Bilingual segmentation for phrasetable pruning in Statistical Machine Translation

Bilingual segmentation for phrasetable pruning in Statistical Machine Translation Germ´an Sanchis-Trilles Daniel Ortiz-Mart´ınez Jes´us Gonz´alez-Rubio Jorge Gonz´alez Francisco Casacuberta Instituto Tecnol´ogico de Inform´atica Departamento de Sistemas Inform´aticos y Computaci´on Universitat Polit`ecnica de Val`encia Valencia, Spain {gsanchis,dortiz,jegonzalez,jgonzalez,fcn}@dsic.upv.es Abstract Statistical machine translation systems have greatly improved in the last years. However, this boost in performance usu- ally comes at a high computational cost, yielding systems that are often not suitable for integration in hand-held or real-time devices. We describe a novel technique for reducing such cost by performing a Viterbi-style selection of the parameters of the translation model. We present results with finite state transducers and phrase- based models showing a 98% reduction of the number of parameters and a 15-fold in- crease in translation speed without any sig- nificant loss in translation quality. 1 Introduction Nowadays, the key step of the process of statisti- cal machine translation (SMT) involves inferring a large table of phrase pairs that are translations of each other from a large corpus of aligned sen- tences. The set of all phrase pairs, together with es- timates of conditional probabilities and other use- ful features, is called phrasetable. Such phrases are applied during the decoding process, combin- ing their target sides to form the final translation. A variety of algorithms to extract phrase pairs has been proposed (Och and Ney, 2000; Marcu and Wong, 2002; Zens et al., 2002; Och and Ney, 2003; Vogel, 2005). Typically, these algorithms heuristi- cally collect a highly redundant set of phrases from each training sentence pair generating phrasetables with a huge number of elements. This bulk comes at a cost. Large phrasetables lead to large data structures that require more re- c⃝ 2011 European Association for Machine Translation. sources and more time to process. More impor- tantly, effort spent in handling large tables could likely be more usefully employed in more features or more sophisticated search processes. Addition- ally, this is the main restriction for the widespread application of SMT techniques in small portable devices like cell phones, PDAs or hand-held game consoles; one can imagine many scenarios that could benefit from a lightweight translation device: tourism, medicine, military, etc. In this paper, we show that is possible to prune phrasetables by removing those phrase pairs that have little influence on the final translation per- formance. Our approach consist in selecting only those phrase pairs extracted from the most proba- ble segmentation of the training sentences. The technique presented here has several advan- tages. It does not depend on the actual algorithm used to extract the phrase pairs, therefore can be applied to every imaginable method that assigns probabilities to phrase pairs. It provides a straight- forward method for pruning the phrasetables, with- out the need of adjusting any additional parameter. It does not significantly affect translation quality, as measured by BLEU or TER scores, while very substantial savings in terms of computational re- quirements are reported. The rest of the paper is organised as follows. Section 2 revised previously published techniques …

Content truncated to 500 words