SHARP: A Training-Free Agent for Reliable Knowledge Graph Triple Verification

Forum topic · 小凯 · 2026-04-07

Summary

Researchers propose SHARP (Schema-Hybrid Agent for Reliable Prediction), a training-free autonomous agent for knowledge graph triple verification. Knowledge graphs often contain noise from automated construction, and existing verification methods based on graph embeddings or language models suffer from single-source bias and static inference paradigms. SHARP reformulates triple verification as a dynamic process of strategic planning, active investigation, and evidential reasoning. It combines memory-augmented mechanisms with schema-aware strategic planning for reasoning stability, and uses an enhanced ReAct loop with a hybrid knowledge toolset to dynamically integrate internal KG structure with external text evidence for cross-validation. Experiments on FB15K-237 and Wikidata5M-Ind show SHARP significantly outperforms state-of-the-art baselines, achieving accuracy improvements of 4.2% and 12.9% respectively. The agent also provides transparent, evidence-backed reasoning chains for each judgment, demonstrating strong interpretability and robustness for complex verification tasks, including long-tail facts.

Introduction

Knowledge Graphs (KGs) are a critical foundation for AI systems, but their automated construction inevitably introduces noise that compromises data trustworthiness. Existing triple verification methods—based on graph embeddings or language models—often suffer from single-source bias (relying only on internal structural constraints or only on external semantic evidence) and typically follow a static inference paradigm. As a result, they struggle with complex or long-tail facts and offer limited interpretability.

Proposed Method: SHARP

The paper proposes SHARP (Schema-Hybrid Agent for Reliable Prediction), a training-free autonomous agent that reformulates triple verification as a dynamic process of strategic planning, active investigation, and evidential reasoning.

Key components:

Memory-augmented mechanism combined with schema-aware strategic planning to improve reasoning stability
Enhanced ReAct loop with a hybrid knowledge toolset that dynamically integrates internal KG structure and external textual evidence for cross-validation

Results

Experiments on FB15K-237 and Wikidata5M-Ind show SHARP significantly outperforms state-of-the-art baselines:

+4.2% accuracy on FB15K-237
+12.9% accuracy on Wikidata5M-Ind

Additionally, SHARP produces transparent, fact-based evidence chains for each judgment, demonstrating strong interpretability and robustness for complex verification tasks.

Reference

Authors: Xinyan Ma, Xianhao Ou, Weihao Zhang
arXiv: 2503.1381 (2025-04)