N-Gram-Based Statistical Machine Translation versus Syntax Augmented Machine Translation: Comparison and System Combination

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 424–432, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics N-gram-based Statistical Machine Translation versus Syntax Augmented Machine Translation: comparison and system combination Maxim Khalilov and José A.R. Fonollosa Universitat Politècnica de Catalunya Campus Nord UPC, 08034 Barcelona, Spain {khalilov,adrian}@talp.upc.edu Abstract In this paper we compare and contrast two approaches to Machine Translation (MT): the CMU-UKA Syntax Augmented Machine Translation system (SAMT) and UPC-TALP N-gram-based Statistical Ma- chine Translation (SMT). SAMT is a hier- archical syntax-driven translation system underlain by a phrase-based model and a target part parse tree. In N-gram-based SMT, the translation process is based on bilingual units related to word-to-word alignment and statistical modeling of the bilingual context following a maximum- entropy framework. We provide a step- by-step comparison of the systems and re- port results in terms of automatic evalu- ation metrics and required computational resources for a smaller Arabic-to-English translation task (1.5M tokens in the train- ing corpus). Human error analysis clari- fies advantages and disadvantages of the systems under consideration. Finally, we combine the output of both systems to yield significant improvements in transla- tion quality. 1 Introduction There is an ongoing controversy regarding whether or not information about the syntax of language can benefit MT or contribute to a hybrid system. Classical IBM word-based models were re- cently augmented with a phrase translation ca- pability, as shown in Koehn et al. (2003), or in more recent implementation, the MOSES MT sys- tem1 (Koehn et al., 2007). In parallel to the phrase- based approach, the N-gram-based approach ap- peared (Mariño et al., 2006). It stemms from 1www.statmt.org/moses/ the Finite-State Transducers paradigm, and is ex- tended to the log-linear modeling framework, as shown in (Mariño et al., 2006). A system follow- ing this approach deals with bilingual units, called tuples, which are composed of one or more words from the source language and zero or more words from the target one. The N-gram-based systems allow for linguistically motivated word reordering by implementing word order monotonization. Prior to the SMT revolution, a major part of MT systems was developed using rule-based algorithms; however, starting from the 1990’s, syntax-driven systems based on phrase hierar- chy have gained popularity. A representative sample of modern syntax-based systems includes models based on bilingual synchronous grammar (Melamed, 2004), parse tree-to-string translation models (Yamada and Knight, 2001) and non- isomorphic tree-to-tree mappings (Eisner, 2003). The orthodox phrase-based model was en- hanced in Chiang (2005), where a hierarchical phrase model allowing for multiple generaliza- tions within each phrase was introduced. The open-source toolkit SAMT2 (Zollmann and Venu- gopal, 2006) is a further evolution of this ap- proach, in which syntactic categories extracted from the target side parse tree are directly assigned to the hierarchically structured phrases. Several publications discovering similarities and differences between distinct translation mod- els have been written over the last few years. In Crego et al. (2005b), the N-gram-based system is contrasted with …

Content truncated to 500 words