This project is dedicated to the problem of Argument Mining and is aimed at reproducing the paper published by Haddadan et al. in 2019. The dataset is comprised of 39 speeches from the presidential debates in the US from the year 1960 to 2016. The corpus sentences are labeled as claims, premises or none of above.
Like the authors, we solve two binary classification tasks: 1) classification of argumentative vs. non-argumentative sentences; and 2) classification of claims vs. premises.
The methodology of the project is entirely reflected by the original paper. We split the practical part into four experimental settings. The splits are based on different methodological approaches.
Setting 1: tf-idf word matrices for unigrams + SVM linear model
Setting 2: tf-idf word matrices for uni-, bi- and trigrams and a set of engineered features + SVM with rbf kernel
Setting 3: fasttext embeddings + LSTM
Setting 4: tf-idf word matrices for uni-, bi- and trigrams and a set of engineered features + Feed Forward Network
The project is carried out as a part of PM "Mining Opinions and Arguments", taught by Prof. Dr. Manfred Stede at the University of Potsdam, Winter Term 2021/22. Our final report can be found here.
Team: Chen Peng, Mariia Poiaganova, Milena Voskanyan
| Folder | Item | Description |
|---|---|---|
| 📁 code | 📁 experiments | Notebooks for 4 experimental settings × 2 tasks |
| linguistic_features.py | Documented functions for feature engineering | |
| preprocessing.py | Code snippet for sentences preprocessing | |
| requirements.txt | Necessary Python frameworks and their versions | |
| 📁 data | 📁 raw | Debates' scripts along with respective annotation files |
| sentence_db_candidate.csv | Main dataset used | |
| 📑 paper | Final report | |
| 📁 presentations | dec-presentation.pdf | Slides for the first presentation |
| feb-presentation.pdf | Slides for the final presentation |
🔹 To reproduce the project, start by cloning the repository:
git clone https://github.com/pc852/Argument-Mining-Presidential-Debates.git
🔹 Next, we recommend running the experiments from Jupyter Notebook environment. If you are new to it, nice guideline can be found here.
🔹 You will need Python version 3.7 or higher, and the following packages:
pandas, scikit-learn, nltk, keras, tensorflow, gensim, numpy, spacy, vadersentiment. Our versions for them can be found here.
🔹 Each notebook is tailored to be run independently, however, you might need some additional preparations for certain experimental settings.
For Settings 2 and 4, with engineered features
In order to compile Part-of-Speech and NER engineering functions, you should download the model en_core_web_sm to your system.
For Setting 3, with LSTM model
You should download the pre-trained fasttext embeddings file wiki-news-300d-1M.vec and move it to project folder. FastText
⚪ Dataset is kept here, but for all the experiments you will need only this file.