Humans try to help computers understand the properties of the real world, and ontologies can be used for this task. Scientists publish their research in papers, and their results should be used to improve existing ontologies to be up-to-date. Manual enhancement of ontologies is highly time-consuming for domain experts. This paper proposes a solution to match a scientific text to the most relevant ontology using artificial neural networks. Our approach selects a paragraph or a sentence, uses representation learning to embed it into a vector space by some embedder, and measures its relevance to embedded textual properties from the selected ontology by a modified version of a Siamese neural network. A modification is based on the extension of one branch of the Siamese network to aggregate inputs from a group of embeddings. We have considered different embedders, in particular two variants of BERT, InferSent, GloVe with TF-IDF weighted mean, Doc2Vec in the distributed memory variant, and the Llama 3.1 with LLM2vec framework. Their quality has been evaluated on a use case with available ontologies from several application domains. The best results were achieved with InferSent and SentenceBERT.