Abstract
Automated English-to-Thai transliteration faces a significant challenge due to the non-monotonous nature of Thai orthography, specifically the "leading vowel" rule. Standard transliteration models often assume a linear, left-to-right alignment between source and target characters. However, in Thai, certain vowels are written before the consonants they phonetically follow. This paper proposes an "Augmented System" that utilizes a vowel-onset transposition step to regulate the Thai script into a monotonous format for model training, followed by a post-processing step to restore correct orthography. This approach achieved a 29% relative improvement in TOP-1 accuracy over standard baseline systems.
Introduction
Transliteration is the phonetic conversion of words between different writing systems. Most state-of-the-art automated systems rely on the "monotonous alignment" assumption, where the sequence of characters in the source language corresponds linearly to the sequence in the target language. While this holds for many language pairs, Thai orthography presents a unique challenge. Certain "leading vowels" in Thai are written to the left of the initial consonant (onset) of a syllable, despite being pronounced after it.
As shown in Figure 2, the transliteration of "Reagan" requires the vowel sound for 'e' to be written before the 'r' sound. To a standard machine learning model, it appears that characters are jumping out of order, which confuses the alignment process and leads to significant errors in output.
Methodology
The authors developed an Augmented System designed to "regulate" the relationship between Thai writing and phonology before the modeling phase. The core innovation is the Vowel-Onset Transposition process.
- Baseline Models: The study utilized two standard approaches: Phrase-based Statistical Machine Translation (pbSMT) using the Moses toolkit, and a Joint Source-Channel (JSC) model using the Sequitur G2P tool.
- Vowel-Onset Transposition: To address the alignment issue, the authors modified the training data. For every Thai syllable containing a leading vowel, the positions of the vowel and the preceding consonant were manually swapped. This created a "regulated" version of the Thai script that matches the phonetic and English character order.
- System Workflow: A standard JSC model was trained on this regulated data. During the decoding (testing) phase, the model produces a regulated output, which is then passed through a post-processing step to swap the vowels and consonants back into their grammatically correct Thai positions.
Results and Discussion
The system was evaluated using the Named Entities Workshop (NEWS) 2016 dataset for English to Thai transliteration. The primary metric was TOP-1 accuracy. The results demonstrated that "regulating" the script significantly outperforms standard models:
- Accuracy Boost: The Augmented System achieved a TOP-1 accuracy of 0.155.
- Relative Improvement: This represents a 29% relative improvement over the pbSMT baseline and a 23% relative improvement over the standard Joint Source-Channel baseline.
These results suggest that incorporating specific linguistic knowledge about syllable structure (Onset-Nucleus-Coda) is more effective than relying solely on complex statistical models to resolve orthographic irregularities. By forcing Thai script into a monotonous format, the authors removed the primary source of noise for alignment algorithms.
Limitations and Future Work
While highly effective for leading vowels, the current system does not yet address "implicit vowels"—vowels that are pronounced but not written in the script. Additionally, the complexities of Thai tonal marking were not explicitly incorporated into the transposition rules. Future work aims to automate this orthography-phonology regulation to handle a wider variety of language-specific quirks without manual rule-writing.
Summary updated