Pusadee Seresangtakul
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Pusadee Seresangtakul has a total of 51 co-authors. Here are the top co-authors by number of shared publications:
- Pongsathon Janyoi
College of Computing, Khon Kaen University,Khon Kaen,Thailand
3 shared publications Arounyadeth Srithirath
Department of Computer Science, Khon Kaen University, Thailand, 40002
2 shared publicationsChitsutha Soomlek
Department of Computer Science Prague University of Economics and Business Czech Republic
2 shared publicationsPaweena Unlee
Department of Innovation and Computer Education, Faculty of Education, Sakon Nakhon Rajabhat University, 47000, Thailand
2 shared publicationsPongsathon Janyoi
Khon Kaen Univ, Dept Comp Sci, Nat Language & Speech Proc Lab, Khon Kaen, Thailand
2 shared publications- Pongsathon Janyoi
College of Computing, Khon Kaen University, Muang, Khon Kaen, Thailand
2 shared publications
Latest Publications
- Open Access
- 23 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 45053-45070
Transformer-Based Grapheme-to-Phoneme Conversion With Knowledge Transfer for Isarn
Past research has demonstrated the significant advantages of the transformer model in grapheme-to-phoneme (G2P) conversion. Transformers typically require a large amount of training data to generalize well, and this poses a challenge in the context of dialects for which limited data are available. In this paper, we propose a transformer-based G2P model for Isarn, which is a dialect of Thai. The proposed model is initially trained on a Thai language dataset to acquire pronunciation-related knowledge, and is then transferred and adapted to the Isarn dialect by using the appropriate representations of graphemes, including syllable boundaries, phonemes, and tones. The proposed transfer learning-based model achieved a word error rate (WER) of 15.63% and phoneme error rate (PER) of 3.98%, surpassing the transformer-based model trained from scratch by using only Isarn data that obtained a WER of 22.96% and a PER of 8.64%, as well as multilingual models such as Multilingual ByT5 G2P (17.86% WER, 5.06% PER). These evaluations, which were carried out on unseen Isarn words and loan words borrowed from Thai, offer compelling evidence that the proposed transfer learning-based model excels not only in comparison with a transformer-based model trained from scratch by using solely Isarn data, but also against other baseline models.
- 23 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 45053-45070
Open AccessTransformer-Based Grapheme-to-Phoneme Conversion With Knowledge Transfer for Isarn
Past research has demonstrated the significant advantages of the transformer model in grapheme-to-phoneme (G2P) conversion. Transformers typically require a large amount of training data to generalize well, and this poses a challenge in the context of dialects for which limited data are available. In this paper, we propose a transformer-based G2P model for Isarn, which is a dialect of Thai. The proposed model is initially trained on a Thai language dataset to acquire pronunciation-related knowledge, and is then transferred and adapted to the Isarn dialect by using the appropriate representations of graphemes, including syllable boundaries, phonemes, and tones. The proposed transfer learning-based model achieved a word error rate (WER) of 15.63% and phoneme error rate (PER) of 3.98%, surpassing the transformer-based model trained from scratch by using only Isarn data that obtained a WER of 22.96% and a PER of 8.64%, as well as multilingual models such as Multilingual ByT5 G2P (17.86% WER, 5.06% PER). These evaluations, which were carried out on unseen Isarn words and loan words borrowed from Thai, offer compelling evidence that the proposed transfer learning-based model excels not only in comparison with a transformer-based model trained from scratch by using solely Isarn data, but also against other baseline models.
- Open Access
- 11 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 40423-40447
A Multitask Learning Framework for Multigranular Causality Detection and Extraction in Indonesian EMRs
Extracting causal relations from Electronic Medical Records (EMRs) is important for transforming unstructured clinical narratives into structured knowledge that supports decision making. However, most existing approaches split causality modeling into two subtasks: causality detection and causality extraction. This causes poorer cross-task learning, error propagation, and missing fine-grained markers. Therefore, we propose a Multitask Causality Learning for Causality Detection and Extraction (MTCL-CDE) framework that simultaneously identifies causal relationships at sentence and token levels. The main idea of our framework is a unified hierarchical multigranular architecture explicitly designed to maximize cross-task signal sharing and capture fine-grained markers at all levels. Our approach utilizes an enhanced hierarchical attention network (HAN) and integrates it into a multigranular architecture. First, each sentence was represented across four hierarchical layers: characters, words, subsentences, and sentences to retain fine-grained markers and contextual signals. Each hierarchical layer produces a specific representation that is fused via a multigranular fusion module into a unified feature representation. Finally, the unified features are fed into a sentence level classifier for detection and a token level classifier for extraction. We evaluated our approach using a real-world Indonesian electronic medical record (EMR) dataset collected from hospitals. The experimental results show that MTCL-CDE achieves an average F1-score of 0.9630 for detection and 0.9003 for extraction, outperforming strong state-of-the-art baseline models including fine-grained-based models, HAN-based models, and other combined multitask frameworks. These results confirm the effectiveness of multitask learning for accurate and comprehensive causality modeling in clinical texts.
