Tamas Orosz
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Tamas Orosz has a total of 55 co-authors. Here are the top co-authors by number of shared publications:
A. Selmeci
Óbudai Egyetem AMK, Székesfehérvár, Hungary
8 shared publications- Easin Arafat
Eötvös Loránd University
7 shared publications Istvan Orosz
Óbuda University/AMK, Székesfehérvár, Hungary
7 shared publications- Attila Márton Putnoki
ELTE Eötvös Loránd University, Budapest, Hungary
5 shared publications - A. Jablonski
Institute of Physical Chemistry, Polish Academy of Sciences
3 shared publications - Beata Lesiak
Institute of Physical Chemistry, Polish Academy of Sciences, Warsaw, Poland
3 shared publications
Latest Publications
- Open Access
- 24 September 2026
- Published by Springer Nature in Machine Learning
- Vol. 115 (10) , 229
SentiMatrix: Parameter-Efficient Fine-Tuning of Encoder-Based Transformers for Multidimensional Sentiment Analysis
Transformer-based architectures have established state-of-the-art benchmarks across Natural Language Processing (NLP) tasks; however, the computational overhead of full fine-tuning remains a significant barrier to scalable deployment. This paper presents SentiMatrix, a systematic evaluation of Low-Rank Adaptation (LoRA)-based Parameter-Efficient Fine-Tuning (PEFT) across four sentiment analysis paradigms: Intent-based, Aspect-based, Fine-grained, and Emotion detection, spanning seven benchmark datasets under a consistent three-stage protocol: inference using task-specific pretrained baseline models (i.e., HuggingFace checkpoints previously fine-tuned on the corresponding benchmarks, serving as informed upper-bound references rather than cold-start zero-shot baselines), Full Fine-Tuning (FFT) as a performance upper bound, and LoRA-based adaptation to quantify the efficiency-performance trade-off. Adaptive Low-Rank Adaptation (AdaLoRA) is benchmarked as a direct comparator. LoRA achieves competitive performance while reducing trainable parameters by up to 99.8%, with lower GPU memory usage and training time in most evaluated settings. It achieves 93.28% accuracy on SST-2, 80.74% on combined Laptop+Restaurant ABSA, 67.63% on E-Commerce 5-class, and 93.22% on emotion detection, while closely approaching FFT on E-Commerce 5-class (68.47%). The results demonstrate an effective efficiency-performance trade-off across diverse sentiment analysis tasks. Ablation experiments examine the effects of adaptation strategy, model capacity, and label granularity. The source code and datasets are available at https://github.com/ELTE-DSED/senti-matrix .
- 24 September 2026
- Published by Springer Nature in Machine Learning
- Vol. 115 (10) , 229
Open AccessSentiMatrix: Parameter-Efficient Fine-Tuning of Encoder-Based Transformers for Multidimensional Sentiment Analysis
Transformer-based architectures have established state-of-the-art benchmarks across Natural Language Processing (NLP) tasks; however, the computational overhead of full fine-tuning remains a significant barrier to scalable deployment. This paper presents SentiMatrix, a systematic evaluation of Low-Rank Adaptation (LoRA)-based Parameter-Efficient Fine-Tuning (PEFT) across four sentiment analysis paradigms: Intent-based, Aspect-based, Fine-grained, and Emotion detection, spanning seven benchmark datasets under a consistent three-stage protocol: inference using task-specific pretrained baseline models (i.e., HuggingFace checkpoints previously fine-tuned on the corresponding benchmarks, serving as informed upper-bound references rather than cold-start zero-shot baselines), Full Fine-Tuning (FFT) as a performance upper bound, and LoRA-based adaptation to quantify the efficiency-performance trade-off. Adaptive Low-Rank Adaptation (AdaLoRA) is benchmarked as a direct comparator. LoRA achieves competitive performance while reducing trainable parameters by up to 99.8%, with lower GPU memory usage and training time in most evaluated settings. It achieves 93.28% accuracy on SST-2, 80.74% on combined Laptop+Restaurant ABSA, 67.63% on E-Commerce 5-class, and 93.22% on emotion detection, while closely approaching FFT on E-Commerce 5-class (68.47%). The results demonstrate an effective efficiency-performance trade-off across diverse sentiment analysis tasks. Ablation experiments examine the effects of adaptation strategy, model capacity, and label granularity. The source code and datasets are available at https://github.com/ELTE-DSED/senti-matrix .
