Neural networks with lstm and gru in application to the task of multiclass classification of text posts of social network users

Abstract

This article discusses two deep learning neural network architectures - long-term short-term memory (LSTM) and closed recurrent units (GRU). These models are proposed to be applied to the problem of multiclass classification of posts of users of social networks to improve the accuracy of automation of the assessment of the severity of psychological functions of users. The aim of the study is to improve the quality of the multiclass classification of user posts through the development and implementation of new models of the second level of the hierarchical classifier. The theoretical significance of the study lies in the construction of new accurate models of class definitions, which will form the basis of models for assessing the severity of personal users. The practical significance lies in improving the automated post classification system, which will complement the existing prototype of the program for analyzing user security. The novelty of the result lies in the creation of a new method for solving the urgent problem of automated classification of posts, which makes it possible to achieve greater accuracy in relation to the existing method earlier. The best classification result was shown by a model based on the LSTM architecture (F1-micro 0.766, F1-macro 0.734, accuracy 0.793).