A Log-Linear Block Transliteration Model based on Bi-Stream HMMs

Proceedings of NAACL HLT 2007, pages 364–371, Rochester, NY, April 2007. c⃝2007 Association for Computational Linguistics A Log-linear Block Transliteration Model based on Bi-Stream HMMs Bing Zhao, Nguyen Bach, Ian Lane, and Stephan Vogel {bzhao, nbach, ianlane, vogel}@cs.cmu.edu Language Technologies Institute School of Computer Science, Carnegie Mellon University Abstract We propose a novel HMM-based framework to accurately transliterate unseen named entities. The framework leverages features in letter- alignment and letter n-gram pairs learned from available bilingual dictionaries. Letter-classes, such as vowels/non-vowels, are integrated to further improve transliteration accuracy. The proposed transliteration system is applied to out-of-vocabulary named-entities in statistical machine translation (SMT), and a significant improvement over traditional transliteration ap- proach is obtained. Furthermore, by incor- porating an automatic spell-checker based on statistics collected from web search engines, transliteration accuracy is further improved. The proposed system is implemented within our SMT system and applied to a real transla- tion scenario from Arabic to English. 1 Introduction Cross-lingual natural language applications, such as in- formation retrieval, question answering, and machine translation for web-documents (e.g. Google translation), are becoming increasingly important. However, current state-of-the-art statistical machine translation (SMT) sys- tems cannot yet translate named-entities which are not seen during training. New named-entities, such as per- son, organization, and location names are continually emerging on the World-Wide-Web. To realize effective cross-lingual natural language applications, handling out- of-vocabulary named-entities is becoming more crucial. Named entities (NEs) can be translated via transliter- ation: mapping symbols from one writing system to an- other. Letters of the source language are typically trans- formed into the target language with similar pronunci- ation. Transliteration between languages which share similar alphabets and sound systems is usually not dif- ficult, because the majority of letters remain the same. However, the task is significantly more difficult when the language pairs are considerably different, for example, English-Arabic, English-Chinese, and English-Japanese. In this paper, we focus on forward transliteration from Arabic to English. The work in (Arbabi et al., 1994), to our knowledge, is the first work on machine transliteration of Arabic names into English, French, and Spanish. The idea is to vow- elize Arabic names by adding appropriate vowels and uti- lizing a phonetic look-up table to provide the spelling in the target language. Their framework is strictly applica- ble within standard Arabic morphological rules. Knight and Graehl (1997) introduced finite state transducers that implement back-transliteration from Japanese to English, which was then extended to Arabic-English in (Stalls and Knight, 1998). Al-Onaizan and Knight (2002) translit- erated named entities in Arabic text to English by com- bining phonetic-based and spelling-based models, and re- ranking candidates with full-name web counts, named en- tities co-reference, and contextual web counts. Huang (2005) proposed a specific model for Chinese-English name transliteration with clusterings of names’ origins, and appropriate hypotheses are generated given the ori- gins. All of these approaches, however, are not based on a SMT-framework. Technologies developed for SMT are borrowed in Virga and Khudanpur (2003) and Ab- dulJaleel and Larkey (2003). Standard SMT alignment models …

Content truncated to 500 words