Meta-level Statistical Machine Translation

International Joint Conference on Natural Language Processing, pages 1151–1157, Nagoya, Japan, 14-18 October 2013. Meta-level Statistical Machine Translation Sajad Ebrahimi†,*, Kourosh Meshgi†† , Shahram Khadivi† and Mohammad Ebrahim Shiri Ahmad Abady* †Human Language Technology Lab, Amirkabir University of Technology, Tehran, Iran ††Graduate School of Informatics, Kyoto University, Kyoto, Japan *Department of Computer Science, Amirkabir University of Technology, Tehran, Iran ebrahimi.sajad@aut.ac.ir, meshgi-k@sys.i.kyoto-u.ac.jp khadivi@aut.ac.ir, shiri@aut.ac.ir Abstract We propose a simple and effective method to build a meta-level Statistical Machine Transla- tion (SMT), called meta-SMT, for system combination. Our approach is based on the framework of Stacked Generalization, also known as Stacking, which is an ensemble learning algorithm, widely used in machine learning tasks. First, a collection of base-level SMTs is generated for obtaining a meta-level corpus. Then a meta-level SMT is trained on this corpus. In this paper we address the issue of how to adapt stacked generalization to SMT. We evaluate our approach on English- to-Persian machine translation. Experimental results show that our approach leads to signifi- cant improvements in translation quality over a phrase-based baseline by about 1.1 BLEU points. 1 Introduction Currently, there exist a number of commercial and research Machine Translation (MT) systems, which are developed under different paradigms such as rule-based, example based, statistical machine translation, trained using different algo- rithms, e.g., phrase-based SMT, hierarchical phrase-based SMT, syntax-based SMT with dif- ferent types and amounts of training data. With the emergence of these various structurally dif- ferent systems, system combination methods have taken a great importance during the past few years. There are several techniques for combining multiple SMT systems to achieve higher transla- tion quality, e.g. sentence-level combination (Hildebrand and Vogel, 2008) simply selects "the best" of the provided translations and phrase-level combination (Matusov et al., 2006; Rosti et al., 2007) can generate new translations differing from all original translations. Most of the state-of-the-art SMT system com- bination methods require multiple SMT systems based on different models. Since it is not easy to have multiple SMT systems, in this work we fo- cus on applying stacking algorithm on a single SMT system rather than multiple SMT systems. We try to increase the performance of an SMT system by introducing a meta-level SMT which can learn how to decrease or modify translation errors on the translation outputs of original SMT system. This task is also known as automatic post-editing (APE) which is a well-studied topic in machine translation community (Simard et al., 2007a; Béchara et al., 2011). To do this, we use stacked generalization which is an ensemble learning algorithm Ensemble Learning is a machine learning par- adigm where multiple learners are trained to solve the same problem. An ensemble is viewed as a collection of learners which are usually called base learners. Base learners are usually generated from training data by a base learning algorithm which can be decision tree, neural network or other kinds of machine learning algo- rithms. Most ensemble methods use a single base learning algorithm to produce homogeneous base learners, but there …

Content truncated to 500 words