Comparison of transformer architecture neural network models based on evaluating the vector representation compactness of semantically similar texts in the European classification skills ESCO

In the process of analyzing short texts of the requirements of the Russian labor market, it was revealed that the same skills may have different formulations in natural language. In this regard, the search for a neural network model capable of effectively identifying semantically similar groups of texts of requirements for further formation of profiles of the skills and professions of the Russian labor market becomes an urgent task. The purpose of the study is to develop a method for evaluating neural network models built on the architecture of transformers by comparing the compactness of vector representations of semantically close short texts of skills of professions from the European classification (European Skills, Competencies, and Occupations). Materials and methods. The article provides an analysis for the original model of the European taxonomy of ESCO skills in English, and the texts of skills translated into Russian by the Yandex Translator and Google Translate automatic translation services. The article also provides a comparison of various methods for obtaining sentence attachments (cls, mean, pooling, Sentence Transformers) for various neural network models built on the transformer architecture. The results of the study show that with the help of the proposed method, it is possible to effectively implement the choice of neural network models for the task of searching for groups of semantically similar texts of requirements from online job texts. Conclusion. The proposed method made it possible to effectively select neural network models for the task of identifying compact groups of semantically similar texts of professional skills, which in turn will make it possible to identify groups of skills when forming profiles of professional skills, including semantically similar formulations, and profiles of entire professions. Such tools will allow you to quickly identify: key changes in the needs of the labor market at the level of individual competencies, will allow you to form an idea of the dynamics and sets of relevant competencies, will increase the effectiveness of management decisions to create digital literacy programs, retraining and advanced training, will allow you to assess competencies, will help all participants in the labor market to more accurately assess the existing trends, supply and demand in the labor market.