The complexity of text is a complex concept consisting of difficultness, readability and comprehensibility and describing the text structure. The determination of text complexity has applied significance in understanding and processing of information and knowledges. Subjective parameters of text include empirical data on the reader’s perception of the text, physical and cognitive abilities, knowledge and education of an individual. Objective parameters are divided into quantitative such as length, frequency of usage or number of tokens, and qualitative which are related to the analysis of linguistic means of categorical language levels and their implementation. The task becomes more complicated with the usage of the large text data. Defining text as a character sequence, the estimating model of complexity can be developed, the choice of the objective parameters, as well as methods of complexity estimation can vary; most of the formulas are universal and based on the linear-regression model. The goal of this paper is the development and implementation of software application in Python and the comparative analysis of basic formulas for English and adapted for the Russian. School textbooks on Social Studies, 5–11 classes (Russian Readability Corpus), make the test sample. The experiments with the text corpus data shows incorrect results what is explained by the fact that the model development based on the texts of different genres and styles and the difference in languages; in addition, the fact, that quantitative parameters may not be sufficient to obtain reliable results, should be taken into account when expanding corpus data.
Modern software produces a lot of events that can be analyzed using process mining techniques. The first step in any process mining pipeline is the collection of event logs. Then, those event logs need to be stored persistently on the disk to be transferred via local and global networks. The problem is that software event logs usually consist of many events, each of which can specify tens of attributes. In such a context, the event log stored in the conventional XML-based XES format consumes a tremendous amount of memory. Moreover, it is not read-friendly, i.e., does not provide tools with any advantage while reading an XES file. In this paper, we present bXES, a binary format for storing and transferring event logs, especially software event logs. We highlight main characteristics of software event logs which are utilized in the format scheme, and next we describe the new format. We developed open-source tools to convert XES logs to bXES and vice versa. Tools are available for C# and Rust ecosystems. Finally, we conduct experiments to demonstrate the bXES compatibility and evaluate it. Based on experimental data, bXES event logs are significantly more compact than conventional XES logs and can compete with logs stored in binary EXI format. This reduces storage volume needs and weakens communication constraints when transferring event logs. The experiments are conducted with data from real-life business processes and software events.
While the literature on digital transformation in education has searched for evidence based practices to improve ICT uptake in school settings, we know little about how schools differ in their approaches. This study aims to overcome the absence of standardized tools that could help to assess the stages and progress of ICT integration in educational settings. By using the example of information literacy development tasks assignment in classroom, we applied a latent class analysis to the survey data obtained from the monitoring the digital transformation of schools in the 2020–21 academic year. Based on the survey data from monitoring the digital transformation of schools, four types of students’ patterns were identified, depending on the information skills tasks assigned to them by their teachers at school. Based on the distribution of students’ patterns of working with information, three typical patterns of schools were identified with the use of multilevel latent class analysis. This study provides evidence for how the development of information literacy differs across schools contexts. As with advent of digital technologies education becomes data intensive domain, new approaches to the big data analysis are encouraged and it can help educators and education policy makers to improve decision-making.
An urgent task is to preserve and maintain the health of the country’s population, including through the promotion of a healthy lifestyle. Since social networks are very popular, especially among young people, it is possible to promote a healthy lifestyle on their basis. Despite the existing research on the influence of social networks on user behaviour, especially to alcohol consumption and smoking, no models are providing personalized recommendations for the user to involve in a healthy lifestyle and quit bad habits. The work aimed to research the young people’s social networks usage indicators and behaviour to a healthy lifestyle and the construction of personalized models to assess the need to change user behaviour. To achieve the aim, experimental research was conducted based on a survey of young people and an assessment of their profiles in social networks. An assessment and analysis of the existence of relationships between indicators of self-assessment of health, the presence of diseases, behaviour to a healthy lifestyle and the behaviour of users in social networks were completed. It was found that self-assessment of health and the presence of chronic diseases are not only interconnected with indicators of behaviour to a healthy lifestyle but also interrelated with respondents’ behaviour indicators in social networks. The theory of cognitive processes and cognitive load can explain these relationships. Based on the presence of interrelationships, regression models were built predicting users’ behaviour to a healthy lifestyle. Using such models embedding in social networks will allow issuing personalized recommendations.
This article discusses the issues of testing large language models. Large language models are the most popular form of generative machine learning models. The simple and clear usage model has led to their enormous popularity. However, like other machine learning models, large language models are susceptible to adversarial attacks. One could even say that the success of large language models has greatly increased interest in the security of machine learning models themselves. This direction immediately turned out to affect all users of machine learning systems. This article discusses the use of ready-made datasets for adversarial testing of large language models.
