Articles

Access the latest knowledge in applied science, electrical engineering, computer science and information technology, education, and health.

Advanced Search
Filter Icon

Filters article

Years

FAQ Arrow
Single Year
Range Year
0
0

Source Title

FAQ Arrow

Authors

FAQ Arrow

30,185 Article Results

IndoBERT for educational assessment: comparative analysis of transformer models in Indonesian question generation

10.11591/ijai.v15.i2.pp1804-1813
Handaru Jati , Yuniar Indrihapsari , Pradana Setialana , Danang Wijaya , Satya Adhiyaksa Ardy , Dhista Dwi Nur Ardiansyah
This study asks whether a monolingual encoder can realistically outperform multilingual and larger transformer models for Indonesian automatic question generation (AQG) when all models share the same training budget. We compare Indonesian bidirectional encoder representations from transformers (IndoBERT), multilingual BERT (mBERT), and BERT-large using a single fine-tuning pipeline with answer highlighting, applied to an Indonesian version of TyDiQA-GoldP and a 20,000 translated subset of SQuAD 2.0. We evaluate model quality using bilingual evaluation understudy score n-gram 4 (BLEU-4), metric for evaluation of translation with explicit ordering (METEOR), and ROUGE-Lincoln (ROUGE-L). IndoBERT consistently achieves the best scores on both datasets (e.g., BLEU-4 of 19.69 on TyDiQA-GoldP and 3.79 on the SQuAD 2.0 subset) while requiring less computation than mBERT and BERT-large. Our results show that language-specific pretraining gives clear advantages for Indonesian AQG, yielding higher accuracy at lower computational cost than multilingual or larger encoders. The work closes a gap in Indonesian AQG benchmarking by providing the first head-to-head comparison of IndoBERT, mBERT, and BERT-large under a shared fine-tuning and evaluation protocol. For educational assessment, the findings offer a practical recipe for building deployable AQG systems on mid-range GPUs that generate higher quality questions without prohibitive training or inference budgets.
Volume: 15
Issue: 2
Page: 1804-1813
Publish at: 2026-04-01

Drone-assisted deep learning weed detection for sustainable agriculture and environmental resilience

10.11591/ijai.v15.i2.pp1428-1440
Agustan Latif , Handaru Jati , Herman Dwi Surjono , Mani Yusuf
Effective weed detection plays a crucial role in sustainable agriculture, boosting crop productivity and supporting environmental conservation. This study compares three deep learning models—YOLOv5, YOLO-NAS, and mask region-based convolutional neural network (Mask R-CNN)-against traditional methods in terms of accuracy, processing speed, and adaptability in tropical agricultural conditions, with Merauke, Indonesia, as the case study. The results show that YOLO-NAS delivers the highest accuracy at 96% with a processing time of 25 ms per image, making it suitable for high precision applications. YOLOv5 balances strong accuracy (94%) with faster processing at 12 ms per image, establishing it as the most effective for real time scenarios. Mask R-CNN also achieves 94% accuracy and provides advanced segmentation capabilities, but its slower processing speed of 31 ms limits large-scale implementation. Traditional methods perform poorly in comparison, with only 85% accuracy and processing time above 50 ms per image. These findings highlight the transformative potential of artificial intelligence (AI)-based weed detection for precision agriculture, particularly in tropical regions like Merauke. Adoption of models such as YOLOv5 reduces manual labor dependence while advancing efficient, eco-friendly weed management. Future research should expand datasets and explore newer models like YOLOv8, YOLO-NAS, vision transformers (ViTs), and hybrid approaches.
Volume: 15
Issue: 2
Page: 1428-1440
Publish at: 2026-04-01

Image feature extraction for road surface damage classification

10.11591/ijai.v15.i2.pp1578-1592
Octaviani Hutapea , Sarifuddin Madenda , Hustinawaty Hustinawaty , Iffatul Mardhiyah
Road surface deterioration poses a critical risk to driving safety and comfort, necessitating timely and accurate detection to support effective maintenance. Manual inspection methods are often inefficient, underscoring the need for automated approaches based on computer vision. This study investigates the integration of feature extraction techniques histogram of oriented gradients (HOG) and local binary pattern (LBP) with convolutional neural network (CNN) architectures ResNet50 and InceptionV3 for the classification of road damage. A dataset of 1,580 images was categorized into five damage types: alligator crack, longitudinal crack, other crack, patching, and potholes. Experimental results indicate that HOG–ResNet50 achieved 79% accuracy, while LBP–InceptionV3 yielded the best performance at 97%. The contributions of this study are threefold: i) an automated framework is proposed that combines texture-based features with deep learning for road damage detection, ii) the LBP–InceptionV3 combination is shown to provide superior accuracy compared to conventional pairings, and iii) the approach offers a scalable and reliable alternative to manual inspection methods, supporting more efficient road maintenance planning.
Volume: 15
Issue: 2
Page: 1578-1592
Publish at: 2026-04-01

