OMOleksandr Marchenko

Oleksandr Marchenko

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Oleksandr Marchenko has a total of 36 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    RAG-BASED DIALOG-STATE CONVERSATIONAL RECOMMENDER FOR UKRAINIAN BOOKS

    by Maksym Shevchenko, Oleksandr Marchenko

    Finding books in digital catalogs through natural-language interaction requires a system to combine verifiable catalog constraints, less formal thematic preferences, and context accumulated across multiple conversational turns. The aim of this study is to develop and experimentally evaluate a retrieval-augmented conversational recommender with an explicit dialog state for Ukrainian-language book requests. The proposed system operates over a normalized research catalog of approximately 5,500 Ukrainian books and separates natural-language understanding (NLU), deterministic dialog policy, structured and vector retrieval, user-profile ranking, and grounded response generation. A strict intermediate frame distinguishes catalog entities from semantic preferences and preserves unresolved mentions, while the dialog manager maintains active constraints, focus, pending actions, feedback, and recommendation history. Depending on the request, the system selects structured, semantic, hybrid, similar-book, author, or profile-based retrieval. Evaluation was conducted at three complementary levels: 90 deterministic regression tests, standalone NLU assessment on 1,000 Ukrainian inputs, and 70 end-to-end (E2E) scenarios comprising 170 turns. All regression tests passed. The large language model (LLM)-based parser achieved 98.3% intent accuracy, 88.7% constraint F1, and 88.0% relaxed full-frame accuracy. In E2E evaluation, the two tested NLU configurations reached 90.59–92.94% relaxed frame accuracy, 92.94–96.47% dialog-manager subset accuracy, and 97.44% recommendation-assertion accuracy. Manual inspection of 185 recommendation-bearing outputs found no book-level mismatch among the lists actually returned. Neither configuration produced duplicate works within a recommendation list or repeated a work across turns of the same scenario. The results demonstrate that explicit state, deterministic retrieval control, and grounded generation can provide reproducible, context-aware book recommendations without delegating catalog membership or item selection to free-form generation. The architecture and evaluation protocol can be reused for other catalog-grounded conversational recommendation tasks.

  • Open Access

    ОПТИМІЗАЦІЯ НАВЧАННЯ ВЕЛИКИХ МОВНИХ МОДЕЛЕЙ ДЛЯ ЗАДАЧ СЕНТИМЕНТ АНАЛІЗУ ТЕКСТІВ УКРАЇНСЬКОЮ МОВОЮ

    by О. О. Марченко, Е. М. Насіров, В. Ю. Тарануха, Д. О. Волошенюк

    У даній роботі досліджується проблема оптимізації гіперпараметрів сучасних мовних моделей на основі трансформерної архітектури, зокрема DistilBERT, BERT, RoBERTa та їхніх похідних, у задачах обробки природної мови. Обґрунтовано доцільність виконання попереднього систематичного пошуку оптимальних значень гіперпараметрів перед початком навчання моделей, оскільки саме ці параметри визначають динаміку процесу оптимізації, стабільність збіжності та узагальнюючу здатність моделі. Показано, що використання стандартних або довільно обраних значень гіперпараметрів часто призводить до погіршення якості результатів, нестабільного процесу навчання та зниження відтворюваності експериментів. У межах роботи проаналізовано вплив ключових гіперпараметрів, зокрема швидкості навчання, коефіцієнта регуляризації, коефіцієнта розігріву, типу планувальника швидкості навчання та кількості епох тренування, на показники якості моделей. Особливу увагу приділено дослідженню поведінки моделей у разі зміни значень зазначених гіперпараметрів, а також виявленню оптимальних діапазонів їхніх значень. Експериментальні результати підтверджують, що некоректний вибір гіперпараметрів може призводити до недонавчання або перенавчання моделі, втрати збіжності алгоритму та суттєвого зниження таких метрик як точність і F1-міра. Водночас ретельний аналіз і обґрунтований добір гіперпараметрів дозволяють досягти стабільного процесу навчання, підвищити якість класифікації та ефективніше використовувати обчислювальні ресурси. Окремо наголошено на актуальності оптимізації гіперпараметрів для моделей, що працюють з україномовними текстами, оскільки обмежений обсяг якісних анотованих датасетів підсилює негативний вплив некоректних налаштувань. Отримані результати демонструють, що навіть без зміни архітектури моделі можна суттєво покращити її продуктивність шляхом належного тюнінгу гіперпараметрів. Практична цінність роботи полягає у формуванні рекомендацій щодо вибору гіперпараметрів для трансформерних моделей, що може бути використано в подальших дослідженнях і прикладних задачах аналізу тональності та класифікації текстів.

