RuBERT Embeddings in the Task of Classifying User Posts on a Social Media

Abstract

This paper presents models for solving the problem of multiclass classification of user posts in a social media. These models are based on embeddings extracted from messages using the RuBert language model and a fully connected neural network built over it. The models presented are compared to a baseline model using long-term short-term memory neurons (LSTM). The results will improve the accuracy of the classification posts, which in turn will improve the accuracy of assessing the psychological characteristics users.