Applying boosting to statistical machine translation

Applying boosting to statistical machine translation ⋆ Antonio L. Lagarda and Francisco Casacuberta Instituto Tecnol´ogico de Inform´atica Universidad Polit´ecnica de Valencia Camino de Vera, s/n, 46022 Valencia, Spain {alagarda,fcn}@iti.upv.es Abstract. Boosting is a general method for improving the accuracy of a given learning algorithm under certain restrictions. In this work, AdaBoost, one of the most popular boosting algorithms, is adapted and applied to statistical machine translation. The appropriateness of this technique in this scenario is evaluated on a real translation task. Results from preliminary experiments confirm that statistical machine transla- tion can take advantage from this technique, improving the translation quality. 1 Introduction State-of-the-art statistical machine translation (SMT) techniques are still far from producing high quality translations. This drawback leads us to introduce an alternative approach to the translation problem. In our work, we will propose an adaptation of boosting [1] to SMT. The purpose of boosting methods is to find a highly accurate rule by com- bining many weak or base hypotheses. The boosting algorithm generates each one of these hypotheses by iteratively calling a weak learning algorithm. Each iteration takes into account the performance of the previous iterations, trying to concentrate on the instances that have not been correctly learned. All these weak hypotheses are then combined into a final hypothesis. AdaBoost (Adaptive Boosting) employs a set of importance weights over the training examples [2, 3]. These weights are used by the learning algorithm to produce a new weak hypothesis with lower error with respect to them. In this way, these weights help the algorithm to concentrate on the examples which are hardest to classify. In machine translation, learning techniques could be considered as weak, due to the low quality of their results. Boosting has previously been applied to machine translation in works like [4] or [5]. In this paper, we propose an adaptation of AdaBoost to a SMT task. Each round, a new translation model ⋆ Work supported by the EC (FEDER) and the Spanish MEC under grant TIN2006-15694-CO2-01, by the Spanish research programme Consolider Ingenio 2010: MIPRCV (CSD2007-00018), by the i3media Cenit project (CDTI 2007-1012), and by VIDI-UPV under project 20070315. 12th EAMT conference, 22-23 September 2008, Hamburg, Germany 88 will be learned, which will produce a new hypothesis. Finally, all these hypotheses will be combined to obtain the final translation. The next sections introduce machine translation and the AdaBoost algo- rithm. After that, we discuss our adaptation proposal of AdaBoost to SMT. Experimental results are presented in section 5. Finally, some conclusions and future work are given in section 6. 2 Machine translation Traditionally, the goal of SMT has been statistically stated as follows [6]. Given a source sentence f J 1 ≡ f1...fj...fJ, we have to find a target sentence eI 1 ≡ e1...ei...eI that maximizes: ˆe ˆI 1 = arg max eI 1 Pr(eI 1|f J 1 ) (1) Using Bayes’ Theorem, and taking into account that Pr(f J 1 ) does not depend on eI 1, we arrive at ˆe ˆI 1 …

Content truncated to 500 words