Method of building embeddings of signs in deep learning problems based on ontologies

Abstract

This paper investigates the problem of embedding features used in datasets for training neural networks. The use of embeddings increases the performance of neural networks, and therefore is an important part of data preparation for deep learning methods. Such a process is based on semantic metrics. It is proposed to use ontologies of the subject areas to which the corresponding feature belongs for embedding. This work developed such a method and investigated its use for the task of categorizing text documents. The research results showed the advantage of the developed method.