Currently, we are witnessing many information wars being waged on the world stage, with a special emphasis on the war in Ukraine. Russia uses information operations as a key element of its hybrid war against Ukraine and directs significant resources to the creation and dissemination of disinformation, the purpose of which is to influence public opinion, undermine trust in the Ukrainian state, its institutions and leaders, as well as to distort the real state of affairs in the international community
In the context of increasing information attacks from Russia, the importance of an accurate and objective analysis of the text sentiment of related to various aspects of the war and its impact on various spheres of life becomes extremely relevant. Automated analysis of the emotional coloring of texts can help detect attempts to manipulate public opinion, assess the general mood in society regarding certain events or policies, and identify changes in the information space that may indicate new directions of information attacks or disinformation campaigns.
The proposed method integrates advanced technologies of natural language processing (NLP), machine learning and linguistic analysis, which allows not only to automate the process of determining the emotional coloring of texts, but also to ensure high accuracy and objectivity of the obtained results. The application of such a complex approach in the conditions of information warfare opens up new opportunities for monitoring the information space, identifying and analyzing information operations, as well as for developing effective strategies for countering disinformation. Given the high dynamics of the information space and the constant change in the tactics and strategies of information warfare, the development and implementation of innovative text analysis methods that allow for prompt monitoring and analysis of attempts at manipulation and disinformation is extremely important.
Further materials are presented according to the following structure. In Chapter 2 the analysis of recent related works is fulfilled and in Chapter 3, the proposed method is described. Chapter 4 is dedicated to implementation and case studies, and Chapter 5 summarizes the received outcomes.
The study of text tonality, which involves the integration of natural language processing (NLP), machine learning, and linguistic analysis, occupies an important place in modern scientific research. In [1], the analysis tool for text sentiment was developed based on the fusion of machine learning algorithms, demonstrating the effectiveness of combining different methods to more accurately determine the emotional coloring of texts. Authors [2] emphasized the importance of using TF-IDF and machine learning methods for sentiment analysis of Twitter data, confirming the importance of these techniques in the field of big data processing. In [3], the impact of the lexical richness of the training corpus on the performance of machine learning in analysis tasks was highlighted, emphasizing the need for a careful approach to data selection and processing. Authors [4] presented an innovative approach to the classification of hierarchical comments using BERT and a specialized naive Bayes classifier, opening up new opportunities for advanced text analysis. In [5], a deep learning-based approach to sentiment analysis is proposed for ranking online products using a set of probabilistic linguistic terms, which demonstrates the potential of using these technologies for commercial purposes. Authors [6] developed a technique for analyzing the sentiment of data from Twitter using NLP and machine learning, which emphasizes the importance of an integrated approach to information processing. The authors [7] demonstrate how sentiment analysis on Twitter can be improved using NLP and machine learning methods for emotion categorization and trend visualization. In [8], the advantages of integrating linguistic analysis for the study of human language are revealed, which opens up new opportunities for accurate thematic categorization. Research [9] focuses on the analysis of online product reviews using advanced deep learning and machine learning techniques [23,24] to improve data classification and extract detailed emotion information [21].
Compared to the analogue [10], which focuses on the sentiment analysis of Twitter data, this work is distinguished by a deep integration of natural language processing (NLP) techniques, machine learning and linguistic analysis. In addition, a distinctive feature of the proposed approach is the development of specialized algorithms for accurate determination of sentiment taking into account contextual semantics and lexical diversity within specific thematic headings. This allows not only to more accurately classify the emotional coloring of texts, but also to detect subtle nuances in emotional expressions, which makes the proposed approach more adaptable to the complexities of natural language and provides a higher accuracy of analysis compared to existing methods.
The proposed method for determining the sentiment of text by thematic headings is a comprehensive approach that integrates natural language processing (NLP), machine learning, and linguistic analysis for automatic classification of text data. The method allows to evaluate the emotional color of the text (positive, neutral or negative) and to express it quantitatively in the form of an expanded percentage scale in the range from -100% to +100%. Let us present this method as a set of the following stages and structurally (Fig. 1):
Stage 1. Text pre-processing [11]:
where P is the number of positive words, N is the number of negative words, Q is the number of neutral words.