In this article, we propose an efficient implementation of operations on metagraphs by using an alternative definition of a metagraph. Definitions of the metagraph structures are given. An alternative definition of a metagraph is proposed. Metagraph operations definitions based on alternative definition of a metagraph are discussed. Operations on hierarchical metagraphs are proposed. The nesting binary relation is discussed. Elementary operations on metagraphs are proposed. Complex operations on metagraphs are discussed. An example of using operations over metagraph is given.
Modern empirical research in machine learning largely relies on developing custom software. Often such software is written by researchers and not professional software engineering. As a result, source code issues and the associated technical debt may accumulate and lead to higher programming effort, obstacles to code reuse, hidden software defects affecting the quality of the research itself. In this paper, we investigate if it is possible to apply automatic tools to prevent or remove these source code issues thus alleviating the need for software engineers in research projects. We analyze the source code of 24 open source research projects in machine learning, identify common issues and propose practical techniques to prevent these issues during coding. We also investigate if an application of an LLM coding assistant can fix common code issues automatically. We found out that 1) frequent source code issues largely the same for different machine learning frameworks 2) most of the issues could be eliminated by following simple coding practices 3) most of the issues could be removed by applying an LLM coding assistant.
We considered two fundamentally different approaches to real-bogus classification within the Zwicky Transient Facility survey data. The first approach is based on neural networks that take sequences of object images as input. The second approach uses features extracted from light curves and classical machine learning methods. Several models for both approaches were tested. Quality metrics were evaluated using k-fold cross-validation. We found that models based on classical machine learning algorithms outperform the neural network approach in both computational performance and quality. The code written during the study is available on https://github.com .
Artificial intelligence tools include software robots that can perform routine intensive processing tasks of data loading and control; conducting of analysis, detecting natural hazards in data streams and making decisions based on climate, forecast and observed data. Artificial intelligence can be used to better understand forecasts and predict the weather. The article presents a wide range of artificial intelligence applications that need to be developed in hydrometeorology. They are related to the collection, data search based on metadata and knowledge graphs, access to data, based on interaction with a chat-bot, forecast of hydrometeorological processes, possible impacts of natural hazards on enterprises and the population; training the population and leaders in behaviour during natural hazards.
The study explores methods for extracting argument relations from texts using large generative language models. Experiments were conducted on a Russian-language corpus of texts related to the field of scientific communication. Prompt-engineering methods were applied, with prompts developed using various tactics. The Mistral-7B was employed as the generative model. The task of extracting argumentative links was formulated as a binary classification problem of the existence/non-existence of a link between two statements. In constructing the dataset, the data were balanced. Positive examples included statements that were part of a single argument (premise, conclusion), while negative examples were generated from statements in the same paragraph for each positive example. Two methods of creating instructions were considered: using ChatGPT and an expert approach using the Chain-of-Thoughts tactic. The best solutions were obtained based on instructions composed by an expert and including context for each statement of one paragraph size. Instructions generated by ChatGPT, while producing comparable results, oftentimes returned incorrect responses. An experimental study was also conducted on an approach, in which the argumentation scheme is predicted immediately, allowing for more precise information about the type of relation to be included in the prompt. This task was also formulated as a binary classification problem. The two most frequent schemes in the examined corpora, “Expert Opinion” and “Example”, were explored.
In recent years, there has been a steady increase in research intensity and complexity in information security, driven by rising cyberattack sophistication. This growth highlights the need for greater accessibility, interoperability, and reusability of research data and methodologies. Addressing these needs requires robust infrastructures that integrate data, tools, and methods while ensuring confidentiality and integrity. This article examines the development of specialized cybersecurity research infrastructures aligned with FAIR (Findable, Accessible, Interoperable, and Reusable) principles. It identifies key challenges, discusses practical solutions, and emphasizes the importance of interdisciplinary integration with research domains like bioinformatics, sociology, and machine learning. Establishing such interconnected infrastructures can accelerate innovation and enhance the effectiveness of cybersecurity research outcomes.
The paper considers the problem of topic modeling and evaluation of topic models, presented in marked up sets of text messages, based on the Word2vec word vector representation model. Clusters constructed as a result of the word vectors analysis can be used for various tasks, including diagnostics of the topic model presented in the marked up collection of text messages. For this purpose, it was proposed to calculate the intersection matrix between the dictionary clusters formed for the entire text corpus and the individual dictionaries of topic subsets in the corpus. The paper presents and discusses the results of a machine experiment with a collection of news messages of one of the regional online media. The results of the experiment demonstrated the feasibility of potential diagnostics for the existing system of topic categories in a collection of text messages and determining the possible directions of its reorganization.
