Transformer-based multilabel classification for identifying hidden psychological conditions in online posts

Abstract

The paper proposes the method of multilabel classification for identifying hidden psychological conditions in online posts was proposed. The method consists of the stages of tokenization, neural network analysis of texts and the formation of conclusions about the presence of hidden psychological conditions. The features of the tokenization stage are the addition of special tokens to fix the boundaries of text fragments, supplement or trim the text to the length of a given dimension. At the stage of text analysis, the presence of each type of hidden psychological conditions is determined by a separate neural network model. The output of the method is the conclusion about the presence of each type of conditions with their numerical measures of manifestations. The created method allows to obtain in the models an improved ability to distinguish specific features for each type of psychological condition, due to training on modified sets of text data, which reduces the probability of confusion between conditions, since the model learns to distinguish their characteristic features. The developed method provides an average value 92.3% of the F1 metric for multilabel classification of hidden psychological conditions, while existing analogues provide an average value 64.5% of the F1 metric for multiclass classification.