Designing Scalable Knowledge Graph Schemas: From Structure to Semantic Modeling

Introduction

This document summarizes a systematic approach to knowledge graph modeling that separates structural representation from semantic meaning, enabling reusable, low‑cost schema design for diverse business scenarios.

Background

Knowledge modeling transforms real‑world information into structured data that can be processed by computers. Traditional relational models lack semantic depth, while ontology‑based methods are often too rigid for large‑scale, heterogeneous data.

Problem

High schema design cost, difficulty in reusing models across domains, and the need to capture multi‑entity, multi‑attribute, and multi‑relational knowledge hinder rapid graph construction.

Solution Overview

The proposed solution introduces a four‑layer framework: (1) strong schema constraints, (2) attribute semantic standardization, (3) concept (meta‑concept) modeling, and (4) multi‑relational/hypergraph representation for events and behaviors.

Schema Design

Core schemas define mandatory fields (id, name, description) and optional extensions for time, location, subjects, and objects. Schemas are reusable via CoreKG inheritance and can be customized per business need.

Entity Definition

Entities represent concrete instances (e.g., persons, companies) with attributes such as id, name, certificate type, birthday, gender, and occupation. Inheritance allows specialized user models (AlipayUser, FortuneUser, etc.) to extend the base Person schema.

Relationship Definition

Binary relationships follow the SPO (Subject‑Predicate‑Object) pattern, e.g., Company‑法人‑Person. Multi‑entity relationships are modeled as hyperedges represented by an event node linked to its participants.

Attribute Semantic Standardization

Attributes are classified as plain properties, standard semantic types (e.g., ID, email, phone), or custom types. Standardization automatically creates virtual edges that propagate semantic connections without increasing storage.

Concept Modeling

Concepts abstract common properties of entities into a hierarchical taxonomy (meta‑concepts). Examples include role, object, organization, brand, and event. Concepts are linked with isA or domain‑specific predicates (e.g., locatedAt for administrative regions).

Multi‑Relational Modeling and Hypergraph

Events and behaviors often involve more than two entities, requiring hypergraph representation. An event node encapsulates all arguments (time, location, subject, object) and connects to each via explicit edges, preserving lossless conversion between SPO triples and hyperedges.

Event Modeling

Events are defined with basic, temporal, spatial, subject, and object elements. Tables illustrate schemas for industrial‑chain events, financial events, and user travel behaviors, showing required fields and example values.

Implementation Steps

Reuse CoreKG schemas where possible.

Design entity‑relationship schemas for the target domain.

Define a meta‑concept hierarchy for semantic classification.

Assign belongTo relations between schemas and concepts.

Produce instance data from structured sources or extracted text, applying attribute standardization and concept chain‑linking.

Build the semantic network by linking concepts with logical rules (e.g., if eventProduct = "汽车整车" and eventIndicator = "销量" then belong to "汽车整车销量事件").

Future Outlook

Integrating large language models (LLMs) with schema‑driven prompts can automate commonsense schema generation, improve extraction quality, and enable interactive knowledge discovery, bridging the gap between symbolic graphs and neural models.

References

Knowledge Graph Survey – Representation, Construction, Reasoning, and Hypergraph Theory.

ASER: Large‑scale commonsense knowledge via higher‑order selectional preference.

Atomic: An Atlas of Machine Commonsense for If‑Then Reasoning.

AliCoCo2: Commonsense Knowledge Extraction, Representation and Application in E‑commerce.