- 21 September 2026
- Published by Springer Nature
- p. 442-456
A Controller-Mediated Multi-agent LLM Framework for Real-Time Digital Banking Workflow Execution
- 21 September 2026
- Published by Springer Nature
- p. 442-456
A Controller-Mediated Multi-agent LLM Framework for Real-Time Digital Banking Workflow Execution
- Open Access
- 6 August 2026
- Published by Szechenyi Istvan University in Acta Technica Jaurinensis
- Vol. 19 (3) , 206-218
Cognitive Information Systems and Visual Interfaces: Enhancing Decisions through Personalization
The purpose of this study is to evaluate how to integrate personalization, cognitive resonance, and science into the design of dashboards to support more efficient decision-making processes and to decrease cognitive load on humans when using Decision-Support Systems (DSS). To accomplish these goals, we used a combination of survey data collected from professional decision makers along with some examples of various dashboards, the study evaluates how user-tailored interfaces; the use of color; and scientifically supported principles of visualizations affect the usability, and engagement of users with dashboards. Findings suggest that personalization aligns dashboards with user-specific needs, supporting clearer and more consistent perceived decision-making, while potentially strengthening cognitive resonance within Cognitive Information Systems (CIS) through effective Human-Computer Interaction (HCI). Furthermore, our findings indicate that scientifically valid dashboards will provide the most effective way to optimize the user experience and foster trust as well as the ability to accommodate varying levels of complexity in different and fast changing operational environments.
- 6 August 2026
- Published by Szechenyi Istvan University in Acta Technica Jaurinensis
- Vol. 19 (3) , 206-218
Open AccessCognitive Information Systems and Visual Interfaces: Enhancing Decisions through Personalization
The purpose of this study is to evaluate how to integrate personalization, cognitive resonance, and science into the design of dashboards to support more efficient decision-making processes and to decrease cognitive load on humans when using Decision-Support Systems (DSS). To accomplish these goals, we used a combination of survey data collected from professional decision makers along with some examples of various dashboards, the study evaluates how user-tailored interfaces; the use of color; and scientifically supported principles of visualizations affect the usability, and engagement of users with dashboards. Findings suggest that personalization aligns dashboards with user-specific needs, supporting clearer and more consistent perceived decision-making, while potentially strengthening cognitive resonance within Cognitive Information Systems (CIS) through effective Human-Computer Interaction (HCI). Furthermore, our findings indicate that scientifically valid dashboards will provide the most effective way to optimize the user experience and foster trust as well as the ability to accommodate varying levels of complexity in different and fast changing operational environments.
- Open Access
- 18 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 125-131
Causal Rule Discovery for Compliance Improvement in Procurement Logs: A Case Study
- 18 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 125-131
Open AccessCausal Rule Discovery for Compliance Improvement in Procurement Logs: A Case Study
- Open Access
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 42214-42223
Decision Support in Cognitive Information Systems: Carbon–Silicon Infocommunication
This paper examines the role of Cognitive Information Systems (CIS) in enterprise decision support, emphasizing Carbon-Silicon agent infocommunication. Cognitive Information Systems enhances decision efficiency by reducing cognitive load and optimizing information processing in complex business scenarios. The study explores cognitive infocommunication, which enables synchronized, adaptive interactions between human (Carbon) and machine (Silicon) agents, supporting context-aware decision-making across operational, tactical, and strategic levels. Cognitive resonance is a crucial factor in ensuring seamless information exchange and strategic alignment and fosters synergy between Carbon and Silicon agents. The paper introduces a structured approach to modeling Carbon-Silicon agent interactions, focusing on cognitive infocommunication principles. A survey of decision-makers validates key aspects of Cognitive Information Systems, reinforcing the importance of semantic alignment, adaptive learning, and personalized decision support.With the dynamic communication adaptable to the evolving business environment, Cognitive Information Systems supports the scalability of the enterprise, ensures the enablement of enterprise optimization, and promotes innovations.
