causal-strength 0.1.0

A package for evaluating causal strength intensity between cause and effect.

EN | 简体中文

causal-strength : Measure the Strength Between Cause and Effect

causal-strength is a Python package for evaluating the causal strength between statements using various metrics such as CESAR (Causal Embedding aSsociation with Attention Rating). This package leverages pre-trained models available on Hugging Face Transformers for efficient and scalable computations.

Table of Contents

• 📜 Citation *
• 🌟 Features *
• 🚀 Installation *
  • Prerequisites
  • Steps
• 🛠️ Usage *
  • Quick Start
  • Evaluating Causal Strength
  • Generating Causal Heatmaps
• 📚 References *

📜 Citation

If you find this package helpful, please star this repository causal-strength and the related repository: defeasibility-in-causality. For academic purposes, please cite our paper:

@inproceedings{cui-etal-2024-exploring,
    title = "Exploring Defeasibility in Causal Reasoning",
    author = "Cui, Shaobo  and
      Milikic, Lazar  and
      Feng, Yiyang  and
      Ismayilzada, Mete  and
      Paul, Debjit  and
      Bosselut, Antoine  and
      Faltings, Boi",
    booktitle = "Findings of the Association for Computational Linguistics ACL 2024",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand and virtual meeting",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-acl.384",
    doi = "10.18653/v1/2024.findings-acl.384",
    pages = "6433--6452",
}

🌟 Features

• Causal Strength Evaluation: Compute the causal strength between two statements using models like CESAR.
• Visualization Tools: Generate heatmaps to visualize attention and similarity scores between tokens.
• Extensibility: Easily add new metrics and models for evaluation.
• Hugging Face Integration: Load models directly from the Hugging Face Model Hub.

🚀 Installation

Prerequisites

• Python 3.7 or higher
• PyTorch (for GPU support, ensure CUDA is properly configured)

Steps

1. Install it directly from PyPI

   pip install causal-strength
   

2. Install it from source code

   git clone https://github.com/cui-shaobo/causal-strength.git
   cd causal-strength
   pip install .
   

🛠️ Usage

Quick Start

Here's a quick example to evaluate the causal strength between two statements:

from causalstrength import evaluate

# Test CESAR Model
s1_cesar = "Tom is very hungry now."
s2_cesar = "He goes to McDonald for some food."

print("Testing CESAR model:")
cesar_score = evaluate(s1_cesar, s2_cesar, model_name='CESAR', model_path='shaobocui/cesar-bert-large')
print(f"CESAR Causal strength between \"{s1_cesar}\" and \"{s2_cesar}\": {cesar_score:.4f}")

This will output the following without errors:

Testing CESAR model:
CESAR Causal strength between "Tom is very hungry now." and "He goes to McDonald for some food.": 0.4482

Evaluating Causal Strength

The evaluate function computes the causal strength between two statements.

1. For the CESAR model.

   from causalstrength import evaluate
   
   # Test CESAR Model
   s1_cesar = "Tom is very hungry now."
   s2_cesar = "He goes to McDonald for some food."
   
   print("Testing CESAR model:")
   cesar_score = evaluate(s1_cesar, s2_cesar, model_name='CESAR', model_path='shaobocui/cesar-bert-large')
   print(f"CESAR Causal strength between \"{s1_cesar}\" and \"{s2_cesar}\": {cesar_score:.4f}")
   

   This will now output the following without errors:

   Testing CESAR model:
   CESAR Causal strength between "Tom is very hungry now." and "He goes to McDonald for some food.": 0.4482
   

2. For the CEQ model

    from causalstrength import evaluate
   
    # Test CEQ Model
    s1_ceq = "Tom is very hungry now."
    s2_ceq = "He goes to McDonald for some food."
    
    print("\nTesting CEQ model:")
    ceq_score = evaluate(s1_ceq, s2_ceq, model_name='CEQ')
    print(f"CEQ Causal strength between \"{s1_ceq}\" and \"{s2_ceq}\": {ceq_score:.4f}")
   

   This will now output the following without errors:

   Testing CEQ model:
   CEQ Causal strength between "Tom is very hungry now." and "He goes to McDonald for some food.": 0.0168
   

Parameters:

• s1 (str): The cause statement.
• s2 (str): The effect statement.
• model_name (str): The name of the model to use ('CESAR', 'CEQ', etc.).
• model_path (str): Hugging Face model identifier or local path to the model.

Generating Causal Heatmaps

Visualize the attention and similarity scores between tokens using heatmaps.

from causalstrength.visualization.causal_heatmap import plot_causal_heatmap

# Statements to visualize
s1 = "Tom is very hungry now."
s2 = "He goes to McDonald for some food."

# Generate heatmap
plot_causal_heatmap(
    s1,
    s2,
    model_name='shaobocui/cesar-bert-large',
    save_path='causal_heatmap.png'
)

This will now output the following without errors:

Testing CESAR model:
Warning: The sliced score_map dimensions do not match the number of tokens.
The causal heatmap is saved to ./figures/causal_heatmap.png

The causal heatmap is as follows:

📚 References

1. Cui, Shaobo, et al. "Exploring Defeasibility in Causal Reasoning." Findings of the Association for Computational Linguistics ACL 2024. 2024.
2. Du, Li, et al. "e-CARE: a New Dataset for Exploring Explainable Causal Reasoning." Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.