Incorporating Pronunciation Variation into Different Strategies of Term Transliteration

Incorporating Pronunciation Variation into Different Strategies of Term Transliteration Jin-Shea KUO 1, 2 Ying-Kuei YANG 2 1Chung-Hwa Telecommunication Laboratories, 12, Lane 551, Sec. 5, Min-Tsu Rd., Yang-Mei, 326, Taoyuan, Taiwan 2Electrical Engineering Dept., National Taiwan University of Science and Technology, 43, Sec. 4, Keelung Rd., 106, Taipei, Taiwan jskuo@cht.com.tw ykyang@mouse.ee.ntust.edu.tw Abstract Term transliteration addresses the problem of converting terms in one language into their phonetic equivalents in the other language via spoken form. It is especially concerned with proper nouns, such as personal names, place names and organization names. Pronunciation variation refers to pronunciation ambiguity frequently encountered in spoken language, which has a serious impact on term transliteration. More than one transliteration variants can be generated by an out-of-vocabulary term due to different kinds of pronunciation variations. It is important to take this issue into account when dealing with term transliteration. Several models, which take pronunciation variation into consideration, are proposed for term transliteration in this paper. They describe transliteration from various viewpoints and utilize the relationships trained from extracted transliterated-term pairs. An experiment in applying the proposed models to term transliteration was conducted and evaluated. The experimental results show promise. These proposed models are not only applicable to term transliteration, but also are helpful in indexing and retrieving spoken document retrieval. 1. Introduction Machine transliteration plays an important role in machine translation. The importance of term transliteration can be realized from an analysis of the terms used in 200 qualifying sentences that were randomly selected from English-Chinese mixed news pages. Each qualifying sentence contained at least one English word. Analysis showed that 17.43% of the English terms were transliterated, and that most of them were content words (words that carry essential meaning, as opposed to grammatical function words such as conjunctions, prepositions, and auxiliary verbs) (Kuo, 2004). Lee (1998) also reported a similar result about the ratio of transliterated-term in Korean KTSET 2.0 corpus. Generally, a transliteration process starts by first examining a pre-compiled lexicon composing of many transliterated-term pairs collected manually or automatically. If a term can be found in the lexicon, one of the transliterated equivalents is selected; otherwise, the transliteration system then deals with this out-of-vocabulary term to try to generate a transliterated term via a sequence of pipelined conversions (Knight, 1998). A phoneme-based conversion approach was used in this pipelined procedure to handle English-Japanese transliteration. Kang (2000) proposed an approach using text-to-text direct correspondence for Korean-English transliteration. By utilizing the combined information of text-to-text and phoneme-based correspondences in converting basic pronunciation units or components of pronunciation units, Lee (1998) and Bilac (2004) proposed approaches to handle English-Korean and Japanese-English transliteration problems, respectively. Pronunciation variation has often been encountered in daily conversations and, therefore, in transliteration. This issue has been discussed in speech recognition (Riley, 1991), but has not been discussed extensively with respect to term transliteration. Two pronunciation variations are encountered when transliterating a source language term into its target language counterpart. One occurs at the source language side, and the other one happens during …

Content truncated to 500 words