The effects of data imbalance on fraud detection model accuracy

10.11591/ijai.v15.i2.pp1402-1408
Rusma Anieza Ruslan , Nureize Arbaiy , Pei-Chun Lin
Machine learning (ML) model performance is often assessed by accuracy, but the quality and balance of data also play crucial roles. Imbalanced datasets, where the minority class has fewer samples than the majority class, can lead to biased predictions favoring the majority class. This study addresses the issue of class imbalance through resampling techniques, including random undersampling (RUS) and random oversampling (ROS), specifically applied to a fraud detection dataset. We classify the resampled datasets using random forest (RF) and gradient boosting (GB) models. Our findings indicate that the RF model, when combined with ROS, achieves an accuracy of 97.4%, surpassing the 96.1% accuracy of the GB model with RUS. This approach demonstrates the importance of addressing class imbalance to improve prediction accuracy in ML.
Volume: 15
Issue: 2
Page: 1402-1408
Publish at: 2026-04-01

An intelligent and explainable IoT-Edge-Cloud architecture for real-time water quality monitoring

10.11591/ijai.v15.i2.pp1109-1120
Sara Bouziane , Badraddine Aghoutane , Aniss Moumen , Anas El Ouali , Ali Essahlaoui , Abdellah El Hmaidi
Continuous and reliable monitoring of water quality is critical for early detection of environmental deterioration, yet conventional monitoring approaches are often slow and lack timely data availability. This study proposes an intelligent and explainable internet of things (IoT)–Edge–Cloud architecture to monitor water quality in real time, using IoT sensing, edge based artificial intelligence (Edge AI), cloud-stream processing, and explainable artificial intelligence (XAI). The system calculates the water quality index (WQI) directly at the edge and predicts its evolution using a stacking ensemble model trained on physicochemical measurements taken from the Moulouya River Basin in Morocco. An explainability module based on Shapley additive explanations (SHAP) values gives a clearer image of the contribution of various parameters to WQI predictions, providing transparency of the features, which builds trust in the model’s output. The proposed architecture was implemented as an end-to-end prototype and validated using a simulation-based experimental that mimicked realistic sensor dynamics and connectivity interruptions. The experimental results show strong predictive performance (R² =0.945), stable system operations, and reliable interpretability highlighting the potential of the proposed approach for scalable, intelligent, and transparent environmental monitoring.
Volume: 15
Issue: 2
Page: 1109-1120
Publish at: 2026-04-01

A hybrid model for enhanced aspect-based sentiment analysis using large language models

10.11591/ijai.v15.i2.pp1825-1838
Mohammed Ziaulla , Arun Biradar
Aspect-based sentiment analysis (ABSA) is a crucial task within natural language processing (NLP), enabling fine-grained opinion mining by identifying sentiments associated with specific aspects of a product or service. While transformer-based models like bidirectional encoder representations from transformers (BERT) have improved sentiment classification, they still struggle with limited contextual adaptability, especially in customer reviews containing complex expressions. Most existing approaches rely heavily on benchmark datasets such as semantic evaluation (SemEval) and multi-aspect multi-sentiment (MAMS), which do not fully capture the diversity of real-world review scenarios. Hence, this research addresses these limitations by proposing a novel hybrid model, called as hybrid-BERT (H-BERT), that integrates span-aware BERT (SpanBERT) with bidirectional long short-term memory (BiLSTM), conditional random field (CRF), and large language models (LLMs). The objective is to enhance aspect extraction and sentiment classification performance using both annotated and synthetic data. The methodology includes preprocessing, hybrid model training, and evaluation using the SemEval 2014 dataset. Experimental results show that H-BERT achieved 90.58% accuracy and 90.56% F-score in the laptop domain and 91.21% accuracy with a 92.03% F-score in the restaurant domain. These results outperform existing models, confirming H-BERT’s robustness and effectiveness. In conclusion, H-BERT improves sentiment understanding in customer reviews.
Volume: 15
Issue: 2
Page: 1825-1838
Publish at: 2026-04-01