The generalized client-server architecture of the text tonality analysis system is presented in Fig. 2. The client initiates the interaction by sending a request to the server, which in turn processes the received information. After processing, the server sends the necessary information back to the client. This process is cyclic, so after transferring information, the client can initiate interaction with the server again. The diagram shows a typical request-response model that is fundamental to a client-server architecture, where the server acts as a provider of resources and services and the client acts as a consumer of those resources and services. To implement the proposed method, a set of data was collected from a wide range of texts of articles from Internet sites of Ukraine, which ensure the representativeness of various styles, genres and topics. The sample covers a variety of emotional contexts to test the algorithm's ability to accurately classify emotional nuances. The analysis of the text will be assigned to one of the following thematic headings, which were determined by experts from the security service of state bodies: Socio-economic, political, military situation in the Russian Federation (attitudes towards the top management, mobilization, deterioration of the economic situation, etc.).
The main functionality of the web interface (Fig. 3) [15] allows users to insert url links of news sites for analysis. After submitting the text, the system uses an algorithm to determine the emotional color, displaying the results in the form of an understandable report. The report includes quantitative indicators of the presence of positive and negative words, as well as an overall text tone index.
As a result of the analysis of the tonality of the text [16], the rubric "Military and political leadership of Ukraine at all levels" was determined, where a positive tone of +80% was indicated. This reflects a high level of positive emotional coloring of the text of the article. Details are given about military aspects, including the acquisition or use of military equipment (the F-16 and Bayraktar are mentioned), and an optimistic attitude towards events related to the war in Ukraine is reflected.
The inversion function (Fig. 4) of the results was included for users who wish to analyze the opposite emotional tone of the text. This can be useful, for example, when analyzing texts containing sarcasm or irony, where the literal meaning of the words can be misleading. The inversion allows you to quickly see how the overall assessment of emotional coloring will change if these stylistic figures are taken into account.
One of the most important aspects of the web interface (Fig. 5a) is the ability to edit keywords for each heading. Users can add new words to dictionaries, remove existing ones, which affects their significance in the analysis algorithm. This allows you to adapt the system to the specific needs of the user and increase the accuracy of determining the tonality for specific topics or writing styles.
In addition, the web interface includes a dictionary management module (Fig. 5c), which allows users to view and update the database of words on the basis of which the analysis is performed. This is especially important to take into account linguistic updates, social and cultural changes that can affect language and its emotional load. a -key words b -dictionary of emotions Figure 5: Example of editing keywords and dictionary A comparative analysis was conducted to assess the effectiveness of the system for determining the tonality of texts (100 news stories) by comparing the automated results of the system with experts' assessments. The analysis took into account such parameters as the assignment of the text to the appropriate rubric by the system and experts, the comparison of the tonality of the texts determined by both the system and experts, and the correspondence of these assessments. Using the URL as a unique identifier for each text ensured accurate tracking of results. In addition to the quantitative evaluations, the experts provided additional notes for a deeper understanding of the reasons for the discrepancies between the expert evaluations and the results of the system, which will contribute to the further improvement of its accuracy and reliability.
Based on the results of filling out Table 1, statistical indicators were calculated, namely the simple percentage of matches between the system and experts.
The obtained results (see Table 1) indicate a fairly high efficiency of the system in terms of classification of texts by headings with a correspondence of 92%. This means that in most cases the system correctly identifies the thematic category of the text, which indicates its reliability in determining the context of news. However, the average tonality deviation of 15% is quite significant and may indicate some shortcomings in the work of the algorithms of the system for evaluating the emotional coloring of the text. This may be due to the incompleteness of the dictionaries used for sentiment analysis. However, dictionaries can be constantly updated, which is a significant advantage of the system, since the language is constantly evolving, and the context and use of the vocabulary can change. The ability to supplement dictionaries by users allows the system to adapt more quickly to novelties in the language and changes in the use of terms, especially in the field of news, where it is important to take into account not only the lexical meaning of words, but also their connotative influence. Thus, the system demonstrates a high accuracy in the classification of thematic headings, but needs improvement in determining tonality. Constant addition and updating of dictionaries, with the possibility of making changes from users, is an important process for increasing the accuracy of the analysis of the emotional coloring of texts.
A method for determining the text sentiment by thematic rubrics is proposed based on an integrated approach that integrates natural language processing, machine learning and linguistic analysis for automatic classification of text data. This allows you to assess the emotional color of the text (positive, neutral or negative) and express it quantitatively in the form of a percentage scale from -100% to +100%.
The text sentiment analysis system implemented based on the method has a high accuracy (92%) of classification of thematic headings. This means that in most cases the system correctly identifies the thematic category of the text, which indicates its reliability in determining the context of news.
However, the average sentiment deviation of 15% is quite significant and may indicate some shortcomings in the work of the algorithms of the system for evaluating
Average tonality deviation 15% the emotional coloring of the text. This may be due to the incompleteness of the dictionaries used for sentiment analysis.
In the future, authors are going to explore the methods [17][18][19][20] for improving the quality and performance of text sentiment analysis.