Statistical analysis of social network messages

Introduction. In recent years, the rapid expansion of digital technologies and online communication platforms has led to an unprecedented growth in textual data. As a result, text mining – also known as text analytics – has become one of the most important tools in both scientific research and real-world applications. A significant portion of human-generated information today is expressed in the form of unstructured text, including news articles, blogs, social media posts, comments, reviews, and official documents. This massive volume of text data contains valuable insights about human behavior, opinions, and trends, which can be discovered only through systematic analysis. The main objective of text mining is to extract meaningful and previously unknown knowledge from large collections of unstructured texts. It involves identifying hidden patterns, semantic relationships, and statistical regularities that can assist in decision-making processes, policy development, and predictive modeling. By integrating techniques from natural language processing (NLP), machine learning, and data mining, text mining allows researchers to automatically process and interpret huge text corpora that would be impossible to analyze manually. With the growth of social networks such as Telegram, Twitter, and YouTube, the importance of text mining has expanded even further. These platforms generate millions of usergenerated messages daily, making it crucial to develop algorithms capable of filtering, categorizing, and evaluating their content. Text mining methods help identify toxic, extremist, or harmful messages, detect sentiment polarity, and reveal communication trends within online communities. Therefore, text mining today serves as a bridge between linguistic research and computational analysis, providing valuable tools for understanding digital communication in modern society. It supports government agencies, media analysts, and academic researchers in discovering knowledge from text, ensuring that large-scale communication data can be effectively used for social, political, and technological insights. [1]. The statistical method converts text into digital form and analyzes it through mathematical and statistical calculations. The advantage of this approach is the ability to process large volumes of text quickly. The main methods are as follows: • Word Frequency – counting the most frequent words; • TF-IDF (Term Frequency – Inverse Document Frequency) – identifying important words in a corpus of texts; • N-gram analysis – studying the repetition system of word combinations; “Fan va texnologiyalar taraqqiyoti” Ilmiy-texnikaviy jurnal №1/2026 163 Computer science and information-communication systems • Clustering – grouping similar texts; • Classification – dividing texts into predefined categories [2–4]. Word Frequency and Zipf’s Law In studies related to word frequency, two main concepts are used: − Absolute frequency (how many times a word occurs in a text) − Relative frequency (the proportion of a word relative to the entire corpus). Based on frequency-based statistical observations, Zipf’s law plays a significant role in linguistics and text mining. It states that in any natural language corpus, the frequency of a word is inversely proportional to its rank in the frequency table. In other words, the most common words occur extremely frequently, while rare words appear only a few times. For example, in English, words such as “the,” “and,” “of,” and “to” dominate most texts, whereas words describing specific concepts, emotions, or events occur infrequently. The implications of Zipf’s law for text mining are profound. It suggests that a relatively small number of high-frequency words carry a large portion of the overall word occurrences, while a vast number of low-frequency words contribute to the richness and diversity of language. This insight allows researchers to compress and process large corpora efficiently by focusing on statistically significant terms while filtering out extremely common or extremely rare words that add noise rather than meaning. However, this efficiency comes with a limitation. Although statistical models like word frequency analysis and TF-IDF enable rapid and automatic processing of large amounts of text, they cannot fully capture the semantic depth, contextual meaning, or pragmatic intent behind words and phrases. For example, the same word may convey different meanings depending on context (“bank” as a financial institution or river edge), which pure frequency-based models cannot differentiate. [5– 6]. Recent studies (2023–2025) clearly demonstrate the effectiveness of this approach. For example, in [5], the method was effectively applied in various fields: in media discourse, word frequency was calculated for over 10,000 news headlines to identify the main topics. The results showed that technological terms such as “Big Tech,” “AI,” and “ChatGPT” were dominant, while social issues such as “jobs,” “diversity,” and “ethics” appeared in only about 6 percent of the material. In another field, [6] attempted to detect traces of LLMs (Large Language Models) in biomedical texts using word frequency analysis. By examining the frequency of certain stylistic units in more than 15 million scientific abstracts, it was concluded that at least 13.5 percent of papers written in 2024 might have been generated with the help of artificial intelligence. TF-IDF Method. The TF-IDF approach is one of the most widely used statistical techniques in text mining and information retrieval. Its primary goal is to identify and quantify how important a particular word is within a document relative to a collection of documents (the corpus). It does so by combining two distinct but complementary measures: Term Frequency (TF) and Inverse Document Frequency (IDF). The Term Frequency (TF) component measures how often a word appears in a single document. Words that occur more frequently within a document are considered more representative of that document’s topic. However, very common words such as “the,” “and,” or “is” often appear in nearly every text, and thus do not provide meaningful distinctions between documents. As a result, words that occur frequently in a document but rarely across the corpus receive a higher weight. The TF-IDF method combines local and global importance of words. Today, TF-IDF still retains its importance in text analysis, and despite the emergence of complex models, it remains a fundamental method due to its simplicity, speed, and interpretability [7–8]. For example, in a study by Chen et al. (2023), comments in medical forums were analyzed using TF-IDF, and the sentiment model achieved 9–12% higher accuracy [7]. In [8], 200,000 tweets about COVID-19 were analyzed using TF-IDF, and logistic regression and SVM models achieved Development of science and technology Scientific and technical journal №1/2026 164 Informatika va axborot-kommunikatsion tizimlar fast and resource-efficient results. Similarly, in [11], the TF-IDF method was used to automatically classify court decisions with 87% accuracy. N-gram. N-gram analysis in text mining examines not only individual words but also the sequential relationships between them. An N-gram represents a contiguous sequence of N linguistic units — which may be words, characters, or syllables — extracted from a text. For example, a bigram (N=2) consists of two consecutive words such as “artificial intelligence”, while a trigram (N=3) might be “text mining analysis”. The choice of N determines the level of contextual detail captured: smaller N-grams (e.g., unigrams, bigrams) represent local word co-occurrence, whereas larger N-grams (e.g., trigrams or 4-grams) capture more complex syntactic and semantic patterns. N-gram models play a central role in language modeling — estimating the probability of a word given its preceding context. This capability makes them foundational for many natural language processing (NLP) tasks, including machine translation, speech recognition, text classification, sentiment analysis, spam detection, and even neurolinguistic research. By examining frequent or rare N-grams, researchers can identify stylistic tendencies, domain-specific expressions, or emotional tones that are not apparent from single-word frequency analysis alone. In practical applications, N-gram analysis helps quantify contextual meaning and identify collocations — words that commonly occur together, forming stable linguistic units such as idioms or technical terms. For instance, the bigram “machine learning” or the trigram “deep neural network” provides far more meaningful information than analyzing “machine,” “learning,” or “deep” separately. This ability to capture contextual dependencies makes N-gram modeling an essential bridge between simple statistical approaches (like TF-IDF) and more sophisticated deep learning-based language models. [9–10]. N-gram analysis has been widely used as a main analytical tool in many studies. For example, [9] used bigram and trigram features to detect hate speech in social networks. Based on more than 50,000 Twitter posts, it was noted that when trigram features were added, accuracy increased by 6 percent, showing that N-gram expanded the context and improved semantic precision. Similarly, [10] analyzed large German newspaper archives using N-gram analysis to observe how key terms in political discourse changed between 2000–2020. The combinations of “Klimawandel” (climate change) bigram and trigram sharply increased since 2015 and were more often used with “Energiewende” (energy transition) in political texts. This demonstrates the effectiveness of this approach in studying historical discourse. N-gram also plays an important role in psycholinguistics. Clustering. In text mining, not only the frequency or statistical weight of words is important, but also the degree of similarity between documents. Understanding these relationships allows researchers to organize large text collections into meaningful groups and uncover hidden thematic structures. For this purpose, clustering methods are applied. Clustering represents an unsupervised learning approach, meaning that it groups documents or textual units based on similarity without relying on predefined categories or labels. Theoretically, clustering algorithms first represent each document as a vector in a highdimensional feature space. These feature vectors can be constructed using various text representations such as TF-IDF, Word2Vec, or BERT embeddings, depending on the desired balance between interpretability and semantic richness. Once the vector representations are obtained, clustering is performed using similarity metrics like cosine similarity, Euclidean distance, or Manhattan distance. These metrics measure how close or distant documents are in the multidimensional space, allowing similar texts to be grouped together and dissimilar ones separated into different clusters. Among the numerous clustering algorithms, several stand out as the most commonly used in text mining. The K-Means algorithm partitions data into a predefined number of clusters by minimizing intra-cluster variance. It is efficient and scalable, making it suitable for large-scale text datasets such as social media messages. Hierarchical clustering, on the other hand, builds a tree-like structure (dendrogram) that represents nested groupings of documents, enabling analysis at various “Fan va texnologiyalar taraqqiyoti” Ilmiy-texnikaviy jurnal №1/2026 165 Computer science and information-communication systems levels of granularity. DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is particularly useful for identifying clusters of arbitrary shape and handling outliers or “noise” texts that do not belong to any group. Gaussian Mixture Models (GMM) offer a probabilistic interpretation of clusters, assuming that data points are generated from a mixture of Gaussian distributions. Meanwhile, Spectral Clustering uses graph-based methods and eigenvalue decomposition to separate data based on similarity graphs, often yielding more flexible results for complex semantic patterns [11–12]. Clustering is widely used in research and practical projects. For example, in [11], hierarchical clustering was applied to a large news article archive, analyzing 120,000 articles from 2010–2022 and identifying political, economic, cultural, and sports topics as natural clusters. Hierarchical clustering provided higher semantic accuracy compared to K-means because its tree structure allowed the identification of smaller subtopics. In [12], the DBSCAN algorithm was used to detect plagiarism in academic papers. By analyzing 30,000 student works, suspiciously similar documents were identified through clustering, determining the degree of plagiarism. DBSCAN was noted to be more effective than traditional similarity metrics in detecting outlier documents. Classification. Classification is one of the most fundamental and widely used approaches in text mining and natural language processing. The main purpose of classification is to automatically assign each text document to one or more predefined categories based on its content. Theoretically, classification represents a supervised learning paradigm, meaning that it requires a labeled dataset — a collection of examples where the correct category (or “label”) for each text is already known. Using these examples, the model learns to recognize patterns and linguistic cues that distinguish one class from another, allowing it to predict the appropriate class for new, unseen texts. Text classification has a broad range of practical applications, including spam filtering, sentiment analysis, topic detection, fake news identification, and harmful content detection in social networks. Depending on the size, complexity, and nature of the dataset, different algorithms can be applied, ranging from traditional statistical models to advanced deep learning architectures [13–14]. Classification methods are effectively applied in various fields. In [13], several classification algorithms were compared to detect hate speech in social networks. On a dataset of 120,000 tweets, Logistic Regression and SVM performed well with TF-IDF, while the BERT-based classifier achieved the highest result with 92% accuracy. In [14], 500,000 articles were collected and classified using Naive Bayes, Random Forest, and Transformer models. The results showed that the Transformer-based RoBERTa model achieved 95% accuracy, while Naive Bayes remained effective in low-resource conditions. Conclusion. Statistical approaches — such as word frequency analysis, TF-IDF weighting, N-gram modeling, clustering, and classification — provide a robust foundation for the quantitative analysis of textual data. These methods are especially effective for exploring patterns within social network messages, where data are often massive in scale yet limited in semantic richness. By focusing on numerical relationships and probabilistic distributions, statistical models make it possible to identify trends, detect topic clusters, and classify messages according to their semantic or emotional tone with relatively low computational cost. Moreover, statistical techniques offer a high level of interpretability. Their outputs — frequencies, similarity measures, and cluster boundaries — are transparent and easily validated, which is crucial for research areas such as social network analysis and digital communication studies. In comparison to deep learning approaches, which often operate as “black boxes,” statistical methods remain valuable for preliminary text analysis, feature extraction, and for projects where explainability is a priority. Therefore, despite the growing popularity of transformer-based and neural models, statistical methods continue to play an important role in natural language processing. Their balance of speed, interpretability, and computational efficiency ensures their relevance in both academic research and applied data analysis of social media communications. Combining these methods with modern Development of science and technology Scientific and technical journal №1/2026 166 Informatika va axborot-kommunikatsion tizimlar machine learning or deep learning models can further enhance the precision and insightfulness of text analytics in the context of social network message evaluation