The TALP-UPC Phrase-Based Translation System for EACL-WMT 2009

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 85–89, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics The TALP-UPC phrase-based translation system for EACL-WMT 2009 José A.R. Fonollosa and Maxim Khalilov and Marta R. Costa-jussà and José B. Mariño and Carlos A. Henríquez Q. and Adolfo Hernández H. and Rafael E. Banchs TALP Research Center Universitat Politècnica de Catalunya, Barcelona 08034 {adrian,khalilov,mruiz,canton,carloshq,adolfohh,rbanchs}@talp.upc.edu Abstract This study presents the TALP-UPC sub- mission to the EACL Fourth Worskhop on Statistical Machine Translation 2009 evaluation campaign. It outlines the ar- chitecture and configuration of the 2009 phrase-based statistical machine transla- tion (SMT) system, putting emphasis on the major novelty of this year: combina- tion of SMT systems implementing differ- ent word reordering algorithms. Traditionally, we have concentrated on the Spanish-to-English and English-to- Spanish News Commentary translation tasks. 1 Introduction TALP-UPC (Center of Speech and Language Applications and Technology at the Universitat Politècnica de Catalunya) is a permanent par- ticipant of the ACL WMT shared translations tasks, traditionally concentrating on the Spanish- to-English and vice versa language pairs. In this paper, we describe the 2009 system’s architecture and design describing individual components and distinguishing features of our model. This year’s system stands aside from the previous years’ configurations which were per- formed following an N-gram-based (tuple-based) approach to SMT. By contrast to them, this year we investigate the translation models (TMs) interpolation for a state-of-the-art phrase-based translation system. Inspired by the work pre- sented in (Schwenk and Estève, 2008), we attack this challenge using the coefficients obtained for the corresponding monolingual language models (LMs) for TMs interpolation. On the second step, we have performed additional word reordering experiments, com- paring the results obtained with a statisti- cal method (R. Costa-jussà and R. Fonollosa, 2009) and syntax-based algorithm (Khalilov and R. Fonollosa, 2008). Further the outputs of the systems were combined selecting the trans- lation with the Minimum Bayes Risk (MBR) al- gorithm (Kumar, 2004) that allowed significantly outperforming the baseline configuration. The remainder of this paper is organized as follows: Section 2 presents the TALP-UPC’09 phrase-based system, along with the translation models interpolation procedure and other minor novelties of this year. Section 3 reports on the ex- perimental setups and outlines the results of the participation in the EACL WMT 2009 evaluation campaign. Section 4 concludes the paper with dis- cussions. 2 TALP-UPC phrase-based SMT The system developed for this year’s shared task is based on a state-of-the-art SMT sys- tem implemented within the open-source MOSES toolkit (Koehn et al., 2007). A phrase-based trans- lation is considered as a three step algorithm: (1) the source sequence of words is segmented in phrases, (2) each phrase is translated into tar- get language using translation table, (3) the target phrases are reordered to be inherent in the target language. A bilingual phrase (which in the context of SMT do not necessarily coincide with their linguistic analogies) is any pair of m source words and n target words that satisfies …

Content truncated to 500 words