KMITL

Permanent URI for this communityhttps://dspace.kmitl.ac.th/handle/123456789/1

Browse

Search Results

Now showing 1 - 7 of 7

Some of the metrics are blocked by your
consent settings
myNER: Contextualized Burmese Named Entity Recognition with Bidirectional LSTM and fastText Embeddings via Joint Training with POS Tagging
(2025-01-01)
Thant, Kaung Lwin
;
Nongpong, Kwankamol
;
Thu, Ye Kyaw
;
Aung, Thura
;
Wai, Khaing Hsu
Named Entity Recognition (NER) involves identifying and categorizing named entities within textual data. Despite its significance, NER research has often overlooked low-resource languages like Myanmar (Burmese), primarily due to the lack of publicly available annotated datasets. To address this, we introduce myNER, a novel word-level NER corpus featuring a 7-tag annotation scheme, enriched with Part-of-Speech (POS) tagging to provide additional syntactic information. Alongside the corpus, we conduct a comprehensive evaluation of NER models, including Conditional Random Fields (CRF), Bidirectional LSTM (BiLSTM)-CRF, and their combinations with fastText embeddings in different settings. Our experiments reveal the effectiveness of contextualized word embeddings and the impact of joint training with POS tagging, demonstrating significant performance improvements across models. The traditional CRF joint-task model with fastText embeddings as a feature achieved the best result, with a 0.9818 accuracy and 0.9811 weighted F1 score with 0.7429 macro F1 score. BiLSTM-CRF with fine-tuned fastText embeddings gets the best result of 0.9791 accuracy and 0.9776 weighted F1 score with 0.7395 macro F1 score.
Some of the metrics are blocked by your
consent settings
ASR Error Correction in Low-Resource Burmese with Alignment-Enhanced Transformers Using Phonetic Features
(2025-01-01)
Lin, Ye Bhone
;
Aung, Thura
;
Thu, Ye Kyaw
;
Oo, Thazin Myint
This paper investigates sequence-to-sequence Transformer models for automatic speech recognition (ASR) error correction in low-resource Burmese, focusing on different feature integration strategies including IPA and alignment information. To our knowledge, this is the first study addressing ASR error correction specifically for Burmese. We evaluate five ASR backbones and show that our ASR Error Correction (AEC) approaches consistently improve word- and character-level accuracy over baseline outputs. The proposed AEC model, combining IPA and alignment features, reduced the average WER of ASR models from 51.56 to 39.82 before augmentation (and 51.56 to 43.59 after augmentation) and improving chrF++ scores from 0.5864 to 0.627, demonstrating consistent gains over the baseline ASR outputs without AEC. Our results highlight the robustness of AEC and the importance of feature design for improving ASR outputs in low-resource settings.
Some of the metrics are blocked by your
consent settings
Enhancing Burmese News Classification with Kolmogorov-Arnold Network Head Fine-Tuning
(2025-01-01)
Aung, Thura
;
Kyaw, Eaint Kay Khaing
;
Thu, Ye Kyaw
;
Oo, Thazin Myint
;
Supnithi, Thepchai
In low-resource languages like Burmese, classification tasks often fine-tune only the final classification layer, keeping pre-trained encoder weights frozen. While Multi-Layer Perceptrons (MLPs) are commonly used, their fixed non-linearity can limit expressiveness and increase computational cost. This work explores Kolmogorov-Arnold Networks (KANs) as alternative classification heads, evaluating Fourierbased FourierKAN, Spline-based EfficientKAN, and Grid-based FasterKAN-across diverse embeddings including TF-IDF, fastText, and multilingual transformers (mBERT, Distil-mBERT). Experimental results show that KAN-based heads are competitive with or superior to MLPs. EfficientKAN with fast-Text achieved the highest F1-score (0.928), while FasterKAN offered the best trade-off between speed and accuracy. On transformer embeddings, EfficientKAN matched or slightly outperformed MLPs with mBERT (0.917 F1). These findings highlight KANs as expressive, efficient alternatives to MLPs for lowresource language classification.
Some of the metrics are blocked by your
consent settings
ThonburianTTS: Enhancing Neural Flow Matching Models for Authentic Thai Text-to-Speech
(2025-01-01)
Aung, Thura
;
Sriwirote, Panyut
;
Thavornmongkol, Thanachot
;
Pipatsrisawat, Knot
;
Achakulvisut, Titipat