- 11 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 40423-40447
Open AccessA Multitask Learning Framework for Multigranular Causality Detection and Extraction in Indonesian EMRs
Extracting causal relations from Electronic Medical Records (EMRs) is important for transforming unstructured clinical narratives into structured knowledge that supports decision making. However, most existing approaches split causality modeling into two subtasks: causality detection and causality extraction. This causes poorer cross-task learning, error propagation, and missing fine-grained markers. Therefore, we propose a Multitask Causality Learning for Causality Detection and Extraction (MTCL-CDE) framework that simultaneously identifies causal relationships at sentence and token levels. The main idea of our framework is a unified hierarchical multigranular architecture explicitly designed to maximize cross-task signal sharing and capture fine-grained markers at all levels. Our approach utilizes an enhanced hierarchical attention network (HAN) and integrates it into a multigranular architecture. First, each sentence was represented across four hierarchical layers: characters, words, subsentences, and sentences to retain fine-grained markers and contextual signals. Each hierarchical layer produces a specific representation that is fused via a multigranular fusion module into a unified feature representation. Finally, the unified features are fed into a sentence level classifier for detection and a token level classifier for extraction. We evaluated our approach using a real-world Indonesian electronic medical record (EMR) dataset collected from hospitals. The experimental results show that MTCL-CDE achieves an average F1-score of 0.9630 for detection and 0.9003 for extraction, outperforming strong state-of-the-art baseline models including fine-grained-based models, HAN-based models, and other combined multitask frameworks. These results confirm the effectiveness of multitask learning for accurate and comprehensive causality modeling in clinical texts.
- 21 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 556-561
Aspect-Based Sentiment Analysis for Thai Language in Restaurant Domain Using Transfer Learning
Aspect-Based Sentiment Analysis (ABSA) provides fine-grained sentiment understanding by identifying and classifying sentiments toward specific aspects of products or services. This study investigates the application of transfer learning with BERT-based models for end-to-end ABSA on Thai restaurant reviews, focusing on four critical aspects: Food, Price, Service, and Environment. We fine-tune multiple pretrained BERT-based models and compare their performance against a BiGRU baseline. Our experiments demonstrate that all BERT-based models achieve substantial performance gains $(76.66-79.80 \%)$ F1-score) over the baseline model (61.85 %), with models pretrained on diverse text styles, including informal language, achieving the best results. This finding validates that pretraining corpora reflecting the linguistic characteristics of target tasks, such as user-generated content, enhance transfer learning effectiveness. We also investigate the impact of training dataset size, revealing that pretrained models fine-tuned with only 50 sentences outperform baseline models trained on the full dataset. Our analysis establishes that 300-500 sentences represent an optimal balance, achieving over 77 % F1-score while requiring only $24-41 \%$ of the full dataset.- 21 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 556-561
Aspect-Based Sentiment Analysis for Thai Language in Restaurant Domain Using Transfer Learning
Aspect-Based Sentiment Analysis (ABSA) provides fine-grained sentiment understanding by identifying and classifying sentiments toward specific aspects of products or services. This study investigates the application of transfer learning with BERT-based models for end-to-end ABSA on Thai restaurant reviews, focusing on four critical aspects: Food, Price, Service, and Environment. We fine-tune multiple pretrained BERT-based models and compare their performance against a BiGRU baseline. Our experiments demonstrate that all BERT-based models achieve substantial performance gains $(76.66-79.80 \%)$ F1-score) over the baseline model (61.85 %), with models pretrained on diverse text styles, including informal language, achieving the best results. This finding validates that pretraining corpora reflecting the linguistic characteristics of target tasks, such as user-generated content, enhance transfer learning effectiveness. We also investigate the impact of training dataset size, revealing that pretrained models fine-tuned with only 50 sentences outperform baseline models trained on the full dataset. Our analysis establishes that 300-500 sentences represent an optimal balance, achieving over 77 % F1-score while requiring only $24-41 \%$ of the full dataset.- 21 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 539-544
Forecasting of Students with Low Academic Performance Using Machine Learning Techniques
This study presents a machine learning framework for predicting first-year undergraduate students at risk of low academic performance using data from 8,726 students (2013–2023) at a public university in Northeastern Thailand. A structured data preparation pipeline—comprising data cleaning, feature engineering, and feature selection—produced and selected the top 15 informative predictors. Five models (Logistic Regression, Decision Tree, Random Forest, XGBoost, and SVM) and a Stacking Ensemble were tuned using stratified 5-fold cross-validation. Logistic Regression achieved the best baseline performance (AUC = 0.8015; F1-at-risk = 0.5281). SMOTE-Tomek, evaluated experimentally, increased recall to 0.7282 but reduced precision, indicating limited suitability for operational use. The findings demonstrate that an interpretable logistic model provides effective early-risk identification to support timely intervention and enhance student retention in Thai higher education.
