Does joint decoding really outperform cascade processing in English-to-Chinese transliteration generation? The role of syllabification

Transliteration is a challengeable task aimed at converting a proper name into another language with phonetic equivalence. Since the conversion relates to the phonetic aspect of a text, syllabification is considered a major factor affecting the performance of a transliteration system. In grapheme-based approaches, there are two routines to transliterate, one is to perform in a pipeline of separate syllabification and other components in generation process step by step, the other is to synchronously segment syllables and generating transliteration options. Usually, joint decoding outperforms the cascade processing in many natural language processing missions, however, syllabification is a special component in transliteration task. Thus in this paper, we investigate the two routines with a systematic analysis and compare their results to illustrate the strength of syllabification. A phrase-based statistical machine translation framework for joint decoding and a conditional random field syllabification system are used in this work for our investigation, which shows a different scenario on the issue of joint decoding versus cascade processing in transliteration.