The KIT-LIMSI Translation System for WMT 2015

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 120–125, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. The KIT-LIMSI Translation System for WMT 2015 †Thanh-Le Ha, ∗Quoc-Khanh Do, †Eunah Cho, †Jan Niehues, ∗Alexandre Allauzen, ∗Franc¸ois Yvon and †Alex Waibel †Karlsruhe Institute of Technology, Karlsruhe, Germany ∗LIMSI-CNRS, Orsay, France †firstname.surname@kit.edu ∗firstname.surname@limsi.fr Abstract This paper presented the joined submis- sion of KIT and LIMSI to the English to German translation task of WMT 2015. In this year submission, we integrated a neu- ral network-based translation model into a phrase-based translation model by rescor- ing the n-best lists. Since the computation complexity is one of the main issues for continuous space models, we compared two techniques to reduce the computation cost. We inves- tigated models using a structured output layer as well as models trained with noise contrastive estimation. Furthermore, we evaluated a new method to obtain the best log-linear combination in the rescoring phase. Using these techniques, we were able to improve the BLEU score of the baseline phrase-based system by 1.4 BLEU points. 1 Introduction In this paper, we present the English→German joint translation system from KIT and LIMSI par- ticipating in the Shared Translation Task of the EMNLP 2015 - Tenth Workshop on Statistical Machine Translation (WMT2015). Our system is the combination of two different approaches. First, a strong phrase-based system from KIT is used to generate a k-best list of translated candi- dates. Second, an n-gram translation model from LIMSI, named SOUL (Structured OUtput Layer), helps to rescore the k-best list by utilizing features extracted from translated tuples. In this year par- ticipation, we also use a version of the neural net- work translation models (Le et al., 2012) trained using NCE algorithm (Gutmann and Hyv¨arinen, 2010) as counterpart to SOUL models. A ListNet- based rescoring method is then applied to integrate two abovementioned approaches. Section 2 describes the KIT phrase-based trans- lation system which is conducted over the phrase pairs. Section 3 describes the LIMSI SOUL and NCE translation models estimated on source-and- target n-gram tuples. We explain the rescoring ap- proach in Section 4. Finally, Section 5 summa- rizes the experimental results of our joint system submitted to WMT2015. 2 KIT Phrase-based Translation System The KIT translation system uses a phrase-based in-house decoder (Vogel, 2003) which finds the best combinations of features in a log-linear framework. The features consist of translation scores, distortion-based and lexicalized reorder- ing scores as well as conventional and non-word language models. In addition, several reorder- ing rules, including short-range, long-range and tree-based reorderings, are applied before decod- ing step as they are encoded as word lattices. The decoder then generates a list of the best candidates from the lattices. To optimize the factors of indi- vidual features on a development dataset, we use minimum error rate training (MERT) (Venugopal et al., 2005). We are going to describe those com- ponents in detail as follows. 2.1 Data and Preprocessing The parallel data mainly used are the corpora ex- tracted …

Content truncated to 500 words