ArgStance

ArgStance is a research project focused on Argumentative Stance Classification using emotion lexicons, with a particular emphasis on the extended NRC Emotion Lexicon. The project lies at the intersection of Argumentation Mining (AM) and Emotion Analysis, with a specific focus on the stance classification subtask within AM.

Extensive preprocessing was conducted to prepare the corpora. The results can be found in the data folder.

Table of Contents

• Environment Setup
• Data
• Usage
  • Training Stance Classifier
  • Qwen-based Classification
• Experiments
• Results

Environment Setup

Prerequisites

• Conda or Anaconda
• Python 3.11+
• CUDA-compatible GPU (optional, but recommended for faster training)

Installation Steps

1. Clone the repository

   git clone <repository-url>
   cd ArgStanceNRC

2. Create conda environment from environment.yml

   conda env create -f environment.yml

3. Activate the environment

   conda activate enrc

Data

The preprocessed datasets are located in the data/ directory:

• dfIBM_stance-v1.json
• dfMT1_stance-v1.json
• dfMT2_stance-v1.json
• dfMT3_stance.json
• dfPE_stance-v2.json
• dfUKB_stance-v1.json

Usage

Training Stance Classifier

The main training script is src/train_stance_classifier.py. It supports various configurations:

Basic Training

# Train on UKB dataset with default settings
cd src
python train_stance_classifier.py --dataset UKB

# Train on PE dataset
python train_stance_classifier.py --dataset PE

Training with NRC Emotion Features

# Train with standard NRC features
python train_stance_classifier.py --dataset PE --use-nrc

# Train with expandNRC features
python train_stance_classifier.py --dataset PE --use-enrc --threshold 0.8

Training without NRC Features

# Train without emotion features (baseline BERT)
python train_stance_classifier.py --dataset PE --no-nrc

Advanced Training Options

# Full configuration example
python train_stance_classifier.py \
    --dataset PE \
    --use-enrc \
    --threshold 0.8 \
    --batch-size 16 \
    --epochs 50 \
    --lr 2e-5 \
    --test-size 0.1 \
    --val-size 0.1 \
    --use-focal-loss \
    --use-data-augmentation \
    --include-topic-in-text

# Freeze BERT weights (train only classifier head)
python train_stance_classifier.py --dataset PE --freeze-bert

# Disable wandb logging
python train_stance_classifier.py --dataset PE --no-wandb

Command-line Arguments

Argument	Type	Default	Description
--dataset	str	UKB	Dataset to use: IBM, IBM_v2, MT1, MT2, MT3, PE, UKB
--use-nrc	flag	False	Use standard NRC emotion features
--use-enrc	flag	False	Use expandNRC emotion features (enhanced)
--no-nrc	flag	False	Disable NRC features (baseline)
--threshold	float	0.8	Threshold for expandNRC features
--batch-size	int	8	Batch size for training
--epochs	int	50	Number of training epochs
--lr	float	1e-5	Learning rate
--test-size	float	0.1	Test set size (fraction)
--val-size	float	0.1	Validation set size (fraction)
--device	str	auto	Device: cpu, cuda, or mps
--freeze-bert	flag	False	Freeze BERT weights
--use-focal-loss	flag	False	Use focal loss for class imbalance
--use-data-augmentation	flag	False	Enable data augmentation
--include-topic-in-text	flag	False	Include topic in input text
--no-wandb	flag	False	Disable wandb logging

Output

Training results are saved in results/<experiment_name>/:

• model_<timestamp>.pt - Trained model checkpoint
• model_<timestamp>_predictions.csv - Test set predictions
• training_history.json - Training metrics history
• training_plots.png - Loss and F1 score plots
• experiment_config.json - Experiment configuration

Qwen-based Classification

For LLM-based stance classification using Qwen:

cd src

# Evaluate on all datasets
python qwen_stance_classification.py --dataset all

# Evaluate on specific dataset
python qwen_stance_classification.py --dataset PE

# With NRC features
python qwen_stance_classification.py --dataset PE --use-nrc

# With expandNRC features
python qwen_stance_classification.py --dataset PE --use-enrc --threshold 0.8

# Custom results directory
python qwen_stance_classification.py --dataset all --results-dir ./my_qwen_results

Note: Qwen classification requires significant GPU memory (recommended: 16GB+ VRAM for 7B model).

Results are saved in qwen_results/:

• qwen_<dataset>_predictions_<timestamp>.csv - Predictions
• qwen_<dataset>_metrics_<timestamp>.json - Metrics
• qwen_<dataset>_report_<timestamp>.txt - Classification report
• qwen_summary_<timestamp>.csv - Summary across all datasets

License

[Specify your license here]

Contact

[Your contact information]