Abstract
Transliteration is the process of converting names or terms between languages while preserving their phonetic properties. Building robust transliteration models typically requires large-scale manually labeled corpora, which are expensive and labor-intensive to produce. This paper explores a semi-supervised approach using multi-view co-training to bootstrap a Phonetic Similarity Model (PSM) from a small seed set of only 100 labeled examples. By treating phoneme-based and grapheme-based paths as distinct, independent views of the same data, the proposed method achieves an F-measure of 0.726, nearing the fully supervised performance of 0.739.
Introduction
As new terminology and proper names emerge daily, static transliteration lexicons quickly become obsolete. Statistical models capable of harvesting transliteration pairs from the web offer a dynamic solution, but they often rely on extensive "gold standard" bilingual corpora for training. The central challenge addressed here is: how can we leverage the vast amount of unlabeled web data to build high-quality transliteration models with minimal manual labeling effort? The authors propose a multi-view framework that exploits the inherent redundancy in linguistic data to bridge the gap between unsupervised and supervised learning.
Methodology: Multi-View Framework
The authors frame transliteration as a multi-view learning problem. They identify four distinct "paths" or views to establish phonetic similarity between an English word (EW) and a Chinese word (CW):
- Phoneme-based (V1): Maps the source word to source phonemes, then to target phonemes, and finally to the target word.
- Grapheme-based (V4): Direct orthographic mapping from source word to target word.
- Hybrid Views (V2 & V3): "Shortcuts" that combine phonemic and graphemic transitions (e.g., Source Phoneme to Target Word).
To train these views with limited supervision, two primary algorithms are employed:
- Co-training: Two learners representing different views are initialized with a tiny seed set. They iteratively label the most confident examples from the unlabeled pool for each other, growing the training set.
- Co-EM: A probabilistic variant of the Expectation-Maximization algorithm where learners exchange label distributions before each iteration to further refine model parameters.
Experimental Setup
The study utilized a 500 MB web corpus, from which 80,094 English-Chinese candidate pairs were extracted based on co-occurrence. A gold standard of 8,898 "Distinct Qualified Transliteration Pairs" (DQTPs) was used for evaluation. Crucially, the semi-supervised models were bootstrapped using a seed set of only 100 randomly selected DQTPs, representing a massive reduction in labeling cost compared to supervised baselines.
Key Results and Insights
The experiments demonstrate that the synergy between different views is highly effective for bootstrapping performance.
- Single View Performance: In an unsupervised setting, the phoneme-based path (V1) is the strongest individual signal, achieving an F-measure of 0.680.
- Multi-View Synergy: Combining views significantly boosts performance. The co-trained V1+V2 model reached an F-measure of 0.726 after several iterations, outperforming any single-view approach.
- Efficiency: This approach outperformed traditional Active Learning methods. While Active Learning required over 8,000 manual labels to reach an F-measure of 0.710, the multi-view co-training method surpassed this threshold with just 100 labels.
The final performance of 0.726 is remarkably close to the fully supervised "upper bound" of 0.739, which utilized 800 times more labeled data.
Conclusion
Multi-view co-training provides a powerful mechanism for building transliteration models with minimal human intervention. The success of the method relies on the views being compatible yet partially uncorrelated—allowing one view to correct the errors of another (e.g., a phoneme model handling irregular spellings via phonetic mapping). Future work may involve incorporating document context as a supplementary view to further enhance extraction accuracy and extending this framework to broader machine translation tasks.
Summary updated