This study introduces an Artificial Intelligence (AI)-based education platform that enables multilingual learning by integrating real-time speech recognition, Speech-to-Text (STT) translation, automated news parsing, and chatbot support functions. By utilizing Whisper-based speech recognition, Google API multilingual translation, Selenium-based news crawling, and an OpenAI chatbot, learners can access translated lecture scripts, stay informed about the latest AI developments in real time, and receive personalized learning support through natural language interaction. The platform is designed to enhance accessibility, eliminate language barriers, and integrate fragmented educational resources into a cohesive system. System evaluation demonstrated high performance, recording a Word Error Rate (WER) of 4% and a Bilingual Evaluation Understudy (BLEU) score of 85, validating the accuracy and reliability of the transcription and translation modules. A System Usability Scale (SUS) evaluation conducted with 30 participants yielded an average score of 79.5, indicating high user satisfaction across diverse age and experience groups. The modular and reproducible system architecture ensures adaptability to different languages and educational environments. This platform contributes to establishing a standard model for global AI-based education systems by supporting scalable and personalized learning. Future research will focus on expanding language support, enhancing context-aware translation accuracy, developing data-driven employment matching capabilities, and implementing effective strategies for broader service deployment. This integrated approach aims to address current limitations in AI education accessibility and foster the development of global digital talent.

This study aims to establish a search service that provides greenwashing pattern detection and infographic of enterprises through AI-based NLP technology. The purpose of it is to provide decision indicators by checking data on greenwashing trends of companies by users, including investment institutions and general consumers. We used a deep learning model to adopted evaluation factors to score the ‘greenwashing’ and collected the title of media data with active monitoring of companies based on those factors. In addition, construct greenwashing sentence discrimination model and a corporate greenwashing scoring model using the BERT model. As a result, a service provided scores using data collected through its keywords through visualization and the process of calculation when users input the name of the company.

Our research specifies the criteria for categorizing comments and supplements the limitations of the existing services that filter out only some of the direct malicious comments centered on specific vocabulary and phrases. In comparison to prior research detecting three sentiments as positive, negative, or neutral, we subdivided expressions into six categories. Our model applying classification downstream tasks to a pre-trained KcBERT model draws results reflecting the meaning and social context of the comments. An improved accuracy compared to previous studies shows a meaningful outcome.

The number of people participating in urban farming and its market size have been increasing recently. However, the technologies that assist the novice farmers are still limited. There are several previously researched deep learning-based crop disease diagnosis solutions. However, these techniques only focus on CNN-based disease detection and do not explain the characteristics of disease symptoms based on severity. In order to prevent the spread of diseases in crops, it is important to identify the characteristics of these disease symptoms in advance and cope with them as soon as possible. Therefore, we propose an improved crop disease diagnosis solution which can give practical help to novice farmers. The proposed solution consists of two representative deep learning-based methods: Image Captioning and Object Detection. The Image Captioning model describes prominent symptoms of the disease, according to severity in detail, by generating diagnostic sentences which are grammatically correct and semantically comprehensible, along with presenting the accurate name of it. Meanwhile, the Object Detection model detects the infected area to help farmers recognize which part is damaged and assure them of the accuracy of the diagnosis sentence generated by the Image Captioning model. The Image Captioning model in the proposed solution employs the InceptionV3 model as an encoder and the Transformer model as a decoder, while the Object Detection model of the proposed solution employs the YOLOv5 model. The average BLEU score of the Image Captioning model is 64.96%, which can be considered to have high performance of sentence generation and, meanwhile, the mAP50 for the Object Detection model is 0.382, which requires further improvement. Those results indicate that the proposed solution allows the precise and elaborate information of the crop diseases, thereby increasing the overall reliability of the diagnosis.

Collecting skin burn image data that is treated as medical data is generally restricted to medical practitioners. Labeling them is even more challenging since it requires professionals to review the whole process of detecting burn area and classifying its degree. A typical machine learning model trained with a small dataset gathered by non-professionals does not guarantee sufficient performance since they are mostly left unlabeled. Data augmentation on images leads to overcoming the limitations of a small dataset. Semi-supervised learning and knowledge distillation enable the application of unlabeled data. This study examines a combination of various augmentation and machine learning methods to maximize performance on detection and classification. Having an F1 score of 38.0

There has been a growing body of research that explores the need for hearing impaired people. However, English Education aids for the hearing-impaired are still lacking. This paper introduces an English pronunciation correction service for the hearing-impaired to help ensure the equal right to education for members of society. The purpose of this research is to allow effective English pronunciation correction by providing user-friendly feedback using visual information such as graphs, images, and personalized corrective voice considering the special needs of hearing-impaired people.

