Political internet discourse: features of text analysis in the context of machine learning

Summary

The article analyzes the concept of discourse in the context of linguistic science and presents the structure and peculiarities of its interpretation, as well as possible misinterpretations that affect its functionality. In addition, the study highlights the identity of political activity (in particular, political (and Internet) discourse) in the context of machine learning tasks and prospects for possible study. In particular, the author focuses on the ontological and axiological ranking of politics as a phenomenon of human existence and highlights the features of political linguistics with its inherent anthropocentrism, expansionism, and functionality. The article also presents the peculiarities of analyzing the texts of political online discourse in the context of machine learning tasks: in particular, the specifics of representation of certain political events, parameterization of emotional support, etc. are highlighted.

Thus, the study of the peculiarities of analyzing political online discourse in the context of machine learning is relevant in the current international situation: in particular, the study of the originality of the process of analyzing the texts of the latter in the context of the specifics of the artificial neural network. In addition, such a study is of particular importance due to the hybrid nature of the russian-Ukrainian confrontation, which is characterized by several fake, distorted, actually false, and other data actualized in the media space.

The author emphasizes the inseparability of analyzing the texts of the above-mentioned discourse and combating the use of such false data (disinformation, misinformation, propaganda) constructed by the russian federation to destabilize the situation in Ukraine. The article proposes a solution to this issue through the development of machine learning algorithms capable of adapting to the fluid political online discourse and fully analyzing communication in it in all its forms and forms. Thus, the study of texts of political online discourse in the context of machine learning is directly related to the possibility of creating artificial neural network models capable of in-depth analysis.

Key words: discourse, political online discourse, analysis of online discourse, text analysis, machine learning, artificial neural networks.

Анотація

political internet discourse

Довгань О. Політичний інтернет-дискурс: особливості аналізу тексту в контексті машинного навчання

Стаття містить аналіз поняття дискурсу у контексті лінгвістичної науки, репрезентує структуру та особливості його тлумачення, а також можливі різночитання, які впливають на функційність. Окрім того, у дослідженні висвітлено самобутність політичної діяльності (зокрема, політичного (та інтернет-) дискурсу в контексті завдань машинного навчання та перспектив можливого вивчення. Зокрема, зосереджено увагу на онтологічному й аксіологічному ранжуванні політики як феномена людського буття та виокремлено ознаки політичної лінгвістики з притаманними для неї антропоцентризмом, експансіонізмом, функційністю. Також у статті репрезентовано особливості аналізу текстів політичного інтернет-дискур- су у контексті завдань машинного навчання: зокрема, висвітлено специфіку репрезентації тих чи тих політичних подій, параметризацію емоційного супроводу тощо.

Таким чином, дослідження особливостей аналізу політичного інтернет-дискурсу в контексті машинного навчання є актуальним в умовах сучасної міжнародної ситуації: зокрема, вивчення самобутності процесу аналізу текстів останнього у розрізі специфіки роботи штучної нейронної мережі. Окрім того, таке дослідження набуває особливої значущості ще й через гібридний характер російсько-українського протистояння, для якого притаманне побутування низки фейкових, викривлених, власне неправдивих та інших даних, актуалізованих у медіапросторі.

Автором підкреслено нерозривність аналізу текстів вищезазначеного дискурсу й боротьби з побутуванням таких неправдивих даних (дезинформації, мізінформації, пропаганди), сконструйованих російською федерацією з метою дестабілізації ситуації в Україні. У статті запропоновано вирішення означеного питання через розвиток алгоритмів машинного навчання, здатних до адаптації до плинного політичного інтернет-дискурсу, повноцінного аналізу комунікації у ньому у всіх видах та формах. Таким чином, вивчення текстів політичного інтернет-дискурсу в контексті машинного навчання безпосередньо пов'язано зі змогою створення штучних нейромережевих моделей, здатних до його ґрунтовного аналізу.

Ключові слова: дискурс, політичний інтернет-дис- курс, аналіз інтернет-дискурсу, текстовий аналіз, машинне навчання, штучні нейронні мережі.