  • Tuning Bert-based Models for Sentiment Analysis of Texts

    by A. V. Anisimov, O. O. Marchenko, E. M. Nasirov, V. Y. Taranukha

  • TUNING BERT-BASED MODELS FOR SENTIMENT ANALYSIS

    by A.V. Anisimov, O.O. Marchenko, E.M. Nasirov, V.Y. Taranukha

    The paper explores the task of determining the emotional tone of Ukrainian texts using large language models. A comparative analysis of the performance of various BERT-based models is conducted. The impact of hyperparameter values on classification accuracy is examined. The paper proposes splitting texts into segments and compares different aggregation methods of representations to improve the accuracy of sentiment detection in long texts. Keywords: artificial intelligence, computational linguistics, neural network.

  • Open Access

    Modern Approaches to Controllable Emotional Speech Synthesis

    by Dmytro Ivashchenko, Олександр Олександрович Марченко, Дмитро Сергійович Іващенко

    The generation of emotionally expressive and controllable speech is one of the most dynamic and technically demanding areas in the intersection of artificial intelligence, natural language processing, and speech synthesis. Recent progress in emotional text-to-speech (TTS) systems has enabled increasingly natural and emotionally nuanced voice generation, shifting from early concatenative methods to advanced neural models. This review provides a structured overview of the state of the art in controllable emotional TTS, highlighting key architectural paradigms. A special focus is placed on emotional control mechanisms, including discrete emotional tagging with categorical or dimensional labels, reference-based control which conditions synthesis on prosodic or stylistic exemplars, and prompt-based techniques that leverage the capabilities of large language models for flexible and intuitive emotional specification.Despite substantial improvements in synthesis quality and emotional expressiveness, several critical challenges remain unresolved. These include the disentanglement of emotional, speaker, and prosodic features, the lack of standardized evaluation metrics for emotional clarity and naturalness, and the significant computational demands associated with training high-fidelity models. Furthermore, the scarcity of diverse and emotion-labeled speech data, especially for low-resource and morphologically rich languages, continues to limit the generalizability of current approaches. This review not only summarizes existing methods and their trade-offs but also outlines promising research directions, aiming to support the development of more robust, efficient, and emotionally expressive speech generation systems.


  • Comparative Analysis of Neural Network Models for Text Classification Problems

    by A. V. Anisimov, O. O. Marchenko, E. M. Nasirov, V. Y. Taranukha


  • TO FERRET OUT A GLOBAL METASCAFFOLDING STANDARD: METASCAFFOLD TAXOFRAME TEMPLATE CASE AS A PARADIGMATIC DEMONSTRATION OF THE APPROACH

