APPLICATION OF GENERATIVE APPROACHES IN SEMANTIC ANALYSIS OF NATURAL LANGUAGE TEXTS

Kuzmenko Andrey V., Kireev Vasiliy S.

It is estimated that about 80% of corporate data is unstructured, which makes it difficult to process and analyze it using traditional methods. Extracting relationships between entities from this data in the form of relational triples allows you to transform unstructured information into a structured form, facilitating access, analysis, and use of knowledge. The paper examines modern approaches to extracting relational triples from natural language texts based on sequence transformation technology. The authors have identified and structured the existing solutions into a set of groups of methods: classical sequence transformation methods, sequence transformation methods into sets, and prompt methods. A comparative analysis of the advantages and disadvantages of these methods is carried out. A methodology for building an auxiliary system for extracting relational triples is proposed.

Keywords:relational triple, neural network, natural language processing, transformers, seq-to-seq, set-to-seq, large language models