Instance Selection for Machine Translation using Feature Decay Algorithms

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 272–283, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics Instance Selection for Machine Translation using Feature Decay Algorithms Ergun Biçici Koç University 34450 Sariyer, Istanbul, Turkey ebicici@ku.edu.tr Deniz Yuret Koç University 34450 Sariyer, Istanbul, Turkey dyuret@ku.edu.tr Abstract We present an empirical study of instance se- lection techniques for machine translation. In an active learning setting, instance selection minimizes the human effort by identifying the most informative sentences for transla- tion. In a transductive learning setting, se- lection of training instances relevant to the test set improves the final translation qual- ity. After reviewing the state of the art in the field, we generalize the main ideas in a class of instance selection algorithms that use feature decay. Feature decay algorithms in- crease diversity of the training set by devalu- ing features that are already included. We show that the feature decay rate has a very strong effect on the final translation quality whereas the initial feature values, inclusion of higher order features, or sentence length normalizations do not. We evaluate the best instance selection methods using a standard Moses baseline using the whole 1.6 million sentence English-German section of the Eu- roparl corpus. We show that selecting the best 3000 training sentences for a specific test sentence is sufficient to obtain a score within 1 BLEU of the baseline, using 5% of the training data is sufficient to exceed the baseline, and a ∼ 2 BLEU improvement over the baseline is possible by optimally selected subset of the training data. In out-of-domain translation, we are able to reduce the train- ing set size to about 7% and achieve a similar performance with the baseline. 1 Introduction Statistical machine translation (SMT) makes use of a large number of parallel sentences, sentences whose translations are known in the target lan- guage, to derive translation tables, estimate param- eters, and generate the actual translation. Not all of the parallel corpus nor the translation table that is generated is used during decoding a given set of test sentences and filtering is usually performed for computational advantage (Koehn et al., 2007). Some recent regression-based statistical machine translation systems rely on a small sized training data to learn the mappings between source and tar- get features (Wang and Shawe-Taylor, 2008; Ser- rano et al., 2009; Bicici and Yuret, 2010). Regres- sion has some computational disadvantages when scaling to large number of training instances. Previous work shows that the more the training data, the better the translations become (Koehn, 2006). However, with the increased size of the parallel corpus there is also the added noise, mak- ing relevant instance selection important. Phrase- based SMT systems rely heavily on accurately learning word alignments from the given parallel corpus. Proper instance selection plays an impor- tant role in obtaining a small sized training set with which correct alignments can be learned. Word- level translation accuracy is also affected by the number of times a word occurs …

Content truncated to 500 words