Named Entity Transliteration with Comparable Corpora

Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 73–80, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Named Entity Transliteration with Comparable Corpora Richard Sproat, Tao Tao, ChengXiang Zhai University of Illinois at Urbana-Champaign, Urbana, IL, 61801 rws@uiuc.edu, {taotao,czhai}@cs.uiuc.edu Abstract In this paper we investigate Chinese- English name transliteration using compa- rable corpora, corpora where texts in the two languages deal in some of the same topics — and therefore share references to named entities — but are not transla- tions of each other. We present two dis- tinct methods for transliteration, one ap- proach using phonetic transliteration, and the second using the temporal distribu- tion of candidate pairs. Each of these ap- proaches works quite well, but by com- bining the approaches one can achieve even better results. We then propose a novel score propagation method that uti- lizes the co-occurrence of transliteration pairs within document pairs. This prop- agation method achieves further improve- ment over the best results from the previ- ous step. 1 Introduction As part of a more general project on multilin- gual named entity identification, we are interested in the problem of name transliteration across lan- guages that use different scripts. One particular is- sue is the discovery of named entities in “compara- ble” texts in multiple languages, where by compa- rable we mean texts that are about the same topic, but are not in general translations of each other. For example, if one were to go through an English, Chinese and Arabic newspaper on the same day, it is likely that the more important international events in various topics such as politics, business, science and sports, would each be covered in each of the newspapers. Names of the same persons, locations and so forth — which are often translit- erated rather than translated — would be found in comparable stories across the three papers.1 We wish to use this expectation to leverage translit- eration, and thus the identification of named enti- ties across languages. Our idea is that the occur- rence of a cluster of names in, say, an English text, should be useful if we find a cluster of what looks like the same names in a Chinese or Arabic text. An example of what we are referring to can be found in Figure 1. These are fragments of two stories from the June 8, 2001 Xinhua English and Chinese newswires, each covering an international women’s badminton championship. Though these two stories are from the same newswire source, and cover the same event, they are not translations of each other. Still, not surprisingly, a lot of the names that occur in one, also occur in the other. Thus (Camilla) Martin shows up in the Chinese version as � � � ma-er-ting; Judith Meulendijks is � � � � � � � yu mo-lun-di-ke-si; and Mette Sorensen is � � � � �mai su-lun-sen. Several other correspondences also occur. While some of the transliterations are …

Content truncated to 500 words