A Phrase-Based Context-Dependent Joint Probability Model for Named Entity Translation

R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 600 – 611, 2005. © Springer-Verlag Berlin Heidelberg 2005 A Phrase-Based Context-Dependent Joint Probability Model for Named Entity Translation Min Zhang1, Haizhou Li1, Jian Su1, and Hendra Setiawan1,2 1 Institute for Infocomm Research, 21 Heng Mui Keng Terrace, Singapore 119613 {mzhang, hli, sujian, stuhs}@i2r.a-star.edu.sg 2 Department of Computer Science, National University of Singapore, Singapore, 117543 hendrase@comp.nus.edu.sg Abstract. We propose a phrase-based context-dependent joint probability model for Named Entity (NE) translation. Our proposed model consists of a lexical mapping model and a permutation model. Target phrases are generated by the context-dependent lexical mapping model, and word reordering is per- formed by the permutation model at the phrase level. We also present a two- step search to decode the best result from the models. Our proposed model is evaluated on the LDC Chinese-English NE translation corpus. The experiment results show that our proposed model is high effective for NE translation. 1 Introduction A Named Entity (NE) is essentially a proper noun phrase. Automatic NE translation is an indispensable component of cross-lingual applications such as machine translation and cross-lingual information retrieval and extraction. NE is translated by a combination of meaning translation and/or phoneme trans- literation [1]. NE transliteration has been given much attention in the literature. Many attempts, including phoneme and grapheme-based methods, various machine learning and rule-based algorithms [2,3] and Joint Source-Channel Model (JSCM) [4], have been made recently to tackle the issue of NE transliteration. However, only a few works have been reported in NE translation. Chen et al. [1] proposed a frequency-based approach to learn formulation and transformation rules for multi- lingual Named Entities (NEs). Al-Onaizan and Knight [5] investigated the transla- tion of Arabic NEs to English using monolingual and bilingual resources. Huang et al. [6] described an approach to translate rarely occurring NEs by combining pho- netic and semantic similarities. In this paper, we pay special attention to the issue of NE translation. Although NE translation is less sophisticated than machine translation (MT) in gen- eral, to some extent, the issues in NE translation are similar to those in MT. Its chal- lenges lie in not only the ambiguity in lexical mapping such as <副(Fu),Deputy> and <副(Fu),Vice> in Fig.1 in the next page, but also the position permutation and fertility of words. Fig.1 illustrates two excerpts of NE translation from the LDC corpus [7]: A Phrase-Based Context-Dependent Joint Probability Model 601 (a) Regional office of science and technology for Africa 非洲(FeiZhou) 地区(DiQu) 科技(KeJi) 办事处(BanShiChu) (b) Deputy chief of staff to office of the vice president 副(Fu) 总统(ZongTong) 办公室(BanGongShi)副(Fu)主任(ZhuRen) Fig. 1. Example bitexts with alignment where the italic word is the Chinese pinyin transcription. Inspired by the JSCM model for NE transliteration [4] and the success of statistical phrase-based MT research [8-12], in this paper we propose a phrase-based context- dependent joint probability model for NE translation. It decomposes the NE transla- tion problem into two cascaded steps: 1) Lexical mapping step, using the phrase-based context-dependent joint prob- …

Content truncated to 500 words