Accurate stroke area classification using extreme gradient boosting with multi-feature extraction

10.11591/ijai.v15.i2.pp1390-1401
Kavikondala Praveen Kumar Rao , Maha Lakshmi Bondla , Bommaraju Srinivasa Rao , Ambidi Naveena , K. V. Balaramakrishna , Srinivasarao Goda
Stroke, one of the most common neurological disorders leading to long-term disability and mortality, requires accurate detection of affected brain regions for timely treatment planning. However, conventional deep learning models face challenges in achieving precise segmentation and robust classification due to noisy inputs, weak feature representation, and poor generalization. To address these gaps, this study introduces a hybrid framework that integrates the ConvNeXt architecture for stroke region segmentation with XGBoost based classification, strengthened through three complementary feature extraction methods: local binary patterns (LBP), adaptive threshold directional binary gradient matrix (AT-DBGM), and wavelet packet transform (WPT). These methods capture textural, directional, and multi resolution features, which are concatenated into a stacked vector and classified using XGBoost. Preprocessing steps, including normalization and resizing, ensure improved input consistency. Experimental evaluations on benchmark stroke imaging datasets show that the proposed framework achieves 98.56% Dice similarity coefficient (DSC), 12.96 mm Hausdorff distance (HD), 99.12% accuracy, 98.69% sensitivity, 99.06% specificity, 98.98% precision, and 98.85% F1-score.
Volume: 15
Issue: 2
Page: 1390-1401
Publish at: 2026-04-01

Blockchain-enabled framework using diversity mutation with siberian tiger optimization for offloading in fog computing

10.11591/ijai.v15.i2.pp1371-1380
Srikanta Murthy Rajini , Reginald Shilpa
Fog computing has developed as a promising framework to support latency sensitive internet of things (IoT) applications for mobile devices operating in dynamic environments. During the offloading process, malicious activities interrupt the existing methods, which increases the execution time. Therefore, this research proposes a diversity mutation with siberian tiger optimization (DM-STO) for computation offloading in blockchain based fog computing. The blockchain is used to secure offload and attain quality of service (QoS) mobile users with less energy consumption and execution time. The DM-STO can balance workloads among local devices and fog servers. The diversity mutation operation improves the exploration ability to dynamic network conditions, leading to efficient computational offloading in fog computing. The execution time, service cost and energy consumption are evaluated to calculate the performance of the proposed DM-STO with varying numbers of IoT requests such as 50, 100, 200, and 300. For 50 IoT requests with a fixed fog server of 10, the DM-STO achieves an execution time of 18 s, a service cost of 10$ and energy consumption of 5 mJ compared to the BAT algorithm.
Volume: 15
Issue: 2
Page: 1371-1380
Publish at: 2026-04-01

Structured data collection and deep learning for retinal OCT image-to-text translation: a comprehensive framework

10.11591/ijai.v15.i2.pp1050-1061
Uday Mande , Shafi Pathan , Pankaj Chandre , Sharvari Mande
This paper presents a comprehensive framework for structured data collection and deep learning (DL)-based translation of retinal optical coherence tomography (OCT) images into diagnostic text. The suggested approach guarantees high-quality OCT data for model training through the use of sophisticated image processing methods like edge detection, noise suppression, and contrast improvement. The study utilizes 84,484 retinal images from the OCT dataset available on Kaggle. The research utilizes various preprocessing techniques, such as median and Gaussian filtering, along with data augmentation strategies like translation, rotation, and scaling, to mitigate class imbalances and improve model performance. The system automatically identifies and categorizes retinal diseases such as drusen, diabetic macular edema (DME), and choroidal neovascularization (CNV) by integrating feature extraction and selection with DL techniques. The research highlights the importance of effective data handling and model scalability to address the increasing need for automated diagnostic tools in ophthalmology. This framework aims to support ophthalmologists in managing the increasing incidence of diabetic retinopathy (DR) and other retinal conditions by enhancing the efficiency of retinal image analysis, thereby improving patient results through early detection and treatment.
Volume: 15
Issue: 2
Page: 1050-1061
Publish at: 2026-04-01

Deep neural network classification in chatbot system family health counseling services