A general statement of the problem and its connection with important scientific or practical tasks. Discourse (from the late Latin discursus - reasoning, proof) is a multivalent term that denotes speech in the general sense, as well as its use together with the system of concepts that it is conditioned by [1]. That is why the Cambridge Dictionary defines the above concept as “the use of language for communication in oral or written form” [2], while the Explanatory Dictionary of the Ukrainian Language positions it as “a coherent text, a superior unity” [3]. Therefore, within the framework of Wittgenstein's “language game” [4, p. 5], discourse can be positioned as a complex object, the nature of which is in the mutual projection of linguistic and behavioral rules.

At the same time, linguistic research is characterized by an original understanding of discourse: it is usually correlated with such concepts as speech, utterance, text, and the activity mediated by it (monologic/dialogic), represented by a communicative act. This produces an interpretation of the concept of discourse in the context of several phenomena with an empirical nature of existence and a simultaneous distortion of the understanding of its essence in terms of localized or fragmented representation of the latter.

Thus, the concept of discourse is positioned in linguistic science as a way of communication and data exchange mediated by textual structures. In turn, this marks its understanding as a socio-cultural product that is representative of the specifics of speech activity in a particular context. That is why, in our opinion, it is productive to interpret discourse as a text in the process of becoming before the interpreter's mental view in all the differentiated nature of this concept (we are talking about the complex structure of the latter: it consists of sentences or their fragments) [5]. In this sense, it is advisable to understand Internet discourse as all forms of communication on the Internet, represented by blogs, forums, e-mail, social networks, various messengers, etc.

Frequently, the content of the above discourse is concentrated around a kind of “cornerstone” - the core concept (or topical discourse), and the logical content of individual sentences (discursive components) is called propositions. The role of the latter is to build connections that ensure the functioning of logical relations within and the stability of their existence for the sake of the unity of all elements: disjunction - “or”, conjunction - “and”, implication - “if-then”, etc. The mechanism of realization of the aforementioned discourse is as follows: the interpreter (a native speaker of the language poly system and, accordingly, of the discourse) accumulates the generalized sense of elementary sentences while updating new data segments. The latter becomes a kind of “announcement” of the next information (the sentence that will be interpreted shortly) within the actualized information (the result of the previous, current, or intermediate interpretation).

In the context of the above, political Internet discourse is noteworthy as a special type of discourse, namely, the originality of the connection between politics and language, which is the communication basis of any activity. This means, first of all, that no political regime can exist without communication, i.e., the language polysystem in the political function. In this sense, politics is considered a certain communicative modus of human activity, in which the language poly system plays a functional role (relatively speaking, a toolkit for political action) because politicians actualize the language poly system for informing (propaganda: appeals, instructions, persuasion, etc.; lawmaking: creation, amendment, etc. of regulatory documents, etc.).

At the same time, the originality of political activity about other types lies in its predominantly discursive nature: in turn, it is associated with the linguistic essence of political actions. We are talking about the growing influence of communicative and semantic components on the existence of political processes, which, in turn, produces an increasing role of semiotic parameterization. Thus, the latter has become the basis for the representation of political activity as exclusively linguistic, and for modern political science, it has given grounds for considering language not as a representative of such activity, but as a component of the political field. The latter has naturally led to the possibility of a reverse perspective, in which politics is seen as a specific form of speech activity, realized by institutionalized speech acts of adapting ontological reality to words.

The above is connected with the ontological and axiological ranking of politics as a phenomenon of human existence: it is not without reason that the disclosure of its essence is represented in several special (sociological, philosophical, political science) dictionaries. Modern political linguistics is characterized by the following features: a) anthropocentrism (linguistic personality becomes a point of reference when studying phenomena);

b) expansionism (the linguistic field of research includes related issues, expanding it); c) functionality (language is studied in the process of everyday life, i.e. in action); d) explanatory (the desire to describe linguistic phenomena along with explaining the essence of what is happening) [6], etc.