Twisted light beams such as optical angular momentum (OAM) with numerous possible orthogonal states have drawn the prodigious contemplation of researchers. OAM multiplexing is a futuristic multi-access technique that has not been scrutinized for optical satellite communication (OSC) systems thus far, and it opens up a new window for ultra-high-capacity systems. This paper presents the 4.8 Tbps (5 wavelengths × 3 OAM beams × 320 Gbps) ultra-high capacity OSC system by incorporating polarization division multiplexed (PDM) 256-Quadrature amplitude modulation (256-QAM) and OAM beams. To realize OAM multiplexing, Laguerre Gaussian (LG) transverse mode profiles such as LG00, LG140, and LG400 were used in the proposed study. The effects of the receiver’s digital signal processing (DSP) module were also investigated, and performance improvement was observed using DSP for its potential to compensate for the effects of dispersion, phase errors, and nonlinear effects using the blind phase search (BPS), Viterbi phase estimation (VPE), and the constant modulus algorithm (CMA). The results revealed that the proposed OAM-OSC system successfully covered the 22,000 km OSC link distance and, out of three OAM beams, fundamental mode LG00 offered excellent performance. Further, a detailed comparison of the proposed system and reported state-of-the-art schemes was performed.

We introduced a crop disease diagnosis mobile application service that improves the weakness of the previous services. We utilized image captioning model with Inception V3-based encoder and transformer-based decoder for detailed explanation. Moreover, object detection model with YOLOv5 was used to display bounding boxes to indicate the damaged region to increase the reliability.

This study presents and discusses the home delivery services in stochastic queuing-inventory modeling (SQIM). This system consists of two servers: one server manages the inventory sales processes, and the other server provides home delivery services at the doorstep of customers. Based on the Bernoulli schedule, a customer served by the first server may opt for a home delivery service. If any customer chooses the home delivery option, he hands over the purchased item for home delivery and leaves the system immediately. Otherwise, he carries the purchased item and leaves the system. When the delivery server returns to the system after the last home delivery service and finds that there are no items available for delivery, he goes on vacation. Such a vacation of a delivery server is to be interrupted compulsorily or voluntarily, according to the prefixed threshold level. The replenishment process is executed due to the (s,Q) reordering policy. The unique solution of the stationary probability vector to the finite generator matrix is found using recursive substitution and the normalizing condition. The necessary and sufficient system performance measures and the expected total cost of the system are computed. The optimal expected total cost is obtained numerically for all the parameters and shown graphically. The influence of parameters on the expected number of items that need to be delivered, the probability that the delivery server is busy, and the expected rate at which the delivery server’s self and compulsory vacation interruptions are also discussed.

We introduce a neural network model detecting Korean malicious bot account in Twitter. We extracted features from Korean malicious bot account in Twitter and implemented a model in the context of classification between human and bot account. This can be seen as an improvement of malicious bot detection, in an area of accuracy and multilingual aspect. The mode also suggests an end-to-end solution framework by offering web service.

This paper presents the SDGs performance status analysis system of universities based on deep learning techniques. Using BERT, the classification of cases corresponding to individual goals was carried out, and TextRank was used to extract keywords for transition cases between overseas universities and domestic universities. Using TF-IDF and cosine similarities, we present examples of overseas universities’ SDGs activities that domestic universities can refer to. As a result, we analyzed the current status of SDGs at each university and presented overseas examples that could be further referenced, and created a website that would be helpful for SDGs transition at domestic universities.

This research aims to develop a system for short-form formatting Korean-English literature through AI-based NLP technology. Using a deep learning model, we extracted the most semantically relevant core text from the full text of literature and constructed an emotional keyword classification model by training the existing pre-trained natural language processing model with a sentence-emotion pair Corpus DataSet. Next, we constructed a model that uses text-to-image technology to create images from the core texts and subjects. By combining the above models, database is built and distributed through the web, and finally, short-form literary content services are provided to users.

