Increasing use of artificial intelligence (AI) in English language learning has created the need for reliable predictive models that can evaluate learner outcomes in adaptive educational settings. This study introduces a hybrid Bidirectional Encoder Representation from Transformer-Gated Recurrent Unit (BERT- GRU) framework that combines BERT-based contextual semantic embedding with GRU-based temporal sequence modeling to predict personalized learning effectiveness. The model is trained on the Adaptive English Learning Dataset, which contains 4,931 classroom interaction sessions and 26 multimodal features, including learner behavior, instructor attributes, and environmental conditions. The dataset captures various learning-related signals like engagement patterns, performance indicators, and instructional context, allowing for thorough modeling of learner progress. The proposed architecture merges deep contextual language representations with sequential learning dynamics to improve prediction reliability and strength. BERT extracts rich contextual embeddings from learner interactions, while GRU captures time dependencies across sequential learning sessions. The main technical contributions of this study include a hybrid embedding, sequence fusion strategy, a structured feature engineering pipeline, and an optimized BERT, GRU architecture for educational outcome prediction. These elements work together to effectively combine semantic and temporal information, enhancing predictive performance in adaptive learning environments. Experimental results show strong performance with an accuracy of 95.17
The demand for personalized, visually rich content in digital advertising has led to the increased adoption of AI technologies such as Generative Adversarial Networks and Natural Language Processing. Yet, finding an accurate match of textual descriptions with high-quality, contextually relevant images is still a challenge, particularly where emotions and creativity are part of the process in advertising. This paper proposes a new framework that integrates DCGANs with natural language processing techniques-sentiment analysis by Text CNN-and generates high-quality, contextually appropriate images from text descriptions for digital advertising. It utilizes the “Stable Diffusion—Image to Prompts” dataset from Kaggle, together with real and AI-generated images. Text CNN conducts sentiment analysis to capture the emotional tones in the input text. The evaluation metrics used for the model include but are not limited to Inception Score (IS), Frechet Inception Distance, and Latent Space Interpolation. The Inception Score demonstrates better performance with the proposed framework, coming to 4.0 against the score of 2.5 in AttnGAN; it has a lower FID of 80 versus 125.98 in AttnGAN, and also presents high-quality and smooth transitions in latent space interpolation. These improvements manifest better emotional alignment in generated images and creativity, outperforming the performance of traditional models. This research points out the scope of integrating GANs and NLP for automating the creation of personalized, high-quality digital ads and thereby offering a scalable solution for modern marketing needs.
Traditional university Physical Education (PE) faces problems such as limited visibility of movement demonstrations and delayed feedback. This research examines the employment of multimedia network teaching platforms in university PE. According to the teaching needs, it gives an opportunity to access through an Internet/Intranet hybrid architecture and selects a High Efficiency Video Coding (H.265/HEVC) standard for coding. A layered platform structure is established, which includes the user access layer, network transmission layer, and core service layer. A multimedia teaching information transmission coding module is developed, which includes functions such as sports multimedia resource information and teaching information transmission coding. This module utilizes a multimedia teaching resource encoding rate control method based on Deep Neural Networks (DNN) to adaptively adjust the resource encoding rate to ensure the transmission quality of teaching resources. The sports multimedia teaching module uses a multimedia vision-based motion recognition method to extract human three-dimensional coordinate motion features and input them into a hybrid kernel Support Vector Machine (SVM) model for motion discrimination, thereby realizing intelligent teaching management of university sports multimedia networks. Experimental results indicate that, under the 5G network environment, the average Quality of Experience (QoE) of the platform’s multimedia teaching videos remains stable at above 0.90, with minimal fluctuation throughout the entire period. The bitrate control method based on DNN can keep the bitrate basically stable at above 3000 Kb · s−1 when the number of online users reaches 70, thus avoiding blurring and freezing of multimedia teaching images. The motion recognition technology can achieve quantitative error correction of student movements.
