The ISL Phrase-Based MT System for the 2007 ACL Workshop on Statistical Machine Translation

Proceedings of the Second Workshop on Statistical Machine Translation, pages 197–202, Prague, June 2007. c⃝2007 Association for Computational Linguistics The ISL Phrase-Based MT System for the 2007 ACL Workshop on Statistical Machine Translation M. Paulik1,2, K. Rottmann2, J. Niehues2, S. Hildebrand 1,2 and S. Vogel1 1Interactive Systems Laboratories, Carnegie Mellon University, Pittsburgh, PA, USA 2Institut f¨ur Theoretische Informatik, Universit¨at Karlsruhe (TH), Karlsruhe, Germany {paulik|silja|vogel}@cs.cmu.edu ; {jniehues|rottmann}@ira.uka.de Abstract In this paper we describe the Interactive Sys- tems Laboratories (ISL) phrase-based ma- chine translation system used in the shared task ”Machine Translation for European Languages” of the ACL 2007 Workshop on Statistical Machine Translation. We present results for a system combination of the ISL syntax-augmented MT system and the ISL phrase-based system by combining and rescoring the n-best lists of the two systems. We also investigate the combination of two of our phrase-based systems translating from different source languages, namely Spanish and German, into their common target lan- guage, English. 1 Introduction The shared task of the ACL 2007 Workshop on Sta- tistical Machine Translation focuses on the auto- matic translation of European language pairs. The workshop provides common training sets for trans- lation model training and language model training to allow for easy comparison of results between the participants. Interactive Systems Laboratories participated in the English ↔ Spanish Europarl and News Commen- tary task as well as in the English ↔ German Eu- roparl task. This paper describes the phrase-based machine translation (MT) system that was applied to these tasks. We also investigate the feasibility of combining the ISL syntax-augmented MT system (Zollmann et al., 2007) with our phrase-based sys- tem by combining and rescoring the n-best lists pro- duced by both systems for the Spanish → English Europarl task. Furthermore, we apply the same com- bination technique to combine two of our phrase- based systems that operate on different source lan- guages (Spanish and German), but share the same target language (English). The paper is organized as follows. In section 2 we give a general description of our phrase-based sta- tistical machine translation system. Section 3 gives an overview of the data and of the final systems used for the English ↔ Spanish Europarl and News Commentary tasks, along with corresponding per- formance numbers. Section 4 shows the data, final systems and results for the English ↔ German Eu- roparl task. In Section 5, we present our experiments involving a combination of the syntax-augmented MT system with the phrase-based MT system and a combination of the Spanish → English and German → English phrase-based systems. 2 The ISL Phrase-Based MT System 2.1 Word and Phrase Alignment Phrase-to-phrase translation pairs are extracted by training IBM Model-4 word alignments in both di- rections, using the GIZA++ toolkit (Och and Ney, 2000), and then extracting phrase pair candidates which are consistent with these alignments, start- ing from the intersection of both alignments. This is done with the help of phrase model training code provided by University of Edinburgh during the NAACL 2006 Workshop …

Content truncated to 500 words