Chimera – Three Heads for English-to-Czech Translation

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 92–98, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Chimera – Three Heads for English-to-Czech Translation Ondˇrej Bojar and Rudolf Rosa and Aleˇs Tamchyna Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´amˇest´ı 25, Prague, Czech Republic surname@ufal.mff.cuni.cz Abstract This paper describes our WMT submis- sions CU-BOJAR and CU-DEPFIX, the lat- ter dubbed “CHIMERA” because it com- bines on three diverse approaches: Tec- toMT, a system with transfer at the deep syntactic level of representation, factored phrase-based translation using Moses, and finally automatic rule-based correction of frequent grammatical and meaning errors. We do not use any off-the-shelf system- combination method. 1 Introduction Targeting Czech in statistical machine transla- tion (SMT) is notoriously difficult due to the large number of possible word forms and com- plex agreement rules. Previous attempts to resolve these issues include specific probabilistic models (Subotin, 2011) or leaving the morphological gen- eration to a separate processing step (Fraser et al., 2012; Mareˇcek et al., 2011). TectoMT (CU-TECTOMT, Galuˇsˇc´akov´a et al. (2013)) is a hybrid (rule-based and statistical) MT system that closely follows the analysis-transfer- synthesis pipeline. As such, it suffers from many issues but generating word forms in proper agree- ments with their neighbourhood as well as the translation of some diverging syntactic structures are handled well. Overall, TectoMT sometimes even ties with a highly tuned Moses configuration in manual evaluations, see Bojar et al. (2011). Finally, Rosa et al. (2012) describes Depfix, a rule-based system for post-processing (S)MT out- put that corrects some morphological, syntactic and even semantic mistakes. Depfix was able to significantly improve Google output in WMT12, so now we applied it on an open-source system. Our WMT13 system is thus a three-headed creature where, hopefully: (1) TectoMT provides missing word forms and safely handles some non- parallel syntactic constructions, (2) Moses ex- ploits very large parallel and monolingual data, and boosts better lexical choice, (3) Depfix at- tempts to fix severe flaws in Moses output. 2 System Description TectoMT Moses cu-tectomt Depfix cu-bojar cu-depfix = Chimera Input Figure 1: CHIMERA: three systems combined. CHIMERA is a sequential combination of three diverse MT systems as depicted in Figure 1. Each of the intermediate stages of processing has been submitted as a separate primary system for the WMT manual evalution, allowing for a more thor- ough analysis. Instead of an off-the-shelf system combination technique, we use TectoMT output as synthetic training data for Moses as described in Section 2.1 and finally we process its output using rule-based corrections of Depfix (Section 2.2). All steps di- rectly use the source sentence. 2.1 Moses Setup for CU-BOJAR We ran a couple of probes with reduced training data around the setup of Moses that proved suc- cessful in previous years (Bojar et al., 2012a). 2.1.1 Pre-processing We use a stable pre-processing pipeline that in- cludes normalization of quotation marks,1 tok- enization, tagging and lemmatization with tools 1We do …

Content truncated to 500 words