The purpose of this study is to provide an API for converting the speech of people with dysarthria into a text form by constructing a model that learns the speech characteristics of Korean speakers with dysarthria. A speech recognition model was constructed by using the Korean speech recognition open-source toolkit (Kospeech) which embodies DeepSpeech2 model. Ten thousand voice files which recorded the speech of people with dysarthria and the corresponding transcription files were also collected. Both files were augmented and used for model training. By performing the WER/CER evaluation, it was confirmed that the model constructed in this study recognized the speech of the Korean speakers with dysarthria better than the existing speech recognition model.

This article analyses a four-dimensional stochastic queueing-inventory system with multiple server vacations and a state-dependent arrival process. The server can start multiple vacations at a random time only when there is no customer in the waiting hall and the inventory level is zero. The arrival flow of customers in the system is state-dependent. Whenever the arriving customer finds that the waiting hall is full, they enter into the infinite orbit and they retry to enter the waiting hall. If there is at least one space in the waiting hall, the orbital customer enters the waiting hall. When the server is on vacation, the primary (retrial) customer enters the system with a rate of λ1(θ1). If the server is not on vacation, the primary (retrial) arrival occurs with a rate of λ2(θ2). Each arrival rate follows an independent Poisson distribution. The service is provided to customers one by one in a positive time with the rate of μ, which follows exponential distribution. When the inventory level drops to a fixed s, reorder of Q items is triggered immediately under (s,Q) ordering policy. The stability of the system has been analysed, and using the Neuts matrix geometric approach, the stationary probability vectors have been obtained. Moreover, various system performance measures are derived. The expected total cost analysis explores and verifies the characteristics of the assumed parameters of this model. The average waiting time of a customer in the waiting hall and orbit are investigated using all the parameters. The monotonicity of the parameters is verified with its characteristics by the numerical simulation. The discussion about the fraction time server being on vacation suggests that as the server’s vacation duration reduces, its fraction time also reduces. The mean number of customers in the waiting hall and orbit is reduced whenever the average service time per customer and average replenishment time are reduced.

In recent times, internet of things (IoT) applications on the cloud might not be the effective solution for every IoT scenario, particularly for time sensitive applications. A significant alternative to use is edge comput-ing that resolves the problem of requiring high bandwidth by end devices. Edge computing is considered a method of forwarding the processing and communication resources in the cloud towards the edge. One of the consid-erations of the edge computing environment is resource management that involves resource scheduling, load balancing, task scheduling, and quality of service (QoS) to accomplish improved performance. With this motivation, this paper presents new soft computing based metaheuristic algorithms for resource scheduling (RS) in the edge computing environment. The SCBMA-RS model involves the hybridization of the Group Teaching Optimization Algorithm (GTOA) with rat swarm optimizer (RSO) algorithm for optimal resource allocation. The goal of the SCBMA-RS model is to identify and allocate resources to every incoming user request in such a way, that the client???s necessities are satisfied with the minimum number of possible resources and optimal energy consumption. The problem is formulated based on the availability of VMs, task characteristics, and queue dynamics. The integration of GTOA and RSO algorithms assist to improve the allocation of resources among VMs in the data center. For experimental validation, a comprehensive set of simulations were performed using the CloudSim tool. The experimental results showcased the superior performance of the SCBMA-RS model interms of different measures.

Breast cancer is the major cause behind the death of women worldwide and is responsible for several deaths each year. Even though there are several means to identify breast cancer, histopathological diagnosis is now considered the gold standard in the diagnosis of cancer. However, the difficulty of histopathological image and the rapid rise in workload render this process time-consuming, and the outcomes might be subjected to pathologists’ subjectivity. Hence, the development of a precise and automatic histopathological image analysis method is essential for the field. Recently, the deep learning method for breast cancer pathological image classification has made significant progress, which has become mainstream in this field. This study introduces a novel chaotic sparrow search algorithm with a deep transfer learning-enabled breast cancer classification (CSSADTL-BCC) model on histopathological images. The presented CSSADTL-BCC model mainly focused on the recognition and classification of breast cancer. To accomplish this, the CSSADTL-BCC model primarily applies the Gaussian filtering (GF) approach to eradicate the occurrence of noise. In addition, a MixNet-based feature extraction model is employed to generate a useful set of feature vectors. Moreover, a stacked gated recurrent unit (SGRU) classification approach is exploited to allot class labels. Furthermore, CSSA is applied to optimally modify the hyperparameters involved in the SGRU model. None of the earlier works have utilized the hyperparameter-tuned SGRU model for breast cancer classification on HIs. The design of the CSSA for optimal hyperparameter tuning of the SGRU model demonstrates the novelty of the work. The performance validation of the CSSADTL-BCC model is tested by a benchmark dataset, and the results reported the superior execution of the CSSADTL-BCC model over recent state-of-the-art approaches.

