PhraseFix: Statistical Post-Editing of TectoMT

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 141–147, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics PhraseFix: Statistical Post-Editing of TectoMT Petra Galušˇcáková, Martin Popel, and Ondˇrej Bojar Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostranské námˇestí 25, Prague, Czech Republic {galuscakova,popel,bojar}@ufal.mff.cuni.cz Abstract We present two English-to-Czech systems that took part in the WMT 2013 shared task: TECTOMT and PHRASEFIX. The former is a deep-syntactic transfer-based system, the latter is a more-or-less stan- dard statistical post-editing (SPE) applied on top of TECTOMT. In a brief survey, we put SPE in context with other system com- bination techniques and evaluate SPE vs. another simple system combination tech- nique: using synthetic parallel data from TECTOMT to train a statistical MT sys- tem (SMT). We confirm that PHRASEFIX (SPE) improves the output of TECTOMT, and we use this to analyze errors in TEC- TOMT. However, we also show that ex- tending data for SMT is more effective. 1 Introduction This paper describes two submissions to the WMT 2013 shared task:1 TECTOMT – a deep- syntactic tree-to-tree system and PHRASEFIX – statistical post-editing of TECTOMT using Moses (Koehn et al., 2007). We also report on exper- iments with another hybrid method where TEC- TOMT is used to produce additional (so-called synthetic) parallel training data for Moses. This method was used in CU-BOJAR and CU-DEPFIX submissions, see Bojar et al. (2013). 2 Overview of Related Work The number of approaches to system combination is enormous. We very briefly survey those that form the basis of our work reported in this paper. 2.1 Statistical Post-Editing Statistical post-editing (SPE, see e.g. Simard et al. (2007), Dugast et al. (2009)) is a popular method 1http://www.statmt.org/wmt13 for improving outputs of a rule-based MT sys- tem. In principle, SPE could be applied to any type of first-stage system including a statistical one (Oflazer and El-Kahlout, 2007; Béchara et al., 2011), but most benefit could be expected from post-editing rule-based MT because of the com- plementary nature of weaknesses and advantages of rule-based and statistical approaches. SPE is usually done with an off-the-shelf SMT system (e.g. Moses) which is trained on output of the first-stage system aligned with reference trans- lations of the original source text. The goal of SPE is to produce translations that are better than both the first-stage system alone and the second-stage SMT trained on the original training data. Most SPE approaches use the reference trans- lations from the original training parallel corpus to train the second-stage system. In contrast, Simard et al. (2007) use human-post-edited first- stage system outputs instead. Intuitively, the lat- ter approach achieves better results because the human-post-edited translations are closer to the first-stage output than the original reference trans- lations. Therefore, SPE learns to perform the changes which are needed the most. However, cre- ating human-post-edited translations is laborious and must be done again for each new (version of the) first-stage system in order to preserve its full advantage …

Content truncated to 500 words