English · Deutsch

LLM Fact Auditor

A post-processing pipeline to fact-check, entity-link, and verify answers from Large Language Models, built for the Web Data Processing Systems course at Vrije Universiteit Amsterdam.

The Challenge

LLMs are prone to hallucination — generating plausible-sounding but factually incorrect statements. This project addresses the verification problem by building an automated auditing layer that sits downstream of any LLM, cross-referencing its claims against authoritative knowledge sources.

Pipeline Architecture

1. Entity Recognition: Uses spaCy and Stanza to identify named entities (people, places, organisations, dates) within LLM-generated answers.
2. Entity Linking: Maps recognized entities to Wikidata entries, resolving ambiguity using contextual scoring.
3. Claim Extraction: Identifies factual claims in the text using dependency parsing and semantic role labelling.
4. Fact Verification: Cross-references extracted claims against the Wikidata knowledge graph via SPARQL queries, assigning a confidence score to each claim.
5. LLM Re-ranking: Uses a Meta Llama model to re-rank candidate answers based on factual accuracy scores, selecting the most reliable response.

Key Technologies

• PyTorch + Hugging Face Transformers: For LLM inference and semantic similarity scoring
• spaCy & Stanza: Multi-framework NLP for robust entity recognition across languages
• Wikidata API: The primary knowledge base for fact verification
• Docker: Containerized pipeline ensuring reproducible execution environments