    by Danylo Dvoichenkov, Oleksandr Marchenko

    The article is presented in the OnlineFirst format, which involves early posting of materials on the journal website before their final printing in paper form. The article has passed the review procedure and has been accepted for publication. The article is available for search and citation by its DOI, which remains unchanged. After the issue is printed in the printing house, a new citation can be made using information about the volume number and page number, using the same DOI reference. Background. The proliferation of Large Language Model (LLM)-based AI systems has led to widespread adoption of scaffolded reasoning frameworks that enhance model capabilities through structured methodologies including chained prompts, multi-agent coordination, and external tool integration. However, the ad-hoc nature of current scaffolding design processes creates significant challenges in interoperability, scalability, and reproducibility across diverse AI workflows. This paper introduces the MetaScaffold paradigm – a higher-level abstraction for structuring and orchestrating AI reasoning processes that provides reusable, systematic, domain-agnostic principles for organizing knowledge and reasoning steps. We present the MetaScaffold TaxoFrame Template (MSTFT) as a practical instantiation of this paradigm, specifically designed for taxonomic processing tasks, demonstrating how standardized meta-structures can improve AI system modularity and effectiveness while maintaining flexibility across different domains and applications. Methods. We formalize MetaScaffolds as structured orchestration mechanisms that dynamically integrate meta-level prompting, agent-based reasoning, self-reflective loops, and external knowledge retrieval through a central decision engine. The MSTFT framework implements this paradigm with core components including: (1) Meta-prompts and Recursive Meta-prompts for iterative refinement cycles, (2) modular agent systems for specialized reasoning capabilities, (3) memory and context management for semantic consistency, and (4) configurable model parameters for task optimization. Our implementation follows a structured procedure involving preprocessing and retrieval, meta-prompt assembly, dynamic model selection, and iterative reasoning with validation. We evaluated MSTFT against Single-Pass LLM (SP-LLM) and Iterative Prompting (IP) baselines across multiple domains (Software, Wood, Biology, Disease) using reference taxonomies constructed from WordNet hypernyms and ConceptNet derived terms, with evaluation focused on conceptual relevance, hierarchical coherence, instruction adherence, and generation efficiency. Results. Experimental validation demonstrates MSTFT's superior performance across all evaluation metrics. In the Software domain, MSTFT generated 80-90 well-structured and semantically distinct subconcepts compared to SP-LLM (~30 concepts) and IP (~50 concepts), which exhibited issues with redundancy and ambiguity. MSTFT consistently maintained three-level hierarchical depth as specified, while baseline methods produced flat or inconsistent structures. The framework achieved optimal balance between quality and cost efficiency (~$0.03 per taxonomy) compared to SP-LLM (~$0.01) and IP (~$0.10). Cross-domain evaluation confirmed MSTFT's generalizability, producing more differentiated and structured hierarchies across Wood, Biology, and Disease domains. Qualitative analysis revealed MSTFT's outputs demonstrated superior conceptual relevance, hierarchical coherence, and minimal logical errors, with approximately 25-30 valid subconcepts per level across three hierarchical levels. Conclusions. This work introduces the MetaScaffold paradigm and demonstrates its practical implementation through MSTFT, addressing critical gaps in existing AI scaffolding methodologies. MSTFT successfully integrates meta-level reasoning, iterative refinement, and multi-agent orchestration into a unified, interoperable framework that significantly improves taxonomy construction quality while maintaining computational efficiency. The framework's modular design, semantic validation mechanisms, and standardized approach provide a foundation for broader applications in AI reasoning tasks including ontology alignment, long-term planning, and multi-step question answering. Future research directions include enhanced semantic validation, multi-modal reasoning support, and expanded empirical evaluation across diverse domains and languages. While limitations exist regarding LLM dependencies and computational overhead, MSTFT establishes a promising paradigm for standardized, reusable AI scaffolding frameworks that balance flexibility with systematic structure.

  • Open Access

    Dialog style transfer with parameter-efficient fine-tuning

    by Ruslan Pravosud, Oleksandr Marchenko

    This work explores the application of modern fine-tuning methods, characterized by high parameter efficiency, to the task of generating dialogue responses in the style of a specific fictional or real-life character. The main goal of the study is to develop an approach that allows existing language models to be adapted to the specific communication or speech style of a particular personality using a minimal number of new parameters. This avoids the need for full retraining of large models and reduces computational costs. One of the key objectives is to enable the implementation of a compact yet functional language model that can be run on a standalone computer without requiring servers or powerful graphics processors. Such an approach opens new possibilities for personalized or entertainment applications, such as creating chatbots capable of interacting with users in the persona of a character from a book, movie, or even a historical figure. To implement this approach, we propose the use of the LoRA (Low-Rank Adaptation) method, which allows efficient fine-tuning of existing transformer architectures — in particular, the GPT-2 model — without retraining all of its layers. Using LoRA makes it possible to add stylistic features to the model’s speech without compromising its core ability to generate logical and coherent texts. To evaluate how well the model has learned to imitate the target speech style, we additionally train a BERT model as a classifier to distinguish between texts written in the desired style and ordinary, neutral responses. This provides an objective quality metric for generation and allows us to assess the effectiveness of the fine-tuning. The dataset used for training was generated using the GPT-4 mini model, which enables the rapid creation of a large number of examples in the desired style.