This research examines modern methods for automating information verification, specifically focusing on narrative texts containing dated content. We introduce three new techniques—CHECK-S, CHECK-V, and CHECK-U—for analyzing texts, along with a new approach to contrastive learning, “Hierarchical Contrastive Learning,” which has been evaluated in competitive environments. The findings demonstrate significant improvements over traditional methods, confirming the potential of these techniques in enhancing automated information verification for narrative texts.
The Glagolitic script, one of the oldest known Slavic scripts, presents a substantial challenge for historical manuscript decryption due to its intricate glyph forms and limited existing digital resources. This paper introduces a novel dataset of Glagolitic letters aimed at facilitating the application of machine learning algorithms in the decipherment of historical documents. The dataset creation process comprised several critical stages: collection of raw data, preparation of images, application of neural networks for letter extraction, clustering of images, training of models to discern noise, and manual validation and annotation of rare letters. The resultant dataset stands as the first publicly accessible Glagolitic script resource tailored for deep learning applications in historical document analysis.
The paper considers the task of automatic expansion of the empirical states set by non-identical states from the collection in W@DIS information system. The relevance of this task is caused by long terms of preparation of new versions of empirical states, as a result of which new collections data cannot be automatically evaluated. The content and structure of collections of states and transitions for the H2O molecule are discussed. An algorithm for filtering data sources from collections and some rules for creating a set of unique quasi-empirical states are described, as well as the effects of filtering collections of transitions and states on their quality. As a result of such filtering, the number of inconsistent identical states and transitions in the collections was significantly reduced.
The article is devoted to the problem of constructing a semantic library of resources on mathematics and mathematical physics based on classical encyclopedias. The mechanism of integration of encyclopedias into the content of the semantic library and the unification of the mathematical encyclopedia edited by academician I.M. Vinogradov and the encyclopedia of mathematical physics edited by academician L.D. Faddeev are investigated. During the integration process, intersections of multiple articles in these encyclopedias are discovered, as well as mutual enrichment of descriptions of their terms. The library’s tools made it possible to form a knowledge graph into which both encyclopedias were integrated. Thanks to the ontological approach, the knowledge graph of the semantic library is saturated with new nodes and links, which in turn leads to the enrichment of the subject areas of the semantic library itself and the subject areas of integrated scientific publications. The library search is accompanied by navigation based on a knowledge graph, which allows you to rely on reliable information from classic encyclopedic sources. The work is addressed to specialists in the field of semantic modeling of scientific subject areas.
Terminologies of specialized problem domains present an important part of knowledge to be extracted for various applications, such as construction of thesauri, ontologies, glossaries and so on. Meanwhile, widely-used automatic term extraction (ATE) methods are mainly statistics-based and show quite average quality, so ways to leverage modern deep learning techniques are currently studied. The paper addresses the task of term recognition based on BERT classifier of term candidates previously extracted from text; cross-domain settings are considered for training BERT models. The dataset constructed for experiments is presented, which contains samples taken from scientific texts in Russian. The results of the experiments with cross-domain term recognition are described, demonstrating comparable or slightly better quality than the most known ATE methods.
Neural networks have become an integral part of everyday life, finding applications in various domestic and industrial tasks. Generative models based on the Transformer architecture play a particularly significant role in natural language processing. These models have achieved, and in some cases surpassed, human-level performance in several tasks. However, despite their high performance, generative models can sometimes produce unexpected results. Understanding the principles behind the decisions of such models is an important and relevant challenge. In this article, we investigate how effectively the T5 model explains its answers in classification tasks. We also compare its interpretative capabilities with those of the BERT model using well-known interpretation methods such as SHAP, LIME, and the attention mechanism.
The paper describes an approach to developing a machine learning ontology, based on the methodology for constructing ontologies of scientific subject domains, developed in A.P. Ershov Institute of Informatics Systems. A brief overview of the basic concepts and terms of machine learning (ML) and known developed ontologies related to this field is given. The paper also provides a brief description of the methodology for constructing ontologies of scientific subject domains, and describes the ontology design patterns developed within the framework of this methodology to represent the basic concepts of the ML subject domain. The developed ML ontology will be used to build an intelligent scientific Internet resource on machine learning that will provide content-based access to systematized knowledge and data in the field of ML, helping users in choosing methods, models and data sets necessary to solve their practical problems.
Keyphrase selection plays a pivotal role within the domain of scholarly texts, facilitating efficient information retrieval, summarization, and indexing. In this work, we explored how to apply fine-tuned generative transformer-based models to the specific task of keyphrase selection within Russian scientific texts. We experimented with four distinct generative models, such as ruT5, ruGPT, mT5, and mBART, and evaluated their performance in both in-domain and cross-domain settings. The experiments were conducted on the texts of Russian scientific abstracts from four domains: mathematics computer science, history, medicine, and linguistics. The use of generative models, namely mBART, led to gains in in-domain performance (up to 4.9