We are talking about the aforementioned performativity (the speech essence of political actions) and autoreferentiality (selfpresentation, presentation of data about oneself) of institutional facts. The latter constructs a social ontology, causing the text to shape it under the guise of representing reality. This is in line with the approach to the language poly system formed in the twentieth century within linguistics and philosophy, in which language was considered not only as a tool for describing ontological reality but also as a mechanism and form of its construction. The latest russian disinformation and manipulative campaigns unfolding in the temporarily occupied territories are illustrative in this regard: In particular, a) monopolization of the religious space (delivery of propaganda literature to churches in Zaporizhzhia region); b) inspection of our fellow citizens' phones (checking Telegram channels subscribed by compatriots in Donetsk region)

c) the formation of pseudo-volunteer movements (enrollment of Ukrainian children in the “Eaglets of russia” - a children's division of the kremlin-funded organization in the Kherson region);

d) the creation of a network of free “district newspapers” (which are looking for local collaborators in Luhansk region to fill the position of chief editors) [7], etc.

Thus, the linguosemantic characteristics of the language poly system in the political function can be supplemented by linguocommunicative ones if we add another element to R. Jakobson's scheme - the political one. The latter can be positioned as an inverse magic function, within which an absent or inanimate “third party” will turn not only into the addressee but also the addressee of the message. At the same time, in the process of political communication, not only the addressee and the addressee are institutionalized but also the communicative process itself, which acquires the features of a semiotic counterpart of itself. The choice of the latter as the basis for the magic function is justified by similar processes of deployment. Similarly, by default, we expect that a speech act will cause a change in ontological reality, reconstructing it by the sense it contains. Therefore, its (the act's) participants are endowed with the appropriate power, the source of which is not myths, constructs, etc., but the social stratification of society.

Thus, the process of communication can be positioned as one of the pivotal ones for human ontology, since it is characterized by a special type of causal relationship with the language polysystem. We are talking about the actualization of several functions of the linguistic phenomenon: communicative, magical, identification, epistemological, thought-creating, and others. That is why, in the context of the above, political Internet discourse is the most representative of the way a society lives, and the specifics of the functioning of narratives in it, which is related to the universal nature of political communication. We are talking about the actualization of relevant spheres of human ontology within it, worldview role in terms of interests and target beliefs, etc. [8].

The main sources of studying political language, given the number of correlations between the latter and all forms of communication on the Internet (including the press, radio, television, etc.) and directly institutional (including leaflets, parliamentary debates, speeches at rallies, political party documents, etc.) discourse, are political Internet discourse in all its diversity of forms and types. The latter is a syncretic construct that arose as a result of the merger of two discourses - political discourse and Internet discourse itself, which led to the originality of its nature.

Analysis of recent research and publications on this topic, highlighting previously unresolved parts of the general problem to which this article is devoted. The problem of the peculiarities of analyzing political online discourse in the context of machine learning has a multilayered nature, which, first of all, produces its actualization in several interdisciplinary and integrated works. Thus, O. Bolkariova [8] argues that the analysis of political discourse (including political Internet discourse) is a new way of studying politics, which, in turn, creates the need for new tools (methods, criteria, models of text, and context analysis). We agree with the researcher: as stated above, machine learning and artificial neural networks as a means of solving the problem of new tools for studying political discourse (including political online discourse) are the solutions to the problem of new tools for studying political discourse.

The analysis of the linguistic and pragmatic features of pleonastic units formed based on the principle of redundancy in political speech is devoted to the work of B. Zayniddinovna [9]. The scientist highlights the specifics of this speech, studying in detail the originality of pleonastic combinations actualized within its framework. The researcher emphasizes that, despite the apparent redundancy and semantic repetition, pleonastic combinations of political speech have a specific methodological task from a pragmatic linguistic point of view.

The study of the possibilities of using artificial intelligence in the context of the development of Internet discourse is contained in A. Robertson, M. Maccarone [10], where the authors update the narrative analysis used for the main trends in the use of such intelligence. The researchers argue that narratives are critical because they resist the impulse of rapid and, in some cases, unlimited technological progress, and also offset apocalyptic narratives about artificial intelligence familiar from mass culture. The study contributes to the understanding of the socio-technical perceptions of a category of actors that have been insufficiently developed to date. In contrast, the work of G. Longo [5] is aimed at studying the correlations between Internet discourse, family, and related concepts, which allows us to explore how rethinking the Internet as a social institution produces a deeper understanding of the everyday life of this phenomenon.

