Data and code for

Kristof, V., Grossglauser, M., Thiran, P., War of Words: The Competitive Dynamics of Legislative Processes, The Web Conference, April 20-24, 2020, Taipei, Taiwan

Note: The repo for War of Words II: Enriched Models of Law-Making Processes is here.

From the root of the repo, install the requirements and local library:

pip install -r requirements.txt
pip install -e lib

Download the (raw) data from https://zenodo.org/record/3757714.

Before training the models, you first need to process the data and format them for input to the model. Run the following commands to filter out dossiers and MEPs with less than 10 edits, split the data into 70/30 for training/testing, and fix the random seed:

# Process dataset for the WoW model.
python mep_only.py \
  path/to/war-of-words-ep7.txt \
  output/to/meponly-ep7.pkl \
  --threshold 10 \
  --split 0.7 \
  --seed 0

# Process dataset for the WoW(R) model.
python rapporteur_advantage.py \
  path/to/war-of-words-ep7.txt \
  output/to/rapadv-ep7.pkl \
  --threshold 10 \
  --split 0.7 \
  --seed 0

Do the same for EP8 by replacing ep7 by ep8 in the above commands.

This should have generated eight pickle files:

• output/to/meponly-ep7-train.pkl
• output/to/meponly-ep7-test.pkl
• output/to/rapadv-ep7-train.pkl
• output/to/rapadv-ep7-test.pkl
• output/to/meponly-ep8-train.pkl
• output/to/meponly-ep8-test.pkl
• output/to/rapadv-ep8-train.pkl
• output/to/rapadv-ep8-test.pkl

You can now use these files to train (and evaluate) the models (see below).

We analyze the parameters after fitting the model on a whole legislature period, i.e., using the whole dataset for one legislature. This means that you need to generate another processed dataset, without splitting. We will do it only for the rapporteur-advantage model as this is our best model, hence providing the best estimation of the parameters. Run the following command:

python rapporteur_advantage.py \
  path/to/war-of-words-ep7.txt \
  output/to/rapadv-ep7.pkl \
  --threshold 10 \
  --seed 0

And do the same for EP8 by replacing ep7 by ep8 above.

This should have generated two additional pickle files:

• output/to/rapadv-ep7.pkl
• output/to/rapadv-ep8.pkl

You can now use these files to analyze the parameters.

To train the models (WoW and WoW(R)), run the models/train.py script. It takes three arguments as parameters:

• --data_path: Path to a processed data set (see above)
• --model_path: Path where the trained model will be saved
• --regularizer: L2-regularizer for the model log-likelihood

Sorry, I forgot to report the value of the best regularizers (found by 10-fold cross-validation) in the paper. Here they are:

You can then run the following command to train a model:

python train.py \
    --data_path ../data/processed/meponly-ep7-train.pkl \
    --model_path meponly-ep7.predict \
    --regularizer 0.32

I provide a bash script (models/train-all-models.sh) to train all the models in one go (make sure the path of the data files match).

As explained above, you nee to train the models on the whole dataset to run the parameter analysis. Run this command to train the model on the whole EP7 legislature. (As a notation, I use the extensions .predict and .fit to differentiate between the models trained for evaluation (when you need to make predictions) and the models trained for parameter analysis (when you care only about fitting the model on all the data that you have).)

python train.py \
    --data_path ../data/processed/meponly-ep7.pkl \
    --model_path meponly-ep7.fit \
    --regularizer 0.32

Run the evaluation.ipynb notebook to reproduce the results of Table 2.

Run the parameter-analysis.ipynb notebook to reproduce the results of Table 3.

Run the edit-graph.ipynb notebook to reproduce the results of Figure 3.

This project requires Python 3.6.

To cite this work, use:

@inproceedings{kristof2020war,
  author = {Kristof, Victor and Grossglauser, Matthias and Thiran, Patrick},
  title = {War of Words: The Competitive Dynamics of Legislative Processes},
  year = {2020},
  booktitle = {Proceedings of The Web Conference 2020},
  pages = {2803–2809},
  numpages = {7},
  location = {Taipei, Taiwan},
  series = {WWW '20}
}