Word Sense Disambiguation (WSD) is a longstanding challenge in natural language processing, particularly in morphologically rich and low-resource languages such as Persian. The inherent ambiguity of Persian named entities exacerbated by domain-specific contexts and limited labeled data complicates both semantic interpretation and information extraction. In this study, we introduce the PWNC corpus, a large-scale, integrated dataset designed for both Named Entity Recognition (NER) and WSD in Persian. The corpus was automatically constructed through a semi-supervised framework, incorporating contextual similarity measures and clustering algorithms to annotate ambiguous entities across ten semantic categories. Utilizing a semi-supervised framework, the proposed homograph semantic categorization method achieved robust performance, with a precision of 83%, recall of 81%, and an F1-score of 82% across over 305K annotated paragraphs. Detailed error analysis revealed challenges in disambiguating closely related senses and weak entities, which were mitigated through contextual embedding strategies. This work provides the first publicly available dual-task corpus for Persian NER and WSD, offering a scalable solution for disambiguation in low-resource tasks and laying the baseline for future research in Persian semantic processing.
This study investigates the effectiveness of machine learning algorithms, including Neural Networks, Bayesian Networks, Support Vector Machines, and Random Forests, in predicting football match outcomes using data from the English Premier League (2018–2022).By incorporating user-generated probabilities for home win, away win, and draw alongside conventional features, the models were evaluated under binary and multi-class classification scenarios. The Support Vector Machine achieved the highest accuracy (69%) in the win-loss scenario, while the Neural Network reached 51% in the win-draw-loss scenario. Results indicate that user-derived features enhance predictive performance, though user predictions show a bias toward home teams, especially in uncertain cases. These findings highlight the potential of integrating user perspectives into predictive modeling and underscore the importance of addressing cognitive bias in sports analytics.
With rapid advancements in information and communication technology, recommender systems have become vital tools across a wide range of online activities and e-commerce processes. Collaborative recommender systems, which utilize user data and contributions to provide suggestions, represent a significant innovation in this field. In this paper, we conduct an analysis of collaborative recommender systems and evaluate their impact on enhancing the efficiency and accuracy of recommendations. To this end, we propose a deep learning approach using a Graph Convolutional Network (GCN), as a special type of Graph Neural Network (GNN). By assigning weights to edges between nodes, scores are calculated for these edges. The importance of the edges varies based on the number of neighboring nodes and their proximity to the target node. The higher the edge score, the more significant the path. To calculate edge weights, we leverage metrics such as Jaccard similarity, cosine similarity, LHN index, and Salton cosine similarity. This approach improves the identification of relationships between nodes and enhances the accuracy of the recommender system. For implementation, we utilized the well-known MovieLens dataset. Ultimately, users were clustered into 18 clusters, with a large number of nodes within each cluster. By clustering users, we increased the number and diversity of recommendations. This significantly improved the performance of the recommender system, yielding promising results.
Medical image analysis, crucial for disease diagnosis and treatment, often suffers from the challenge of class imbalance, where the area of normal tissue significantly outweighs that of abnormal regions. Furthermore, the varying class ratios across different images within a dataset complicate the application of uniform loss adjustments. To address these issues and advance automated segmentation, this study proposes a novel deep learning model integrating the strengths of YOLO Version 8's efficient feature extraction modules (SPPF and C2F) within a U-shaped architecture enhanced by a Receptive Field Enhancement (RFE) module. The RFE module, acting as an advanced skip connection, strategically fuses multi-scale features from corresponding and subsequent encoder layers processed through SPPF and C2F to enrich feature transfer and improve receptive field. To specifically tackle the class imbalance and the diversity of class distributions across images, we introduce a novel Adapt Exponential Loss function. This pixel-level loss dynamically adjusts class weights for each image based on its individual lesion-to-total-pixel ratio (k). We evaluated our proposed model and loss function on challenging skin lesion datasets: ISIC 2018, ISIC 2017, and PH2. Our method achieved significant segmentation performance with IoU scores of 86.47%, 85.67%, and 93.13%, and Dice scores of 91.63%, 90.19%, and 96.02% on ISIC 2018, ISIC 2017, and PH2, respectively, demonstrating its effectiveness in accurately delineating skin lesions despite class imbalance and varying lesion proportions. This work contributes a robust framework for medical image segmentation, facilitating more reliable diagnostic tools in dermatology.
