Universal Sentence Encoder for Russian

This master's thesis describes the development of a universal semantic text encoder for the Russian language. The primary objective of the study was to create a model that surpasses existing analogs in the quality of semantic text encoding. The work includes a review of existing text encoding methods and the data used for this purpose, as well as their validation methods. Approaches based on BERT models, loss functions, the use of instructions, and task-specific fine-tuning were considered. Models were developed and trained using collected and generated datasets, including both positive and negative text pairs. Special attention was paid to creating hard negative pairs and validating the models on various tasks such as semantic similarity assessment, text classification, and relevant information retrieval. As a result, a model was created that shows improved performance on target metrics compared to its counterparts, as confirmed by testing on several benchmarks.