Transliteration as Constrained Optimization

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 353–362, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Transliteration as Constrained Optimization Dan Goldwasser Dan Roth Department of Computer Science University of Illinois Urbana, IL 61801 {goldwas1,danr}@uiuc.edu Abstract This paper introduces a new method for iden- tifying named-entity (NE) transliterations in bilingual corpora. Recent works have shown the advantage of discriminative approaches to transliteration: given two strings (ws, wt) in the source and target language, a classifier is trained to determine if wt is the translitera- tion of ws. This paper shows that the translit- eration problem can be formulated as a con- strained optimization problem and thus take into account contextual dependencies and con- straints among character bi-grams in the two strings. We further explore several methods for learning the objective function of the opti- mization problem and show the advantage of learning it discriminately. Our experiments show that the new framework results in over 50% improvement in translating English NEs to Hebrew. 1 Introduction Named entity (NE) transliteration is the process of transcribing a NE from a source language to some target language based on phonetic similarity be- tween the entities. Identifying transliteration pairs is an important component in many linguistic appli- cations which require identifying out-of-vocabulary words, such as machine translation and multilingual information retrieval (Klementiev and Roth, 2006b; Hermjakob et al., 2008). It may appear at first glance that identifying the phonetic correlation between names based on an orthographic analysis is a simple, straight-forward Figure 1: Named entities transliteration pairs in English and Hebrew and the character level mapping between the two names. The Hebrew names can be romanized as ee- ta-l-ya and a-ya task; however in many cases a consistent deter- ministic mapping between characters does not ex- ist; rather, the mapping depends on the context the characters appear in and on transliteration conven- tions which may change across domains. Figure 1 exhibits two examples of NE transliterations in En- glish and Hebrew, with the correct mapping across the two scripts. Although the two Hebrew names share a common prefix1, this prefix can be mapped into a single English character or into two differ- ent characters depending on the context it appears in. Similarly, depending on the context it appears in, the English character a can be mapped into different characters or to an “empty” character. 1In all our example the Hebrew script is shown left-to-right to simplify the visualization of the transliteration mapping. 353 In recent years, as it became clear that solutions that are based on linguistics rules are not satisfac- tory, machine learning approaches have been de- veloped to address this problem. The common ap- proach adopted is therefore to view this problem as a classification problem (Klementiev and Roth, 2006a; Tao et al., 2006) and train a discriminative classifier. That is, given two strings, one in the source and the other in the target language, extract pairwise features, and train a classifier that deter- mines if one …

Content truncated to 500 words