[Paper Review] SCEF: A Support-Confidence-aware Embedding Framework for Knowledge Graph Refinement

Yu Zhao, Ji Liu|arXiv (Cornell University)|Feb 18, 2019

Advanced Graph Neural Networks47 references4 citations

TL;DR

SCEF is a novel knowledge graph embedding framework that jointly performs KG completion and error correction by modeling triple support and confidence. It integrates structural entailment and external textual evidence into a translation-based energy function, achieving state-of-the-art performance on real-world datasets.

ABSTRACT

Knowledge graph (KG) refinement mainly aims at KG completion and correction (i.e., error detection). However, most conventional KG embedding models only focus on KG completion with an unreasonable assumption that all facts in KG hold without noises, ignoring error detection which also should be significant and essential for KG refinement.In this paper, we propose a novel support-confidence-aware KG embedding framework (SCEF), which implements KG completion and correction simultaneously by learning knowledge representations with both triple support and triple confidence. Specifically, we build model energy function by incorporating conventional translation-based model with support and confidence. To make our triple support-confidence more sufficient and robust, we not only consider the internal structural information in KG, studying the approximate relation entailment as triple confidence constraints, but also the external textual evidence, proposing two kinds of triple supports with entity types and descriptions respectively.Through extensive experiments on real-world datasets, we demonstrate SCEF's effectiveness.

Motivation & Objective

• Address the limitation of conventional KG embedding models that assume noise-free facts and ignore error detection.
• Simultaneously improve knowledge graph completion and correction by incorporating both support and confidence signals.
• Enhance robustness by leveraging internal structural patterns (e.g., relation entailment) and external textual evidence (e.g., entity types and descriptions).
• Develop a unified embedding framework that models both the reliability and plausibility of triples.

Proposed method

• Formulate a model energy function by extending translation-based KG embedding with support and confidence components.
• Define triple confidence using approximate relation entailment derived from structural patterns in the KG.
• Introduce two types of triple support: one based on entity types and another based on entity descriptions from external text.
• Optimize the energy function end-to-end to jointly learn representations for completion and error detection.
• Use a unified framework to balance support and confidence signals during training and inference.
• Apply the framework to real-world KGs to evaluate both completion and correction performance.

Experimental results

Research questions

• RQ1 Can a unified embedding framework effectively combine knowledge graph completion and error correction?
• RQ2 How do support and confidence signals improve the robustness of KG embeddings?
• RQ3 To what extent do external textual evidence (types and descriptions) enhance triple support modeling?
• RQ4 How does incorporating relation entailment affect triple confidence estimation?
• RQ5 Does the proposed framework outperform existing models in both completion and correction tasks?

Key findings

• SCEF achieves state-of-the-art performance on knowledge graph completion benchmarks by jointly modeling support and confidence.
• The integration of external textual evidence (entity types and descriptions) significantly improves triple support estimation.
• Modeling relation entailment as confidence constraints enhances the reliability of predicted facts.
• SCEF demonstrates superior error detection capability compared to conventional models that ignore noise.
• Extensive experiments on real-world datasets confirm the effectiveness and robustness of the proposed framework.
• The framework maintains strong performance across diverse KGs, indicating generalization potential.