CimS – The CIS and IMS joint submission to WMT 2014 translating from English into German

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 71–78, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics CimS – The CIS and IMS joint submission to WMT 2014 translating from English into German Fabienne Cap:, Marion Weller:�, Anita Ramm�, Alexander Fraser: : CIS, Ludwig-Maximilian University of Munich – (cap|fraser)@cis.uni-muenchen.de � IMS, University of Stuttgart – (wellermn|ramm)@ims.uni-stuttgart.de Abstract We present the CimS submissions to the 2014 Shared Task for the language pair EN→DE. We address the major problems that arise when translating into German: complex nominal and verbal morphol- ogy, productive compounding and flex- ible word ordering. Our morphology- aware translation systems handle word formation issues on different levels of morpho-syntactic modeling. 1 Introduction In our shared task submissions, we focus on the English to German translation direction: we ad- dress different levels of productivity of the Ger- man language, i.e., nominal and verbal inflec- tion and productive word formation, which lead to data sparsity and thus confuse classical SMT systems. Our basic goal is to make the two languages as morphosyntactically similar as possible. We use a parser and a morphological analyser to re- move linguistic features from German that are not present in English and reorder the English input to make it more similar to the German sen- tence structure. Prior to training, all words are lemmatised and compounds are split into single words. This is not only beneficial for word align- ment, but it also allows us to generalise over in- flectional variants of the same lexemes and over single words which could occur in one place as a standalone word and in another place as part of a compound. Translation happens in two steps: first, we translate from English into split, lemma- tised German and then, we perform compound merging and generation of inflection as a post- processing step. This way, we are able to cre- ate German compounds and inflectional vari- ants that have not been seen in the parallel train- ing data. In this paper, we investigate the performance of well-established source-side reordering, nomi- nal re-inflection and compound processing sys- tems on an up-to-date shared task. In addition, we present experimental results on a verbal in- flection component and a syntax-based variant including source-side reordering. 2 Related Work Re-Inflection The two-step translation ap- proach we use was described by e.g. Toutanova aet al. (2008) and Jeong et al. (2010), who use number of morphological and syntactic features derived from both source and target language. More recently, Fraser et al. (2012) describe a similar approach for German using different CRF-based feature prediction models, one for each of the four grammatical features to be predicted for German words in noun phrases, namely number, gender, case and definiteness. This approach also handles word- formation issues such as portmanteau splitting and compounding. Weller et al. (2013) added subcategorization information in combination with source-side syntactic features in order to improve the prediction of case. De Gispert and Mariño (2008) generate verbal inflection for translation from …

Content truncated to 500 words