We introduce ThonburianTTS, a finetuned Thai text-to-speech (TTS) system based on the E2-TTS and F5-TTS architectures, designed to improve pronunciation accuracy, alignment robustness, and zero-shot speaker adaptation for the Thai language. Our models are trained on both Thai script and International Phonetic Alphabet (IPA) transcriptions to evaluate the impact of phonetic input on synthesis quality. We evaluate performance using objective metrics, including Word Error Rate (WER), Syllable Error Rate (SylER), Character Error Rate (CER), Speech Naturalness (MOSNet), Speaker Similarity (SIM-O) and Synthesis Speed (RTF). Our best model, F5-TTS trained on Thai script, achieves a WER of 25.72 %, SylER of 14.17 %, CER of 8.70 %, a MOSNet score of 3.9451, and a SIM-O score of 88.30 %. While IPA-based models yield comparable or higher scores in naturalness and speaker similarity, they underperform in accuracy-related metrics such as WER, SylER, and CER. We also show that increasing the Number of Function Evaluations (NFE) leads to improved model accuracy. ThonburianTTS outperforms strong baselines such as MMS-TTS and PyThaiTTS in both intelligibility and speaker similarity, highlighting the effectiveness of flow matching-based architectures for high-quality TTS in tonal, low-resource languages like Thai. The code and pretrained models are available at https://github.com/biodatlab/thonburian-tts.
Some of the metrics are blocked by your
consent settings
Myanmar Text Grade-Level Prediction Using Statistical and Linguistics Features
(2025-01-01)
Wai, Khaing Hsu
;
Nwe, Hlaing Myat
;
Aung, Thura
;
Thu, Kaung Khant Si
;
Mon, Hsu Yee
Readability assessment supports curriculum design and adaptive learning by estimating the difficulty of a text for readers at different proficiency levels. However, research on text readability for the Myanmar language remains unexplored, mainly due to the absence of labeled resources and established computational approaches. This paper presents the first comprehensive study on Myanmar text readability classification. We construct a grade-annotated corpus from official Myanmar school textbooks (Grades 1-12) and extract linguistic, statistical, and Myanmar-specific indicators. We then evaluate regression and classification baselines, text-only embeddings, and ensembles. We also adapt three classic readability formulas (LIX, Dale-Chall, Flesch) to Myanmar and empirically show that their score distributions overlap heavily across educational levels. Experimental results show that ensemble-based models achieved the best performance in predicting grade levels, demonstrating the effectiveness of our feature design and modeling framework. This work introduces the first readability dataset, modeling approaches, and benchmark results for the Myanmar language, providing a strong foundation for future research in readability prediction, low-resource Natural language processing (NLP), and educational text analysis.
Some of the metrics are blocked by your
consent settings
myContradict: Semi-supervised Contradictory Sentence Generation for Myanmar language
(2024-01-01)
Thu, Ye Kyaw
;
Nwe, Ei Myat
;
Aung, Thura
In this paper, we proposed a semi-supervised low-resource approach for generating contradictory sentences in the Myanmar language using the self-training perspective. This semi-supervised learning method leverages Bidirectional Long Short-Term Memory neural network (BiLSTM) and Transformer encoder-decoder models. All of the experiments were done in both syllable- and word-level tokens. We also investigated the importance of linguistic features in the Myanmar language for generating contradictory sentences. Our approach found improvement in BiLSTM models with and without POS Tag features for both tokenization schemes. The best contradictory sentence generation model is the syllable-level Transformer model using the self-training approach with only one iteration, which gained the chrF++ score of 0.75. It also gained 0.86 Precision, 0.85 Recall, and 0.85 F1-score of ROUGE-L scores and 0.94 Precision, 0.94 Recall, and 0.94 F1-score of BERTScore.
Some of the metrics are blocked by your
consent settings
myOCR: Optical Character Recognition for Myanmar language with Post-OCR Error Correction
(2024-01-01)
Aung, Thura
;
Thu, Ye Kyaw
;
Oo, Myat Noe
This paper presents the Myanmar Optical Character Recognition (OCR), named myOCR. It utilizes a synthetic text image dataset with 14 different font styles that contains 25,790 text images. The system includes Convolutional Neural Networks (CNN) for feature extraction, Bidirectional Long-Short Term Memory (BiLSTM) networks for sequence modeling, and Connectionist Temporal Classification (CTC) for decoding, evaluated across various iterations (3,000, 6,000, 9,000) and hidden states (64, 128, 256). Statistical Post-OCR correction methods involve N(3,4,5)-grams and edit distances with the Symmetric Delete Spelling correction algorithm (SymSpell). For Neural Machine Translation-based correction, BiLSTM and Transformer models are employed, while the mT5-base and mBART-50 models are used for LLM-based correction. The best base (optical) model is the model with 9,000 iterations that achieved a chrF++ score of over 97.90 and a Word Error Rate (WER) of 9.18%. Transformer correction improved its chrF++ to 99.31 and reduced the WER to 0.66%.

KMITL

Browse

Filters

Settings

Sort By

Results per page

Search Results