Ontology Learning refers to the task of (semi-)automatically building ontologies from text via NLP, in order to avoid the extremely labor-intensive and time-consuming process of creating them manually.
NLP may be applied at different levels of the creation of an ontology: learning of new concepts, learning of semantic relations between concepts (is-a, part-whole, etc.), and ontology population, i.e., identification of instances of concepts (individuals) and their classification with respect to the classes of an existing legal ontology, e.g., LKIF.
Learning of new concepts is usually done via statistical NLP techniques such as Topic Modelling and Keyphrase Extraction, which are able to identify most recurring (groups of) linguistic patterns in text. The grouped patterns are domain-specific terms and their synonyms, which are assumed to correspond to relevant concepts in the domain.
Learning of semantic relations and ontology population may be done via either statistical or rule-based NLP procedures. The latter uses symbolic methods that are based on lexico-syntactic patterns, which are manually crafted or deducted automatically. They are generally more precise than statistical approaches but they do not tend to scale on the high variability of how a relation can be expressed in natural language (lower recall).