CimS - The CIS and IMS Joint Submission to WMT 2015 addressing morphological and syntactic differences in English to German SMT

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 84–91, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. CimS - The CIS and IMS Joint Submission to WMT 2015 addressing morphological and syntactic differences in English to German SMT Fabienne Cap1, Marion Weller1,2, Anita Ramm1 and Alexander Fraser1 1 CIS, Ludwig-Maximilian University of Munich – (cap|ramm|fraser)@cis.uni-muenchen.de 2 IMS, University of Stuttgart – wellermn@ims.uni-stuttgart.de Abstract We present the CimS submissions to the WMT 2015 Shared Task for the transla- tion direction English to German. Simi- lar to our previous submissions, all of our systems are aware of the complex nomi- nal morphology of German. In this pa- per, we combine source-side reordering and target-side compound processing with basic morphological processing in order to obtain improved translation results. We also report on morphological processing for English to French. 1 Introduction This paper presents our submissions to the WMT shared task 2015. We use customised solutions to address morphological challenges in the English to German translation direction. Our goal is to make German and English as similar as possible in order to obtain better word alignments and hence an improved translation quality. We base our work on three main components, which we have care- fully investigated separately in the past. (i) Nominal Inflection We use context-based prediction of German inflectional endings. This improves fluency and enables the creation of mor- phological forms which have not occurred in the training data. (ii) Source-side Reordering We reorder the En- glish source text in order to make it more sim- ilar to the German word order. This improves word alignment and thus translation quality. It also makes the reordering task in decoding easier. (iii) Compound Processing We split German compounds into simple words for training. In decoding, we translate only simple words, some of which are re-combined into compounds after- wards in post-processing. This allows us to create compounds which have not occurred in the train- ing data. This year, our main focus is on combining nominal inflection prediction and source-side reordering. We investigated both of these components sepa- rately in the past and expect an additive positive effect on translation quality when combined. We then added compound processing, which we al- ready have investigated in combination with nomi- nal inflection before, but not together with source- side reordering. Here, we also expect the com- bination to outperform the single components in terms of translation quality. 2 Methodology The underlying idea of all of our systems is to improve translation quality by making the source and target languages more similar than they usu- ally are. We address three common problems in English to German SMT: morphological richness in terms of inflectional variants, productive com- pounding and different word orders. In Figure 1, we illustrate the latter two of these problems using an example sentence which contains both a Ger- man compound (“Mehrheitsvotum” = “majority vote”) and different word orders. The methods we use to solve all three of these problems are implemented as …

Content truncated to 500 words