The research problem is defined – the need for automating the narration and visualization of the complicated business processes for people with different types of sensory learning. The research methodology is created and discussed; a literature review is made to help find possible solution; diagram sources are chosen; a conceptual solution is discussed and proposed. A prototype experiment that uses LLMs is set up as proof of concept. The diagrams are analyzed using the prototype and the results are discussed. Conclusions on the applicability of this type of technology are made, and it is found that there is benefit from using the chosen or similar approaches for these purposes. Future directions for related research are proposed.
Graded logic is a human-centric continuum-valued propositional logic of natural human commonsense logical reasoning and decision making. It is a seamless generalization of the classical bivalent Boolean logic, fuzzy logic propositional calculus, and non-classical continuum-valued logics. This paper is an extended abstract which summarizes basic properties of Graded Logic and its applications in decision-support systems.
In this paper we suggest a flexible model for software support system that is able to automate data collection and analysis of academic and research management information. We follow on the experimental implementation, tested at the Plovdiv University to assess the strengths and challenges that face introduction of intelligent data processing and application of LLMs in support for higher education. Usability and quality of various support tools is also assessed, based on feedback and use statistics collected over the complete experiment period.
Strategic implementation of artificial intelligence (AI) in higher education (HE) is becoming essential as institutions seek to remain relevant and effective in a rapidly evolving digital landscape. The expansion of AI tools presents both transformative opportunities and significant challenges to conventional educational paradigms. This paper investigates how AI can be strategically integrated within HE by using the PDCA cycle as a part of decision support systems (DSS). Beginning with an overview of DSS, the study examines its practical applications in higher education institutions (HEIs) through different models for AI adoption. Using a bibliometric analysis of literature on "artificial intelligence", "strategy", and "university," and grounded in a conceptual framework developed from a systematic review, the research highlights key strategic priorities. These include the necessity of AI adoption to follow phased roadmaps and aligns with institutional mission, often requiring dedicated leadership, planning, and governance structures as well as institutional change management.
In the Bulgarian school education system, the use of electronic grade-books has become widespread, yet the full analytical potential of the data they collect remains largely untapped. Despite significant technological advancements, educational institutions still underutilize automated tools and artificial intelligence (AI) for educational diagnostics and administrative decision-making. This paper presents an approach based on Intuitionistic Fuzzy Evaluations (IFE) to assess student academic performance and detect early signs of learning difficulties.
Context-aware clustering of texts, particularly short texts, is a challenging task. Although metadata can provide valuable contextual cues to enhance clustering quality, such information is often incomplete or inconsistently available in real-world datasets. In this paper, we propose a novel clustering strategy that integrates both textual and metadata-based similarities, even when metadata is partially missing. Our method employs the Ordered Weighted Averaging (OWA) aggregator to fuse multiple similarity scores into a single aggregated value for each pair of texts. To handle missing metadata, we adapt the OWA mechanism by renormalising weights based only on available information, thereby avoiding potentially unreliable imputation or complete exclusion of certain metadata. We further introduce a confidence score that quantifies the reliability of each aggregated similarity, reflecting the proportion of missing metadata. Clustering is then performed using the K-Medoids algorithm on the resulting dissimilarity matrix. We demonstrate this approach on a real-world dataset of short Byzantine poems, where orthographic similarity is complemented with sparse metadata. The final clusters, stored in a graph database along with their confidence scores, enable meaningful interpretation and visualisation of the results, including the identification of uncertain cluster assignments due to missing contextual information.
Artificial Intelligence (AI) influences the practice of individuals and organizations, providing tools for higher efficiency and performance. While AI is helping software engineers in software design, some gaps exist for knowledge and document management. The paper presents recent research on knowledge management (KM) in software engineering (SE). It considers the role of KM in SE and the main tools used in software processes. Next, the authors present a model of knowledge domains in SE, on which base a concept is developed for an automated document management system integrating AI tools to facilitate end-user activities.
The role of context in database querying is explored, with a brief reference to its broader relevance. Context, pervasive in human activity, significantly influences intentions, preferences, interpretation, decision-making, and action. Despite its intuitive appeal, context remains a multifaceted concept requiring systematic investigation.
