Improving Statistical Machine Translation Performance by Training Data Selection and Optimization

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 343–350, Prague, June 2007. c⃝2007 Association for Computational Linguistics Improving Statistical Machine Translation Performance by Training Data Selection and Optimization Yajuan Lü, Jin Huang and Qun Liu Key Laboratory of Intelligent Information Processing Institute of Computing Technology Chinese Academy of Sciences P.O. Box 2704, Beijing 100080, China {lvyajuan, huangjin,liuqun}@ict.ac.cn Abstract Parallel corpus is an indispensable resource for translation model training in statistical machine translation (SMT). Instead of col- lecting more and more parallel training corpora, this paper aims to improve SMT performance by exploiting full potential of the existing parallel corpora. Two kinds of methods are proposed: offline data optimi- zation and online model optimization. The offline method adapts the training data by redistributing the weight of each training sentence pairs. The online method adapts the translation model by redistributing the weight of each predefined submodels. In- formation retrieval model is used for the weighting scheme in both methods. Ex- perimental results show that without using any additional resource, both methods can improve SMT performance significantly. 1 Introduction Statistical machine translation relies heavily on the available training data. Typically, the more data is used to estimate the parameters of the translation model, the better it can approximate the true trans- lation probabilities, which will obviously lead to a higher translation performance. However, large corpora are not easily available. The collected cor- pora are usually from very different areas. For example, the parallel corpora provided by LDC come from quite different domains, such as Hongkong laws, Hangkong Hansards and Hongkong news. This results in the problem that a translation system trained on data from a particular domain(e.g. Hongkong Hansards) will perform poorly when translating text from a different domain(e.g. news articles). Our experiments also show that simply putting all these domain specific corpora together will not always improve translation quality. From another aspect, larger amount of training data also requires larger computational resources. With the increasing of training data, the improvement of translation quality will become smaller and smaller. Therefore, while keeping collecting more and more parallel corpora, it is also important to seek effective ways of making better use of available parallel training data. There are two cases when we train a SMT system. In one case, we know the target test set or target test domain, for example, when building a specific domain SMT system or when participating the NIST MT evaluation1. In the other case, we are unaware of any information of the testing data. This paper presents two methods to exploit full potential of the available parallel corpora in the two cases. For the first case, we try to optimize the training data offline to make it match the test data better in domain, topic and style, thus improving the translation performance. For the second case, we first divide the training data into several do- mains and train submodels for each domain. Then, in the translation process, we …

Content truncated to 500 words