- 21 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 539-544
Forecasting of Students with Low Academic Performance Using Machine Learning Techniques
This study presents a machine learning framework for predicting first-year undergraduate students at risk of low academic performance using data from 8,726 students (2013–2023) at a public university in Northeastern Thailand. A structured data preparation pipeline—comprising data cleaning, feature engineering, and feature selection—produced and selected the top 15 informative predictors. Five models (Logistic Regression, Decision Tree, Random Forest, XGBoost, and SVM) and a Stacking Ensemble were tuned using stratified 5-fold cross-validation. Logistic Regression achieved the best baseline performance (AUC = 0.8015; F1-at-risk = 0.5281). SMOTE-Tomek, evaluated experimentally, increased recall to 0.7282 but reduced precision, indicating limited suitability for operational use. The findings demonstrate that an interpretable logistic model provides effective early-risk identification to support timely intervention and enhance student retention in Thai higher education.
- Open Access
- 26 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 114696-114715
AEMS: Adaptive Ensemble GNNs for Multibehavior Stream Recommendation
Traditional systems often focus on a singular action type, such as a purchase, while overlooking the rich behavioral pluralism that is inherent in user engagement, such as the transient clicks or latent purchase intents signaled by cart additions. This narrow focus leads to sparse datasets and hinders recommendation accuracy. In response to the evolving landscape of real-time recommendation systems, contemporary frameworks are increasingly leveraging dynamic user engagement streams that span actions, such as purchasing, item retention, and exploratory browsing, to refine suggestions. To bridge this gap, we introduce a novel and revolutionary model: adaptive ensemble graph neural networks (GNNs), for multibehavior stream recommendation (AEMS). The AEMS synergizes long-term preference patterns (derived from historical interactions) with real-time user intents and item attributes (captured through multibehavior signals), integrating them via an adaptive ensemble neural gating mechanism. This design effectively fuses multifaceted behaviors and addresses the challenges of data sparsity and heterogeneity. This architecture is further enhanced by a behavior-aware training paradigm that dynamically weights heterogeneous interactions, such as clicks, add-to-cart actions, likes, and other behavioral signals, to increase its recommendation precision. Extensive experiments conducted on real-world datasets demonstrate that the AEMS significantly outperforms the state-of-the-art methods, demonstrating its ability to decode complex, high-order synergies in streaming environments.
- 26 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 114696-114715
Open AccessAEMS: Adaptive Ensemble GNNs for Multibehavior Stream Recommendation
Traditional systems often focus on a singular action type, such as a purchase, while overlooking the rich behavioral pluralism that is inherent in user engagement, such as the transient clicks or latent purchase intents signaled by cart additions. This narrow focus leads to sparse datasets and hinders recommendation accuracy. In response to the evolving landscape of real-time recommendation systems, contemporary frameworks are increasingly leveraging dynamic user engagement streams that span actions, such as purchasing, item retention, and exploratory browsing, to refine suggestions. To bridge this gap, we introduce a novel and revolutionary model: adaptive ensemble graph neural networks (GNNs), for multibehavior stream recommendation (AEMS). The AEMS synergizes long-term preference patterns (derived from historical interactions) with real-time user intents and item attributes (captured through multibehavior signals), integrating them via an adaptive ensemble neural gating mechanism. This design effectively fuses multifaceted behaviors and addresses the challenges of data sparsity and heterogeneity. This architecture is further enhanced by a behavior-aware training paradigm that dynamically weights heterogeneous interactions, such as clicks, add-to-cart actions, likes, and other behavioral signals, to increase its recommendation precision. Extensive experiments conducted on real-world datasets demonstrate that the AEMS significantly outperforms the state-of-the-art methods, demonstrating its ability to decode complex, high-order synergies in streaming environments.
