The Society of Convergence Knowledge Transactions ISSN:2287-8920(Print) 융복합지식학회논문지

This paper proposes a method to improve abstractive summarization performance by incorporating the structural characteristics of news articles. Since the CNN/DailyMail dataset does not provide sentence-level 5W1H structural labels, this study automatically generated news sentence labels by applying 5W1H-based prompting to GPT-5 mini. The generated labels were not treated as manually annotated gold labels, but were used as weakly supervised pseudo-labels to train a BERT-BiGRU-CRF sequence labeling model. The trained model classifies news sentences into seven categories: LEAD, EVENT, REACTION, OUTLOOK, CONSEQUENCE, QUOTE, and NOISE. Sentences-assigned core labels were then selected and used as inputs for abstractive summarization models, including BART, BERTSUMABS, and T5. Experimental results show that the sequence labeling model achieved an accuracy of 73.6% and an F1-score of 72.2% on the GPT-5 mini based pseudo-labels. In the summarization experiments, the LEQ and LEQC methods generally outperformed the baseline models. In particular, LEQC+BART achieved the highest performance, with ROUGE-1, ROUGE-2, and ROUGE-L scores of 47.0%, 25.6%, and 36.8%, respectively. These results indicate that selecting structurally important sentences is effective in improving abstractive summarization performance.

Emotional speech data is difficult to collect in sufficient quantities due to privacy and security constraints, and its construction requires significant cost owing to the need for expert annotators. Furthermore, the absence of standardized test datasets for validating speech quality evaluation metrics makes it difficult to quantitatively assess the reliability of existing metrics. To address both of these challenges simultaneously, this paper proposes QC-MetricGAN, a MetricGAN-based emotional speech augmentation framework with controlled quality-level generation. QC-MetricGAN defines five discrete quality levels based on target STOI values, and trains the Generator to produce speech at each target quality level by applying level-wise differentiated quality degradation inputs and an asymmetric loss function. In addition, an adaptive training strategy is employed to detect the Generator's structural tendency toward over-enhancement and automatically strengthen the input quality degradation configuration, thereby guiding stable convergence toward the target quality level. Experiments conducted on four datasets — EMO-DB, RAVDESS, and ESD (English and Chinese, evaluated independently) demonstrate that 113 out of 125 experimental trials (90.4%) satisfy the convergence criterion and a consistent trend is observed in which larger dataset scales lead to improved convergence rates and generation stability. The multi-quality datasets generated by the proposed framework can serve as standardized benchmarks for evaluating speech quality assessment metrics, and can further contribute to improving the robustness and generalization performance of Speech Emotion Recognition (SER) models when applied as training data under diverse real-world acoustic conditions.

This study analyzes the “Google AI Startup School Gyeongbuk Watch Party” as a hybrid model for regional startup education. While startup resources are heavily concentrated in Seoul, this program utilized a mix of global online sessions and local offline networking to bridge the gap. Using an exploratory mixed-methods design with 139 participants over six weeks, including students and entrepreneurs, the study found high satisfaction, a high re-engagement rate, and perceived gains in entrepreneurial self-efficacy through hands-on practice with Gemini and AI Studio. In particular, offline networking (forming an average of 2.7 new relationships) showed relatively high sustained interest and participation from participants. The study suggests that public-private-academic partnerships using global AI tools can effectively decentralize startup ecosystems and foster regional talent. As a single-case exploratory study, however, the findings should be generalized with caution.

This study conducted an empirical analysis using the Extended Technology Acceptance Model (ETAM) to examine the factors influencing corporate acceptance of generative AI-based data analytics solutions. The research model included organizational fit, productivity, convenience, reliability, and perceived risk as external variables reflecting the characteristics of generative AI environments. In addition, usage intention and acceptance intention were distinguished to analyze the technology acceptance process. The results showed that productivity and convenience had significant positive effects on perceived usefulness, while perceived ease of use also strongly influenced perceived usefulness. In contrast, organizational fit and reliability did not show significant effects. Furthermore, perceived usefulness significantly affected usage intention, and usage intention had a significant effect on acceptance intention. This study empirically demonstrates that practical usability and perceived usefulness are key factors in the acceptance of generative AI-based analytics solutions and provides practical implications for establishing corporate AI adoption strategies.

