [--device {cuda,cpu}]
[--seed SEED]
optional arguments:
-h, --help show this help message and exit
-r REPRESENTATIONS_PATH, --representations-path REPRESENTATIONS_PATH
The folder with the representations created by
"create_representations.py"
-p PROBLEM, --problem PROBLEM
A "csv" file created by "create_problem.py"
-o OUTPUT_FILENAME, --output-filename OUTPUT_FILENAME
--epochs EPOCHS
--load-model LOAD_MODEL
Load model to continue training.
--save-model-path SAVE_MODEL_PATH
Path to save the model.
--vocabulary-size VOCABULARY_SIZE
--device {cuda,cpu}
--seed SEED
Please use the following bibtex entry:
@inproceedings{barlas2020cross,
title={Cross-Domain Authorship Attribution Using Pre-trained Language Models},
author={Barlas, Georgios and Stamatatos, Efstathios},
booktitle={IFIP International Conference on Artificial Intelligence Applications and Innovations},
pages={255--266},
year={2020},
organization={Springer}
}
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent