AUTOMATIC DOCUMENTATION AND MATHEMATICAL LINGUISTICS最新文献

LLMs are penetrating increasingly deeply into everyday use. However, the differences between the linguistic mechanisms of LLMs and natural human text processing processes have not well thoroughly investigated, which prevents LLMs from being used in the most efficient way. The aim of this paper is to analyze the linguistic functioning of LLMs and to determine the discourse specific characteristics of machine processing. In this paper, several methods are being applied, i.e., an experimental method, computer and linguistic analysis, as well as quantitative methods for the statistical analysis and the interpretation of linguistic text characteristics. The research materials include the Brown linguistic corpus of human texts and the corpora of artificially generated texts, using Claude Sonnet 3.7 and Grok-3. The paper finds differences in the mechanisms of language processing between neural networks and humans, i.e., the peculiarities of the use of discursive markers and characteristic linguistic parameters that LLMs do not reproduce.

The article considers the possibilities of optimization and modification of system and network software applications in order to prevent the impact of a cyber-attack of the man-in-the-middle type (MITM). The difficulties of recognizing and blocking such attacks, the features of their impact on mobile systems are described. Based on analytical reports of several well-known companies and systems, a statistical analysis of the MITM attack was performed with distribution by types of network data traffic against the background of their functional efficiency and direction of impact.

Based on the analysis of the functional capabilities and algorithmic solutions of several existing morphological analyzers of the Russian language and of the composition and structure of Russian dictionaries, a concept has been developed and design solutions for declarative tools for the next-generation MetaFraz morphological analyzer have been proposed. The concept envisions the creation of an optimized complex of declarative tools to ensure high performance and accuracy in processing text data. The key element is a balanced configuration of lexical resources combined with effective search algorithms. The use of modern data structures and algorithmic solutions allows for a significant increase in performance while maintaining high accuracy in morphological analysis. The proposed solutions form the basis for creating a scalable system capable of processing significant volumes of text information while meeting strict quality requirements for linguistic analysis.

The development of digital technologies and artificial intelligence is contributing to the evolution of decision support systems in medicine, where there is a growing interest in health recommendation systems (HRSs), which today requires an expansion of the knowledge base concerning the surrounding world to improve the explainability of the recommendations provided. In the paper, the potential of three ChatGPT models (GPT-3.5, GPT-4, and GPT-4o) for generating explanations of cardiovascular risk factors has been evaluated. The results of the analysis show that these models have significant potential to generate understandable, safe, and correct explanations of disease risk factors, which can form the basis for the development of promising HRSs to automate the preparation of medical reports and providing patients with explainable recommendations.

Phishing attacks are among the most common cyberattacks aimed at stealing user personal information. Machine learning methods have successfully been used to detect phishing attacks. One such method is fuzzy systems. The purpose of this article is to address the problem of spam detection using fuzzy classifiers optimized by the crow search algorithm. The efficiency of the constructed classifiers was tested with four data sets containing descriptions of phishing attacks. The obtained results were compared with the results of other classifiers. With comparable accuracy, fuzzy classifiers use fewer features.

The dynamics of Russian papers in the Web of Science database for the period 2008–2023 were studied. The results showed that the share of Russian papers on hard science in this international information platform has held steady at around 2% during this period. The number of Russian authors publishing annually in this field increased by 2.8 times from 2008 to 2023. Moreover, in the last two years (2022 and 2023), more than 150 thousand Russian researchers have annually become authors of papers in this field. The total number of authors (both Russian and non-Russian) of Russian papers in hard science in 2023 exceeded 320 thousand people. At the same time, on average, each year there were 1.7 non-Russian authors per Russian author in each paper. In the case of the SSs, this ratio was also in favor of non-Russian authors—1.5 non-Russian authors for every Russian author. It was only in the humanities and arts that this ratio was in favor of Russian authors: for every Russian author there were slightly more than 0.3 non-Russian authors.

We consider two well-known river crossing problems, one concerning jealous husbands and the other missionaries and cannibals. We show that a symmetry group action on the state set of the jealous husbands problem allow us to simplify it to the missionaries and cannibals problem. Then category theory is used to describe the relationship between the two problems. It is also noted that the missionaries and cannibals problem arose (more than a thousand years after the jealous husbands problem) precisely when the group approach began to be widely spread and popularized.

The development of an artificial intelligence-based language model for classifying English-language scientific articles by SRSTI codes is described. This improves the processes of reviewing and indexing scientific publications. A pre-processed dataset of scientific articles was used for training and testing the models. An architecture for cascade classification was developed, and the performance of models with various parameters was evaluated. As a result of this research, a console application for automatic classification of large flows of scientific articles was created.

As game services that require constant content updates to retain players, automating the generation of adaptive playable characters has become an urgent task. This article examines existing approaches to character generation, including evolutionary algorithms and in-session adaptation systems. Current solutions are limited by their inability to provide sufficient long-term adaptation to individual player styles and their reliance on manual design. To address these limitations, we propose a three-component system that integrates: player action modeling based on gameplay replays using reinforcement learning (RL) agents, character generation through combinatorial mechanics and parameter balancing, and automatic validation via simulations to assess balance and alignment with a player’s individual style. This work synthesizes contemporary research, highlighting the potential that generative methods have to reduce development costs for game services. The results could accelerate prototyping and enhancing the long-term viability of live-service projects.

This article is devoted to the study of users’ gaming experience and the formation of a cognitive model of their interaction with the new video game Koshcheiskie Prodelki (Koshchei’s Tricks), which is in an early stage of development. The user experience (UX) research methodology proposed by the authors and its results demonstrate how qualitative data on players’ perception, emotions, feelings, behavioral patterns, and achievements help optimize the process of creating video games, making them not only entertaining but also psychologically effective. The results have practical significance for indie developers who, with limited resources, strive to create high‑quality game products.