TÜBİTAK-BİLGEM German-English Machine Translation Systems for W13

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 109–113, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics T ¨UB˙ITAK-B˙ILGEM German-English Machine Translation Systems for WMT’13 ˙Ilknur Durgar El-Kahlout and Cos¸kun Mermer T ¨UB˙ITAK-B˙ILGEM Gebze 41470, Kocaeli, TURKEY {ilknur.durgar,coskun.mermer}@tubitak.gov.tr Abstract This paper describes T ¨UB˙ITAK-B˙ILGEM statistical machine translation (SMT) sys- tems submitted to the Eighth Work- shop on Statistical Machine Transla- tion (WMT) shared translation task for German-English language pair in both di- rections. We implement phrase-based SMT systems with standard parameters. We present the results of using a big tun- ing data and the effect of averaging tun- ing weights of different seeds. Addition- ally, we performed a linguistically moti- vated compound splitting in the German- to-English SMT system. 1 Introduction T ¨UB˙ITAK-B˙ILGEM participated for the first time in the WMT’13 shared translation task for the German-English language pairs in both directions. We implemented a phrase-based SMT system by using the entire available training data. In the German-to-English SMT system, we performed a linguistically motivated compound splitting. We tested different language model (LM) combina- tions by using the parallel data, monolingual data, and Gigaword v4. In each step, we tuned systems with five different tune seeds and used the average of tuning weights in the final system. We tuned our systems on a big tuning set which is generated from the last years’ (2008, 2009, 2010, and 2012) development sets. The rest of the paper describes the details of our systems. 2 German-English 2.1 Baseline All available data was tokenized, truecased, and the maximum number of tokens were fixed to 70 for the translation model. The Moses open SMT toolkit (Koehn et al., 2007) was used with MGIZA++ (Gao and Vogel, 2008) with the stan- dard alignment heuristic grow-diag-final (Och and Ney, 2003) for word alignments. Good-Turing smoothing was used for phrase extraction. Sys- tems were tuned on newstest2012 with MERT (Och, 2003) and tested on newstest2011. 4- gram language models (LMs) were trained on the target side of the parallel text and the mono- lingual data by using SRILM (Stolcke, 2002) toolkit with Kneser-Ney smoothing (Kneser and Ney, 1995) and then binarized by using KenLM toolkit (Heafield, 2011). At each step, systems were tuned with five different seeds with lattice- samples. Minimum Bayes risk decoding (Kumar and Byrne, 2004) and -drop-unknown parameters were used during the decoding. This configuration is common for all of the ex- periments decribed in this paper unless stated oth- erwise. Table 1 shows the number of sentences used in system training after the clean-corpus pro- cess. Data Number of sentences Europarl 1908574 News-Commentary 177712 Commoncrawl 726458 Table 1: Parallel Corpus. We trained two baseline systems in order to as- sess the effects of this year’s new parallel data, commoncrawl. We first trained an SMT system by using only the training data from the previ- ous WMT shared translation tasks that is europarl and news-commentary (Baseline1). As the second baseline, we also included the new parallel data commoncrawl only in the …

Content truncated to 500 words