A statistical approach to machine aided translation of terminology banks
Jyun-Sheng Chang, Andrew Chang, Tsuey-Fen Lin, Sur-Jin Ker · 1992
This paper reports on a new statistical approach to machine aided translation of terminology bank. The text in the bank is hyphenated and then dissected into roots of 1 to 3 syllables. Both hyphenation and dissection are done with a set of initial probabilities of syllables and roots. The probabilities are repeatedly revised using an EM algorithm. After each iteration of hyphenation or dissection, the resulting syllables and roots are counted subsequently to yield more precise estimation of probability. The set of roots rapidly converges to a set of most likely roots. Preliminary experiments have shown promising results. From a terminology bank of more than 4, 000 terms, the algorithm extracts 223 general and chemical roots, of which 91% are actually roots. The algorithm dissects a word into roots with around 86% hit rate. The set of roots and their hand-translation are then used in a compositional translation of the terminology bank. One can expect the translation of terminology bank using this approach to be more cost-effective, consistent, and with a better closure.