LLM Fact Auditor：为大语言模型答案构建事实核查流水线

Core Introduction LLM Fact Auditor is a post-processing fact-checking system for large language model (LLM) answers, developed by students from the Web Data Processing Systems course at Vrije Universiteit Amsterdam. It addresses the critical 'hallucination' issue in LLMs (generating plausible but factually incorrect content) through a three-stage pipeline: entity linking, answer extraction, and fact verification.

Basic Info

• Original author/maintainer: dettinjo
• Source: GitHub (https://github.com/dettinjo/LLM-Fact-Auditor)
• Release/update time: 2026-06-02

This system aims to enhance the reliability of AI-generated content, making it suitable for high-accuracy scenarios like news writing and academic research.

LLMs excel in natural language tasks but suffer from 'hallucination'—generating content that seems reasonable but contains factual errors or unverifiable info. This is fatal in fields requiring precision (e.g., news, academic research, medical advice).

To tackle this, the Amsterdam Free University student team created LLM Fact Auditor: a post-processing pipeline to verify LLM outputs and improve their credibility.

The system uses a modular three-step process:

1. Entity Linking: Identify named entities (people, places, organizations) in LLM answers and link them to Wikipedia entries (via spaCy/Stanza tools). This provides a factual foundation for verification.
2. Answer Extraction: Distill long LLM responses into concise answers (e.g., 'yes/no' for binary questions, specific entities for entity queries) using Hugging Face Transformers models.
3. Fact Verification: Cross-check the extracted answer against structured knowledge sources like Wikidata and linked Wikipedia pages. Results are marked as correct, incorrect, or unverifiable.

Key technologies used:

• Python: Main development language.
• Docker: Containerization for consistent environments.
• Llama 2/3: Base LLMs for answer generation.
• PyTorch: Deep learning framework for model inference.
• Hugging Face Transformers: Pre-trained models for answer extraction.
• spaCy: NER and text processing.

The modular design allows easy replacement of components (e.g., switching entity linking tools or adding new knowledge bases).

Command Line Usage

• Default (Llama 2): python3 main.py < ./test_data/input.txt > ./test_data/output.txt
• Llama3 (higher quality/faster): python3 main.py --llama_ver=3 < ./test_data/input.txt > ./test_data/output.txt

Output Fields

• R: Raw LLM response
• A: Extracted concise answer
• C: Correctness result (correct/incorrect/unverifiable)
• E: Linked entities with Wikipedia URLs

Example For the question 'Is Managua the capital of Nicaragua?', output includes:

• R: Original Llama answer
• A: 'yes'
• C: 'correct'
• E: Links to Managua and Nicaragua Wikipedia pages.

LLM Fact Auditor offers a practical way to improve LLM reliability by using external knowledge verification (complementary to model internal improvements).

Applications:

• News/content审核: Verify AI-generated news facts.
• Education: Help students spot errors in learning materials.
• Enterprise KM: Check accuracy of internal docs/reports.
• Search engines: Add credibility scores to results.

Future Plans

1. Add relation extraction to verify entity relationships.
2. Support more knowledge bases beyond Wikidata.
3. Develop a web interface for easier use.
4. Improve handling of nuanced/complex claims.

Summary LLM Fact Auditor is a pragmatic solution to LLM hallucination. It provides a reusable, extensible framework for fact-checking, which will become increasingly important as LLMs are used in critical fields. It complements model improvements to build more reliable AI systems.