ИСТИНА

In this paper we study the use of meta-embeddings approaches, which combine several source embeddings, for the taxonomy class prediction of new terms. We test the proposed approach in the information-security domain in the task of enriching the Ontology on Natural Sciences and technologies (OENT). We show that autoencoder-based meta-embeddings with triplet loss achieve the best results in the task. The highest results are obtained on combination of in-domain and out-of-domain embeddings.