Towards Effective Use of Training Data in Statistical Machine Translation

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 317–321, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Towards Effective Use of Training Data in Statistical Machine Translation Philipp Koehn and Barry Haddow University of Edinburgh Edinburgh, United Kingdom {pkoehn,bhaddow}@inf.ed.ac.uk Abstract We report on findings of exploiting large data sets for translation modeling, language mod- eling and tuning for the development of com- petitive machine translation systems for eight language pairs. 1 Introduction We report on experiments carried out for the devel- opment of competitive systems on the datasets of the 2012 Workshop on Statistical Machine Translation. Our main focus was directed on the effective use of all the available training data during training of translation and language models and tuning. We use the open source machine translation sys- tem Moses (Koehn et al., 2007) and other standard open source tools, hence all our experiments are straightforwardly replicable1. Compared to all single system submissions by participants of the workshop we achieved the best BLEU scores for four language pairs (es-en, en-es, cs-en, en-cs), the 2nd best results for two language pairs (fr-en, de-en), as well as a 3rd place (en-de) and a 5th place (en-fr) for the remaining pairs. We improved upon this in the post-evaluation period for some of the language pairs by more systematically applying our methods. During the development of our system, we saw most gains from using large corpora for translation model training, especially when using subsampling techniques for out-of-domain sets, using large cor- pora for language model training, and larger tuning sets. We also observed mixed results with alternative tuning methods. We also experimented with hierar- chical models and semi-supervised training, but did not achieve any improvements. 1Configuration files and instructions are available at http: //www.statmt.org/wmt12/uedin/. LP Baseline +UN fr-en 28.2 28.4 (+.2) es-en 29.1 28.9 (–.2) en-fr 28.8 28.7 (–.1) en-es 31.0 30.9 (–.1) LP Baseline +GigaFrEn fr-en 28.7 29.1 (+.4) en-fr 29.3 30.3 (+1.0) Table 1: Gains from larger translation models: UN (about 300 million English words), GigaFrEn (about 550 million English words). We report all results in case-sensitive BLEU (mt- eval13a) on the newstest2011 test set (Callison- Burch et al., 2011). Please also note that base- line scores vary throughout the paper, since different methods were investigated at different time points. 2 Better Translation Models 2.1 Using Large Training Sets The WMT evaluation campaign works with the largest training sets in the field. Our French-English systems are trained on a parallel corpus with 1,072 million French and 934 million English words. Training a system on this amount of data takes about two weeks. The basic data sets for the language pairs are the Europarl and NewsCommentary corpora consist of about 50 million words and 3 million words, respec- tively. These corpora are quite close to the target domain of news reports, and give quite good results. Table 1 shows the gains from using the much larger UN (about 300 million words) and GigaFrEn cor- pora (about 550 million words). …

Content truncated to 500 words