DeepFCA

Abstract

Biomedical ontologies contain target domain knowledge. In many cases, multiple ontologies are created independently for different purposes in the same biomedical domain. To fuse and extend existing knowledge, we need to find the corresponding entities (i.e. classes and properties) from different ontologies. Formal Concept Analysis (FCA) is a mature mathematical tool for biomedical ontology matching tasks and has achieved competitive performance. The FCA-based method mainly matches the ontologies through lexical tokens and structural information. This method ignores the inherent semantics of entities. On the other hand, representation learning techniques are widely used in different NLP tasks to capture the semantic similarity of words. In this paper, we propose a novel biomedical ontology matching method which we dub DeepFCA. We use pre-trained word vectors to initialize the vector representations onto which semantic information is inscribed. FCA embedding techniques are used to refine these vectors. DeepFCA combines FCA and word2vec methods to enhance the performance of biomedical ontology matching. To the best of our knowledge, this is the first attempt to apply FCA embedding techniques to biomedical ontology matching. Experiments on real-world biomedical ontologies show that DeepFCA improves the recall and F1-measure compared with the traditional FCA-based algorithm. It also achieves competitive performance compared with several state-of-the-art systems.