ANALYZING KNOWLEDGE GRAPH RELIABILITY AND SEMANTIC COHERENCE THROUGH EMBEDDING CLUSTERING

<p dir="ltr">Effective knowledge representation is fundamental for computer systems to understand and process information seamlessly. One fundamental task in this domain is to construct knowledge bases, such as knowledge graphs and ontologies, that are intrinsically related to textual data. However, several factors, including background (e.g., culture and language), purpose, and construction methods, contribute to knowledge graph completion. Consequently, some skepticism remains about the reliability of knowledge graphs. Previous research has explored related areas of reliability, such as refinement, managing quality, or detecting uncertainty. However, they usually approach it at a triple level, and evaluating the reliability of the knowledge graph itself remains limited. Motivated by those aspects, this study asks: To what extent do knowledge graphs reliably structure semantic coherence, and can they be considered sufficiently expressive and complete under practical resource and format constraints? To examine this, the study performs clustering on embeddings derived from verbalized triples in knowledge graphs, using three diverse relation-level types (Full Relation, Preposition-Stripped Relation, and Masked Relation), to explore semantic consistency of the relation use in graph construction. This study leverages two language model-based embeddings, BERT and Sentence-BERT, to capture semantics at both the relational and sentence levels. The results show that relation-level BERT embeddings achieve the highest semantic coherence, while sentence-level embeddings are sensitive. These findings offer insights into the reliability of the knowledge graph and relation-based semantic coherence.</p>