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 42214-42223
Open AccessDecision Support in Cognitive Information Systems: Carbon–Silicon Infocommunication
This paper examines the role of Cognitive Information Systems (CIS) in enterprise decision support, emphasizing Carbon-Silicon agent infocommunication. Cognitive Information Systems enhances decision efficiency by reducing cognitive load and optimizing information processing in complex business scenarios. The study explores cognitive infocommunication, which enables synchronized, adaptive interactions between human (Carbon) and machine (Silicon) agents, supporting context-aware decision-making across operational, tactical, and strategic levels. Cognitive resonance is a crucial factor in ensuring seamless information exchange and strategic alignment and fosters synergy between Carbon and Silicon agents. The paper introduces a structured approach to modeling Carbon-Silicon agent interactions, focusing on cognitive infocommunication principles. A survey of decision-makers validates key aspects of Cognitive Information Systems, reinforcing the importance of semantic alignment, adaptive learning, and personalized decision support.With the dynamic communication adaptable to the evolving business environment, Cognitive Information Systems supports the scalability of the enterprise, ensures the enablement of enterprise optimization, and promotes innovations.
- Open Access
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 42794-42808
Cognitive Information System Alignment-Aware Decision-Support Framework
In dynamic business environments, effective decision-making depends not only on computational performance but also on the alignment between system-generated recommendations and the cognitive processes of decision-makers. While Business Intelligence and Artificial Intelligence enhance data-driven analysis within Cognitive Information Systems, insufficient attention has been given to formally modelling this alignment dimension. This paper introduces the Cognitive Information System Alignment-Aware Decision-Support Framework, a structured multi-layer architecture that explicitly separates normative evaluation, explicit preference representation, and alignment assessment.Within this architecture, Cognitive Resonance is formalised as a diagnostic alignment construct that quantifies the correspondence between recommended decision alternatives and decision-maker preferences. The formulation preserves analytical independence between ranking logic and alignment measurement, thereby preventing conflation of decision quality with preference correspondence. An illustrative simulation-based use case demonstrates the structural feasibility of embedding the alignment model into a dashboard-based Cognitive Information Systems environment and integrating normative scoring with an optional temporal learning component. The results show that ranking outcomes and alignment values may vary independently under changing evaluation configurations, highlighting the diagnostic value of explicit alignment modelling. The proposed architecture contributes a configurable and interpretable alignment-aware layer that complements existing multi-criteria and learning-based decision-support approaches, providing a structured foundation for preference-sensitive and transparent decision-making in organisational contexts.
- 17 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 42794-42808
Open AccessCognitive Information System Alignment-Aware Decision-Support Framework
In dynamic business environments, effective decision-making depends not only on computational performance but also on the alignment between system-generated recommendations and the cognitive processes of decision-makers. While Business Intelligence and Artificial Intelligence enhance data-driven analysis within Cognitive Information Systems, insufficient attention has been given to formally modelling this alignment dimension. This paper introduces the Cognitive Information System Alignment-Aware Decision-Support Framework, a structured multi-layer architecture that explicitly separates normative evaluation, explicit preference representation, and alignment assessment.Within this architecture, Cognitive Resonance is formalised as a diagnostic alignment construct that quantifies the correspondence between recommended decision alternatives and decision-maker preferences. The formulation preserves analytical independence between ranking logic and alignment measurement, thereby preventing conflation of decision quality with preference correspondence. An illustrative simulation-based use case demonstrates the structural feasibility of embedding the alignment model into a dashboard-based Cognitive Information Systems environment and integrating normative scoring with an optional temporal learning component. The results show that ranking outcomes and alignment values may vary independently under changing evaluation configurations, highlighting the diagnostic value of explicit alignment modelling. The proposed architecture contributes a configurable and interpretable alignment-aware layer that complements existing multi-criteria and learning-based decision-support approaches, providing a structured foundation for preference-sensitive and transparent decision-making in organisational contexts.
