Detecting AI-Generated Text

This folder contains Jupyter Notebooks for experiments and analysis on distinguishing human-written texts from AI-generated texts.

Contents

• 1_Data_Preprosessing.ipynb: Data preprocessing and cleaning (tokenization, tokenization to vectors, etc.).
• 2_multi_models_stacking.ipynb: Building and evaluating multiple models (KNN, Random Forest, stacking ensemble, etc.).
• 3_CNN_Visualization.ipynb: Text-CNN implementation and visualization (filters / feature maps).

Quick Start

1. Install dependencies (use a virtual environment):

   pip install -r requirements.txt

2. Open and run the notebooks in order: 1_Data_Preprosessing.ipynb, then 2_multi_models_stacking.ipynb, and finally 3_CNN_Visualization.ipynb.

Data

• Source: Kaggle competition (processed CSV included in the project).
• Example size: ~29,133 English texts (see notebooks for exact splits and files).
• Labels: 0 = human-written, 1 = LLM-generated.

Models & Methods

• Traditional ML: KNN and Random Forest as base learners, Logistic Regression as a meta-learner for stacking.
• Deep Learning: Text-CNN (Embedding → Conv → Pool → Dense) for end-to-end training and visualization.
• Text processing: tokenization, frequency-based tokenizers, and vectorization methods to transform text into numerical features.

Sample Results

• Logistic Regression: Accuracy ≈ 0.60
• KNN: Accuracy ≈ 0.67
• Random Forest: Accuracy ≈ 0.72
• Stacking: Accuracy ≈ 0.71
• Text-CNN: Accuracy ≈ 0.66

(For detailed experimental settings, hyperparameters and evaluation metrics, see the experiment sections inside each notebook.)

Environment Recommendations

• Python 3.8+. Use a virtual environment (venv or conda).
• If using a GPU, install a compatible deep learning framework (TensorFlow or PyTorch) and CUDA drivers.

Requirements Check & Example Run Scripts

This project includes a small helper script to verify that packages listed in requirements.txt are installed with compatible versions, plus example scripts to execute the notebooks end-to-end.

• check_requirements.py: Verify installed packages against requirements.txt and report missing or incompatible packages.
• run_all_notebooks.sh: Bash script to install deps and execute all notebooks (creates executed/ outputs).
• run_all_notebooks.ps1: PowerShell equivalent for Windows users.

Usage examples:

1. Quick requirements check (preferred before running notebooks):

python check_requirements.py

2. Run all notebooks (bash):

bash run_all_notebooks.sh

3. Run all notebooks (PowerShell):

.\run_all_notebooks.ps1

Notes:

• The run scripts use jupyter nbconvert --execute to run notebooks non-interactively and save executed copies under the executed/ folder.
• Running the notebooks may take time and may require additional system resources (RAM/GPU) depending on model choices.

---

This project is a CUHKSZ-DDA4210 course assignment.