Publication Type

Conference Proceeding Article

Version

acceptedVersion

Publication Date

8-2023

Abstract

Knowledge graphs are repositories of facts about a world. In this work, we seek to distill the set of entities or nodes in a knowledge graph into a specified number of constitutive nodes, whose embeddings would be retained. Intuitively, the remaining accessory nodes could have their original embeddings “forgotten”, and yet reconstitutable from those of the retained constitutive nodes. The constitutive nodes thus represent the semantically constitutive entities, which retain the core semantics of the knowledge graph. We propose a formulation as well as algorithmic solutions to minimize the reconstitution errors. The derived constitutive nodes are validated empirically both in quantitative and qualitative means on three well-known publicly accessible knowledge graphs. Experiments show that the selected semantically constitutive entities outperform those selected based on structural properties alone.

Keywords

embeddings, knowledge graph, semantically constitutive

Discipline

Artificial Intelligence and Robotics | Databases and Information Systems

Research Areas

Data Science and Engineering

Publication

DEXA 2023: Database and Expert Systems Applications

ISBN

9783031398469

Identifier

10.1007/978-3-031-39847-6_36

City or Country

Penang, Malaysia

Download

Share

COinS