Improving Translation of Unknown Proper Names Using a Hybrid Web-based Translation Extraction Method

Improving Translation of Unknown Proper Names Using a Hybrid Web-based Translation Extraction Method Min-Shiang Shia Jiun-Hung Lin Scott Yu Wen-Hsiang Lu Department of Computer Science and Information Engineering National Cheng Kung University, Taiwan, R.O.C. {foreverdream, jhlin, scottyu}@csie.ncku.edu.tw, whlu@mail.ncku.edu.tw Abstract Recently, we have proposed several effective Web-based term translation extraction methods exploring Web resources to deal with translation of Web query terms. However, many unknown proper names in Web queries are still difficult to be translated by using our previous Web-based term translation extraction methods. Therefore, in this paper we propose a new hybrid translation extraction method, which combines our pervious Web-based term translation extraction method and a new Web-based transliteration method in order to improve translation of unknown proper names. In addition, to efficiently construct a good quality transliteration model, we also present a mixed-syllable-mapping transliteration model and a Web-based semi-supervised learning algorithm to explore search-result pages further for collecting large amounts of English-Chinese transliteration pairs from the Web. 1 Introduction In machine translation (MT) (Brown et al. 1993) or cross-language information retrieval (CLIR) (Jaleel and Larkey 2003; Pirkola et al. 2003), unknown term translation are still problematic and remain to be solved. Conventionally, most of the existing MT or CLIR systems rely mainly on general-purpose bilingual dictionaries, which usually lack translations of proper names or technical terms, and thus are unable to deal with such problems. We have proposed an effective Web-based approach to exploring abundant language-mixed texts on the Web like anchor texts and search-result pages for alleviating the difficulty of unknown query term translation (Lu et al. 2002, 2004; Cheng et al. 2004). However, the approach employing statistical techniques still suffers from the problem of data sparseness and indirect association errors in finding translations of low-frequency unknown terms (Melamed, 2000). According to the report in previous research (Davis et al. 1998), around 50% of unknown terms are proper names. To improve translation of unknown proper names, in this paper, we propose a hybrid translation extraction method, which is composed of our pervious search-result-based term translation extraction method (Section 3.2) and a new Web-based transliteration method (Section 3.3). Transliteration is the process that converting a sequence of substrings or characters in the source language (e.g., English) into a pronunciation-approximate substring/character sequence in the target language (e.g., Chinese). Many researchers have proposed phoneme-based mapping techniques for proper name transliteration (Jung et al. 2000; Knight & Graehl 1998; Lin & Chen 2002; Meng et al. 2001; Virga & Khudanpur 2003), but converting an English word from phonemic representation to Chinese Pinyin and from Pinyin to Chinese characters may cause double errors. Taking this problem into consideration, we thus try to adopt direct orthographical mapping for proper name transliteration and propose a simple mixed-syllable-mapping transliteration model which can effectively increase the correct mapping between an English-Chinese transliteration pair with different number of transliteration unit (syllable), such as “Ericsson” (易利信) with four English transliteration units “e”, ”ri”, “c”, “sson” and three Chinese transliteration units “易”, “利”, “信” (Section 3.3). Additionally, to train …

Content truncated to 500 words