Weakly Labeled News AI

This project explores the application of weak supervision techniques for classifying news titles, aiming to build effective text classification models with minimal hand-labeled data. It leverages programmatic labeling and tools like Snorkel to create training data from multiple noisy heuristics (labeling functions).

Overview

Traditional supervised learning requires large amounts of manually labeled data, which can be expensive and time-consuming to create. This project demonstrates an alternative approach using weak supervision, where we define labeling functions (LFs) that provide noisy, imperfect labels for our data. These LFs are then combined using a Label Model to generate probabilistic training labels, which can be used to train a powerful downstream classification model.

The primary focus is on classifying news titles into predefined categories (e.g., Sports, World, Technology, Business).

Key Features

• Weak Supervision: Core methodology using multiple noisy labeling functions.
• Programmatic Labeling: Defining heuristics and rules as code to generate labels.
• Labeling Functions (LFs): Custom-defined Python functions that vote on the likely class of a news title.
• Snorkel Framework (implied): Likely used for its Label Model to aggregate LF outputs and generate probabilistic labels.
• Majority Vote Baseline: Comparison with a simple majority vote aggregation.
• Data-Centric AI Principles: Focus on improving data quality and generation rather than solely model architecture changes.
• Jupyter Notebook for Analysis: Core experimentation and analysis in NewsTitleWeakSupervision.ipynb.
• Poetry for Dependency Management: Ensuring reproducible environments.

Getting Started

Follow these instructions to set up and run the project locally.

Prerequisites

• Python (version specified in pyproject.toml)
• Poetry for Python packaging and dependency management.

Installation

1. Clone the repository:

   git clone <your-repository-url>
   cd weakly-labeled-news-ai

2. Set up the environment: The create-env.sh script is likely responsible for creating a virtual environment and installing dependencies using Poetry.

   bash create-env.sh

   Alternatively, if create-env.sh just runs Poetry commands, you might do:

   poetry install

   Make sure to activate the Poetry shell or environment:

   poetry shell

3. Environment Variables: If the project uses an .env file for configuration (e.g., paths to data, API keys), create one by copying a template if provided (e.g., .env.example) or by creating it manually:

   # .env file example
   # DATA_PATH="data/your_dataset.csv"

Usage

The main workflow, including data loading, defining labeling functions, training the Label Model, and evaluating performance, is expected to be in the Jupyter Notebook:

1. Ensure your environment is activated:

   poetry shell

2. Launch Jupyter Lab or Jupyter Notebook:

   jupyter lab
   # or
   # jupyter notebook

3. Open and run the cells in NewsTitleWeakSupervision.ipynb.

The notebook should guide you through:

• Loading the news title dataset.
• Defining and applying labeling functions (LFs).
• Using a Label Model (e.g., from Snorkel) to aggregate LF outputs.
• Analyzing the performance of the Label Model and potentially training a downstream classifier.

Dependencies

This project uses Poetry to manage dependencies. All required packages are listed in the pyproject.toml file, and specific versions are locked in poetry.lock.

Key libraries likely include:

• pandas for data manipulation
• numpy for numerical operations
• scikit-learn for machine learning utilities and metrics
• snorkel (or similar) for weak supervision tools
• jupyterlab or notebook for running the analysis notebook
• NLP libraries like spacy or nltk (if used for LFs)
• matplotlib or seaborn for visualization (if applicable)
• dotenv for environment variable management (if used)

Contributing

Contributions are welcome! If you'd like to contribute, please fork the repository and use a feature branch.

License

This project is licensed under the terms of the LICENSE file. Please see the LICENSE file for more details.