Japanese language learners have a difficult task of grammatical error correction (GEC) because of the complex morphology particle use and proficiency level variations. Non-adaptive rule-based systems and statistical models have in many cases the difficulty to explain or model contextual dependencies and learner specific error patterns. This paper suggests a combined fine-tuned Japanese BERT with Seq2Seq Transformer as proficiency classifier, error detector and with grammatical mender for improving Japanese writing proficiency in University. Data was gathered on the basis of the Japanese learner corpora including the sentences with proficiency annotation and corrected references basing on JFLEG-JP and institutional learner data. Using these inputs, the framework jointly learns proficiency estimation and grammatical correction; Cross-entropy loss is employed to optimize both proficiency classification and sequence generation tasks, enabling effective contextual learning and accurate correction performance. Fine-tuning of the last four layers of BERT helps to detect semantic and syntactic subtleties. The suggested model is assessed with the help of the BLEU, Precision, Recall, and F1-score measures. Experimental outcomes show better results of a BLEU score of 61.9, Precision of 0.81, Recall of 0.78 and an F1-score of 0.79 which beats the current baseline models. The proposed framework enables proficiency-aware contextual grammatical correction with adaptive pedagogical feedback for Japanese university writing support.
Early childhood education plays a crucial role in shaping a child’s future development and accurately predicting developmental outcomes is vital for personalized learning and early intervention strategies. However, existing models often struggle with integrating diverse data sources and ensuring interpretability. This study proposes a hybrid Machine Learning (ML) framework that combines Random Forest (RF) and Extreme Gradient Boosting (XGBoost) to predict preschool children’s developmental outcomes, addressing the need for robust and interpretable models in early childhood education. The SmartLearn Preschool Dataset, which includes multimodal data such as child profiles, interaction logs, behavioral metrics and audio-visual features, is used for model training. The framework employs a stacking ensemble approach, where RF and XGBoost serve as base models and Logistic Regression (LR) is used as the meta-model. Hyperparameter optimization is performed using Optuna and (SHapley Additive exPlanations) SHAP values are incorporated for interpretability, offering insights into feature importance. The model is evaluated using several performance metrics, including accuracy, precision, recall and F1-score. The proposed model achieved an impressive accuracy of 98.78
This paper investigates a multi-source data fusion method for smart substations based on digital twin technology to eliminate the disparity and heterogeneity of multi-source data and enhance operational maintenance capabilities. Sensors at the perception layer collect multi-source data from the smart substation and transmit it to the platform layer. The virtual-physical mapping module at the platform layer employs edge-folding technology to achieve lightweight processing of the smart substation digital twin model, reducing the number of mesh faces while preserving key geometric features. The multi-source data fusion module performs consistency checks using the Wilcoxon rank-sum test to determine data fusibility. The validated data is then fused using the Joint Kalman Filter algorithm. The fusion results are transmitted to the application layer, where they are converted into various operational applications and update the digital twin model of the substation. Experimental results demonstrate that the digital twin model accurately reproduces the physical layout and equipment composition, with the information entropy of the multi-source data fusion consistently exceeding 0.85. The proposed method was benchmarked against particle filtering, Extended Kalman Filtering (EKF), and Unscented Kalman Filtering (UKF), showing a 5
Mobile Ad-hoc Networks are dynamic and self-configuring networks where nodes communicate with each other without require for any fixed infrastructure. Quality of Service (QoS) in such networks is crucial for ensuring reliable and efficient communication. In this research, a novel approach, RS-MANET-DBO-AODV-QoS is proposed, which enhances route selection in MANETs by integrating the Dung Beetle Optimizer (DBO) with Ad hoc On-Demand Distance Vector (AODV) protocol. The proposed mechanism utilizes unique characteristics of DBO along with AODV to select the most optimal route for data delivery. Pheromone values of paths are calculated based on various metrics including congestion, E2E reliability, number of hops, remaining energy of nodes along the path. The route with higher pheromone value is chosen for transmitting data packets, thereby improving QoS in MANETs. Then evaluate performance of proposed method regards data packet delivery ratio, E2E delay, throughput, percentage of surviving nodes, complexity analysis, convergence, energy efficiency and QoS gain. Comparative analysis is conducted against existing techniques such as Energy Aware Routing through Genetic Algorithm and AOMDV in MANET (GA-AOMDV-MANET) Enhanced-Ant-AODV for optimal route selection in mobile ad-hoc network (Ant-AODV-MANET) and Multi-objective based route selection method utilizing AOMDV in MANET (MORS-AOMDV-MANET) respectively.