This study aims to examine patterns of alignment between sentiment and risk discourses in industrial news and safety discourse in Korean language textbooks for foreign workers. News data from the fisheries, construction, and agriculture sectors were analyzed using sentiment analysis and LDA topic modeling, while safety-related content in Korean language textbooks was examined through discourse analysis. The results indicate that news discourse across all three industries showed an overall tendency toward positive sentiment, while construction-related news contained a relatively higher proportion of negative discourse centered on industrial accidents and economic downturns. In contrast, safety discourse in the textbooks exhibited different patterns across industries. The fisheries textbook was dominated by risk and warning-oriented expressions, while the agriculture textbook showed a relatively balanced discourse tendency. Standard Korean Language Textbook 2 emphasized protection- and prevention-oriented discourse. This study highlights the need for Korean language education for foreign workers that reflects industry-specific risk characteristics and suggests directions for textbook development integrating emotional awareness and industrial safety education. The findings may contribute to improving Korean language education and establishing policy foundations for preventing industrial accidents among foreign workers.

Unauthorized obstacles on tactile paving, such as kickboards, bicycles, and stacked objects, are major safety hazards for visually impaired pedestrians. This study proposes a YOLOv8(You Only Look Once version 8)-based platform to detect tactile paving intrusions and support pedestrian safety. The proposed method extracts tactile paving regions using YOLOv8 segmentation and detects electric scooters using YOLOv8 detection, then determines intrusions by analyzing the spatial overlap between the two regions through an overlap ratio. The judgment results and detected images are integrated with a Firebase-based storage architecture and an Android application for mobile verification. A quantitative evaluation using a balanced dataset of 100 intrusion and 100 non-intrusion images showed that the polygon-overlap-based method achieved an F1-score of 88.26%, outperforming the detection-only baseline F1-score of 76.86%. With the final decision rule including region-validity verification, the proposed method achieved 91.00% accuracy and a 91.26% F1-score.

This study proposes a LiDAR-SLAM-based IoT-UAV platform that links condition monitoring in indoor battery storage facilities with indoor localization for a mobile platform. The system consists of a sensing node equipped with an MLX90640 thermal array and an MQ-2 gas sensor, LoRa communication, an SQLite-based control server, and a Cartographer module integrated into an F550 UAV. The sensing node calculates a representative temperature from the thermal array and uses the uncalibrated MQ-2 output as a relative value. Threshold combinations assign the measurements to four functional states, after which the measurement time, sensor values, and state labels are transmitted and stored. Functional tests confirmed the sequential operation of sensor-data acquisition, state classification, LoRa message transmission and reception, database storage, and visualization. In addition, Gazebo simulation and an indoor corridor test demonstrated two-dimensional map generation and relative-position display in the map frame using LiDAR, IMU, and optical-flow information. These results show that the fixed sensing node, communication and data-management layer, and UAV localization module can be organized as a coherent information flow, providing component-level evidence for the proposed platform architecture. The present study focuses on functional implementation rather than fire-detection or suppression performance. Future work should validate the thresholds and sensing response under controlled battery-fire conditions, quantify LoRa and SLAM performance, and implement closed-loop operation from sensing-node localization to autonomous UAV navigation and initial response.

With the recent proliferation of cloud-based work, the increase in telecommuting and remote work, and the expansion of intranet-penetrating attacks such as ransomware, Zero Trust architecture has emerged as a core model of global security strategy. The global market for commercial Zero Trust solutions is also expanding rapidly, including ZTNA, SASE (Secure Access Service Edge), and SSE (Security Service Edge). Domestically, adoption is spreading across the public, financial, and private sectors centered on KISA’s ‘Zero Trust Guideline 2.0’ and KOZETA initiatives. However, compared with foreign solutions, domestic offerings demonstrate strengths in regulatory compliance and support for on‑premises and cloud-hybrid environments, while exhibiting limitations in areas such as SASE integration, automation and operational capabilities, global scalability, and publicly available performance benchmarks. Therefore, to enhance the competitiveness of domestic Zero Trust technologies, it is necessary to identify strengths and weaknesses through objective comparison and diagnosis and to propose directions for future improvement. Accordingly, this study set the core domains of NIST SP 800-207 as common diagnostic criteria and compared and assessed domestic and international Zero Trust solutions from the perspectives of integration, scalability, and regulatory and control compliance. Through this analysis, we aim to propose directions for strengthening the global competitiveness of domestic solutions and to provide baseline material for domestic policy and standardization discussions.

