| Abstract |
An ontology is a set of structural rules that represent concepts of a domain, which can be used to perform logic-based operations for retrieving or inferring new information. Ontologies have with them the knowledge of a particular domain expressed in the form of relational triples (subject, object, predicate). Several rules are expressed within the ontologies that define how the concepts in the domain are related. Rules are often expressed as a triple in the following format (concept, relation, concept). Often, in complex domains, such as smart city or smart home, there are several domains (and hence several ontologies) at play in parallel, hence requiring reasoning/intelligence across several domains and ontologies. Interoperability of information is necessary to find effective ways to align different ontologies. As the number of ontologies grows for a given domain, and as an overlap between ontologies grows proportionally, hence, there is a need to develop accurate and reliable techniques to perform this task automatically. Mapping concepts across ontologies can be challenging considering that there are fundamental differences in the syntax, structure and properties. Extraction of semantic interconnection between the ontologies is the key challenge for extending the knowledge base across ontologies and potentially across related domains. We propose to use a two-system representation learning network of a generator and discriminator for the case of ontology mapping. The generator uses a softmax classifier on the concepts and relations across ontologies to produce triples. It also uses concept embeddings on each of the ontology’s properties such as (1) name, (2) label, (3) comments, (4) properties, (5) instance descriptions, (6) concept attributes and (7) the neighborhood of nearby concepts. The discriminator network uses a convolution neural network with relation attention mechanism augmented by the concept descriptions for asserting the claims the generator has produced. Both the systems are in a feedback cycle to learn from each other. The output of the system will be a collection of triples that enumerate all the related concepts across ontologies. These triples will be subject to scrutiny out of band by domain experts to allow only valid concepts and triples to be selected into the alignment. The aligned ontology allows extensive querying and inference across related ontologies and domains more expressively than using the ontologies in isolation. © 2022 selection and editorial matter, Avadhesh Kumar, Shrddha Sagar, T. Ganesh Kumar and K. Sampath Kumar; individual chapters, the contributors. |