- Open Access
- 10 October 2023
- Published by Prague University of Economics and Business in Acta Informatica Pragensia
- Vol. 12  (2) , 243-259
Multi-Class Text Classification on Khmer News Using Ensemble Method in Machine Learning Algorithms
Peer-reviewed journal on social and business aspects of informatics that covers the theory and application of enterprise information systems, as well as interactions between ICT and people.
- 10 October 2023
- Published by Prague University of Economics and Business in Acta Informatica Pragensia
- Vol. 12  (2) , 243-259
Open AccessMulti-Class Text Classification on Khmer News Using Ensemble Method in Machine Learning Algorithms
Peer-reviewed journal on social and business aspects of informatics that covers the theory and application of enterprise information systems, as well as interactions between ICT and people.
- 24 May 2023
- Published by Springer Nature
- p. 53-62
Predicting Foot and Mouth Disease in Thailand’s Nakhon Ratchasima Province Through Machine Learning
- 24 May 2023
- Published by Springer Nature
- p. 53-62
Predicting Foot and Mouth Disease in Thailand’s Nakhon Ratchasima Province Through Machine Learning
- Open Access
An Efficient Deep Learning for Thai Sentiment Analysis
The number of reviews from customers on travel websites and platforms is quickly increasing. They provide people with the ability to write reviews about their experience with respect to service quality, location, room, and cleanliness, thereby helping others before booking hotels. Many people fail to consider hotel bookings because the numerous reviews take a long time to read, and many are in a non-native language. Thus, hotel businesses need an efficient process to analyze and categorize the polarity of reviews as positive, negative, or neutral. In particular, low-resource languages such as Thai have greater limitations in terms of resources to classify sentiment polarity. In this paper, a sentiment analysis method is proposed for Thai sentiment classification in the hotel domain. Firstly, the Word2Vec technique (the continuous bag-of-words (CBOW) and skip-gram approaches) was applied to create word embeddings of different vector dimensions. Secondly, each word embedding model was combined with deep learning (DL) models to observe the impact of each word vector dimension result. We compared the performance of nine DL models (CNN, LSTM, Bi-LSTM, GRU, Bi-GRU, CNN-LSTM, CNN-BiLSTM, CNN-GRU, and CNN-BiGRU) with different numbers of layers to evaluate their performance in polarity classification. The dataset was classified using the FastText and BERT pre-trained models to carry out the sentiment polarity classification. Finally, our experimental results show that the WangchanBERTa model slightly improved the accuracy, producing a value of 0.9225, and the skip-gram and CNN model combination outperformed other DL models, reaching an accuracy of 0.9170. From the experiments, we found that the word vector dimensions, hyperparameter values, and the number of layers of the DL models affected the performance of sentiment classification. Our research provides guidance for setting suitable hyperparameter values to improve the accuracy of sentiment classification for the Thai language in the hotel domain.
Open AccessAn Efficient Deep Learning for Thai Sentiment Analysis
The number of reviews from customers on travel websites and platforms is quickly increasing. They provide people with the ability to write reviews about their experience with respect to service quality, location, room, and cleanliness, thereby helping others before booking hotels. Many people fail to consider hotel bookings because the numerous reviews take a long time to read, and many are in a non-native language. Thus, hotel businesses need an efficient process to analyze and categorize the polarity of reviews as positive, negative, or neutral. In particular, low-resource languages such as Thai have greater limitations in terms of resources to classify sentiment polarity. In this paper, a sentiment analysis method is proposed for Thai sentiment classification in the hotel domain. Firstly, the Word2Vec technique (the continuous bag-of-words (CBOW) and skip-gram approaches) was applied to create word embeddings of different vector dimensions. Secondly, each word embedding model was combined with deep learning (DL) models to observe the impact of each word vector dimension result. We compared the performance of nine DL models (CNN, LSTM, Bi-LSTM, GRU, Bi-GRU, CNN-LSTM, CNN-BiLSTM, CNN-GRU, and CNN-BiGRU) with different numbers of layers to evaluate their performance in polarity classification. The dataset was classified using the FastText and BERT pre-trained models to carry out the sentiment polarity classification. Finally, our experimental results show that the WangchanBERTa model slightly improved the accuracy, producing a value of 0.9225, and the skip-gram and CNN model combination outperformed other DL models, reaching an accuracy of 0.9170. From the experiments, we found that the word vector dimensions, hyperparameter values, and the number of layers of the DL models affected the performance of sentiment classification. Our research provides guidance for setting suitable hyperparameter values to improve the accuracy of sentiment classification for the Thai language in the hotel domain.