10.11591/ijai.v15.i2.pp1211-1218
Andi Riansyah , Sam Farisa Chaerul Haviana , Ratna Supradewi , Muhammad Ainul Wahib
Mental health problems affect many aspects of life, including physical well being, work productivity, social functioning, and suicide risk. In Indonesia, access to professional mental health services remains very limited: only a small proportion of people with depression receive treatment and the number of mental health professionals per population is far below international recommendations, creating an urgent service gap. This study proposes an artificial intelligence–based chatbot to support family mental health counseling services in Indonesia. The chatbot uses a deep neural network (DNN) to classify user questions into counseling intent categories and to provide appropriate responses. Psychologists compiled and verified a dataset of Indonesian counseling questions and responses, which was then pre processed using standard text processing techniques and encoded with a bag of words (BoW) representation. A fully connected DNN with one input layer, two hidden layers of eight neurons each, and a SoftMax output layer was trained using the Adam optimizer (learning rate 0.01) on 80% of the data and evaluated on the remaining 20%. The best configuration achieved a training accuracy of 96%, with test results of 93% accuracy, 92% precision, 93% recall, and 92% F1-score. These findings indicate that proposed DNN based chatbot can accurately classify counseling intents and generate contextually appropriate responses, suggesting its potential as complementary tool to support initial family mental health counseling in Indonesia.
Volume: 15
Issue: 2
Page: 1211-1218
Publish at: 2026-04-01

TunDC: a public benchmark dataset for sentiment analysis and language modeling in the Tunisian dialect

10.11591/ijai.v15.i2.pp1891-1908
Ahmed Khalil Boulahia , Mourad Mars
The development of natural language processing (NLP) applications has increasingly focused on dialectal variations of languages. The Tunisian dialect (TD), a widely spoken variant of Arabic, poses unique linguistic challenges due to its lack of standardized writing conventions and influences from multiple languages, including French, Italian, Turkish, and Berber. In this work, we introduce TunDC, a dataset of 20,044 labeled comments designed to advance NLP research on the TD. The dataset covers diverse linguistic forms (Arabic, Latin, and mixed scripts), and each comment was manually annotated for positive or negative sentiment by native speakers, achieving high inter-annotator agreement. To evaluate its effectiveness, we fine-tuned various models on TunDC. The bert-base-arabic-TunDC-mixed model achieved an accuracy of 0.84 and a macro-averaged F1-score of 0.83, demonstrating strong generalization across sentiment categories and writing systems. A stratified data-splitting strategy considering both sentiment and script type further improved accuracy by approximately 8% compared to standard splits. As a publicly available resource, TunDC contributes to the computational linguistics community, fostering advancements in language modeling and applications tailored to the TD.
Volume: 15
Issue: 2
Page: 1891-1908
Publish at: 2026-04-01

Sentiment-aware user-item recommendation combining weighted XGBoost and optimized similarity metrics

10.11591/ijai.v15.i2.pp1851-1862
Snehal Bhogan , Vijay S. Rajpurohit , Sanjeev S. Sannakki
User-item recommendation systems play a vital role in enhancing personalized digital experiences across e-commerce and social media platforms. Traditional recommendation approaches, such as collaborative filtering (CF) and content-based filtering (CBF), often suffer from challenges like data sparsity, cold-start issues, and limited contextual understanding. Sentiment-aware recommendation systems have emerged as a promising solution by incorporating emotional insights extracted from user reviews, thereby improving recommendation accuracy and personalization. This study proposes a novel sentiment-aware user-item recommendation system (SAUIRS) framework that integrates optimized term frequency inverse document frequency (O-TF-IDF), parameterized bidirectional encoder representations from transformers (P-BERT), weighted extreme gradient boosting (WXGBoost), and an optimized similarity metrics model. The optimized TF-IDF enhances feature selection, reducing dimensionality while preserving relevant textual information. P-BERT, a fine-tuned BERT model, improves sentiment classification accuracy by leveraging deep contextual embeddings. WXGBoost further refined sentiment predictions, addressing class imbalance and enhancing model robustness. The extracted sentiment information is incorporated into an optimized similarity metrics model to improve recommendation precision by aligning user preferences with sentiment-driven insights. Extensive experiments conducted on Amazon benchmark datasets demonstrate the superior performance in terms of accuracy, root mean square error (RMSE), and mean absolute error (MAE) of the proposed framework compared to state-of-the-art recommendation models.
Volume: 15
Issue: 2
Page: 1851-1862
Publish at: 2026-04-01

