README.md

Argument Clustering

Accompanying repository of our ECIR2021 Paper "Diversity Aware Relevance Learning for Argument Search". Arxiv

We used the dataset of the ECIR2020 paper "A Framework for Argument Retrieval", which is available by request from the authors.

Install requirements:

pip install -U pip
pip install -U -r requirements.txt

Preprocessing:

You have to request the dataset from the authors and adjust the paths in settings.py

Step 1: Extract all claims, premises and the matching

PYTHONPATH=src:$PYTHONPATH python3 executables/preprocessing/read_json.py --input_dir=... --output_dir=output/

Step 2: Precompute bert-features for the claims, premises and claim-premise pairs (choices=['pair', 'claims', 'premises'])

PYTHONPATH=src:$PYTHONPATH python3 executables/preprocessing/generate_features.py --mode=...

Step 3: Generate negative claim-premise pairs either randomly or based on similarity

# Either
PYTHONPATH=src:$PYTHONPATH python3 executables/preprocessing/generate_negative_samples_nn.py
# or
PYTHONPATH=src:$PYTHONPATH python3 executables/preprocessing/generate_negative_samples.py

Step 4: Generate train-/test-/validation split

PYTHONPATH=src:$PYTHONPATH python3 executables/preprocessing/generate_sets.py

Execution:

PYTHONPATH=src:$PYTHONPATH python3 executables/evaluation/evaluate_baselines.py --force > output/output_energy.txt