Biomedical Knowledge Graph Construction Based on Embedding Models and Ontologies

Abstract

Many machine learning and linguistic tools assist biomedical practitioners and researchers in fulfilling their daily missions. Text mining techniques, such as NLP, help process large amounts of biomedical data and perform tasks such as information retrieval, question answering, and document recommendation. Linguistic resources, such as ontologies and terminologies, provide a standardized knowledge reference across all biomedical subdomains to resolve ambiguities and unify the semantic meaning of biomedical words. This paper presents a series of experiments carried out to investigate new ways of exploiting ontologies to improve different aspects of text mining tools. In this scope, we focus on knowledge graph extraction (KGE), a high-level task that requires extracting knowledge from natural language text in the form of nodes and edges to construct a structured graph. We test new theories regarding the 3 fundamental operations in the KGE process: Named Entity Recognition (NER), Entity Linking (EL), and Relation Extraction (RE), to improve trainability and generalization aspects of state-of-the-art tools while preserving decent performance in each task. These experiments involve combining word embedding models, graph embedding models, and ontology data in a unified setup to develop NER, EL, and RE components trained on standard knowledge. Based on our experiments, we designed a state-of-the-art RE model with a 0.75 F1 score and recovered the shortcomings of the NER and EL experiments. Lastly, we apply our work in an ontology enrichment application scenario by integrating the newly conceived RE model into a KGE pipeline to extract ontology-aligned relations from text.