Hyperparameter optimization plays a critical role in improving the performance of deep learning models; however, conventional cloud-based approaches can incur substantial communication overhead, latency, and privacy concerns due to centralized data transmission and processing. To address these limitations, this paper presents a distributed hyperparameter tuning system for deep learning in a fog computing environment. The proposed system employs a master–worker architecture consisting of a fog gateway and multiple fog workers, and integrates Random Search and Bayesian Optimization in a sequential tuning process. It also incorporates resource-aware task allocation, early stopping, and fault-handling mechanisms to improve computational efficiency and system robustness under resource-constrained conditions. The system was evaluated using the CIFAR-10 dataset and a ResNet-based model with 3, 5, 7, and 9 fog workers. Experimental results show that increasing the number of fog workers from 3 to 9 reduced the hyperparameter search time from 76.4 minutes to 40.7 minutes, corresponding to a 46.7% reduction. However, the increase in duplicate searches, fault rate, and average response time indicates that communication overhead and synchronization costs limit linear scalability as the number of fog nodes increases. Therefore, the proposed system demonstrates that efficient distributed hyperparameter exploration is feasible in resource-constrained fog environments, while future work should investigate asynchronous task allocation that accounts for the resource states of heterogeneous nodes and techniques for reducing communication overhead.

This study investigates the impact of information security related organizational environment, such as role and communication, on information security compliance and active participation among employees. The risk of organizational insider information leakage grows in parallel with the adoption of advanced information systems. Using survey data from 475 employees in organizations with established information security policies, this research tests the influence of role and communication ambiguity on IS behaviors. These findings reveal that information security role ambiguity decreases both information security in-role behavior and IS extra-role behavior by undermining employees’ perceptions of information security policy authenticity. Moreover, information security communication ambiguity exacerbates these negative effects, further diminishing information security policy authenticity and information security in-role behavior. The results underscore the detrimental impact of ambiguity in organizational roles and communication on information security compliance and participation, emphasizing the need for clear communication and defined work strategies to enhance internal information security practices.

Remote photoplethysmography (rPPG) enables contactless heart-rate estimation from facial videos. Since cameras sample at a finite frame rate, periodic illumination components can be folded into the heart-rate band through aliasing, causing deep rPPG models to output a fake heart rate. Rather than claiming this occurs frequently in real environments, we analyze its effect under controlled synthetic flicker. Under 30 fps sampling, flickers at 29, 31.5, 32, and 33 Hz fold into fake rates of 60, 90, 120, and 180 BPM, and this lock-in was consistently observed across DeepPhys, TSCAN, and EfficientPhys on VIPL-HR and PURE. The proposed Extended Harmonic Ratio (EHR) detected aliased outputs with AUC scores of 0.930 and 0.958 on the two datasets. A clean-data control found no clear natural aliasing, with false positives explained by harmonic confusion. On samples identified as fake-HR lock-in, an auxiliary mitigation using candidate peak reselection—without prior knowledge of the flicker frequency—reduced the residual lock-in rate to 13%.

This study proposes measures for the mid- to long-term enhancement of preservice teacher training curricula and the establishment of a sustainable feedback framework by analyzing quantitative course evaluation ratings (2,166 courses) and qualitative narrative data (4,852 cases) over the past five years from the college of education and pedagogical certificate programs at a university in the metropolitan area. The results revealed that preservice teachers exhibited higher satisfaction with practical, field-connected learning and creative tasks—such as teaching aid development and microteaching—than with unidirectional theory lectures. Conversely, critical demands were identified for stronger alignment with the teacher recruitment examination and national curriculum, optimization of assignment submission timing, and transparency of evaluation criteria. Based on these findings, a four-pillar virtuous feedback framework is proposed: course and assignment design, evaluation fairness and operational punctuality, recruitment and field connectivity, and feedback support infrastructure. Although limited by its reliance on single-university data, this study holds practical significance by empirically capturing longitudinal trends in preservice teachers' perceptions, providing foundational empirical data and practical insights for curriculum optimization and policy implementation in teacher training institutions.

The purpose of this study was to examine the effects of DanceSport on health-related physical fitness and cardiovascular disease risk factors in obese elderly women, and to provide basic data for exercise prescription for the elderly population. The training group consisted of 12 participants, while the control group consisted of 7 participants. The training group engaged in DanceSport (Rumba, Cha-Cha-Cha, and Jive) at 50–80% of VO₂max (beginner to amateur level), for 1 hour per day, 3 days per week, over a 12-week period. Body fat percentage did not change significantly within either group, and no significant between-group difference was found after the intervention. Serum lipoprotein levels (TC, LDL-C, TG, HDL-C) also showed no significant between-group differences, although the training group showed a decreasing trend in triglycerides and stable HDL-C, while the control group showed a decreasing trend in HDL-C. In conclusion, the 12-week DanceSport program did not produce a statistically significant improvement in body fat percentage among obese elderly women, but showed a tendency to prevent deterioration of the blood lipid profile. Further studies with increased exercise intensity, frequency, or duration are needed to confirm effects on body fat.