Choosing the best machine translation system to translate a sentence by using only source-language information

Choosing the best machine translation system to translate a sentence by using only source-language information Felipe S´anchez-Mart´ınez Dep. de Llenguatges i Sistemes Inform`atics Universitat d’Alacant E-03071, Alacant, Spain fsanchez@dlsi.ua.es Abstract This paper describes a novel approach aimed to identify a priori which subset of machine translation (MT) systems among a known set will produce the most reliable translations for a given source-language (SL) sentence. We aim to select this sub- set of MT systems by using only informa- tion extracted from the SL sentence to be translated, and without access to the inner workings of the MT systems being used. A system able to select in advance, with- out translating, that subset of MT systems will allow multi-engine MT systems to save computing resources and focus on the com- bination of the output of the best MT sys- tems. The selection of the best MT systems is done by extracting a set of features from each SL sentence and then using maximum entropy classifiers trained over a set of par- allel sentences. Preliminary experiments on two European language pairs show a small, non-statistical significant improvement. 1 Introduction Machine translation (MT) has become a viable technology that helps individuals in assimilation —to get the gist of a text written in a language the reader does not understand— and dissemination —to produce a draft translation to be post-edited for publication— tasks. However, none of the differ- ent approaches to MT, whether statistical (Koehn, 2010), example-based (Carl and Way, 2003), rule- based (Hutchins and Somers, 1992) or hybrid (Thur- mair, 2009), always provide the best results. This c⃝ 2011 European Association for Machine Translation. is why some researchers have investigated the de- velopment of multi-engine MT (MEMT) systems (Eisele, 2005; Macherey and Och, 2007; Du et al., 2009; Du et al., 2010) aimed to provide transla- tions of higher quality than those produced by the isolated MT systems in which they are based on. MEMT systems can be classified according to how they work. On one hand, we find systems that combine the translations provided by several MT systems into a consensus translation (Banga- lore et al., 2001; Bangalore et al., 2002; Matusov et al., 2006; Heafield et al., 2009; Du et al., 2009; Du et al., 2010); the output of these MEMT sys- tems may differ from those provided by the indi- vidual MT systems they are based on. On the other hand, we have systems that decide which transla- tion, among all the translations computed by the MT systems they are based on, is the most appro- priate one (Nomoto, 2004; Zwarts and Dras, 2008) and output this translation without changing it in any way. In-between, we find the MEMT systems that build a consensus translation from a reduced set of translations, i.e. systems that first chose the subset with the most promising translations, and then combine these translations to produce a single output (Macherey and Och, 2007). Even though MEMT systems that select the most promising translation and those that …

Content truncated to 500 words