The fundamental role of political Internet discourse in the context of social networks for all spheres of modern life is highlighted in the work of A. Chaves-Montero [11]. This study focuses on the analysis of tweets related to social services published on Twitter (currently - X) during the election campaigns in Spain in 2015-2019. In their work, the researchers used the Gephi tool to observe how these messages are spread on the aforementioned network. The results show that users use Twitter for informal communication during the election period, focusing on messages, condemnation, and positive evaluation to increase their visibility and influence.

Instead, a fundamental analysis of the socio-cultural features of comments under political news that turn into a public forum in the context of online discussions is presented in C. Kuchler et al. [12]. The study of the specifics of the political reaction to the prospect and reality of refugees entering their local communities on the example of comments in political online discourse using a structural and thematic model is presented in the work of C. Kelling, B. Monroe [13]. The results of the linguistic and stylistic analysis of German-language comments in the context of political Internet discourse are presented in D. Kuchina [14], in which the author analyzed and identified the distinctive features inherent in the above genre.

The work of G. Ramya [15] continues the study of commentary texts based on the study of online political and property data for their further moderation. The author highlights the features of training a neural network model using such algorithms as LSTM and BERT (Bidirectional Encoder Representations from Transformers), as well as GPT2 (Generative Pre-Trained Transformer 2), which helps in text generation by increasing the size of the data set for training in various classification models. The above produces the ability to train such a model in such a way that it can give accurate results during text classification, as well as in the process of identifying false/true comments contained in the data source.

The study of the role of the perception of threats from other cultures caused by terrorist attacks and criminal events in public discourse and support for right-wing radical parties is presented in the work of F. Giavazzi et al. [16] The authors propose to divide Twitter users (currently - X) into electoral districts in Germany and use machine learning to calculate the degree of textual similarity between the tweets they produce and the tweets of the accounts of the main German parties. The study of Internet discourse continues with the work of J. Govers et al. [17], in which the authors state that the influence of social networks is directly related to the problem of protecting the security of Internet users and freedom of expression on the Internet. The researchers propose and highlight a unique approach to the contextual analysis of hate speech, which integrates all other studies in this area.

The analysis of Internet discourse texts in the context of hate speech is studied in the work of M. Fazil et al. [18] in which the authors emphasize that social networks are a means of facilitating the process of real-time communication for users, while simultaneously producing problems of hate speech and false data. The researchers present an automatic method for detecting hateful messages, namely a convolutional BiLSTM neural network for classifying the latter. This neural network model updates existing methods for representing words in a multi-channel environment with multiple filters with unequal kernel sizes to localize semantic relationships in different windows.

The study by K. Bagchi, E. Banker, I. Ogunleye [19] is devoted to the representation of the work with the problem of false data mentioned in the previous work, the authors prove the falsehood of the assumption that there is no content moderation by Internet platforms. Thus, scientists have identified three current trends in the dissemination of such content: medical misinformation, hate speech, and child exploitation. In each of these areas, they highlight recent measures taken by policymakers to regulate how online platforms moderate these undesirable forms of content. They then detail actions that platforms are already taking in these areas to combat the challenges posed by online discourse.

The problem of hate speech in political Internet discourse continues to be studied by V Wanniarachchi et al. [20], who propose a methodological framework that uses a new mixed-methods approach to identify hate speech patterns from large text arrays collected from social networks. The authors actualize computer- mediated quantitative methods, including natural language processing (NLP) methods: sentiment analysis, emotion analysis, and topic modeling, as well as qualitative analysis of online discourse.

According to P. Kar and S. Debbarma [21], the main problem of text analysis for hate speech detection using machine learning is the reduction of the problem to a binary classification that does not take into account either the thematic focus or the target nature. Thus, since there is no unified annotated dataset and no scientific study that could give an idea of the relationship between the features of such a language, the available methods usually study only one or two of its features at a time. In addition, these methods are not effective for multilingual environments where most conversations are mixed, so scientists have proposed using a hybrid diagonal recurrent neural network (RNN) with diagonal shunting (FE-DGRNN) to detect hate speech and analyze sentiment in multilingual mixed-code texts. The experience of creating and training a transformational neural network model for analyzing emotions and sentiments in Hungarian political texts is presented in I. Uveges, O. Ring [22], in which the training data were manually annotated texts of parliamentary speeches from 2014 to 2018, rich in various emotions.