The novel context of Big Data has demonstrated that classical relational databases are not suitable: novel platforms for managing an incredible variety of datasets have become necessary, as demonstrated by the popularity of "data lakes" and "data lakehouses". One common issue of modern data platforms is to detect pairs of datasets that concern the same topic. However, a matching that is purely syntactic is not effective: the exploitation of modern AI techniques for Natural-Language Processing, such as word embedding and sentence embedding, promise to address the issue in a (more or less) semantic way. The contribution of the paper is a novel methodology (called "TopicRank") for flexible querying data platforms, so as to find out pairs of datasets that concern the same topic, on the basis of the textual description that accompany datasets as meta-data. The paper presents the results of a preliminary experiment that was conducted on a real pool of datasets.
The actualization of software test by artificial intelligence (AI) introduces transformative capabilities that improve coverage, reduce human error, and enhance efficiency. This paper explores the implications of applying AI in the context of automated test case generation, focusing particularly on the use of generative models to create both positive and negative test scenarios. For the experimental evaluation, a mock server was developed to simulate a real-world API environment, ensuring controlled testing conditions while maintaining realistic API behavior. The AI-generated test cases demonstrated increased defect detection and coverage while significantly reducing test creation time. Various real-world inspired scenarios were explored in the paper, validating the benefits of AI-driven test actualization. The results suggest that AI-assisted automation offers substantial advantages for modern software testing, particularly within DevOps practices and continuous integration pipelines.
Graph databases are powerful tools for representing complex, interconnected knowledge, but their structure and semantics pose significant challenges for natural language interaction. In this work, we present a guided approach based on the Retrieval-Augmented Generation (RAG) schema to assist large language models in generating queries over graph-based databases. By incorporating semantically relevant paths into the prompt, our method reduces the reasoning complexity that the model must handle, enabling more accurate query generation. We evaluate the performance of different language models under two schema interaction paradigms-full schema exploration and path-specific guidance-across queries related to risk assessment in passenger and flight data. Our preliminary results show that path-specific guidance significantly improves model performance, particularly for smaller models. That semantic prefiltering of the graph structure enhances the model's ability to focus on relevant information.
Improving the quality of education is a key priority for higher education institutions, with technological innovations playing a crucial role in this process. In the era of digitalization, universities are increasingly integrating modern technological solutions that facilitate communication between students, faculty, and administration. This study focuses on the automation of tutoring systems and the support of students in their interactions with academic mentors. A model of an intelligent tutoring system is proposed, which can be implemented in higher education institutions. The goal is not to replace academic mentors but to assist them in their work, making them more effective, informed, and engaged in supporting students. The system can be viewed as a virtual assistant that reduces the mentor's workload. The system is flexible and can be upgraded with new functionalities to meet the specific needs of different educational institutions and adapt to evolving educational requirements.
This paper focuses on applying reinforcement learning and parameter efficient fine-tuning methods to train NLP model to generate dialogue utters in a style of a specific character. The goal is to fine-tune relatively small model that will not consume a lot of computing resources and can be deployed and used on a single machine. As a result, this resource efficient method can be used for implementation of character-specific chat bots playing role of a specific character, generating fiction stories or new data sets. We propose using PPO algorithm together with LoRA to fine-tune GPT2 model to generate a character style utterances. For the reward function, BERT model was trained to distinguish between the desired style texts and regular ones, BERTScore and Self-BLUE were used to improve the dialogue flow quality. Dataset for training was generated with GPT4-mini.
This paper presents SmartNutritionDiabetes, a novel smart data model designed to semantically integrate food intake, nutritional goals and glycemic response for enhanced diabetes mellitus management. Developed in alignment with the Smart Data Models initiative and building upon existing frameworks such as SAREF4HEALTH and HL7 FHIR, the proposed model addresses a critical gap in representing nutrition-related data within chronic disease monitoring systems. By leveraging linked data principles and semantic interoperability, the model facilitates real-time, context-aware decision-making in healthcare environments. The model is implemented and tested in a simulated environment, where its performance is evaluated against traditional approaches. Results demonstrate significant improvements in data accessibility, semantic expressiveness, and cross-domain interoperability, making it a valuable foundation for next-generation personalized healthcare applications.