Efficient autonomous robot navigation in unstructured and unknown terrains is critical for applications such as agriculture, disaster response, and exploration. Traditional path planning methods, such as Particle Swarm Optimisation (PSO) and Grey Wolf Optimisation (GWO), face challenges in dynamic environments, as they handle terrain classification and path planning as independent processes, lacking a unified framework that adapts navigation decisions based on real-time terrain feedback. This study presents a novel hybrid optimisation framework combining PSO, GWO, and Evolutionary Algorithms (EA) through a coordinated Exploration-Integration Hybrid Optimisation (EI-HO) pipeline that distinguishes it from conventional standalone methods, while using a Multi-Layer Perceptron (MLP) for accurate terrain classification based on multimodal sensor data, including Inertial Measurement Unit (IMU), vibration signals, and suspension parameters. The proposed system was evaluated using a comprehensive terrain dataset and tested in simulation environments like Robot Operating System (ROS) and Gazebo. The results showed that the proposed method achieved an accuracy of 0.92, a precision of 0.91, a recall of 0.90, and an F1-score of 0.91, outperforming PSO (accuracy = 0.85) and GWO (accuracy = 0.80). The Exploration-Integration Hybrid Optimisation achieved a balanced final objective score of 42.39, compared to 35.6 for Grey Wolf Optimisation and 54.7 for Particle Swarm Optimisation, confirming both navigation accuracy and convergence efficiency gains. Future work will focus on improving real-time learning and optimising computational efficiency.
The swift advancement of the Internet of Things (IoT) along with Artificial Intelligence (AI) technologies has changed the face of smart education by allowing constant supervision and evaluation of student learning habits. Nevertheless, the present-day systems still cannot cope with time-series behavioral data processing efficiently and also do not yield correct real-time engagement predictions. This study aims to develop an IoT-based monitoring and analysis system with a Long Short- Term Memory (LSTM) deep. intelligent education learning model to forecast and improve the learning behaviour of the students. environments. The wearable sensors and smart classroom devices were used to gather time-series data. measuring variables of attention duration, Internet activity, keystroke, and physiological parameters. indicators. Pre-processing of data was done with Moving Average Filter to eliminate noises and Min–Max. Normalization through scaling, although features extraction developed behavioural indicators such as study. duration, consistency of concentration and frequency of participation. The optimized LSTM model, written in Python, was trained on sequential data with hyperparameters optimized to tune sequence. length, batch size, and learning rate, and dropout rate. The experimental assessment proved that the prediction accuracy of the proposed model was 98.7
The ability to innovate and be an entrepreneur is a skill that should be taught to college students who must learn to adapt, innovate and solve problems in a rapidly changing economic system. A reliable prediction and right assessment of these competencies is a must for customized education and entrepreneurship guidance. The majority of present methods of measuring entrepreneurial potential are, however, static and are measured by means of expert opinions or performance indicators which fail to capture the dynamic and multidimensional nature of entrepreneurial potential. This restriction results in poor adaptability, subjectivity and poor prediction of the data. This paper introduces a deep reinforcement learning (DRL) based framework named Innovation-Evaluation DRL Predictor (IEDP) to solve the above challenges and evaluate and predict students capability of innovation and entrepreneurship. For IEDP, the outcomes of creativity are quantified through an IEDP reward function, which combines expected and actual reward functions based on items such as originality deviation from baseline responses, novelty score based on behavioural divergence in decision patterns, and performance gain over iterative learning episodes. The approach incentivizes creativity-oriented results, enhances decision-making strategies, and forecasts the likelihood of entrepreneurial success with greater accuracy. The proposed method can be used for continuous monitoring of student performance, adaptive feedback and directing allocation of resources for innovative education by educators, institutions and policy makers. IEDP has been shown to outperform traditional assessment approaches in terms of predictive power, adaptability to changing surroundings, and correlation with entrepreneurial outcomes in the real world. The study highlights the potential of AI-assisted approaches in developing innovative, future-ready employees.
