The Karlsruhe Institute of Technology Translation Systems for the WMT 2014

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 130–135, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics The Karlsruhe Institute of Technology Translation Systems for the WMT 2014 Teresa Herrmann, Mohammed Mediani, Eunah Cho, Thanh-Le Ha, Jan Niehues, Isabel Slawik, Yuqi Zhang and Alex Waibel Institute for Anthropomatics and Robotics KIT - Karlsruhe Institute of Technology firstname.lastname@kit.edu Abstract In this paper, we present the KIT systems participating in the Shared Translation Task translating between English↔German and English↔French. All translations are generated using phrase-based translation systems, using different kinds of word-based, part-of- speech-based and cluster-based language models trained on the provided data. Additional models include bilingual lan- guage models, reordering models based on part-of-speech tags and syntactic parse trees, as well as a lexicalized reordering model. In order to make use of noisy web-crawled data, we apply filtering and data selection methods for language modeling. A discriminative word lexicon using source context information proved beneficial for all translation directions. 1 Introduction We describe the KIT systems for the Shared Trans- lation Task of the ACL 2014 Ninth Workshop on Statistical Machine Translation. We participated in the English↔German and English↔French translation directions, using a phrase-based de- coder with lattice input. The paper is organized as follows: the next sec- tion describes the data used for each translation direction. Section 3 gives a detailed description of our systems including all the models. The trans- lation results for all directions are presented after- wards and we close with a conclusion. 2 Data We utilize the provided EPPS, NC and Common Crawl parallel corpora for English→German and German→English, plus Giga for English→French and French→English. The monolingual part of those parallel corpora, the News Shuffle corpus for all four directions and additionally the Gigaword corpus for English→French and German→English are used as monolingual train- ing data for the different language models. For optimizing the system parameters, newstest2012 and newstest2013 are used as development and test data respectively. 3 System Description Before training we perform a common preprocess- ing of the raw data, which includes removing long sentences and sentences with a length mismatch exceeding a certain threshold. Afterwards, we nor- malize special symbols, dates, and numbers. Then we perform smart-casing of the first letter of every sentence. Compound splitting (Koehn and Knight, 2003) is performed on the source side of the cor- pus for German→English translation. In order to improve the quality of the web-crawled Common Crawl corpus, we filter out noisy sentence pairs us- ing an SVM classifier for all four translation tasks as described in Mediani et al. (2011). Unless stated otherwise, we use 4-gram lan- guage models (LM) with modified Kneser-Ney smoothing, trained with the SRILM toolkit (Stol- cke, 2002). All translations are generated by an in-house phrase-based translation system (Vo- gel, 2003), and we use Minimum Error Rate Training (MERT) as described in Venugopal et al. (2005) for optimization. The word align- ment of the parallel corpora is generated using the GIZA++ Toolkit (Och and Ney, 2003) …

Content truncated to 500 words