LLM Fact-Checking Pipeline

Evaluating LLM Capabilities of Fact-Checking Political Statements Presentation

Final Project for COMS 3997: Large Language Models: Foundations and Ethics w/ Prof. Smaranda Muresan

Team Members: Shayan Chowdhury, Sunny Fang, Ha Yeon Kim, JP Suh, Anissa Arakal

Through our project Our project is a comprehensive fact-checking pipeline that leverages large language models (LLMs) and various natural language processing (NLP) techniques to verify the factuality of claims or statements, especially related to political speech.

Pipeline Overview

1. Claim Atomization: The input statement is broken down into a set of atomic claims, each focusing on a single idea or concept, to facilitate more granular fact-checking.

2. Question Generation: For each atomic claim, a set of independent questions is generated to guide the web search for relevant information to verify the claim.

3. Web Querying & Scraping: For each question, relevant search results are fetched from the web, and the content of the top search results is scraped for further processing.

4. Retrieval Augmented Generation (RAG): The scraped web content is processed and split into smaller chunks, and the most relevant chunks are retrieved using a Retrieval Augmented Generation (RAG) approach based on the input question.

5. Answer Synthesis: Given the relevant document chunks, answers are synthesized to the generated questions using the language model's capabilities to extract relevant information.

6. Claim Classification: Based on the generated answers, a chain-of-thought reasoning process is employed to classify each atomic claim as true or false.

7. Fact Score Generation: Finally, the verdicts for all atomic claims are combined to generate a fact score label (e.g., True, Mostly True, Half True, Mostly False, Pants on Fire, or Unverifiable) for the original statement.

Models

In this project, we utilize two different classes of models:

• HuggingFace Transformers
  • Specifically, we are using the Mistral-7B-Instruct-v0.2 model.
  • We have future plans to include other models such as Meta AI's Llama-2-70b-chat-hf and Allen AI's OLMo-7B.
• OpenAI GPT-4 Turbo (TO BE ADDED)

Directory Structure

project/
├── utils/
│   ├── __init__.py
│   ├── web_utils.py     # Functions for web scraping
│   ├── code_utils.py    # General code utility functions
│   └── nlp_utils.py     # NLP-related utility functions
├── pipeline_{HF/GPT}/tasks/
│   ├── claim_atomization.py      # Task for breaking down statements into atomic claims
│   ├── question_generation.py    # Task for generating questions to verify claims
│   ├── web_querying.py           # Task for fetching and scraping web search results
│   ├── rag_retrieval.py          # Task for retrieving relevant document chunks using RAG
│   ├── answer_synthesis.py       # Task for synthesizing answers from relevant documents
│   ├── claim_classification.py   # Task for classifying claims as true or false
│   └── fact_score.py             # Task for generating a fact score label for the original statement
├── main_pipeline.py              # Orchestrates the entire fact-checking pipeline
└── main.py                       # Entry point for running the pipeline

Usage

1. Run pip install -r requirements.txt to install the required Python packages and dependencies.
2. Replace any placeholders or configurations (e.g., API keys) in .env with your actual values.
3. Run the main_pipeline.py file and provide the input statement(s) to fact-check.

from pipeline import verify_statement

if __name__ == "__main__":
    statement = "President Biden claimed the economy created a record 15 million jobs in the first three years of his term."
    fact_score_label, results = verify_statement(statement)

This will run the entire fact-checking pipeline for the provided statement with the following outputs:

• fact_score: the final fact score label
• results: detailed results for each atomic claim (including questions, answers, sources, and verdicts)