The Karlsruhe Institute of Technology Translation Systems for the WMT 2013

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 104–108, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics The Karlsruhe Institute of Technology Translation Systems for the WMT 2013 Eunah Cho, Thanh-Le Ha, Mohammed Mediani, Jan Niehues, Teresa Herrmann, Isabel Slawik and Alex Waibel Karlsruhe Institute of Technology Karlsruhe, Germany firstname.lastname@kit.edu Abstract This paper describes the phrase-based SMT systems developed for our partici- pation in the WMT13 Shared Translation Task. Translations for English↔German and English↔French were generated us- ing a phrase-based translation system which is extended by additional models such as bilingual, fine-grained part-of- speech (POS) and automatic cluster lan- guage models and discriminative word lexica (DWL). In addition, we combined reordering models on different sentence abstraction levels. 1 Introduction In this paper, we describe our systems for the ACL 2013 Eighth Workshop on Statistical Ma- chine Translation. We participated in the Shared Translation Task and submitted translations for English↔German and English↔French using a phrase-based decoder with lattice input. The paper is organized as follows: the next sec- tion gives a detailed description of our systems including all the models. The translation results for all directions are presented afterwards and we close with a conclusion. 2 System Description The phrase table is based on a GIZA++ word alignment for the French↔English systems. For the German↔English systems we use a Discrim- inative Word Alignment (DWA) as described in Niehues and Vogel (2008). For every source phrase only the top 10 translation options are con- sidered during decoding. The SRILM Toolkit (Stolcke, 2002) is used for training SRI language models using Kneser-Ney smoothing. For the word reordering between languages, we used POS-based reordering models as described in Section 4. In addition to it, tree-based reordering model and lexicalized reordering were added for German↔English systems. An in-house phrase-based decoder (Vogel, 2003) is used to perform translation. The trans- lation was optimized using Minimum Error Rate Training (MERT) as described in Venugopal et al. (2005) towards better BLEU (Papineni et al., 2002) scores. 2.1 Data The Europarl corpus (EPPS) and News Commen- tary (NC) corpus were used for training our trans- lation models. We trained language models for each language on the monolingual part of the training corpora as well as the News Shuffle and the Gigaword corpora. The additional data such as web-crawled corpus, UN and Giga corpora were used after filtering. The filtering work for this data is discussed in Section 3. For the German↔English systems we use the news-test2010 set for tuning, while the news- test2011 set is used for the French↔English sys- tems. For testing, news-test2012 set was used for all systems. 2.2 Preprocessing The training data is preprocessed prior to train- ing the system. This includes normalizing special symbols, smart-casing the first word of each sen- tence and removing long sentences and sentence pairs with length mismatch. Compound splitting is applied to the German part of the corpus of the German→English system as described in Koehn and Knight (2003). 3 Filtering of Noisy Pairs The filtering …

Content truncated to 500 words