Statistical Post Editing and Dictionary Extraction: Systran/Edinburgh Submissions for ACL-WMT2009

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 110–114, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics Statistical Post Editing and Dictionary Extraction: Systran/Edinburgh submissions for ACL-WMT2009 Loic Dugast*,** and Jean Senellart* *SYSTRAN S.A. La Grande Arche 1, Parvis de la D´efense 92044 Paris La D´efense Cedex France Philipp Koehn** **School of Informatics University of Edinburgh 10 Crichton Street, Edinburgh United Kingdom Abstract We describe here the two Sys- tran/University of Edinburgh submissions for WMT2009. They involve a statistical post-editing model with a particular han- dling of named entities (English to French and German to English) and the extraction of phrasal rules (English to French). 1 Introduction Previous results had shown a rather satisfying per- formance for hybrid systems such as the Statis- tical Phrase-based Post-Editing (SPE) (Simard et al., 2007) combination in comparison with purely phrase-based statistical models, reaching simi- lar BLEU scores and often receiving better hu- man judgement (German to English at WMT2007) against the BLEU metric. This last result was in accordance with the previous acknowledgment (Callison-Burch et al., 2006) that systems of too differing structure could not be compared reliably with BLEU. We participated in the recent Work- shop on Machine Translation (WMT’09) in the language pairs English to French and German to English. On the one hand we trained a Post- Editing system with an additional special treat- ment to avoid the loss of entities such as dates and numbers. On the other hand we trained an addi- tional English-to-French system (as a secondary submission) that made use of automatically ex- tracted linguistic entries. In this paper, we will present both approaches. The latter is part of on- going work motivated by the desire to both make use of corpus statistics and keep the advantage of the often (relative to automatic metrics’s scores) higher rank in human judgement given to rule- based systems on out-of-domain data, as seen on Figure 1: Translation with PBMT post-editing the WMT 2008 results for both English to French and German to English (Callison-Burch et al., 2008). 2 Statistical Post Editing systems 2.1 Baseline The basic setup is identical to the one described in (Dugast et al., 2007). A statistical translation model is trained between the rule-based transla- tion of the source-side and the target-side of the parallel corpus. This is done separately for each parallel corpus. Language models are trained on each target half of the parallel corpora and also on additional in-domain corpora. Figure 1 shows the translation process. Here are a few additional details which tend to improve training and limit unwanted statistical ef- fects in translation: • Named entities are replaced by special tokens on both sides. By reducing vocabulary and combined with the next item mentioned, this should help word alignment. Moreover, en- tity translation is handled more reliably by the rule-based engine. • The intersection of both vocabularies (i.e. vo- 110 cabularies of the rule-based output and the reference translation) is used to produce an …

Content truncated to 500 words