Neuro-symbolic AI aims to reduce hallucinations in neural networks by applying the rules and logic of symbolic reasoning, often structured as ontologies. An ontology provides the high level knowledge framework that can be used to govern the model and underpin the hybrid approach.
When a neuro-symbolic AI model needs to be reused, made more explainable (XAI), and unified with other systems, ontologies are essential.
What is neuro-symbolic reasoning?
Neuro-symbolic reasoning seeks to connect neural perception, which is unstructured and data-driven, with logical reasoning, which is structured and rule-based. It can be applied to everything from scientific discovery, to robots and domain intelligent systems.
The neural side of the model enables it to learn from raw data, by recognising objects in an image for example. The symbolic side enables it to figure out the relationship that exists between those objects, by understanding that a flowerpot is supported by the table underneath it, say.
What is an ontology?
Ontologies set out the precise meaning of key concepts in a domain or ontology application, along with how they relate to one another, in machine-readable format. AI services can then draw logical conclusions from the information and present them as human-friendly answers using a large language model (LLM). This is not something that a simple taxonomy or folder system can do.
A medical ontology might define what ‘disease’, ‘symptom’ and ‘treatment’ mean as part of a controlled vocabulary. It will also define their relationships, such that diseases ’cause’ symptoms, and treatments ‘are associated with’ diseases, enabling AI to reason with the knowledge.
Ontologies are generally built using standard languages like OWL (Web Ontology Language), sometimes with the help of ontology software.
Is it worth developing an ontology for neuro-symbolic AI?
Alternatives to ontologies can be used in domains where the meanings of words (semantics) are less important. For example, DeepMind used symbolic geometry reasoning instead of an ontology for its AlphaGeometry tool. In this case, logic and rules guided by neural learning was enough to govern the model.
However, when a neuro-symbolic system needs to reason using interrelated domain concepts, or integrate with other semantically rich systems, an ontology appears to be the most effective solution at present. For example, a neuro-symbolic system for cancer was developed in 2024 to transform unstructured clinical notes into structured terms, using the the UMLS medical vocabulary as a case study.