In recent years, the application of deep learning techniques has revolutionized various domains, including the realm of sports analytics. The analysis of ball tracking and trajectory in sports has become an increasingly vital area of research, driven by advancements in technology and the growing demand for data-driven insights in athletic performance. In volleyball, a sport characterized by rapid movements and strategic play, the ability to accurately track the trajectory of the ball is crucial for both training and competitive analysis. This paper proposes novel deep learning models for accurate volleyball ball detection and tracking. By incorporating attention mechanisms into the YOLOv8 and YOLOv10 architecture, our models significantly improve performance, particularly in challenging situations involving occlusions and fast movements. The proposed models across several metrics compared to baseline and other models. Specifically, achieved precision (94.2% and 94.7%, respectively) and recall (88.1% and 87.6%, respectively) and real-time processing speeds, making them suitable for various sports analytics applications.
Dialogue understanding for low-resource languages like Persian remains challenging due to limited annotated data, which constrains supervised training at scale. We propose a simple yet effective training-free method that combines machine translation, retrieval-based example selection, and prompting with a large language model (GPT-4o) to improve zero-shot cross-lingual performance. Given a Persian utterance translated into English, our method retrieves semantically and lexically similar English examples using a hybrid similarity function, translates them back into Persian, and constructs a few-shot prompt tailored to the input. This input-sensitive strategy enhances the quality of the examples, helping the model align more effectively with each instance. Experimental results on the Persian-ATIS dataset show that our approach improves intent detection and achieves competitive slot filling performance, outperforming state-of-the-art baselines without requiring any supervision in the target language. The modular pipeline is easy to reproduce and, in future work, can be extended to other low-resource languages, tasks, or retrieval configurations. The repository of our work is available at https://anonymous.4open.science/r/Persian_Language_Understanding-FDF4.
Brain tumors are among the most life-threatening neurological conditions, requiring precise and early diagnosis for effective treatment planning. Traditional deep learning models, such as Convolutional Neural Networks (CNNs) and ResNet-based architectures, have demonstrated promising results in brain tumor classification. However, these models often struggle to capture long-range dependencies within MRI images, which are crucial for accurate classification. To overcome this limitation, we propose a Hybrid CNN-ViT model, combining the strengths of Vision Transformers (ViT) and CNNs to achieve high-precision brain tumor classification. The CNN component effectively extracts local spatial features, while the ViT module captures global contextual relationships within MRI scans. The model is evaluated on a four-class dataset of Glioma, Meningioma, Pituitary tumors, and non-tumor images, achieving an impressive accuracy of 98.37%, surpassing conventional CNN-based methods. By leveraging transfer learning, the approach enhances classification performance while reducing reliance on large-scale labeled datasets. The proposed Hybrid CNN-ViT model offers a scalable, robust, and efficient solution for real-world neuro-oncological diagnostics, significantly improving the accuracy of MRI-based brain tumor detection.
Nowadays, the recognition of emotions using speech signals has gained popularity because of its vast number of applications in different fields such as medicine, online marketing, online search engines, education systems, criminal investigations, traffic collisions, and more. Many researchers have adopted different methodologies to improve emotion classification accuracy using speech signals. This study presents a novel time-series-to-graph transformation framework for speech emotion recognition. Speech signals were segmented into overlapping windows, each converted into graphs, from which 16 structural features were extracted. Significant features were then selected via Minimum Redundancy Maximum Relevance (mRMR) and used to train four classifiers: random forest (RF), linear discriminant analysis (LDA), support vector machine (SVM), and k-nearest neighbors (KNN). Finally, a soft-voting ensemble strategy was employed to integrate their predictions, yielding improved classification performance. The proposed method achieved the highest sensitivity, specificity, and accuracy for the SAVEE database: 83.57%, 98.93%, and 98.16%, respectively. Similarly, for the EmoDB database, the highest values were 94.47%, 99.09%, and 98.40%, respectively. We also compared our results with other methods and found that our method outperformed state-of-the-art techniques in emotion classification.
