Unsupervised Named Entity Transliteration Using Temporal and Phonetic Correlation

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 250–257, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Unsupervised Named Entity Transliteration Using Temporal and Phonetic Correlation Tao Tao, Su-Youn Yoon, Andrew Fister, Richard Sproat and ChengXiang Zhai University of Illinois at Urbana-Champaign {syoon9,afister2,rws}@uiuc.edu, {taotao,czhai}@cs.uiuc.edu Abstract In this paper we investigate unsuper- vised name transliteration using compara- ble corpora, corpora where texts in the two languages deal in some of the same top- ics — and therefore share references to named entities — but are not translations of each other. We present two distinct methods for transliteration, one approach using an unsupervised phonetic translit- eration method, and the other using the temporal distribution of candidate pairs. Each of these approaches works quite well, but by combining the approaches one can achieve even better results. We believe that the novelty of our approach lies in the phonetic-based scoring method, which is based on a combination of care- fully crafted phonetic features, and empiri- cal results from the pronunciation errors of second-language learners of English. Un- like previous approaches to transliteration, this method can in principle work with any pair of languages in the absence of a train- ing dictionary, provided one has an esti- mate of the pronunciation of words in text. 1 Introduction As a part of a on-going project on multilingual named entity identification, we investigate unsu- pervised methods for transliteration across lan- guages that use different scripts. Starting from paired comparable texts that are about the same topic, but are not in general translations of each other, we aim to find the transliteration correspon- dences of the paired languages. For example, if there were an English and Arabic newspaper on the same day, each of the newspapers would likely contain articles about the same important inter- national events. From these comparable articles across the two languages, the same named enti- ties such as persons and locations would likely be found. For at least some of the English named entities, we would therefore expect to find Ara- bic equivalents, many of which would in fact be transliterations. The characteristics of transliteration differ ac- cording to the languages involved. In particular, the exact transliteration of say, an English name is highly dependent on the language since this will be influenced by the difference in the phonological systems of the language pairs. In order to show the reliability of a multi-lingual transliteration model, it should be tested with a variety of different lan- guages. We have tested our transliteration meth- ods with three unrelated target languages — Ara- bic, Chinese and Hindi, and a common source lan- guage — English. Transliteration from English to Arabic and Chinese is complicated (Al-Onaizan and Knight, 2002). For example, while Arabic or- thography has a conventional way of writing long vowels using selected consonant symbols — ba- sically <w>, <y> and <?>, in ordinary text short vowels are rarely written. When transliter- ating English names there is the option …

Content truncated to 500 words