Abstract
Transliteration converts words in a source language (e.g., English) into words in a target language (e.g., Vietnamese). This conversion considers the phonological structure of the target language, as the transliterated output needs to be pronounceable in the target language. For example, a word in Vietnamese that begins with a consonant cluster is phonologically invalid and thus would be an incorrect output of a transliteration system. Most statistical transliteration approaches, albeit being widely adopted, do not explicitly model the target language's phonology, which often results in invalid outputs. The problem is compounded by the limited linguistic resources available when converting foreign words to transliterated words in the target language. In this work, we present a phonology-augmented statistical framework suitable for transliteration, especially when only limited linguistic resources are available. We propose the concept of pseudo-syllables as structures representing how segments of a foreign word are organized according to the syllables of the target language's phonology. We performed transliteration experiments on Vietnamese and Cantonese. We show that the proposed framework outperforms the statistical baseline by up to 44.68% relative, when there are limited training examples (587 entries).
Introduction
Machine transliteration—the process of converting a word from a source language (e.g., English "Manchester") to a target language (e.g., Vietnamese "man chét sờ tơ")—is a critical component for tasks such as Named Entity Recognition (NER) and cross-lingual information retrieval. However, existing approaches often fall into two categories, both with significant drawbacks in low-resource settings.
Statistical models are powerful and scalable but are often "phonology-blind." When training data is scarce, these models frequently produce invalid character sequences—such as illegal consonant clusters—that are unpronounceable in the target language. Conversely, symbolic or rule-based systems produce valid outputs with minimal data but are brittle, expensive to construct, and fail to improve significantly as more data becomes available. This research addresses this "valley of death" by introducing a framework that combines the structural "common sense" of rule-based systems with the flexibility of statistical modeling.
Methodology: The Pseudo-Syllable Framework
The core innovation of this work is the Pseudo-Syllable. Rather than treating transliteration as a direct string-to-string mapping, the framework decomposes the process into three phonologically-aware stages:
- Pseudo-syllable Formulation: The model assigns functional "roles" to source language letters. For instance, in "DISNEY," 'D' might be labeled as an Onset (O) and 'I' as a Nucleus (N). This forces the source word into a structural template (Onset-Nucleus-Coda) compatible with the target language's phonology.
- Pseudo-syllable-to-Phoneme Mapping: Labeled units are mapped to target phonemes using an n-gram language model. By constraining the search space to valid syllabic roles, the model avoids phonologically impossible outputs.
- Tone Assignment: For tonal languages like Vietnamese and Cantonese, lexical tones are assigned in a final pass based on the phonetic context of the entire syllable.
This approach is largely language-independent because the basic syllable structure (O-N-Cd) is a near-universal linguistic property, even if specific rules vary by language.
Experimental Evaluation
The framework was evaluated on two language pairs representing limited-resource scenarios:
- English-to-Vietnamese: Utilizing the NIST OpenKWS13 dataset, with training sizes ranging from 100 to 587 words.
- English-to-Cantonese: Utilizing a dataset of 473 English loanwords from Hong Kong Polytechnic University.
By forcing the model to organize source letters into pseudo-syllables, the system effectively uses linguistic structure as a regularizer. This prevents the exploration of impossible phonetic outputs, a common failure mode for standard statistical models when data is sparse.
Results and Discussion
The proposed framework consistently outperformed the standard Joint Source-Channel statistical baseline. In Vietnamese transliteration, the model achieved a 23.71% relative improvement in Token Error Rate (TER) at the smallest training size (100 words). At the largest training size (587 words), it improved the Tone Error Rate by 44.68%.
As shown in the performance analysis, while symbolic systems maintain a steady but mediocre error rate, the proposed model starts with lower error and continues to improve as data increases, eventually surpassing both baselines. This demonstrates that incorporating structural linguistic constraints can significantly boost the performance of statistical transliteration in low-resource environments.
Conclusion
Future work will address limitations such as pipeline error propagation—where an error in syllable labeling affects subsequent steps—by exploring end-to-end modeling. Additionally, the authors aim to incorporate inter-syllable context to better handle phenomena like tone sandhi, which are common in many Asian languages.
Summary updated