The work of Y Lupu et al. [6] classified the types of hate speech and hostility on mainstream and marginal online resources: the authors conducted a supervised machine learning analysis of 7 types of hate speech in online discourse on 6 interconnected online platforms. The researchers found that offline events, such as protests and elections, are often followed by an increase in the number of types of hate speech in online discourse that seem to have little to do with the main event. The urgency of localizing hate speech in political online discourse is evidenced by the study by F. Jafri et al. [23], in which the authors present a new dataset called IEHate, which contains 11,457 manually annotated Hindi tweets related to the Indian Assembly election campaign from November 1, 2021, to March 9, 2022. The above dataset, according to the authors, is a valuable resource for researchers and practitioners working to develop and evaluate methods for detecting hate speech in languages with limited resources.

One of the important components of machine learning (at the beginning) and natural language processing by an artificial neural network (later) with preliminary linguistic analysis is the aforementioned sentiment analysis, the features of which are presented in V Derbentsev et al. [24]. The authors argue for the relevance of the latter in the context of working with Internet discourse, presenting a comparison of deep learning models of deep artificial neural networks (convolutional (CNN), with long short-term memory (LSTM) layers and bidirectional LSTM with CNN layers (BiLSTM-CNN)) within their study. E. del Valle, L. de la Fuente [25] study political debates in online discourse (in particular, social media), in which the authors consider the above phenomenon in the context of its derivatives: first of all, hate speech. The scholars note that such derivative problems are of pivotal importance for social development due to the degree of their influence on public opinion and, accordingly, the communication strategies of political parties, state institutions, media corporations, and lobbyists.

The problem of an automated methodology for analyzing political online discourse is also addressed in the work ofE. Fournier- Tombs, G. di Marzo Serugendo [26], in which the authors used a quantitative approach based on machine learning to measure the quality of discourse in political discussions on the Internet. The researchers used the DelibAnalysis framework, which made their research methodology accessible, representative, and reproducible. This system updated a simplified version of the DQI to train a classifier, which was then used by the researchers to predict the discourse quality of any uncoded comment in a particular political discussion on the Internet.

The work of P. Tornberg [27] presents an innovative toolkit, namely, large language models (LLM), which the author positions as a universal method of text analysis in the social sciences. The latter, according to the scientist, is due to the simplicity, speed, accessibility, etc. of their actualization and wide possibilities for text analysis tasks: from annotation to classification of sentiment analysis and critical analysis of online discourse. The researcher in his paper discusses in detail the stages of the above analysis (software installation, API setup, data upload, development of analysis prompts, text analysis, and verification of results). The results and the course of the study are illustrated by the author's accomplishment of the difficult task of identifying populism in political texts.

Therefore, the study of the peculiarities of analyzing political online discourse in the context of machine learning is relevant in the current international situation: in particular, the study of the originality of the process of analyzing the latter's texts in the context of the specifics of the artificial neural network. In addition, such a study is of particular importance due to the hybrid nature of the Russian-Ukrainian confrontation, which is characterized by the existence of several fake, distorted, actually false, and other data actualized in the media space.

In turn, the existence of the above data (misinformation, propaganda, etc.) in political online discourse is productive due to the favorable nature of the latter. Thus, political Internet discourse is by its nature an environment in which various groups, individuals, and others express their position on certain events, and subjectively interpret them, thus influencing public opinion [18]. First of all, this raises the question of the authenticity (reliability, truthfulness, veracity, etc.) ofthe data that exists in this type ofdiscourse, becoming the core problem of its functioning and research. In turn, the latter requires the development of innovative tools for processing its data to identify unreliable segments and protect against manipulative paradigms of the enemy, which constitutes the unresolved tasks of the problem we are studying and the relevance of studying text analysis in the context of machine learning.

Formation of the purpose of the article (statement of the task). The purpose of the article is to consider the peculiarities of analyzing the texts of political Internet discourse as an object of research. The subject is the specifics of the above phenomenon in the context of machine learning and the work of artificial neural networks as an innovative tool of linguistic science.