In this paper, we propose a business analysis method that automatically collects review information related to a specific product by using a web crawler, analyzes the customer's emotional response to the product, and supports marketing activities. In order to process data collected from web pages and social networks into information that can be used for marketing, several levels of data processing and text mining techniques are needed. Although various studies have been carried out for this purpose, the data collected through lots of effort and cost contain more extensive information than needed for marketing. So the usefulness of the information obtained through data processing and analysis is not really good. In this paper, we propose a system that automatically receives data from interested sites, interested fields, interested keywords, and target period information from a user and crawls the data. In addition, the reviews that can be used only in marketing can be selected to judge whether or not they are positive, thus it improves the accuracy of the analysis. For the experiment, we collected the reviews of the books sold by specific publishers over the past three years and conducted reputation analysis. The proposed classifier distinguishes through supervisor whether the data collected is a proper review or not. The accuracy of the review classifier is 98.7%. The reputation analyzer judges whether the review is positive or negative with the 86.1% accuracy. The results of this study can be used directly in various industries. And we plan to develop the reputation analyzer, improving the accuracy and extracting the reputation factors affecting customers.

It is a non-deterministic challenge on a fog computing network to schedule resources or jobs in a manner that increases device efficacy and throughput, diminishes reply period, and maintains the system well-adjusted. Using Machine Learning as a component of neural computing, we developed an improved Task Group Aggregation (TGA) overflow handling system for fog computing environments. As a result of TGA usage in conjunction with an Artificial Neural Network (ANN), we may assess the model's QoS characteristics to detect an overloaded server and then move the model's data to virtual machines (VMs). Overloaded and underloaded virtual machines will be balanced according to parameters, such as CPU, memory, and bandwidth to control fog computing overflow concerns with the help of ANN and the machine learning concept. Additionally, the Artificial Bee Colony (ABC) algorithm, which is a neural computing system, is employed as an optimization technique to separate the services and users depending on their individual qualities. The response time and success rate were both enhanced using the newly proposed optimized ANN-based TGA algorithm. Compared to the present work's minimal reaction time, the total improvement in average success rate is about 3.6189 percent, and Resource Scheduling Efficiency has improved by 3.9832 percent. In terms of virtual machine efficiency for resource scheduling, average success rate, average task completion success rate, and virtual machine response time are improved. The proposed TGA-based overflow handling on a fog computing domain enhances response time compared to the current approaches. Fog computing, for example, demonstrates how artificial intelligence-based systems can be made more efficient.

Demand forecasting is the activity of predicting the future using historical data and establishing a model that can grasp trends.Demand forecasting is widely used in a variety of business areas, including production and inventory planning as well as process management.The goal of a company that publishes and sells books is to accurately predict sales volume, thereby increasing book sales, generating more revenue, and reducing losses from inventory management.Using data analysis to predict accurate publishing demand and establishing countermeasures against factors that may cause returns can reduce the amount of losses incurred due to inventory control and returns.The purpose of this study is to identify the factors affecting the sale and return of specific books and to create a model to forecast sales demand.For this purpose, we used the sales data of the books sold for 5 years (2012 ~ 2016) by A publishing company.In addition, we collected the data related to books in the Internet portal system and SNS site.We hypothesized the factors that affect the sale and return of books and collected the variables needed for hypothesis testing from web pages and SNS sites.As a result of this study, it was possible to identify the factors affecting the return and sales of a specific book, and it was possible to establish a sales order prediction model.Because the available data is limited in the study, the scope of this study was limited to forecasting the sales demand of some books.If we apply the proposed analytical procedure and method directly from the company, we can expect better prediction results.It is also expected to be applicable to various business processes of book publishing or sales companies.