Automated SQL grading systems continue to face challenges in achieving high accuracy. While large language models (LLMs) offer new possibilities for semantic understanding, their probabilistic nature and occasional inconsistency limit their reliability for high-stakes, fully automated assessment, especially at scale. This paper proposes a hybrid grading framework that decomposes the SQL evaluation process into modular, interpretable stages–strategically combining traditional techniques, targeted LLM prompting, and a formal model of uncertainty. Rather than relying solely on LLMs for direct grading, we leverage their strengths in supporting subtasks such as test data generation and query equivalence suggestion–tasks where LLM errors are more easily identified and corrected. To adequately express partial correctness, we incorporate intuitionistic fuzzy sets (IFS) as a foundation, capturing distinct degrees of correctness and incorrectness. Our framework enables selective manual verification in cases with high uncertainty, while substantially reducing the overall human effort required. In contexts where full automation is not yet acceptable–such as high-stakes courses–the method provides a practical path toward scalable grading without compromising accuracy. This work aims not for marginal LLM accuracy gains, but for a robust, human-in-the-loop solution that balances automation with trustworthiness–paving the way for more scalable and interpretable educational technologies.
This article examines the improvement of information and analytical systems in local healthcare to ensure good storage, processing, and analysis of medical data, leading to better decision-making. The theory of relational databases, fuzzy logic, and framework networks form the work's methodological foundation. The creation of a new formal apparatus is made possible by the development of semantic models. A hierarchical structure for knowledge representation lowers computational resource consumption and increases the flexibility of the input data for inference. Unlike prior approaches, we provide a fuzzy inference method that makes it easier to locate starting points by using a database based on a hierarchical structure, which speeds up the final stage.
This paper presents a deep learning-based API designed for automated brain tumour classification from MRI scans, addressing the need for accessible diagnostic tools in clinical and resource-limited environments. Leveraging two state-of-the-art models, YOLO for real-time object detection and Roboflow for multi-label image classification, the study develops and evaluates an AI-powered diagnostic API implemented with FastAPI. The models were trained on a publicly available dataset containing glioma, meningioma, pituitary tumours, and non-tumorous images. Evaluation metrics include accuracy, validation accuracy, and confusion matrices. Roboflow achieved superior classification accuracy (96.1
This paper applies InterCriteria Analysis (ICrA), grounded in intuitionistic fuzzy logic, to investigate the systemic inconsistencies among key indicators related to youth classified as NEET (Not in Employment, Education, or Training) within the European Union. Using Eurostat data from 2002 to 2023, the study focuses on multidimensional parameters such as urban-rural disparities, gender differences, educational attainment, and labor market participation. The intuitionistic fuzzy approach enables the detection of logical confirmation (μ), contradiction (ν), and hesitation (π) among criteria, offering a nuanced semantic interpretation beyond linear correlations. The results show a predominant concentration of country-pair evaluations in the dissonance zones, with over 88
Ontology matching, particularly when applied to large and noisy web-based datasets, presents significant challenges related to both volume and veracity. In this paper, we introduce a novel hybrid methodology designed to address these complexities effectively. Our approach integrates traditional text matching techniques, string similarity metrics (Levenshtein distance and TF-IDF cosine similarity), Sentence-BERT embeddings, and a Siamese neural network based on LSTM architectures. We focus on the practical task of matching company names extracted from online job postings to formal company registries in Italy and Germany. To assess our methodology, we conducted extensive empirical experiments matching distinct company names from Lightcast job postings with entries from the Orbis dataset. After initial normalization procedures, we combined exact matching and similarity-based approaches, enhancing robustness through deep learning-driven embeddings from our Siamese network. Our evaluation shows that while traditional methods (particularly TF-IDF-based matching) performed reliably, the integration of deep learning provided additional discriminative power, especially beneficial in handling noisy, inconsistent naming conventions. Our proposed hybrid methodology significantly improves ontology matching accuracy, thus offering an effective solution to large-scale, real-world entity resolution tasks.
Knowledge graphs have emerged as a widely used method for representing and organizing information in a structured, graph-based form that facilitates understanding, navigation, and utilization by both machines and humans. The graphs model classes, entities and their relationships as nodes and edges in a graph. Natural logics are logics where sentences are expressed in a stylized form of natural language and where computational reasoning is conducted directly on the natural logic phrases, rather than on the underlying formal logic. NATURALOG is a dialect of natural logics that comes with a graph form, enabling knowledge in the knowledge base to be visualized and processed as graphs. NATURALOG thus offers an approach to natural logic graphs and additionally incorporates logical quantifiers, compound terms, and deductive reasoning, thereby going beyond the capabilities of traditional knowledge graph models. Here we specifically discuss the affinity between knowledge graphs and the natural logic graphs of NATURALOG.