Personalized learning path recommendation is a key issue in intelligent education systems, aiming to generate learning sequences that conform to teaching principles based on learners’ knowledge status and learning objectives. However, English learning has a significant hierarchical structure and prerequisite dependence, making it difficult for traditional recommendation methods to simultaneously characterize knowledge structure constraints and learner dynamic evolution. Therefore, this paper proposes a personalized English learning path recommendation model based on Knowledge Graphs (KG) and learner profiles. First, an English learning knowledge graph covering knowledge points, skills, resources, assessments, and CEFR levels is constructed to explicitly model multi-type educational semantic relationships. Second, a dynamic learner profile composed of knowledge mastery vectors and skill ability vectors is designed to characterize learning states at multiple granularities. Furthermore, a knowledge tracing method integrating a knowledge tracing model and knowledge graph neighborhood propagation is proposed to improve the stability and generalization ability of state prediction. Based on this, candidate learning paths are generated and ranked under the pre-requirement constraints of the knowledge graph, achieving a balance between teaching rationality and personalized needs. The proposed model is evaluated using standard performance metrics, including Precision (0.381), Recall (0.357), and NDCG (0.401). Experimental results show that the proposed method significantly outperforms the comparative models in terms of recommendation accuracy, learning gain, and path completion rate, while also possessing good interpretability and adaptability.
The development of English listening and speaking skills is essential for effective academic communication and global interaction, yet many learners face challenges because traditional learning environments fail to provide them with personalized feedback and objective evaluation methods. The goal of this research is to create an Artificial intelligence framework which automatically evaluates English speech performance through its assessment of multiple speech components including content and fluency and pronunciation. The proposed work introduces a hybrid model which combines Harris Hawks Optimization (HHO) for optimal acoustic feature selection with a Self-Attention Deep Neural Network (SA-DNN) for speech performance prediction. The Speech Content Fluency and Pronunciation Scores dataset provides speech recordings which undergo pre-processing through noise removal and silence trimming and normalization and resampling. The extraction process involves key acoustic features which include Mel-Frequency Cepstral Coefficients and pitch and energy and speech rate and pause duration and these features undergo optimization through HHO to eliminate any redundant elements. The SA-DNN model receives the optimized feature subset which uses Self-Attention to identify essential speech patterns that occur across different time intervals to determine multiple speech scoring dimensions. Experimental results demonstrate that the proposed SA-DNN with HHO model significantly outperforms baseline approaches which achieve Mean Absolute Error of 0.0307 and Mean Squared Error of 0.0019 and Root Mean Squared Error of 0.0431 and Coefficient of Determination (R2) of 0.975 which indicates high prediction accuracy and reliability. The proposed framework functions as an efficient and accurate automatic speech evaluation system which also enables intelligent language learning systems to operate.