- Open Access
- 5 February 2026
- Published by Szechenyi Istvan University in Acta Technica Jaurinensis
- Vol. 19 (2) , 93-102
Integrating Explainable AI into Model-Driven and Low-Code Enterprise Applications
Explainable artificial intelligence has become increasingly important in enterprise settings, as organisations require transparent and trustworthy decision-support tools. At the same time, low-code and model-driven platforms are widely adopted for building business applications, yet their high level of abstraction often hides the reasoning behind automated recommendations. This study examines how explainability can be systematically incorporated into such environments by introducing a modular approach that separates predictive functions from the generation of human-interpretable explanations. The proposed concept builds on an external reasoning layer that provides both predictive outputs and concise, user-oriented justifications through a unified interface, allowing enterprise systems to present explanations without modifying existing development workflows. To demonstrate the feasibility of the approach, the study applies it to a representative enterprise scenario involving personalised recommendations. The proof-of-concept implementation shows that explanations can be delivered in real time and integrated seamlessly into standard business user interfaces. The results highlight that the proposed solution can enhance transparency, support user trust, and increase the adoption of data-driven features in low-code and model-driven applications. The study contributes a practical architectural pattern that can serve as a foundation for future explainable enterprise systems and provides initial evidence that explanation services can operate effectively alongside contemporary development paradigms.
- 5 February 2026
- Published by Szechenyi Istvan University in Acta Technica Jaurinensis
- Vol. 19 (2) , 93-102
Open AccessIntegrating Explainable AI into Model-Driven and Low-Code Enterprise Applications
Explainable artificial intelligence has become increasingly important in enterprise settings, as organisations require transparent and trustworthy decision-support tools. At the same time, low-code and model-driven platforms are widely adopted for building business applications, yet their high level of abstraction often hides the reasoning behind automated recommendations. This study examines how explainability can be systematically incorporated into such environments by introducing a modular approach that separates predictive functions from the generation of human-interpretable explanations. The proposed concept builds on an external reasoning layer that provides both predictive outputs and concise, user-oriented justifications through a unified interface, allowing enterprise systems to present explanations without modifying existing development workflows. To demonstrate the feasibility of the approach, the study applies it to a representative enterprise scenario involving personalised recommendations. The proof-of-concept implementation shows that explanations can be delivered in real time and integrated seamlessly into standard business user interfaces. The results highlight that the proposed solution can enhance transparency, support user trust, and increase the adoption of data-driven features in low-code and model-driven applications. The study contributes a practical architectural pattern that can serve as a foundation for future explainable enterprise systems and provides initial evidence that explanation services can operate effectively alongside contemporary development paradigms.
- Open Access
- 28 January 2026
- Published by Springer Nature in Artificial Intelligence Review
- Vol. 59 (3)
Cognitive and artificial intelligence evaluation framework
The Cognitive and Artificial Intelligence Evaluation (CAIE) framework provides a structured and domain-independent methodology for assessing the intelligence of artificial and information systems in a broader perspective. The primary achievement of this research is the categorization of over ninety cognitive features into six evaluation zones, supported by a two-stage scoring model that combines detailed feature-level analysis with higher-level structural interpretation. This approach has proven effective in identifying system maturity and developmental potential, offering systematic insights into both strengths and weaknesses across cognitive domains. The practical validation through use-case analysis demonstrates that CAIE is adaptable to diverse technological contexts, enabling consistent comparison between AI and non-AI systems. By treating cognitive features as measurable and comparable attributes, the framework introduces a coherent mechanism for benchmarking, scalability, and strategic development. The main contribution of this work lies in advancing both academic research and real-world implementation by delivering a cognitively informed, practically relevant tool that bridges theoretical evaluation concepts with actionable methods for designing and improving intelligent systems.
