Statistical Post-Editing on SYSTRAN‘s Rule-Based Translation System

Proceedings of the Second Workshop on Statistical Machine Translation, pages 220–223, Prague, June 2007. c⃝2007 Association for Computational Linguistics Statistical Post-Editing on SYSTRAN’s Rule-Based Translation System Loïc Dugast, Jean Senellart SYSTRAN SA La Grande Arche 1, Parvis de la Défense 92044 Paris La Défense Cedex France dugast@systran.fr senellart@systran.fr Philipp Koehn School of Informatics University of Edinburgh 2 Buccleuch Place Edinburgh, EH8 9LW United Kingdom pkoehn@inf.ed.ac.uk Abstract This article describes the combination of a SYSTRAN system with a “statistical post- editing” (SPE) system. We document qualitative analysis on two experiments performed in the shared task of the ACL 2007 Workshop on Statistical Machine Translation. Comparative results and more integrated “hybrid” techniques are dis- cussed. 1 Introduction The evolution of SYSTRAN’s architecture over the last years has been to « open » the system to enable interaction between the internal system’s rules and the external input – see Senellart (2003), Attnas et al. (2005). Based on this architecture, several directions are explored to introduce the use of « corpus-based » approaches at several levels of the process: - use of corpus-based tools to validate and enrich linguistic resources (detection of forbidden se- quences, bilingual terminology extraction), - auto- matic recognition of the text domain, - use of a corpus-based decision mechanism within « word boundary » (Chinese word identification), disam- biguation… - use of word sense disambiguation techniques – and the use of a language model in the generation phase to select alternative transla- tions, prepositions, and local reordering (adjective positioning). These tools have been presented in Senellart (2006) and most of them will be integrated in SYSTRAN version 7 systems. Independently, two experiments were carried out for the shared task of the ACL 2007 Workshop on Statistical Machine Translation to combine a raw SYSTRAN system with a statistical post- editing (SPE) system. One experiment was run by NRC using the language pair English<>French in the context of « Automatic Post-Edition » systems using the PORTAGE system as described in Si- mard et al. (2007). The second experiment based on the same principle was run on the Ger- man>English and Spanish>English1 language pairs using the Moses system (Koehn et al. 2007). The objective was to train a SMT system on a parallel corpus composed of SYSTRAN translations with the referenced source aligned with its referenced translation. Beyond both (a) the huge (and expected) im- provement of the BLEU score for the combined system compared to raw translation output (for German-English, around 10 BLEU points for the Europarl test set of WMT2007) and (b) the (ex- pected) corresponding improvement of the transla- tion fluency, we provide qualitative analysis on the contributions (positive and negative) of the SPE layer imposed on the SYSTRAN translation output in this paper. For this analysis we classifiy the dif- ferent types of “post-editing” changes and point out the alternative isolated statistical components that could achieve the same results. We conclude with two possible approaches: breaking down the “statistical layer” into different components/tools each specialized in a narrow and …

Content truncated to 500 words