Training Data in Statistical Machine Translation - the More, the Better?

Proceedings of Recent Advances in Natural Language Processing, pages 551–556, Hissar, Bulgaria, 12-14 September 2011. Training Data in Statistical Machine Translation – The More, the Better? – Monica Gavrila University of Hamburg, Germany gavrila@informatik. uni-hamburg.de Cristina Vertan University of Hamburg, Germany cristina.vertan@uni-hamburg.de Abstract Current statistical machine translation (SMT) systems are stated to be dependent on the availability of a very large training data for producing the language and trans- lation models. Unfortunately, large paral- lel corpora are available for a limited set of language pairs and for an even more lim- ited set of domains. In this paper we investigate the behavior of an SMT system exposed to training data of different sizes and types. Our experimen- tal results show that even parallel corpora of modest sizes can be used for training purposes without lowering too much the evaluation scores. We consider two lan- guage pairs in both translation directions for the experiments: English-Romanian and German-Romanian. 1 Introduction Statistical machine translation (SMT) is the most frequently used paradigm, especially when a trans- lation system has to be implemented for a new (less researched) language pair. The pure statisti- cal approach has the advantage that no additional bilingual linguistic expertise is required. Once the training data is available, open-source, language independent systems can be reused. However, the quality of the results is strongly influenced by the size and type of the available training data. State-of-the-art literature tends to share the opinion that the larger the data, the better the re- sults. (Suresh, 2010) shows that a larger corpus size for training increases the quality of a Moses- based SMT system, for the Europarl corpus for English-French. The same conclusion appears also in (Koehn et al., 2003), for German-English. In (Brants et al., 2007) experiments for Arabic- English data with billions of tokens are presented and a dependency between the output quality and the size of the training data is also demonstrated. Unfortunately, large amount of parallel train- ing data is available only for a restricted num- ber of language pairs and domains. Addition- ally, the training step on large corpora is time and (computing-) resources consuming. On the other hand, smaller corpora can be more easily achieved and have the advantage of requiring less time for training. They also offer the possibility of manu- ally correcting and creating the data. Experiments with smaller data for Serbian- English (approx. 2.6K sentences) are presented in (Popovic and Ney, 2006). In the same paper also experimental results for Spanish-English, with dif- ferent data sizes are reported. The systems trained on smaller data give acceptable results. However, the trend remains the same: larger data provides better results. For English-Romanian, SMT systems are pre- sented in (Cristea, 2009) and (Ignat, 2009), with BLEU results of 0.5464 and 0.3208, respec- tively. Although both systems use as training and test data parts of the JRC-Acquis corpus, the architecture described in (Cristea, 2009) in- volves the use of linguistic resources and the sys- tem implemented in (Ignat, 2009) uses …

Content truncated to 500 words