Megat Norulazmi Megat Mohamed Noor
Research Subjects
Megat Norulazmi Megat Mohamed Noor has a total of 11 co-authors. Here are the top co-authors by number of shared publications:
- Ahmad Sanmorino
Universitas Indo Global Mandiri
1 shared publication - An'NIsa Pratama Putri
Universitas Muslim Indonesia
1 shared publication - Herdianti Darwis
Universitas Muslim Indonesia Makassar Indonesia
1 shared publication Herman
1 shared publication- Ihwana As'AD
Universitas Muslim Indonesia
1 shared publication - Irawati
Universitas Muslim Indonesia
1 shared publication
Latest Publications
- Open Access
- 24 November 2025
- Published by Manuscript Technomedia LLP in Information Research Communications
- Vol. 2 (2) , 175-181
Evaluating Small BERT Based Models on Automated Essay Scoring Task
Automated Essay Scoring (AES) systems address the limitations of manual grading, such as inefficiency, subjectivity, and scalability issues in educational assessments, and this study evaluates the performance of lightweight BERT-based models for AES tasks, aiming to identify the most effective and computationally efficient variant for integration into a proposed AI Essay Score Bot by focusing on smaller, distilled transformer models to balance high accuracy with reduced resource demands, building on advancements in Natural Language Processing (NLP) from models like BERT, RoBERTa, and their variants. The publicly available ASAP-AES dataset was used, encompassing essays from prompts 1 to 8, with eleven small-scale BERT-based models tested: DistilBERT-base-uncased, DistilBERT-base-uncased-distilled-SQuAD, ALBERT-base-v1, ALBERT-base-v2, DistilRoBERTa-base, SqueezeBERT-uncased, SqueezeBERT-MNLI, SqueezeBERT-MNLI-headless, BERT-base-uncased, RoBERTa-base, and BORT; training involved a 5-fold cross-validation with 80% training and 20% validation splits, hyperparameter tuning across batch sizes (8, 16, 20), learning rates (1e-4, 3e-4, 3e-5, 4e-5, 5e-5), and epochs (5, 10, 15, 20), while the EXPATS toolkit facilitated model implementation, training, and evaluation under an in-domain schema, with performance measured using the Quadratic Weighted Kappa (QWK) metric. DistilBERT-base-uncased achieved the highest average QWK of 0.926 (batch size 16, 10 epochs), outperforming others across most prompts, with improvements noted in hyperparameter configurations (e.g., learning rate 3e-5), while ALBERT-base-v1 followed closely with a maximum QWK of 0.920 (batch size 8, 20 epochs), despite GPU memory constraints limiting batch sizes; smaller models like DistilRoBERTa-base (average QWK 0.907) and SqueezeBERT-uncased (0.910) surpassed larger counterparts such as BERT-base-uncased (0.903) and RoBERTa-base (0.860), with prompts 1 and 7 consistently challenging, where SqueezeBERT-uncased scored highest on Prompt 1 (QWK 0.880) and SqueezeBERT-MNLI on Prompt 7 (0.780), and underperforming models included BORT (average 0.770) and ALBERT-base-v2 (0.830), affected by architectural simplifications. The superior performance of distilled models like DistilBERT and ALBERT underscores the benefits of knowledge distillation and parameter optimization techniques, enabling better QWK scores with lower computational overhead compared to full-sized BERT and RoBERTa, with batch size increases (up to 20) enhancing DistilBERT variants, particularly on difficult prompts, while ALBERT’s efficiency stemmed from factorized embeddings and cross-layer sharing; challenges on Prompts 1 and 7 suggest dataset-specific complexities, such as varied essay structures, indicating that lightweight models are ideal for resource-limited applications, though further distillation of ALBERT could yield even more compact variants without sacrificing precision. Distilled transformer models, especially DistilBERT-base-uncased and ALBERT-base-v1, offer an optimal trade-off between accuracy and efficiency for AES tasks, making them suitable for real-world deployment in web-based AI tools, and future research should investigate distilling ALBERT to enhance compactness and explore transfer learning for cross-domain AES improvements.