Presentation of the main research material with full justification of the scientific results obtained. The above- mentioned is the pivotal role of the language poly system in the existence of ontological reality, within which it is a mechanism and form of constructing the latter. Thus, the modes of actualization of the linguistic poly system corresponding to various social functions naturally produce the creation of new types of reality, or rather, ideas about the latter. At the same time, it is obvious that the above-mentioned representations of reality (or ideas about reality) are axiologically determined: first of all, it means that such a representation does not exist outside the language poly system, and cannot carry a socially significant sense under such conditions.

In this context, political Internet discourse can be positioned as a special type of discourse precisely because, in our opinion, it includes all the others. In turn, this is due to the specific functionality of such discourse: for example, participants (voters, candidates, politicians, etc.) within the communication process actualize the language poly system mainly for some kind of influence, often manipulative [16]. Thus, the specificity of this construct allows us to talk about the productivity of using machine learning as a powerful tool for its analysis. However, the success of updating the latter naturally correlates with the thoroughness and quality of the preliminary linguistic analysis of the text, which must be done manually to update its algorithm in machine learning, and finally in the work of an artificial neural network.

It is advisable to take into account the following components inherited from the political discourse itself: a) subject matter (representation of the life of a group or person); b) original function (the core function is considered to be informational, while among others we can mention realization-forming (formation of specific political realities), protest (expression of disagreement with the actions of the current government and calls for its replacement) and magic (formation of belief in the fundamental “otherness” of the next candidate, party, etc;) c) emotional engagement and influence of texts (we are talking about the impact on public consciousness and political processes: It is actualized through the mediating role of political Internet discourse as a specific environment for interaction between political actors and the target audience (voters, politicians, etc.).

At the same time, such an influence can be both constructive (promoting discussion and understanding of complex political issues) and destructive (increasing political polarization and a sense of distrust in the political system as a whole through critical pathos, since one of the tasks of Internet texts is to assess the actions of the authorities, which produces the use of several linguistic cliches (cliches, terminology, etc.)). The above suggests that the analysis of political online discourse texts in the context of machine learning should be built on three possible levels:

1. Stylistic, which, along with the analyzed data (since the content aspect is pragmatic, it is a description of certain facts of ontological reality), also contains its axiological ranking. Thus, we are talking about assessment, modality, etc.: for example, reports on the course of the russian-Ukrainian war can be positioned as an attack by fascist extremists on the “friendly” russian military forces defending the representatives of the “russian world” and the national liberation struggle against the occupiers).

2. Manipulative (or rhetorical), in which a certain desire by the addressee's idea of it is imposed under the guise of information (the above-mentioned construction of ontological reality by the language poly system). In the context of this level, it is advisable to talk about the existence of both direct and indirect (disguised) false information: misinformation, disinformation, etc. For example, the diplomatic fake used in the video with the Commander-inChief of the Armed Forces of Ukraine Valeriy Zaluzhnyi illustrates the existence of such data quite vividly. In essence, they are something that has taken place in the process of description or representation, transformed in such a way as to relay Russian narratives. In this case, by narrative, we mean unsubstantiated stories that aim to reproduce a fundamentally different image of an event or an image of another event by the addressee [28].

The above, according to L. de Saussure, is a special type of actualization of the linguistic poly system, which can be determined by such grammatical characteristics as goal setting, intentions, etc. The latter, in turn, are linguistically formalized in the quantitative predominance of linguistic sense that serves for the correct realization of the addressee's goal [29, p. 119]. It is noteworthy that such levels (stylistic and manipulative) imply the existence of a certain reality autonomous from the linguistic poly system, while the role of the latter is reduced to a kind of formal design (seemingly objective or deliberately distorted), which is inherent in political Internet discourse.

1. Rajandran K., Lee C. Politics in Malaysia : A discourse perspective. Discursive approaches to politics in Malaysia : Legitimising governance. Singapore : Springer, 2023. P. 1-15. Open Acess : website. URL: https://goo.su/dxqjyV (date of application: 20.11.23).

2. Discourse. Cambridge Dictionary : website. URL: https://goo.su/ RGvpm (date of application: 20.11.23).

3. Дискурс. Горох : онлайн-бібліотека. URL: https://goo.su/OxlSp2 (дата звернення: 20.11.23).