Unsupervised Constraint Driven Learning For Transliteration Discovery

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 299–307, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Unsupervised Constraint Driven Learning For Transliteration Discovery Ming-Wei Chang Dan Goldwasser Dan Roth University of Illinois at Urbana ChampaignYuancheng Tu Urbana, IL 61801 {mchang21,goldwas1,danr,ytu}@uiuc.edu Abstract This paper introduces a novel unsupervised constraint-driven learning algorithm for iden- tifying named-entity (NE) transliterations in bilingual corpora. The proposed method does not require any annotated data or aligned cor- pora. Instead, it is bootstrapped using a simple resource – a romanization table. We show that this resource, when used in conjunction with constraints, can efficiently identify translitera- tion pairs. We evaluate the proposed method on transliterating English NEs to three differ- ent languages - Chinese, Russian and Hebrew. Our experiments show that constraint driven learning can significantly outperform existing unsupervised models and achieve competitive results to existing supervised models. 1 Introduction Named entity (NE) transliteration is the process of transcribing a NE from a source language to some target language while preserving its pronunciation in the original language. Automatic NE transliteration is an important component in many cross-language applications, such as Cross-Lingual Information Re- trieval (CLIR) and Machine Translation(MT) (Her- mjakob et al., 2008; Klementiev and Roth, 2006a; Meng et al., 2001; Knight and Graehl, 1998). It might initially seem that transliteration is an easy task, requiring only finding a phonetic mapping between character sets. However simply matching every source language character to its target lan- guage counterpart is not likely to work well as in practice this mapping depends on the context the characters appear in and on transliteration conven- tions which may change across domains. As a result, current approaches employ machine learning meth- ods which, given enough labeled training data learn how to determine whether a pair of words consti- tute a transliteration pair. These methods typically require training data and language-specific expertise which may not exist for many languages. In this pa- per we try to overcome these difficulties and show that when the problem is modeled correctly, a sim- ple character level mapping is a sufficient resource. In our experiments, English was used as the source language, allowing us to use romanization ta- bles, a resource commonly-available for many lan- guages1. These tables contain an incomplete map- ping between character sets, mapping every charac- ter to its most common counterpart. Our transliteration model takes a discriminative approach. Given a word pair, the model determines if one word is a transliteration of the other. The features used by this model are character n-gram matches across the two strings. For example, Fig- ure 1 describes the decomposition of a word pair into unigram features as a bipartite graph in which each edge represents an active feature. We enhance the initial model with constraints, by framing the feature extraction process as a struc- tured prediction problem - given a word pair, the set of possible active features is defined as a set of latent binary variables. …

Content truncated to 500 words