Can we Relearn an RBMT System?

Proceedings of the Third Workshop on Statistical Machine Translation, pages 175–178, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Can we relearn an RBMT system? Loïc Dugast (1,2) dugast@systran.fr Jean Senellart (1) senellart@systran.fr Philipp Koehn (2) pkoehn@inf.ed.ac.uk (1) SYSTRAN S.A. La Grande Arche 1, Parvis de la Défense 92044 Paris La Défense Cedex France (2) School of Informatics University of Edinburgh 2 Buccleuch Place Edinburgh, EH8 9LW United Kingdom Abstract This paper describes SYSTRAN submissions for the shared task of the third Workshop on Statistical Machine Translation at ACL. Our main contribution consists in a French-English statistical model trained without the use of any human-translated parallel corpus. In substitu- tion, we translated a monolingual corpus with SYSTRAN rule-based translation engine to produce the parallel corpus. The results are provided herein, along with a measure of error analysis. 1 Introduction Current machine translation systems follow two different lines of research: (1) manually written rules associated with bilingual dictionaries (rule- based systems), (2) a statistical framework (statis- tical machine translation) based on large amount of monolingual and parallel corpora. The first line uses linguistically generalized information based on what humans understand from what happens in a given language (source and target) and what hap- pens in the translation process. The translation process is building a translation from a given source sentence based on this knowledge. The sec- ond line exploits implicit information present in already translated corpora and more generally any text production in the target language to automati- cally find the most likely translation for a given source sentence. This approach has proven to be competitive with the rule-based approach when provided with enough resources on a specific do- main. Though based on fundamentally different paradigms and exploiting different types of infor- mation, these two research lines are not in opposi- tion and may be combined to produce improved results. For instance, serial combination of the two approaches has produced very good results in WMT07 (Simard, 2007), (Dugast, 2007) and NIST07 (Ueffing, 2008). (Schwenk et al., 2008) also combines both approaches and resources to build a better system. The SYSTRAN’s R&D team actually works to merge these two approaches, drawing benefit from their respective strengths. Initially, the SYSTRAN system was a pure rule-based system that in recent years began integrating statistical features and cor- pus-based model (Senellart, 2006). It must be noted that, for sake of simplification of the ex- periment and its interpretation, the base system mentioned in this paper is a purely rule-based ver- sion. In the framework of this research effort, vari- ous exploratory experiments are being run which aim both at finding efficient combination setups and at discriminating strengths and weaknesses of rule-based and statistical systems. We had performed a first analysis on a statistical post-editing system (Dugast, 2007). The system submitted for Czech-English follows this setup. We present also here an original French-English statistical model which doesn’t make use of the target side of the parallel data to train its phrase- table, but rather …

Content truncated to 500 words