Mood at work: Ramses versus Pharaoh

Proceedings of the Workshop on Statistical Machine Translation, pages 126–129, New York City, June 2006. c⃝2006 Association for Computational Linguistics Mood at work: Ramses versus Pharaoh Alexandre Patry, Fabrizio Gotti and Philippe Langlais RALI — DIRO Universit´e de Montr´eal {patryale,gottif,felipe}@iro.umontreal.ca Abstract We present here the translation system we used in this year’s WMT shared task. The main objective of our participation was to test RAMSES, an open source phrase- based decoder. For that purpose, we used the baseline system made available by the organizers of the shared task1 to build the necessary models. We then carried out a pair-to-pair comparison of RAMSES with PHARAOH on the six different translation directions that we were asked to perform. We present this comparison in this paper. 1 Introduction Phrase-based (PB) machine translation (MT) is now a popular paradigm, partly because of the relative ease with which we can automatically create an ac- ceptable translation engine from a bitext. As a mat- ter of fact, deriving such an engine from a bitext con- sists in (more or less) gluing together dedicated soft- ware modules, often freely available. Word-based models, or the so-called IBM models, can be trained using the GIZA or GIZA++ toolkits (Och and Ney, 2000). One can then train phrase-based models us- ing the THOT toolkit (Ortiz-Mart´ınez et al., 2005). For their part, language models currently in use in SMT systems can be trained using packages such as SRILM (Stolcke, 2002) and the CMU-SLM toolkit (Clarkson and Rosenfeld, 1997). 1www.statmt.org/wmt06/shared-task/ baseline.html Once all the models are built, one can choose to use PHARAOH (Koehn, 2004), an efficient full- fledged phrase-based decoder. We only know of one major drawback when using PHARAOH: its licensing policy. Indeed, it is available for non- commercial use in its binary form only. This severely limits its use, both commercially and sci- entifically (Walker, 2005). For this reason, we undertook the design of a generic architecture called MOOD (Modular Object- Oriented Decoder), especially suited for instantiat- ing SMT decoders. Two major goals directed our design of this package: offering open source, state- of-the-art decoders and providing an architecture to easily build these decoders. This effort is described in (Patry et al., 2006). As a proof of concept that our framework (MOOD) is viable, we attempted to use its functionalities to implement a clone of PHARAOH, based on the com- prehensive user manual of the latter. This clone, called RAMSES, is now part of the MOOD distribu- tion, which can be downloaded freely from the page http://smtmood.sourceforge.net. We conducted a pair-to-pair comparison between the two engines that we describe in this paper. We provide an overview of the MOOD architecture in Section 2. Then we describe briefly RAMSES in Sec- tion 3. The comparison between the two decoders in terms of automatic metrics is analyzed in Section 4. We confirm this comparison by presenting a man- ual evaluation we conducted on an random sample of the translations produced by both decoders. This is reported in Section 5. We conclude …

Content truncated to 500 words