Aspect-Based Sentiment Analysis (ABSA) has become a critical tool for extracting fine-grained insights from user opinions. This paper introduces DeGF-ABSA (DeBERTa-Gated Fusion for Aspect-Based Sentiment Analysis), a novel architecture that addresses key limitations in existing approaches by dynamically balancing global contextual features and aspect features. Unlike traditional methods that rigidly combine context and aspect representations—or transformer-based models lacking explicit mechanisms to disentangle aspect-specific signals—DeGF-ABSA leverages DeBERTa’s disentangled attention mechanism, which excels at modeling positional dependencies in technical texts, paired with a gated fusion layer. This layer adaptively weights the contributions of the context features that come from the [CLS] token, and the aspect-specific features come from the mean of aspect tokens. This helps in accurately determining the sentiment in complex sentences. Experiments on SemEval 2014 datasets achieve state-of-the-art results: 86.68% accuracy (84.50% F1) for Laptops and 91.43% accuracy (86.83% F1) for Restaurants.Cross-domain generalization is critical for aspect-based sentiment analysis, as domain-specific aspects and vocabulary vary significantly. Sentiment expressions also differ across domains, such as 'delicious' for food reviews versus 'fast performance' for electronics, requiring adaptable models to capture contextual nuances. Evaluating the DeGF-ABSA model's performance on datasets from domains beyond laptops and restaurants would provide valuable insights into its ability to generalize and its potential for broader applicability.
Farsi optical character recognition remains challenging due to the script’s cursive structure, positional glyph variations, and frequent diacritics. This study conducts a comparative evaluation of five foundational deep learning architectures widely used in OCR—two lightweight CRNN based models aimed at efficient deployment and three Transformer based models designed for advanced contextual modeling—to examine their suitability for the distinct characteristics of Farsi script. Performance was benchmarked on four publicly available datasets: Shotor and IDPL PFOD2 for printed text, and Iranshahr and Sadri for handwritten text, using word level accuracy, parameter count, and computational cost as evaluation criteria. CRNN based models achieved high accuracy on word level datasets—99.42% (Shotor), 97.08% (Iranshahr), 98.86% (Sadri)—while maintaining smaller model sizes and lower computational demands. However, their accuracy dropped to 78.49% on the larger and more diverse line level IDPL PFOD2 dataset. Transformer based models substantially narrowed this performance gap, exhibiting greater robustness to variations in font, style, and layout, with the best model reaching 92.81% on IDPL PFOD2. To the best of our knowledge, this work is among the first comprehensive comparative studies of lightweight CRNN and Transformer based architectures for Farsi OCR, encompassing both printed and handwritten scripts, and establishes a solid performance baseline for future research and deployment strategies.
Identifying and classifying anomalies in textual data from social networks is challenging due to the linguistic complexity and diverse user expressions. While deep learning and machine learning techniques offer promise in tackling this problem, their effectiveness is limited by insufficient data. The effect of Generative Adversarial Networks (GANs) on anomaly detection and Classification is assessed in this paper, along with their relevance for generating synthetic text data. Combining synthetic and real data enhances classification accuracy, especially in settings of limited data. In this paper, Lasso and Ridge regression techniques are used for anomaly detection and classification. Experimental results reveal the superior performance of the proposed model in identifying and classifying anomalies under new datasets generated by GAN. By combining statistical methods with generative techniques, the solution becomes not only more interpretable and scalable but also better suited for advanced text analysis in fast-changing environments like social media platforms.
Fault prediction in power transformers is pivotal for safeguarding operational reliability and reducing system disruptions. Leveraging dissolved gas analysis (DGA) data, AI‑driven techniques have recently been employed to enhance predictive performance. This paper introduces a novel machine-learning framework that integrates Hist Gradient Boosting (HGB) with a metaheuristic Particle Swarm Optimization (PSO) algorithm for hyperparameter tuning, thereby guaranteeing classifier robustness. The proposed method underwent a two‑stage evaluation: first, Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), and HGB were benchmarked, revealing HGB as the most effective method; second, PSO was applied to optimize HGB’s hyperparameters, yielding further performance improvements. Experimental results demonstrate that the hybrid HGB‑PSO model achieves an accuracy of 97.85 %, precision of 98.90 %, recall of 97.33 %, and an F1‑score of 98.99 %. All simulations and comparative analyses against state‑of‑the‑art methods were implemented in Python, and confusion‑matrix analysis was employed to assess predictive performance comprehensively. These findings demonstrate that the hybrid HGB‑PSO method achieves superior accuracy and robustness in transformer fault prediction.
