反向異文字音譯相似度評量方法與跨語言資訊檢索

本文首先根據音譯的方向是否跨不同文字系統，將機器音譯分成「正向同文字」、「正向異文字」、「反向同文字」與「反向異文字」等四種來討論。接著以相似度的比較作為音譯系統的基礎，將語音相似度分為物理聲音、音素、和形素三個層級，並討論計算語音相似度的方式。最後，提出一個以音素相似度為基礎的方法，以中文和英文的音譯為例，進行反向異文字的音譯。實驗結果顯示在音素上的比較，比在形素上的比較來得有效。在一個1,261 個人名的候選名單中，執行配偶配對實驗，平均排名是7.80，其中57.65%的排名為第一名。

This paper classifies the problem of machine transliteration into four types, i.e., forward/backward transliteration between same/different character sets, based on transliteration direction and character sets. A phoneme-based similarity measure is proposed to deal with backward transliteration between different character sets. Chinese-English information retrieval is taken as an example. The experiments show that phoneme-based approach is better than grapheme-based approach. In a mate matching of 1,261 candidates, the average rank is 7.80 and 57.65% of candidates are ranked as number one.