- 24 November 2025
- Published by Manuscript Technomedia LLP in Information Research Communications
- Vol. 2 (2) , 175-181
Open AccessEvaluating Small BERT Based Models on Automated Essay Scoring Task
Automated Essay Scoring (AES) systems address the limitations of manual grading, such as inefficiency, subjectivity, and scalability issues in educational assessments, and this study evaluates the performance of lightweight BERT-based models for AES tasks, aiming to identify the most effective and computationally efficient variant for integration into a proposed AI Essay Score Bot by focusing on smaller, distilled transformer models to balance high accuracy with reduced resource demands, building on advancements in Natural Language Processing (NLP) from models like BERT, RoBERTa, and their variants. The publicly available ASAP-AES dataset was used, encompassing essays from prompts 1 to 8, with eleven small-scale BERT-based models tested: DistilBERT-base-uncased, DistilBERT-base-uncased-distilled-SQuAD, ALBERT-base-v1, ALBERT-base-v2, DistilRoBERTa-base, SqueezeBERT-uncased, SqueezeBERT-MNLI, SqueezeBERT-MNLI-headless, BERT-base-uncased, RoBERTa-base, and BORT; training involved a 5-fold cross-validation with 80% training and 20% validation splits, hyperparameter tuning across batch sizes (8, 16, 20), learning rates (1e-4, 3e-4, 3e-5, 4e-5, 5e-5), and epochs (5, 10, 15, 20), while the EXPATS toolkit facilitated model implementation, training, and evaluation under an in-domain schema, with performance measured using the Quadratic Weighted Kappa (QWK) metric. DistilBERT-base-uncased achieved the highest average QWK of 0.926 (batch size 16, 10 epochs), outperforming others across most prompts, with improvements noted in hyperparameter configurations (e.g., learning rate 3e-5), while ALBERT-base-v1 followed closely with a maximum QWK of 0.920 (batch size 8, 20 epochs), despite GPU memory constraints limiting batch sizes; smaller models like DistilRoBERTa-base (average QWK 0.907) and SqueezeBERT-uncased (0.910) surpassed larger counterparts such as BERT-base-uncased (0.903) and RoBERTa-base (0.860), with prompts 1 and 7 consistently challenging, where SqueezeBERT-uncased scored highest on Prompt 1 (QWK 0.880) and SqueezeBERT-MNLI on Prompt 7 (0.780), and underperforming models included BORT (average 0.770) and ALBERT-base-v2 (0.830), affected by architectural simplifications. The superior performance of distilled models like DistilBERT and ALBERT underscores the benefits of knowledge distillation and parameter optimization techniques, enabling better QWK scores with lower computational overhead compared to full-sized BERT and RoBERTa, with batch size increases (up to 20) enhancing DistilBERT variants, particularly on difficult prompts, while ALBERT’s efficiency stemmed from factorized embeddings and cross-layer sharing; challenges on Prompts 1 and 7 suggest dataset-specific complexities, such as varied essay structures, indicating that lightweight models are ideal for resource-limited applications, though further distillation of ALBERT could yield even more compact variants without sacrificing precision. Distilled transformer models, especially DistilBERT-base-uncased and ALBERT-base-v1, offer an optimal trade-off between accuracy and efficiency for AES tasks, making them suitable for real-world deployment in web-based AI tools, and future research should investigate distilling ALBERT to enhance compactness and explore transfer learning for cross-domain AES improvements.
