AN APPROACH TO THEMATIC CLASSIFICATION OF TEXTUAL INFORMATION BY MEANS OF NATURAL LANGUAGE PROCESSING

The present work proposes method of thematic classification of textual information, which improves the accuracy and relevance of the results of thematic analysis, contributing to more informed decisions based on textual data. The created method differs from the existing ones in the possibility of flexible definition of topics due to the application of thematic modeling, as well as in the expanded set of keywords due to the inclusion as keywords found with the help of LDA, as well as by adding to the keyword lists of target objects of the subject area.The method of thematic classification of textual information is intended for automatic determination of the number of topics and grouping of text documents according to their main topics, and differs from analogues in the possibility of flexible definition of topics due to use of thematic modeling, as well as the component of expanding keywords set due to inclusion in set of target volumes due to inclusion of target objects of subject area in the set.. The method transforms input data in the form of studied set of texts into output data in the form of classes (topics) number in sample, the identified dominant topic for each text, and presentation of each topic as extended keywords lists. The field of developed method application is thematic classification of text information, which is necessary, for example, when solving tasks of determining sentiment tone of texts in socially-oriented services, analyzing reviews in electronic commerce systems, detecting disinformation in news aggregators, etc. Improving the quality of thematic classification of textual information directly ensures increase in efficiency of solving these applied problems. This approach allows to efficiently organize text data large volumes, providing structured access to information. Thus, applying the proposed method to social media data containing many opinions, comments, and discussions can help in understanding the main topics discussed by users and identifying sentiments in the community. The method is implemented by software and investigated on English-language dataset. From thematic modeling in dataset, the optimal topics number was determined to be 14, and the cross-validation test yielded result of 0.86, which is improvement of 0.15 compared to use of LDA in its pure form for classification. Future scientific research will be focused on increasing topic classification percentage under conditions of uneven distribution of data in classes and researching topic modeling algorithms to improve proposed method. Keywords: text classification, machine learning, NLP, NER, LDA, keywords, target objects.