Enrichment of ontologies using machine learning and summarization

Biomedical ontologies are structured knowledge systems in biomedicine. They play a major role in enabling precise communications in support of healthcare applications, e.g., Electronic Healthcare Records (EHR) systems. Biomedical ontologies are used in many different contexts to facilitate information and knowledge management. The most widely used clinical ontology is the SNOMED CT. Placing a new concept into its proper position in an ontology is a fundamental task in its lifecycle of curation and enrichment. A large biomedical ontology, which typically consists of many tens of thousands of concepts and relationships, can be viewed as a complex network with concepts as nodes and relationships as links. This large-size node-link diagram can easily become overwhelming for humans to understand or work with. Adding concepts is a challenging and time-consuming task that requires domain knowledge and ontology skills. IS-A links (aka subclass links) are the most important relationships of an ontology, enabling the inheritance of other relationships. The position of a concept, represented by its IS-A links to other concepts, determines how accurately it is modeled. Therefore, considering as many parent candidate concepts as possible leads to better modeling of this concept. Traditionally, curators rely on classifiers to place concepts into ontologies. However, this assumes the accurate relationship modeling of the new concept as well as the existing concepts. Since many concepts in existing ontologies, are underspecified in terms of their relationships, the placement by classifiers may be wrong. In cases where the curator does not manually check the automatic placement by classifier programs, concepts may end up in wrong positions in the IS-A hierarchy. A user searching for a concept, without knowing its precise name, would not find it in its expected location. Automated or semi-automated techniques that can place a concept or narrow down the places where to insert it, are highly desirable. Hence, this dissertation is addressing the problem of concept placement by automatically identifying IS-A links and potential parent concepts correctly and effectively for new concepts, with the assistance of two powerful techniques, Machine Learning (ML) and Abstraction Networks (AbNs). Modern neural networks have revolutionized Machine Learning in vision and Natural Language Processing (NLP). They also show great promise for ontology-related tasks, including ontology enrichment, i.e., insertion of new concepts. This dissertation presents research using ML and AbNs to achieve knowledge enrichment of ontologies. Abstraction networks (AbNs), are compact summary networks that preserve a significant amount of the semantics and structure of the underlying ontologies. An Abstraction Network is automatically derived from the ontology itself. It consists of nodes, where each node represents a set of concepts that are similar in their structure and semantics. Various kinds of AbNs have been previously developed by the Structural Analysis of Biomedical Ontologies Center (SABOC) to support the summarization, visualization, and quality assurance (QA) of biomedical ontologies. Two basic kinds of AbNs are the Area Taxonomy and the Partial-area Taxonomy, which have been developed for various biomedical ontologies (e.g., SNOMED CT of SNOMED International and NCIt of the National Cancer Institute). This dissertation presents four enrichment studies of SNOMED CT, utilizing both ML and AbN-based techniques