Ontology

How it works

An ontology defines three main things. First, the kinds of things it deals with (e.g. 'customer', 'order', 'product'). Second, the properties each thing has (e.g. an order's 'date' and 'amount'). Third, the relationships between things (e.g. 'a customer places an order', 'an apple is a kind of fruit'). Once these are defined, a computer can derive facts that are not written down from the facts that are, or check whether any pieces of information contradict each other.

How it has developed

As a standard for sharing ontologies on the web, the W3C (the international body that develops web standards) published OWL (Web Ontology Language) in 2004, the revised OWL 2 in 2009, and its second edition in 2012. OWL is a logic-based language, so programs can use it to check whether knowledge is consistent or to surface hidden facts. In the same year, 2009, the W3C also made SKOS a standard, a lightweight format designed to make relatively simple vocabularies such as thesauri and classification schemes easy to share on the web. Since large language models (LLMs) appeared, research connecting the two has grown. The 2023 'LLMs4OL' study tested several models on how well LLMs, simply prompted with questions and without additional training, can perform ontology-building tasks such as identifying term types, finding broader and narrower concepts, and extracting relationships between concepts. A paper from the same year, 'Unifying Large Language Models and Knowledge Graphs: A Roadmap', organized research into three directions: enhancing LLMs with structured knowledge, using LLMs to build and populate knowledge structures, and having the two help each other.

Why it matters

LLMs speak fluently, but their inner workings are hard to inspect and they sometimes state facts incorrectly. Knowledge that clearly records concepts and relationships, like an ontology, becomes a reliable reference that AI can lean on when answering. Also, when different departments or systems call the same thing by different names, a shared ontology makes it easier to link the data together. That is why it often comes up alongside technologies that use 'structured knowledge', such as enterprise AI agents and GraphRAG.

Things to know

Building a good ontology from scratch takes a lot of time and effort from domain experts. Real-world knowledge keeps changing, so an ontology is not built once and finished; it needs to be revised continually. LLMs can help with the building work, but they can also create incorrect relationships, so human review is still needed. For reference, 'ontology' was originally a philosophical term dealing with 'what exists', and information science borrowed the name to refer to concept systems like the one described above.

Example

The data analytics platform company Palantir promotes the 'Ontology' as a core feature of its product Foundry. The company's documentation describes it as the 'operational layer for the organization' that connects scattered data to real-world objects such as factories, equipment and products, or to concepts such as customer orders and financial transactions, and states that in many cases it serves as the organization's digital twin. It organizes data into 'objects, properties and links', and on top of that defines 'actions' that people or AI can execute, so it can be used for decision-making.

Related terms

Knowledge graph GraphRAG LLM Semantic search Hallucination AI agent

References

• LLMs4OL: Large Language Models for Ontology Learning (2023)
• Unifying Large Language Models and Knowledge Graphs: A Roadmap (2023)
• Ontology (information science) - Wikipedia
• OWL - Semantic Web Standards
• Ontology overview (Palantir Foundry documentation)
• SKOS Simple Knowledge Organization System Reference