The UPV-PRHLT combination system for WMT 2011

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 140–144, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics The UPV-PRHLT combination system for WMT 2011 Jes´us Gonz´alez-Rubio and Francisco Casacuberta Instituto Tecnol´ogico de Inform´atica Departamento de Sistemas Inform´aticos y Computaci´on Universitat Polit`ecnica de Val`encia {jegonzalez|fcn}@dsic.upv.es Abstract This paper presents the submissions of the pat- tern recognition and human language technol- ogy (PRHLT) group to the system combina- tion task of the sixth workshop on statistical machine translation (WMT 2011). Each sub- missions is generated by a multi-system mini- mum Bayes risk (MBR) technique. Our tech- nique uses the MBR decision rule and a linear combination of the component systems’ prob- ability distributions to search for the minimum risk translation among all the sentences in the target language. 1 Introduction The UPV-PHRLT approach to machine translation (MT) system combination is based on the mini- mum Bayes risk system combination (MBRSC) al- gorithm (Gonzlez-Rubio et al., 2011). A multi- system MBR technique that computes consensus translations over multiple component systems. MBRSC operates directly on the outputs of the component models. We perform an MBR decod- ing using a linear combination of the component models’ probability distributions. Instead of re- ranking the translations provided by the component systems, we search for the hypothesis with the min- imum expected translation error among all the pos- sible finite-length strings in the target language. By using a loss function based on BLEU (Papineni et al., 2002), we avoid the hypothesis alignment prob- lem that is central to standard system combination approaches (Rosti et al., 2007). MBRSC assumes only that each translation model can produce expec- tations of n-gram counts; the latent derivation struc- tures of the component systems can differ arbitrary. This flexibility allows us to combine a great variety of MT systems. 2 Minimum Bayes risk Decoding SMT can be described as a mapping of a word se- quence f in a source language to a word sequence e in a target language; this mapping is produced by the MT decoder D(f). If the reference translation e is known, the decoder performance can be mea- sured by the loss function L(e, D(f)). Given such a loss function L(e, e′) between an automatic transla- tion e′ and a reference e, and an underlying proba- bility model P(e|f), MBR decoding has the follow- ing form (Goel and Byrne, 2000; Kumar and Byrne, 2004): ˆe = arg min e′∈E R(e′) (1) = arg min e′∈E � e∈E P(e|f) · L(e, e′) , (2) where R(e′) denotes the Bayes risk of candidate translation e′ under loss function L, and E repre- sents the space of translations. If the loss function between any two hypotheses can be bounded: L(e, e′) ≤ Lmax, the MBR de- coder can be rewritten in term of a similarity func- tion S(e, e′) = Lmax − L(e, e′). In this case, in- stead of minimizing the Bayes risk, we maximize the Bayes gain G(e′): ˆe = arg max e′∈E G(e′) (3) …

Content truncated to 500 words