Scene text segmentation in natural images remains challenging because text instances often appear curved, irregular, multi-oriented, and embedded in cluttered backgrounds. Accurate boundary localization is particularly important for reliable text-region extraction, while heavy architectures may limit real-time applicability. This study aims to develop a compact boundary-aware framework for real-time scene text segmentation and localization. To this end, an edge-aware transformer architecture is proposed based on a U-Net-like encoder-decoder structure. The model combines convolutional local feature extraction with a transformer bottleneck for global-context modeling and uses dual output heads to predict both text-region segmentation masks and edge masks. Polygon annotations from the Total-Text dataset are converted into rasterized segmentation and boundary masks, and the model is optimized using a composite loss that combines region-level and boundary-level supervision. On the Total-Text validation set, the proposed model achieved a Dice Score of 0.6176, an IoU of 0.4472, a pixel accuracy of 93.57
Point-of-Interest (POI) recommendations in Location-Based Social Networks (LBSNs) are susceptible to data sparsity, dynamic user preferences, and the need to strike a balance between exploration and exploitation. In this paper, we present a Multi-Agent Hierarchical Q-Learning (MAHQL) framework that combines adversarial learning, meta-learning, and multi-agent coordination to overcome these issues. The three-phase model first captures user preferences across global, local, and temporal dimensions, thereby addressing data sparsity and cold-start issues. The second phase posits a Q-policy generator that integrates the multi-agent coordination with adaptive exploration strategies to strike a balance between exploration and exploitation. In the third phase, a meta-learning–enhanced hierarchical discriminator provides diverse recommendations that reduce the echo-chamber effect. MAHQL has been tested on real-world datasets: Gowalla, Foursquare, and Weeplaces. It has achieved 71.84
Cybersecurity plays a crucial role in safeguarding digital infrastructures from ever-changing threats. Besides, it guarantees that critical data and systems are always protected, secure, and available. An Intrusion Detection System (IDS) keeps an eye on network traffic and system operations for signs of unauthorized access or malware activity. However, outdated methods typically generate false alarms and have quite low detection accuracy. To address these issues, this work proposes a robust cybersecurity Framework with Progressive Graph Convolutional Networks for Advanced Intrusion Detection and Classification of Sophisticated Cyber Attacks (CSF-PGCN-AID-SCA). The framework integrates an Adaptive Two-Stage Unscented Kalman Filter (ATSUKF) for data pre-processing, Tasmanian Devil Optimization (TDO), a meta-heuristic inspired by the hunting and foraging behavior of Tasmanian devils, is applied to select the most relevant feature and Progressive Graph Convolutional Networks (PGCN) for accurate intrusion detection and attack classification. The framework is implemented in Python and evaluated utilizing the NSL-KDD dataset. The proposed method achieves 8.9–22.5
Online news media and social networking sites have become frequent sources of information for updates on global events and general information. To attract more attention and generate more clicks, many online platforms use attention-grabbing and deceptive headlines for their articles, known as clickbait. The current research presents a hybrid method that combines features extracted through an autoencoder model and handcrafted linguistic features to identify clickbait posts. The proposed model is assessed for the optimal number of linguistic and autoencoder features and validated using three benchmark datasets that differ in platform and language. The integration of four linguistic features and fifty autoencoder-based features outperformed the current state-of-the-art models. The proposed model achieved an accuracy of 0.90 and a weighted F _1 -score of 0.90 on the Clickbait Challenge dataset, surpassing all previously reported results. Similarly, on the Kaggle News dataset, it attained an accuracy of 0.89 and a weighted F _1 -score of 0.89, outperforming prior benchmarks. For the Bengali News dataset, it achieved an accuracy of 0.76 and a weighted F _1 -score of 0.72.