Fetal organ detection using feature enhancement with attention and residual block

10.11591/ijai.v15.i2.pp1593-1604
Nuswil Bernolian , Siti Nurmaini , Ade Iriani Sapitri , Annisa Darmawahyuni , Muhammad Naufal Rachmatullah , Bambang Tutuko , Firdaus Firdaus
The rapid advancements in fetal ultrasonography have significantly enhanced prenatal diagnosis in recent years. Deep learning (DL) architectures have further streamlined the process of organ detection, improved diagnostic accuracy, and reduced observer dependency. This study proposes a computer-aided DL approach for fetal organ segmentation using the you only look once (YOLO) algorithm, a state-of-the-art method for object detection and image segmentation. This study identified and classified 15 fetal organs, including the umbilical vein, stomach, abdomen, brain (trans-cerebellum, trans-thalamic, and trans-ventricular regions), femur, head, thorax (chest cavity), heart (circumference, left atrium, left ventricle, right atrium, right ventricle), and aorta. We compared the performance of YOLOv7, YOLOv8, YOLOv9, and YOLOv11 architectures. The results showed that YOLOv9 outperformed YOLOv7, YOLOv8, and YOLOv11 achieving mAP50 and mAP95 scores of 91.90% and 94.50%, respectively. This performance surpasses previous studies that focused on classifying only a limited number of fetal organs.
Volume: 15
Issue: 2
Page: 1593-1604
Publish at: 2026-04-01

Automated bacteria and fungi classification using convolutional neural network on embedded system

10.11591/ijai.v15.i2.pp1132-1142
Tarik Bouganssa , Maryem Ait Moulay , Samar Aarabi , Abedelali Lasfar , Abdelatif EL Afia
In this study, we created and applied novel concepts for hardware-based image identification and categorization. For artificial intelligence (AI) and image recognition applications, this includes putting algorithms for recognizing colors, textures, and shapes into practice. Our contribution uses an embedded device with a camera and a microcomputer (Raspberry-Pi4 type) to replace the optical assessment of Petri dishes. Our object recognition system processes images efficiently by using a state-of-the-art kernel function and a new neighborhood architecture. Using the well-known convolutional neural network (CNN) architecture, YOLOv8, as a pre-trained model, we evaluated the proposed CNN-based method for object recognition in a number of demanding scenarios. Several Petri plates, uncontrolled settings, and different backgrounds and illumination were used to evaluate the technology. Our dynamic mode integrates a CNN network with an attention mask to highlight the traits of bacteria and fungi, ensuring robust recognition. We implemented our algorithm on a Raspberry Pi 400, connected to a CMOS 3.0 camera sensor and a human-machine interface (HMI) for instant display of results.
Volume: 15
Issue: 2
Page: 1132-1142
Publish at: 2026-04-01

Identification of chemical markers for species differentiation in Aquilaria essential oils using self-organizing maps

10.11591/ijai.v15.i2.pp1339-1348
Nur Athirah Syafiqah Noramli , Muhammad Ikhsan Roslan , Noor Aida Syakira Ahmad Sabri , Nurlaila Ismail , Zakiah Mohd Yusoff , Mohd Nasir Taib
This study analyzes the chemical diversity of essential oils from four Aquilaria species, A. beccariana, A. malaccensis, A. crassna, and A. subintegra, which are important sources of agarwood used in perfumery and traditional medicine. Despite their economic and ecological value, the chemical profiles of these species remain insufficiently characterized, hindering accurate species differentiation and resource management. This research aims to identify distinctive chemical patterns to improve species classification. Self-organizing maps (SOMs) were employed to analyze complex chemical composition data and to identify significant compounds responsible for species separation. The analysis revealed several compounds with strong discriminatory power and species-specific distribution patterns, with compounds C, D, and E identified as the most significant markers. These findings demonstrate substantial biochemical diversity among Aquilaria species and confirm the effectiveness of SOM for essential oil profiling. The results support improved species identification and have important implications for ecological conservation, sustainable agarwood management, and pharmacological development.
Volume: 15
Issue: 2
Page: 1339-1348
Publish at: 2026-04-01
1 2 3
...
2012 2013
Show 32 of 2013

Discover Our Library

Embark on a journey through our expansive collection of articles and let curiosity lead your path to innovation.

Explore Now
Library 3D Ilustration