Augmented words to improve a deep learning-based Indonesian syllabification.

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Machine Learning

Background:

• Deep learning syllabification models excel in high-resource languages but struggle with low-resource ones.
• Indonesian, a low-resource language, presents challenges for accurate syllabification due to limited data.
• Existing models often yield high error rates for specific linguistic contexts like named entities.

Purpose of the Study:

• To improve the performance of deep learning-based syllabification models for the Indonesian language.
• To address the limitations of current models in low-resource language scenarios.
• To enhance accuracy in syllabifying both formal Indonesian words and named entities.

Main Methods:

• Proposed two key procedures: massive data augmentation and phonotactic-based validation.
• Implemented a deep learning architecture combining Bidirectional Long Short-Term Memory (BiLSTM), Convolutional Neural Networks (CNN), and Conditional Random Fields (CRF).
• Employed four data augmentation techniques: transposing nuclei, swapping consonant-graphemes, flipping onsets, and creating acronyms.

Main Results:

• Massive data augmentation significantly expanded the Indonesian dataset by 12.8 million valid words based on phonotactic rules.
• The enhanced BiLSTM-CNN-CRF model demonstrated significant improvements across formal words and named-entities datasets using 5-fold cross-validation.
• Augmenting the training data effectively reduced the word error rate (WER), particularly for long formal words and named entities.

Conclusions:

• Massive data augmentation and validation are effective strategies for improving syllabification in low-resource languages like Indonesian.
• The proposed methods enhance the robustness and accuracy of deep learning models for complex linguistic data.
• This approach offers a viable solution for overcoming data scarcity challenges in natural language processing tasks.