Selectional Restriction Extraction for Frame-Based Knowledge Graph Augmentation

The Semantic Web is an ambitious project aimed at creating a global, machine-readable web of data, to enable intelligent agents to access and reason over this data. Ontologies are a key component of the Semantic Web, as they provide a formal description of the concepts and relationships in a particular domain. Exploiting the expressiveness of knowledge graphs together with a more logically sound ontological schema can be crucial to represent consistent knowledge and inferring new relations over the data. In other words, constraining the entities and predicates of knowledge graphs leads to improved semantics. The same benefits can be found for restrictions over linguistic resources, which are knowledge graphs used to represent natural language. More specifically, it is possible to specify constraints on the arguments that can be associated with a given frame, based on their semantic roles (selectional restrictions). However, most of the linguistic resources define very general restrictions because they must be able to represent different domains. Hence, the main research question tackled by this thesis is whether the use of domain-specific selectional restrictions is useful for ontology augmentation, ontology definition and neuro-symbolic tasks on knowledge graphs. To this end, we have developed a tool to empirically extract selectional restrictions and their probabilities. The obtained constraints are represented in OWL-Star and subsequently mapped into OWL: we show that the mapping is information preserving and invertible if certain conditions hold. The OWL ontologies are inserted inside Framester, an open lexical-semantic resource for the English language, resulting in an improved and augmented language resource hub. The use of selectional restrictions is also tested for ontology documentation and neuro-symbolic tasks, showing how they can be exploited to provide meaningful results.