A Joint Source-Channel Model for Machine Transliteration

A Joint Source-Channel Model for Machine Transliteration Li Haizhou, Zhang Min, Su Jian Institute for Infocomm Research 21 Heng Mui Keng Terrace, Singapore 119613 {hli,sujian,mzhang}@i2r.a-star.edu.sg Abstract Most foreign names are transliterated into Chinese, Japanese or Korean with approximate phonetic equivalents. The transliteration is usually achieved through intermediate phonemic mapping. This paper presents a new framework that allows direct orthographical mapping (DOM) between two different languages, through a joint source-channel model, also called n-gram transliteration model (TM). With the n-gram TM model, we automate the orthographic alignment process to derive the aligned transliteration units from a bilingual dictionary. The n-gram TM under the DOM framework greatly reduces system development effort and provides a quantum leap in improvement in transliteration accuracy over that of other state-of-the-art machine learning algorithms. The modeling framework is validated through several experiments for English-Chinese language pair. 1 Introduction In applications such as cross-lingual information retrieval (CLIR) and machine translation, there is an increasing need to translate out-of-vocabulary words from one language to another, especially from alphabet language to Chinese, Japanese or Korean. Proper names of English, French, German, Russian, Spanish and Arabic origins constitute a good portion of out-of-vocabulary words. They are translated through transliteration, the method of translating into another language by preserving how words sound in their original languages. For writing foreign names in Chinese, transliteration always follows the original romanization. Therefore, any foreign name will have only one Pinyin (romanization of Chinese) and thus in Chinese characters. In this paper, we focus on automatic Chinese transliteration of foreign alphabet names. Because some alphabet writing systems use various diacritical marks, we find it more practical to write names containing such diacriticals as they are rendered in English. Therefore, we refer all foreign-Chinese transliteration to English-Chinese transliteration, or E2C. not Transliterating English names into Chinese is straightforward. However, recalling the original from Chinese transliteration is even more challenging as the E2C transliteration may have lost some original phonemic evidences. The Chinese-English backward transliteration process is also called back-transliteration, or C2E (Knight & Graehl, 1998). In machine transliteration, the noisy channel model (NCM), based on a phoneme-based approach, has recently received considerable attention (Meng et al. 2001; Jung et al, 2000; Virga & Khudanpur, 2003; Knight & Graehl, 1998). In this paper we discuss the limitations of such an approach and address its problems by firstly proposing a paradigm that allows direct orthographic mapping (DOM), secondly further proposing a joint source-channel model as a realization of DOM. Two other machine learning techniques, NCM and ID3 (Quinlan, 1993) decision tree, also are implemented under DOM as reference to compare with the proposed n-gram TM. This paper is organized as follows: In section 2, we present the transliteration problems. In section 3, a joint source-channel model is formulated. In section 4, several experiments are carried out to study different aspects of proposed algorithm. In section 5, we relate our algorithms to other reported work. Finally, we conclude the study with some discussions. 2 Problems in transliteration Transliteration is a …

Content truncated to 500 words