Challenges in Building an Arabic-English GHMT System with SMT Components

Challenges in Building an Arabic-English GHMT System with SMT Components Nizar Habash†, Bonnie Dorr‡, Christof Monz§ †Center for Computational Learning Systems, Columbia University habash@cs.columbia.edu ‡Department of Computer Science, University of Maryland bonnie@umiacs.umd.edu §Department of Computer Science, Queen Mary, University of London christof@dcs.qmul.ac.uk Abstract The research context of this paper is de- veloping hybrid machine translation (MT) systems that exploit the advantages of linguistic rule-based and statistical MT systems. Arabic, as a morphologically rich language, is especially challenging even without addressing the hybridiza- tion question. In this paper, we describe the challenges in building an Arabic- English generation-heavy machine trans- lation (GHMT) system and boosting it with statistical machine translation (SMT) components. We present an extensive evaluation of multiple system variants and report positive results on the advantages of hybridization. 1 Introduction The research context of this work is developing hy- brid machine translation (MT) systems that exploit the advantages of linguistic rule-based and statisti- cal MT systems. Arabic, as an example of a mor- phologically rich language, is especially challeng- ing even without addressing the hybridization ques- tion. In this paper, we describe the challenges in building an Arabic-English generation-heavy ma- chine translation (GHMT) system (Habash, 2003a) and extending it with statistical machine translation (SMT) components. A major challenge for working with Arabic is the proliferation of inconsistent morphological repre- sentations in different resources and tools for Arabic natural language processing (NLP) (Habash, 2006). This inconsistency is heightened when one attempts to combine techniques used in linguistically-aware MT approaches with those of surface-based statis- tical MT approaches, where the level of represen- tation of the phrase (beyond the word) is different for each of these two approaches. We describe how we address this issue in our system and present an extensive evaluation addressing its various strengths and weaknesses. We show positive improvements when extending our basic GHMT system with SMT components. The remainder of this paper is organized as fol- lows: the next section (Section 2) discusses previous work on hybridization in MT. This is followed by a discussion of Arabic-specific challenges for MT implementations in Section 3. Section 4 describes the Arabic components of our basic GHMT system. Section 5 describes the extensions we made to in- tegrate SMT components into the GHMT system. Section 6 presents three evaluations of multiple MT system variants. 2 Previous Work We discuss research related to our approach in the areas of generation-heavy MT and MT hybridiza- tion. 2.1 Generation-Heavy MT GHMT is an asymmetrical hybrid approach that addresses the issue of MT resource poverty in source-poor/target-rich language pairs by exploiting symbolic and statistical target-language resources (Habash and Dorr, 2002; Habash, 2003a; Habash, 56 Proceedings of the 7th Conference of the Association for Machine Translation in the Americas, pages 56-65, Cambridge, August 2006. ©2006 The Association for Machine Translation in the Americas 2003b). Expected source-language resources in- clude a syntactic parser and a simple one-to-many translation dictionary. No transfer rules or complex interlingual representations are used. Rich target- language symbolic resources such as …

Content truncated to 500 words