The KIT-LIMSI Translation System for WMT 2014

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 84–89, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics The KIT-LIMSI Translation System for WMT 2014 ∗Quoc Khanh Do, †Teresa Herrmann, ∗†Jan Niehues, ∗Alexandre Allauzen, ∗Franc¸ois Yvon and †Alex Waibel ∗LIMSI-CNRS, Orsay, France †Karlsruhe Institute of Technology, Karlsruhe, Germany ∗surname@limsi.fr †firstname.surname@kit.edu Abstract This paper describes the joined submis- sion of LIMSI and KIT to the Shared Translation Task for the German-to- English direction. The system consists of a phrase-based translation system us- ing a pre-reordering approach. The base- line system already includes several mod- els like conventional language models on different word factors and a discriminative word lexicon. This system is used to gen- erate a k-best list. In a second step, the list is reranked using SOUL language and translation models (Le et al., 2011). Originally, SOUL translation models were applied to n-gram-based translation sys- tems that use tuples as translation units instead of phrase pairs. In this article, we describe their integration into the KIT phrase-based system. Experimental re- sults show that their use can yield sig- nificant improvements in terms of BLEU score. 1 Introduction This paper describes the KIT-LIMSI system for the Shared Task of the ACL 2014 Ninth Work- shop on Statistical Machine Translation. The sys- tem participates in the German-to-English trans- lation task. It consists of two main components. First, a k-best list is generated using a phrase- based machine translation system. This system will be described in Section 2. Afterwards, the k- best list is reranked using SOUL (Structured OUt- put Layer) models. Thereby, a neural network lan- guage model (Le et al., 2011), as well as several translation models (Le et al., 2012a) are used. A detailed description of these models can be found in Section 3. While the translation system uses phrase pairs, the SOUL translation model uses tu- ples as described in the n-gram approach (Mari˜no et al., 2006). We describe the integration of the SOUL models into the translation system in Sec- tion 3.2. Section 4 summarizes the experimen- tal results and compares two different tuning al- gorithms: Minimum Error Rate Training (Och, 2003) and k-best Batch Margin Infused Relaxed Algorithm (Cherry and Foster, 2012). 2 Baseline system The KIT translation system is an in-house imple- mentation of the phrase-based approach and in- cludes a pre-ordering step. This system is fully described in Vogel (2003). To train translation models, the provided Eu- roparl, NC and Common Crawl parallel corpora are used. The target side of those parallel corpora, the News Shuffle corpus and the GigaWord cor- pus are used as monolingual training data for the different language models. Optimization is done with Minimum Error Rate Training as described in Venugopal et al. (2005), using newstest2012 and newstest2013 as development and test data, respectively. Compound splitting (Koehn and Knight, 2003) is performed on the source side (German) of the corpus before training. Since the web-crawled Common Crawl corpus is noisy, this corpus is first filtered using an …

Content truncated to 500 words