Dataset and experiments for the COLING2022 Argument Mining Workshop Paper "Boundary Detection and Categorization of Argument Aspects via Supervised Learning"
This repository contains code and data for the paper Argument Mining Workshop Paper Boundary Detection and Categorization of Argument Aspects via Supervised Learning presented during the 9th Workshop on Argument Mining at COLING2022.
In the datasets folder, you can find three files with connl formatted data for each topic that we have described in the paper. In the experiments folder are test confgurations for reproducing the papers results. You can inspect the config files to see hyperparameters and the models that we have used. This repository may also be a starting point to experiment with argument aspect with other models or hyperparameters You can replace models with other available models from huggingface.
The Corpus is also available on Zenodo. When using the corpus, please cite the zenodo repository:
Ruckdeschel, Mattes, & Wiedemann, Gregor. (2023). Argument Aspect Corpus (1.0) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.7525183
you need pipenv for dependency
installation. Then, run pipenv sync
in order to create the environment.
Note: We needed a non-default pytorch version for compatability issues with our GPU, so there's a
custom source in the pipfile for pytorch. Familiarize yourself with your GPU
and how to install torch and cuda for it. The code needs a cuda-compatible GPU
in order to run.
Start a shell in your virtual environment using pipenv shell
before you use
any commands.
In the experiments folder, you find sub-folder for all experiments you need to run in order to reproduce our results, with a prefix for the coresponding dataset. run the experiments using
python run_experiment.py run experiments/<foldername>
Experiments can take a long time, depending on your GPU.
Note: There's an issue with FLAIR regarding data pair evaluation, which we
fixed locally. You need to replace
<your_home_directory>/.local/share/virtualenvs/argument-aspect-corpus-v1-XXXXXXXX/lib/python3.9/site-packages/flair/data.py
with the file from this repository if you get an error saying: AttributeError:
'DataPair' object has no attribute 'to_original_text'
You can get evaluations of the models using
python run_experiment.py results experiments/<foldername>
For Sentence prediction and nervaluate use
python sentence_predictions.py report experiments/<foldername>
and
python sentence_predictions.py nervaluate experiments/<foldername>
Please cite our paper when using this repository.
Mattes Ruckdeschel and Gregor Wiedemann. 2022. Boundary Detection and Categorization of Argument Aspects via Supervised Learning. In Proceedings of the 9th Workshop on Argument Mining, pages 126–136, Online and in Gyeongju, Republic of Korea. International Conference on Computational Linguistics.