In recent years, sign language recognition has emerged as a major challenge in the fields of image processing and machine learning. People with hearing impairments use sign language to communicate, but the lack of automated tools to translate it has created significant communication barriers. This study presents a hybrid model based on convolutional neural networks (CNNs), transformers, and hidden Markov models (HMMs) to accurately recognize sign language gestures using the MNIST sign language dataset. The model first extracts image features from handwritten images using CNNs and then feeds these features into the Transformer model to process complex and long-term dependencies in the feature sequence. In the next step, to smooth the predictions and improve accuracy, a hidden Markov model is employed, which adjusts the final predictions based on previous sequences. The results show that the proposed model utilizing HMM achieves an accuracy of 99% and a sign error rate of 0.0098, demonstrating its high efficiency in recognizing hand gestures. This research represents an important step toward developing assistive devices for the deaf and enhancing human interaction.
In this study, an intelligent deep learning–based system is proposed for automated detection of surface defects in copper cathode blanks used in the electrorefining process. The proposed pipeline combines a YOLOv8-based segmentation model with an EfficientNetV2-S classifier to localize and analyze defect-relevant regions of each blank. The segmentation module identifies the main copper regions, edge strips, and defect-prone areas associated with surface anomalies such as scratches, dents, misalignment, and discoloration, effectively reducing background interference and improving classification reliability. The dataset includes 5,266 labeled images with a significant class imbalance, addressed using focal loss and class weighting during training. Experimental results on the test set demonstrate strong performance, achieving 98.32% accuracy, 96.71% precision, 95.67% recall, an F1-score of 96.19%, and an AUC of 0.9953. Grad-CAM visualizations and error analysis further confirm that the model consistently focuses on meaningful defect regions while remaining robust to background and illumination variations. These results highlight the effectiveness of the proposed approach for reliable quality control in industrial copper electrorefining lines.
Teleoperation systems are increasingly deployed in critical applications such as robotic surgery, industrial automation, and hazardous environment exploration. However, these systems are highly susceptible to network-induced delays, cyber-attacks, and system uncertainties, which can degrade performance and compromise safety. This paper proposes a Graph Neural Network (GNN)-based Digital Twin (DT) framework to enhance the cyber-resilience and predictive control of teleoperation systems. The GNN-based anomaly detection mechanism accurately identifies cyber-attacks, such as false data injection (FDI) and denial-of-service (DoS) attacks, with a detection rate of 24.3% and a false alarm rate of only 1.8%, significantly outperforming conventional machine learning methods. Furthermore, the predictive digital twin model, integrated with model predictive control (MPC), effectively compensates for latency and dynamic uncertainties, reducing control errors by 14.12% compared to traditional PID controllers. Simulation results in a robotic teleoperation testbed demonstrate a 24.4% improvement in trajectory tracking accuracy under variable delay conditions, ensuring precise and stable operation.
Accurate and reliable stock price prediction is both a formidable and essential task in financial markets, requiring the use of advanced techniques. This paper presents an innovative approach that integrates Long Short-Term Memory (LSTM) networks with Modified Complex Variational Mode Decomposition (MCVMD) for preprocessing and the Secretary Bird Optimization Algorithm (SBOA) for hyperparameter optimization. In the preprocessing phase, MCVMD decomposes stock price time series into intrinsic mode functions, effectively capturing complex patterns and reducing noise. To enhance predictive performance, SBOA optimizes both the hyperparameters of the LSTM network and the decomposition parameters of MCVMD. The proposed methodology is evaluated on datasets from six companies: Ferrari NV (RACE) and Intesa Sanpaolo (ISP) from Italy, Amadeus IT (AMA) and Repsol (REP) from Spain, and Hitachi Ltd (6501) and Chugai Pharmaceutical Co., Ltd. (4519) from Japan. Results show that the LSTM-MCVMD-SBOA model achieves lower error values compared with conventional benchmarks including ARIMA-GARCH, vanilla LSTM, Long Short-Term Memory-Particle Swarm Optimization (LSTM-PSO), and Long Short-Term Memory-Sine Cosine Algorithm (LSTM-SCA). Compared with these alternatives, SBOA was selected because of its superior balance between exploration and exploitation, inspired by secretary bird hunting and evasion behavior, which enables efficient search in complex optimization landscapes. Overall, the proposed model demonstrates significantly improved predictive accuracy over conventional methods, highlighting the efficacy of combining advanced decomposition with nature-inspired optimization for stock market forecasting.
