Top-Down vs Bottom-Up Ontology Construction
Two approaches to building ontological structures with LLM assistance, reflecting different assumptions about where structure should come from.
Top-Down: LLMs as Ontology Assistants
The traditional knowledge engineering approach, augmented by LLM capabilities. Human experts define semantic requirements, competency questions, domain descriptions, user stories, and LLMs help translate these into formal ontologies.
Priorities: Semantic consistency, structural completeness, logical rigor, expert oversight at checkpoints.
The LLM’s role: Co-modeler that converts natural language specifications into OWL or similar formalisms. It accelerates the translation process but doesn’t fundamentally change who’s driving.
When it works well: Domains where precision matters and the schema has downstream implications (regulatory contexts, scientific ontologies, enterprise systems with multiple stakeholders).
Bottom-Up: KGs for LLMs
The emerging approach, optimized for LLM consumption. Schemas are induced automatically from data rather than designed by experts. The knowledge graph serves as grounding infrastructure for language models rather than as a resource for human interpretation.
Priorities: Factual coverage, scalability, maintainability, continuous adaptation.
The LLM’s role: Both constructor and consumer. It extracts knowledge, abstracts patterns into schema elements, and later uses that structure for reasoning.
When it works well: Open domains, rapidly evolving knowledge, systems where the primary “user” of the knowledge graph is another AI system rather than a human analyst.
The Key Difference
Top-down emphasizes getting the schema right according to human understanding. Bottom-up emphasizes getting the schema useful for machine reasoning.
This isn’t just a methodological choice, it reflects different answers to the question of what knowledge graphs are for.
When Each Applies
Choose top-down when: Domain expertise captures distinctions not visible in text; schema errors have significant consequences; multiple human stakeholders need to agree on meaning; formal validation is required.
Choose bottom-up when: Scale matters more than perfect consistency; the knowledge is for machine consumption; the domain is well-represented in existing corpora; continuous updates outweigh initial precision.
The hybrid path: Use top-down methods to establish high-stakes core concepts, then bottom-up methods to expand coverage around that stable core.
Related: 06-atom—llm-kg-paradigm-inversion, 06-atom—emergent-vs-designed-schemas, 06-atom—three-bottlenecks-kg-construction