Devising a combined method for analyzing the emotional coloring of texts

One of the key tasks of natural language processing (NLP) is to analyze the emotional coloring of text, which plays an important role in numerous application areas, including marketing, sociology, psychology, public opinion analysis and information security. Text sentiment analysis systems allow you to quickly obtain structured information about the emotional mood of society, predict reactions to certain events and identify potential threats or destructive content. Despite significant progress in the field of NLP, existing methods for determining emotional coloring of text have a number of limitations that reduce their effectiveness. In particular, traditional methods often do not take into account the contextual meaning of words, which is critical for accurate emotional coloration recognition. In addition, some methods have difficulty analyzing polysemous words, sarcasm, irony, and slang expressions. Therefore, it is an urgent task to further improve methods for analyzing text tone, in particular by combining several methods and using deep learning models. The aim of the study is to increase the accuracy of classifying the emotional coloring of natural language texts by using lexical, statistical, and contextual methods that will allow taking into account both superficial lexical features and deep semantic connections in the text. The combined method, which proposed in this paper combines statistical (TF-IDF) and contextual (BERT) vector representations of the text. This combination allows taking into account frequency patterns and deep semantic dependencies between words. The use of the Random Forest ensemble classifier allowed to build a robust model, that can effectively classify short English texts with a high level of accuracy. Experimental results have shown, that the proposed combined method has a higher classification accuracy (89 %) of texts compared to the basic – TF-IDF + RF and BERT + RF (78 % and 82 % respectively). The use of the combined method will increase the efficiency of context analysis and recognition of complex language structures, which makes it promising for analyzing public opinion in social networks, media, and chatbots; for use in customer support services; for determining the emotions of users of web services, websites and web applications.