Natural Language Processing Strategies for Fine-Grained Knowledge Graph Verification

• Abi cit

  Lautech

  Author

Knowledge graphs (KGs) have become foundational infrastructure for search, question answering, and retrieval-augmented generation, yet the automated triples populating them remain error-prone due to extraction noise and large language model (LLM) hallucination. Existing verification methods predominantly return binary support/refute labels that neither localize which component of a triple is faulty nor provide traceable evidence for human verification. This study addresses this gap by developing and evaluating a fine-grained, slot-level verification framework that recasts triple validation as a question-driven, corpus-grounded reasoning task. Using a design-based research methodology integrating retrieval-augmented generation (RAG), slot-specific question generation, and independent slot-level answering, the framework was tested on 12,847 triples extracted from Wikipedia and news corpora across three domains. Results demonstrate that slot-level verification achieved 89.4% overall accuracy, a 14.7 percentage point improvement over single-label baseline approaches, with fault localization accuracy reaching 83.2%. Evidence-grounded reasoning contributed a 9.3-point accuracy gain over memorization-based verification. The findings establish that decomposing triple verification into head, relation, and tail slots with explicit evidence retrieval significantly improves both accuracy and interpretability of KG quality assurance. Practically, the framework enables scalable, evidence-backed verification workflows that reduce human review costs while maintaining auditability.

• pdf

Copyright (c) 2026 Abi cit (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.