- Open Access
- 31 March 2025
- Published by Yocto Brain in Indonesian Journal of Data and Science
- Vol. 6 (1) , 123-133
A Comperative Study on Efficacy of CNN VGG-16, DenseNet121, ResNet50V2, And EfficientNetB0 in Toraja Carving Classification
Passura', or Toraja carvings, are a significant part of the culture in Toraja and have deep symbolic significance and beautiful patterns. Inspired by folklore and natural elements, these motifs capture the harmonious coexistence of nature, the divine, and humans. In order to classify Toraja carving motifs, this study uses Convolutional Neural Networks (CNN), specifically the architectures VGG-16, DenseNet121, ResNet50V2, and EfficientNetB0. Seven motif types were represented in the 700 photos that made up the dataset, which was split between 80% training and 20% validation data. Despite showing signs of overfitting, the EfficientNetB0 model had the greatest validation accuracy of 98%. ResNet50V2 was the most successful model for this job, exhibiting strong performance with 95.33% accuracy. These findings demonstrate how CNN architectures, with precise motif classification, can be used to preserve and record cultural heritage. This study advances the field of digital cultural preservation and lays the groundwork for further research using bigger datasets to improve model dependability and reduce overfitting.
- 31 March 2025
- Published by Yocto Brain in Indonesian Journal of Data and Science
- Vol. 6 (1) , 123-133
Open AccessA Comperative Study on Efficacy of CNN VGG-16, DenseNet121, ResNet50V2, And EfficientNetB0 in Toraja Carving Classification
Passura', or Toraja carvings, are a significant part of the culture in Toraja and have deep symbolic significance and beautiful patterns. Inspired by folklore and natural elements, these motifs capture the harmonious coexistence of nature, the divine, and humans. In order to classify Toraja carving motifs, this study uses Convolutional Neural Networks (CNN), specifically the architectures VGG-16, DenseNet121, ResNet50V2, and EfficientNetB0. Seven motif types were represented in the 700 photos that made up the dataset, which was split between 80% training and 20% validation data. Despite showing signs of overfitting, the EfficientNetB0 model had the greatest validation accuracy of 98%. ResNet50V2 was the most successful model for this job, exhibiting strong performance with 95.33% accuracy. These findings demonstrate how CNN architectures, with precise motif classification, can be used to preserve and record cultural heritage. This study advances the field of digital cultural preservation and lays the groundwork for further research using bigger datasets to improve model dependability and reduce overfitting.
- Open Access
- 27 August 2022
- Published by Universitas Atma Jaya Yogyakarta in Indonesian Journal of Information Systems
- Vol. 5 (1) , 64-73
Machine Learning for Clustering Regencies-Cities Based on Inflation and Poverty Rates in Indonesia
The COVID-19 pandemic has increased inflation and poverty rates in many cities, thus requiring considerable attention from the government as a policymaker. Therefore, this study aims to cluster regencies/cities that need mitigation priorities from the Indonesian government based on inflation and poverty rates in 2021. Four machine learning methods, namely k-Means (KM), Partitioning around medoids (PAM), Ward, and Divisive analysis (Diana) are utilized and compared to achieve that purpose. Clustering 90 regencies/cities in Indonesia produced five optimal clusters. Furthermore, the clustering results were validated using the Silhouette width (SW) and Dunn index (DI). The results showed that the k-means method produced the most compact cluster. Hence, this study's results can be utilized as a reference for the government in determining the steps and priorities of economic policy in Indonesia.
- 27 August 2022
- Published by Universitas Atma Jaya Yogyakarta in Indonesian Journal of Information Systems
- Vol. 5 (1) , 64-73
Open AccessMachine Learning for Clustering Regencies-Cities Based on Inflation and Poverty Rates in Indonesia
The COVID-19 pandemic has increased inflation and poverty rates in many cities, thus requiring considerable attention from the government as a policymaker. Therefore, this study aims to cluster regencies/cities that need mitigation priorities from the Indonesian government based on inflation and poverty rates in 2021. Four machine learning methods, namely k-Means (KM), Partitioning around medoids (PAM), Ward, and Divisive analysis (Diana) are utilized and compared to achieve that purpose. Clustering 90 regencies/cities in Indonesia produced five optimal clusters. Furthermore, the clustering results were validated using the Silhouette width (SW) and Dunn index (DI). The results showed that the k-means method produced the most compact cluster. Hence, this study's results can be utilized as a reference for the government in determining the steps and priorities of economic policy in Indonesia.