Feature Decay Algorithms for Fast Deployment of Accurate Statistical Machine Translation Systems

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 78–84, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Feature Decay Algorithms for Fast Deployment of Accurate Statistical Machine Translation Systems Ergun Bic¸ici Centre for Next Generation Localisation, Dublin City University, Dublin, Ireland. ergun.bicici@computing.dcu.ie Abstract We use feature decay algorithms (FDA) for fast deployment of accurate statistical machine translation systems taking only about half a day for each translation direc- tion. We develop parallel FDA for solving computational scalability problems caused by the abundance of training data for SMT models and LM models and still achieve SMT performance that is on par with us- ing all of the training data or better. Par- allel FDA runs separate FDA models on randomized subsets of the training data and combines the instance selections later. Parallel FDA can also be used for selecting the LM corpus based on the training set selected by parallel FDA. The high qual- ity of the selected training data allows us to obtain very accurate translation outputs close to the top performing SMT systems. The relevancy of the selected LM corpus can reach up to 86% reduction in the num- ber of OOV tokens and up to 74% reduc- tion in the perplexity. We perform SMT experiments in all language pairs in the WMT13 translation task and obtain SMT performance close to the top systems us- ing significantly less resources for training and development. 1 Introduction Statistical machine translation (SMT) is a data in- tensive problem. If you have the translations for the source sentences you are translating in your training set or even portions of it, then the trans- lation task becomes easier. If some tokens are not found in your training data then you cannot trans- late them and if some translated word do not ap- pear in your language model (LM) corpus, then it becomes harder for the SMT engine to find their correct position in the translation. Current SMT systems also face problems caused by the proliferation of various parallel cor- pora available for building SMT systems. The training data for many of the language pairs in the translation task, part of the Workshop on Ma- chine translation (WMT13) (Callison-Burch et al., 2013), have increased the size of the available par- allel corpora for instance by web crawled corpora over the years. The increased size of the training material creates computational scalability prob- lems when training SMT models and can increase the amount of noisy parallel sentences found. As the training set sizes increase, proper training set selection becomes more important. At the same time, when we are going to trans- late just a couple of thousand sentences, possibly belonging to the same target domain, it does not make sense to invest resources for training SMT models over tens of millions of sentences or even more. SMT models like Moses already have filter- ing mechanisms to create smaller parts of the built models that are relevant to the test set. In this …

Content truncated to 500 words