KEY APPROACHES TO AUTOMATIC INDEXING OF SCIENTIFIC TEXTS WITH KEY TERMS BASED ON ADVANCED METHODS AND VECTOR REPRESENTATION MODELS

Komarov Ivan Dmitrievich

This article describes the features of automatic indexing of scientific texts with key terms based on methods and models of vector representations. Within the framework of the analysis, such classes of methods as classical vector representations of words, contextual vector structures, graph-weighted schemes and solutions adapted to specialized corpora are identified, allowing to compare semantic elements of scientific publications and form semantic descriptions of texts. In addition, the following models are considered: pre–trained language structures, contextual transformer architectures, models of vector representation of sentences and documents, as well as generative encoder-decoder solutions. According to the results of the study, four key approaches to automatic indexing of scientific texts with key terms have been identified: embedding-ranking, contextual-transformative with sequence markup, generative approach based on encoder–decoder architecture and large-scale models, structural-embedding hybrid.

Keywords:automatic indexing, keywords, vector representations, methods, models, approaches, scientific texts.