This paper proposes a real-time, stage-based posture correction system for conventional deadlift exercises, aiming to prevent injuries through predictive movement analysis and immediate corrective feedback. The novelty of this work lies in its ability to continuously monitor joint-specific angles across distinct deadlift stages—Set-Up, Lifting, and Lock-Out—and proactively detect improper postures before they result in incorrect execution or potential injury. The system leverages MediaPipe and OpenCV to extract 33 body keypoints and analyzes critical joint angles involving the shoulders, hips, knees, ankles, and trunk to ensure proper alignment and maintenance of a neutral spine throughout the exercise. A rule-based angle evaluation framework is applied in real time, triggering multi-modal feedback—including visual indicators, textual instructions, and audio guidance via text-to-speech—whenever predefined safety thresholds are exceeded. Additionally, the system tracks exercise repetitions and classifies each repetition as correct or incorrect based on joint-specific criteria, enabling users to quantitatively monitor training quality. The system operates at an average frame rate of 24–30 FPS on a standard CPU, ensuring smooth real-time performance without specialized hardware. Experimental evaluations were conducted using both real exercise performances validated by a professional trainer and multiple professional deadlift training videos under diverse conditions, including varying lighting, complex backgrounds, and crowded scenes. Results demonstrate the system’s effectiveness in detecting stage-specific posture errors and delivering timely, actionable feedback. Identified limitations include sensitivity to camera misalignment, keypoint occlusions, and variations in body types, which are addressed as directions for future improvement. Overall, the proposed system offers a practical and accessible solution for predictive deadlift posture correction, enhancing exercise safety and effectiveness in real-world training environments.
This article introduces Kepler-aSI, a novel matching approach designed to address potential semantic gaps in tabular data by utilising a Knowledge Graph. The task poses significant challenges for machines, necessitating additional cognitive capabilities incorporated into the matching methods. Our primary objective is to devise a rapid and effective technique for annotating tabular data with relevant features extracted from a Knowledge Graph. This approach combines search and filter services with advanced text pre-processing techniques. We conducted experimental assessments within the SemTab 2021 and SemTab 2022 challenges to evaluate its performance. The obtained results demonstrate promising performance and commendable rankings, signifying the effectiveness of our proposed method.
Robust classification of neurodegenerative diseases, such as Alzheimer’s Disease (AD) at the Mild Cognitive Impairment (MCI) stage, requires machine learning models resilient to common data complexities. In this paper, a new distribution-agnostic classification method called the Class Median Min–Max Threshold (CM3T) is presented. CM3T applies a modality-specific threshold and aggregates decisions using a majority voter classifier using volumetric characteristics of brain tissues from 220 MRI samples. Its core computational novelty is deriving decision boundaries from class-specific medians and standard deviations, offering inherent robustness against skewed distributions and class imbalance. Furthermore, CM3T allows independent control of margins for each classification group, an advantage over traditional techniques. The CM3T model significantly outperforms conventional machine learning methods, achieving superior performance with 96.3
Firstly, a new version of fuzzy sets called bi-quadratic fuzzy sets is presented in this work. We checked the capability of bi-quadratic fuzzy sets based on suitable examples and compared them with intuitionistic fuzzy sets, Pythagorean fuzzy sets, and Fermatean fuzzy sets. We also stated some set-theoretic operations, score, and accuracy functions. Further, to strengthen and better handle real-life optimization problems, a new concept and model called the bi-quadratic fuzzy optimization technique with bi-quadratic membership and bi-quadratic non-membership functions are introduced in this article. Many modifications of fuzzy optimization techniques have been presented using ordinary linear membership and non-membership functions. However, membership and non-membership can be bi-quadratic instead of linear. Therefore, in the domain of uncertainty and hesitation, the bi-quadratic membership and non-membership functions may play a vital role in real-life optimization problems rather than ordinary ones. Here, we construct the bi-quadratic membership and non-membership functions in a developed computational algorithm. Secondly, it is interesting to note that in real-life problems, the value of data changes very fast, so linear membership does not work properly. For example, the income of the person in a country, in such a case, piecewise linear membership may play a perfect role. So, to deal with problems with piecewise linearity, a piecewise linear optimization technique is constructed based on it, and a computational algorithm designed to show its application is also discussed in this paper. Finally, the outcomes of the developed algorithms are compared via the degree of closeness with the outcomes of the existing algorithms.
