Abstract: In Argument Mining (AM), the integral sub-task of argument component classification refers to the classification of argument components as claims or premises. In this context, the content of the component alone does not actually suffice to predict it’s corresponding class accurately. In fact, additional lexical, contextual, and structural features are needed. Here, we propose a unified model for argument component classification based on BERT. Our model incorporates the component itself together with contextual and structural features – given as text – instead of the usual numerical form. This new technique enables BERT to build a customized and enriched representation of the component. We evaluate our model on three datasets that reflect a diversity of written and spoken discourses. We achieve state-of-art results on two datasets and 95% of the best results on the third. Our approach shows that BERT is capable of exploiting non-textual information given in a textual form. To the best of our knowledge, our work is the first to investigate and implement a features as text, BERT-based model for text classification.
Our model incorporates contextual and structural features of the argument component to build enriched BERT based representation for argument classification. More details about BERT and other transformer models can be found here: https://github.com/huggingface/transformers
We experiment on our model with 3 datasets:
- Persuasive Essays: https://tudatalib.ulb.tu-darmstadt.de/handle/tudatalib/2422
- Yes We Can: https://github.com/ElecDeb60To16/Dataset
- Change My View: https://github.com/chridey/change-my-view-modes
The contextual and structural features of all three datasets have been pre-computed separately. All three datasets in the PyTorch ready .pt format are given in the /datasets folder.
We train our model for two tasks:
- BERT fine-tune on the three datasets
- comparing numerical concatenation versus our structural features as text
More details about BERT fine-tuning can be found here: https://huggingface.co/docs/transformers/training
The notebooks used for the experiments are given in the /notebooks folder. More detailed explanation about the usage is provided in the README file in the folder.
We use the following versions of the packages:
torch==1.10
transformers==4.12.5
datasets==1.15.1
tensorboard==2.6.0
All experiments have been performed on Paperspace's Gradient cloud AI/ML platform: https://gradient.run/