Ontology Learning

Ontology learning is the semi-automatic construction of ontologies from data—usually text—extracting concepts, hierarchies and relations. First it helps to recall what an ontology is: in computing, Tom Gruber's canonical definition is “an explicit specification of a conceptualization,” that is, a formal description of a domain's concepts, their properties and the relations among them, expressed in languages such as OWL or RDF.

From the lexical to the logical

Ontology learning is organized in layers that rise from the linguistic to the formal, in what is often called the ontology-learning “layer cake”: first the relevant terms are extracted, then their synonyms are grouped into concepts, organized into a hierarchy of “is-a” relations (the taxonomy), further non-hierarchical relations are added and, at the top, axioms and rules are induced. Each layer builds on the previous one.

What it is not

It should be distinguished from three concepts it is confused with. Ontology engineering is its manual construction by experts; ontology learning partly automates it. Relation extraction is just one of its subtasks, not the whole process. And a knowledge graph is not a synonym for an ontology: the ontology is the schema—the types, the relations and the rules—whereas the knowledge graph is the data, the concrete entities and their connections; an ontology can serve as the schema for a knowledge graph.

Methods

The methods combine linguistic and statistical techniques with patterns such as Hearst's, which detect “is-a” relations in text (for example, the phrase “birds such as the ostrich” reveals that the ostrich is a bird). Today, large language models open a new paradigm for extracting concepts, inducing taxonomies and populating ontologies, though with open challenges such as logical consistency and hallucinations.