- 28 January 2026
- Published by Springer Nature in Artificial Intelligence Review
- Vol. 59 (3)
Open AccessCognitive and artificial intelligence evaluation framework
The Cognitive and Artificial Intelligence Evaluation (CAIE) framework provides a structured and domain-independent methodology for assessing the intelligence of artificial and information systems in a broader perspective. The primary achievement of this research is the categorization of over ninety cognitive features into six evaluation zones, supported by a two-stage scoring model that combines detailed feature-level analysis with higher-level structural interpretation. This approach has proven effective in identifying system maturity and developmental potential, offering systematic insights into both strengths and weaknesses across cognitive domains. The practical validation through use-case analysis demonstrates that CAIE is adaptable to diverse technological contexts, enabling consistent comparison between AI and non-AI systems. By treating cognitive features as measurable and comparable attributes, the framework introduces a coherent mechanism for benchmarking, scalability, and strategic development. The main contribution of this work lies in advancing both academic research and real-world implementation by delivering a cognitively informed, practically relevant tool that bridges theoretical evaluation concepts with actionable methods for designing and improving intelligent systems.
- Open Access
- 1 January 2026
- Published by Wiley in International Journal of Intelligent Systems
- Vol. 2026 (1)
DSIQ‐GEN: Automatic Generation and Classification of Data Science Interview Questions
Domain‐specific question generation and classification systems are increasingly needed in technical fields like data science for education, recruitment, and adaptive learning. However, existing large language models lack domain‐specific knowledge and provide limited control over question attributes. This study develops a comprehensive framework for automatic generation and classification of data science interview questions using parameter‐efficient fine‐tuning (PEFT) methods. We evaluate three PEFT approaches, Low‐Rank Adaptation (LoRA), P‐tuning, and Prefix tuning, across 105 checkpoints on Llama‐3.2‐1B. LoRA achieves optimal performance with 85% diversity, 87.1% domain relevance, and only 0.9% trainable parameters. Conditional generation expands the 167‐question corpus to 1011 questions (505.4% increase), achieving balanced difficulty distribution and reducing topic standard deviation from 6.9% to 2.5% in the fully merged dataset. Classification experiments show Transformers achieve 87.7% difficulty and 96.7% topic accuracy on extended data, with keyword tokenization (198 tokens) matching full vocabulary (357 tokens) while reducing training time by approximately 15%. Unsupervised clustering validates quality with 77.3% purity (+8.1 points). An independent human evaluation ( n = 104 generated questions, 3 raters) corroborates these automatic metrics, yielding substantial inter‐rater agreement (weighted κ = 0.602) and directly addressing the risk of circular validation inherent in classifier‐only evaluation pipelines. Critical findings reveal conditioning fails below 8%–10% class representation, and topic semantic overlap reduces effectiveness. This research demonstrates PEFT as a practical solution for synthetic data generation in specialized domains under data scarcity constraints.
- 1 January 2026
- Published by Wiley in International Journal of Intelligent Systems
- Vol. 2026 (1)
Open AccessDSIQ‐GEN: Automatic Generation and Classification of Data Science Interview Questions
Domain‐specific question generation and classification systems are increasingly needed in technical fields like data science for education, recruitment, and adaptive learning. However, existing large language models lack domain‐specific knowledge and provide limited control over question attributes. This study develops a comprehensive framework for automatic generation and classification of data science interview questions using parameter‐efficient fine‐tuning (PEFT) methods. We evaluate three PEFT approaches, Low‐Rank Adaptation (LoRA), P‐tuning, and Prefix tuning, across 105 checkpoints on Llama‐3.2‐1B. LoRA achieves optimal performance with 85% diversity, 87.1% domain relevance, and only 0.9% trainable parameters. Conditional generation expands the 167‐question corpus to 1011 questions (505.4% increase), achieving balanced difficulty distribution and reducing topic standard deviation from 6.9% to 2.5% in the fully merged dataset. Classification experiments show Transformers achieve 87.7% difficulty and 96.7% topic accuracy on extended data, with keyword tokenization (198 tokens) matching full vocabulary (357 tokens) while reducing training time by approximately 15%. Unsupervised clustering validates quality with 77.3% purity (+8.1 points). An independent human evaluation ( n = 104 generated questions, 3 raters) corroborates these automatic metrics, yielding substantial inter‐rater agreement (weighted κ = 0.602) and directly addressing the risk of circular validation inherent in classifier‐only evaluation pipelines. Critical findings reveal conditioning fails below 8%–10% class representation, and topic semantic overlap reduces effectiveness. This research demonstrates PEFT as a practical solution for synthetic data generation in specialized domains under data scarcity constraints.