Accurate prediction of drug–target interactions is essential for advancing drug discovery and repositioning efforts. This study introduces a comprehensive framework that effectively addresses key challenges in DTI prediction, including dataset imbalance and high-dimensional feature representations. The approach integrates multiple protein descriptors—specifically, nine statistical and sequence-based features—and drug molecular fingerprints encoded via Morgan algorithms, with optimal feature combinations selected through validation to capture diverse biological and chemical information. To mitigate dataset imbalance, a one-class SVM-based undersampling method (One-SVM-US) models the distribution of positive interactions to guide the selective reduction of the majority class, thereby effectively balancing positive and negative samples. Furthermore, a supervised, classification-oriented variational autoencoder is employed to compress the high-dimensional features into a lower-dimensional space while preserving class-discriminative information relevant to interaction prediction. The refined features are then classified using machine learning models to predict potential drug–target pairs. Experimental evaluations on benchmark datasets demonstrate the effectiveness of the proposed framework, with results showing perfect AUC-ROC scores of 1.00 on the EN, GPCR, and NR datasets, and a score of 0.9731 on the IC dataset, indicating performance improvements over existing methods. These findings confirm the robustness and potential of the approach as a reliable tool for drug–target interaction prediction.
Magnetic Resonance Imaging (MRI) often suffers from noise and Intensity Non-Uniformity (INU), making segmentation a challenging task. The Fuzzy C-Means (FCM) algorithm, a widely used clustering method for image segmentation, is highly sensitive to noise and its convergence rate depends on data distribution. FCM employs the Euclidean distance metric, which fails to adapt to variations in data point distributions within compact and similarly shaped clusters. Additionally, this metric is not locally adaptive to different cluster shapes. This paper introduces a Conditional Spatial Gustafson-Kessel Clustering Algorithm based on Information Theory (CSGKIT) to address these challenges. First, information theory is incorporated to enhance the algorithm's robustness against noise and improve segmentation accuracy. Second, the Mahalanobis distance replaces the Euclidean distance to better accommodate cluster shapes during the clustering process. Finally, a conditional spatial approach uses a fuzzy-weighted membership matrix to incorporate local spatial interactions between neighboring pixels. The proposed CSGKIT algorithm is evaluated on two datasets: the BrainWeb simulated dataset and the Open Access Series of Imaging Studies (OASIS) dataset. Experimental results indicate that CSGKIT outperforms other FCM-based algorithms in segmentation accuracy across various tissue types.
The Convolutional Restricted Boltzmann Machine (CRBM) is a generative model that extracts representations from unlabeled data, achieving success in various applications. However, its unsupervised nature may yield suboptimal representations for specific classification tasks. This paper proposes adapting k-means clustering to enhance CRBM parameters, aligning features with informative cluster centers. A novel criterion combining generative and soft-K-Means objectives optimizes both cluster centers and CRBM parameters, allowing for continued unsupervised feature learning.Experiments on MNIST, CIFAR10, and three facial expression datasets (JAFFE, KANADE, BU) show that the proposed method enhances the learning process and offers a more informative representation compared to standard and classification CRBM.
Detecting driver distraction is critically important, as it remains a major contributor to road accidents and traffic-related injuries worldwide. This study introduces a novel hybrid deep learning model that integrates Spatio-Temporal Graph Convolutional Networks (ST-GCN) with a Transformer Encoder and Attention mechanisms to effectively detect distracted driving behaviors. The ST-GCN component captures spatial and temporal dependencies in 3D skeletal motion data, modeling the dynamic body movements of the driver. Following this, a Transformer Encoder is employed to further refine temporal representations by leveraging global attention, allowing the model to understand long-range dependencies and subtle behavioral patterns over time. In addition, an Attention mechanism is applied to emphasize the most informative joints and time frames. To address class imbalance in the dataset, the model uses a focal loss function, which helps focus training on more difficult-to-classify examples. The proposed approach is validated on the 3D skeletal Drive&Act dataset, where it achieves a high accuracy of 97.47%, outperforming existing models, particularly under challenging conditions such as poor lighting and complex driving environments. The system demonstrates strong potential for real-time driver monitoring, offering an intelligent solution to enhance road safety and reduce accident risks through early detection of driver distraction.
