Multilingual Transliteration Using Feature based Phonetic Method

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 112–119, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Multilingual Transliteration Using Feature based Phonetic Method Su-Youn Yoon, Kyoung-Young Kim and Richard Sproat University of Illinois at Urbana-Champaign {syoon9,kkim36,rws}@uiuc.edu Abstract In this paper we investigate named entity transliteration based on a phonetic scoring method. The phonetic method is computed using phonetic features and carefully designed pseudo features. The proposed method is tested with four languages – Arabic, Chinese, Hindi and Korean – and one source language – English, using comparable corpora. The proposed method is developed from the phonetic method originally proposed in Tao et al. (2006). In contrast to the phonetic method in Tao et al. (2006) constructed on the basis of pure linguistic knowledge, the method in this study is trained using the Winnow machine learning algorithm. There is salient improvement in Hindi and Arabic compared to the previous study. Moreover, we demonstrate that the method can also achieve comparable results, when it is trained on language data different from the target language. The method can be applied both with minimal data, and without target language data for various languages. 1 Introduction. In this paper, we develop a multi-lingual transliteration system for named entities. Named entity transliteration is the process of producing, for a name in a source language, a set of one or more transliteration candidates in a target language. The correct transliteration of named entities is crucial, since they are frequent and important key words in information retrieval. In addition, requests in retrieving relevant documents in multiple languages require the development of the multi-lingual system. The system is constructed using paired comparable texts. The comparable texts are about the same or related topics, but are not, in general, translations of each other. Using this data, the transliteration method aims to find transliteration correspondences in the paired languages. For example, if there were an English and Arabic newspaper on the same day, each of the newspapers would contain articles about the same important international events. From these comparable articles across the paired languages, the same named entities are expected to be found. Thus, from the named entities in an English newspaper, the method would find transliteration correspondences in comparable texts in other languages. The multi-lingual transliteration system entails solving several problems which are very challenging. First, it should show stable performance for many unrelated languages. The transliteration will be influenced by the difference in the phonological systems of the language pairs, and the process of transliteration differs according to the languages involved. For example, in Arabic texts, short vowels are rarely written while long vowels are written. When transliterating English names, the vowels are disappeared or written as long vowels. For example London is transliterated as lndn نﺪ�����ﻨﻟ, and both vowels are not represented in the transliteration. However, Washington is often transliterated as wSnjTwn نوطﻧــــــــــــﺷاو , and the final vowel is realized with long vowel. Transliterations in Chinese are very different from …

Content truncated to 500 words