As an important data structure model, ontology has become one of the core contents in information science. Multi-dividing ontology algorithm combines the advantages of graph structure and learning algorithms proved to have high efficiency. In this paper, in terms of multi-dividing proper loss functions, we propose new multi-dividing ontology learning algorithms for similarity measure and ontology mapping construction. Several theoretical statistical characteristics supporting the new learning model are given. Finally, four experiments on different scientific fields verify that our multi-dividing ontology algorithm has high accuracy and efficiency in application implements.