- Open Access
- 10 September 2021
- Published by Prague University of Economics and Business in Acta Informatica Pragensia
- Vol. 10  (2) , 155-171
Sentiment Analysis for Thai Language in Hotel Domain Using Machine Learning Algorithms
Peer-reviewed journal on social and business aspects of informatics that covers the theory and application of enterprise information systems, as well as interactions between ICT and people. Recenzovan? ?asopis o soci?ln?ch a byznys aspektech informatiky zahrnuj?c? teorie a aplikace podnikov?ch informa?n?ch syst?m? a interakce mezi ICT a lidmi.
- 10 September 2021
- Published by Prague University of Economics and Business in Acta Informatica Pragensia
- Vol. 10  (2) , 155-171
Open AccessSentiment Analysis for Thai Language in Hotel Domain Using Machine Learning Algorithms
Peer-reviewed journal on social and business aspects of informatics that covers the theory and application of enterprise information systems, as well as interactions between ICT and people. Recenzovan? ?asopis o soci?ln?ch a byznys aspektech informatiky zahrnuj?c? teorie a aplikace podnikov?ch informa?n?ch syst?m? a interakce mezi ICT a lidmi.
- Open Access
Multi-Stroke Thai Finger-Spelling Sign Language Recognition System with Deep Learning
Sign language is a type of language for the hearing impaired that people in the general public commonly do not understand. A sign language recognition system, therefore, represents an intermediary between the two sides. As a communication tool, a multi-stroke Thai finger-spelling sign language (TFSL) recognition system featuring deep learning was developed in this study. This research uses a vision-based technique on a complex background with semantic segmentation performed with dilated convolution for hand segmentation, hand strokes separated using optical flow, and learning feature and classification done with convolution neural network (CNN). We then compared the five CNN structures that define the formats. The first format was used to set the number of filters to 64 and the size of the filter to 3 × 3 with 7 layers; the second format used 128 filters, each filter 3 × 3 in size with 7 layers; the third format used the number of filters in ascending order with 7 layers, all of which had an equal 3 × 3 filter size; the fourth format determined the number of filters in ascending order and the size of the filter based on a small size with 7 layers; the final format was a structure based on AlexNet. As a result, the average accuracy was 88.83%, 87.97%, 89.91%, 90.43%, and 92.03%, respectively. We implemented the CNN structure based on AlexNet to create models for multi-stroke TFSL recognition systems. The experiment was performed using an isolated video of 42 Thai alphabets, which are divided into three categories consisting of one stroke, two strokes, and three strokes. The results presented an 88.00% average accuracy for one stroke, 85.42% for two strokes, and 75.00% for three strokes.
Open AccessMulti-Stroke Thai Finger-Spelling Sign Language Recognition System with Deep Learning
Sign language is a type of language for the hearing impaired that people in the general public commonly do not understand. A sign language recognition system, therefore, represents an intermediary between the two sides. As a communication tool, a multi-stroke Thai finger-spelling sign language (TFSL) recognition system featuring deep learning was developed in this study. This research uses a vision-based technique on a complex background with semantic segmentation performed with dilated convolution for hand segmentation, hand strokes separated using optical flow, and learning feature and classification done with convolution neural network (CNN). We then compared the five CNN structures that define the formats. The first format was used to set the number of filters to 64 and the size of the filter to 3 × 3 with 7 layers; the second format used 128 filters, each filter 3 × 3 in size with 7 layers; the third format used the number of filters in ascending order with 7 layers, all of which had an equal 3 × 3 filter size; the fourth format determined the number of filters in ascending order and the size of the filter based on a small size with 7 layers; the final format was a structure based on AlexNet. As a result, the average accuracy was 88.83%, 87.97%, 89.91%, 90.43%, and 92.03%, respectively. We implemented the CNN structure based on AlexNet to create models for multi-stroke TFSL recognition systems. The experiment was performed using an isolated video of 42 Thai alphabets, which are divided into three categories consisting of one stroke, two strokes, and three strokes. The results presented an 88.00% average accuracy for one stroke, 85.42% for two strokes, and 75.00% for three strokes.