The UPV-PRHLT Combination System for WMT 2010

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 296–300, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics The UPV-PRHLT Combination System for WMT 2010 Jes´us Gonz´alez-Rubio and Jes´us Andr´es-Ferrer and Germ´an Sanchis-Trilles Guillem Gasc´o and Pascual Mart´ınez-G´omez and Martha-Alicia Rocha Joan-Andreu S´anchez and Francisco Casacuberta Instituto Tecnol´ogico de Inform´atica Departamento de Sistemas Inform´aticos y Computaci´on Universidad Polit´ecnica de Valencia {jegonzalez|jandres|gsanchis}@dsic.upv.es {ggasco|pmartinez|mrocha}@dsic.upv.es {jandreu|fcn}@dsic.upv.es Abstract UPV-PRHLT participated in the System Combination task of the Fifth Workshop on Statistical Machine Translation (WMT 2010). On each translation direction, all the submitted systems were combined into a consensus translation. These consen- sus translations always improve transla- tion quality of the best individual system. 1 Introduction The UPV-PRHLT approach to MT system combi- nation is based on a refined version of the algo- rithm described in (Gonz´alez-Rubio and Casacu- berta, 2010), with additional information to cope with hypotheses of different quality. In contrast to most of the previous approaches to combine the outputs of multiple MT sys- tems (Bangalore et al., 2001; Jayaraman and Lavie, 2005; Matusov et al., 2006; Schroeder et al., 2009), which are variations over the ROVER voting scheme (Fiscus, 1997), we consider the problem of computing a consensus translation as the problem of modelling a set of string patterns with an adequate prototype. Under this frame- work, the translation hypotheses of each of the MT systems are considered as individual patterns in a set of string patterns. The (generalised) me- dian string, which is the optimal prototype of a set of strings (Fu, 1982), is the chosen prototype to model the set of strings. 2 System Combination Algorithm The median string of a set is defined as the string that minimises the sum of distances to the strings in the set. Therefore, defining a distance between strings is the primary problem to deal with. The most common definition of distance be- tween two strings is the Levenshtein distance, also known as edit distance (ED). This metric computes the optimal sequence of edit operations (insertions, deletions and substitutions of words) needed to transform one string into the other. The main problem with the ED is its dependence on the length of the compared strings. This fact led to the definition of a new distance whose value is inde- pendent from the length of the strings compared. This normalised edit distance (NED) (Vidal et al., 1995) is computed by averaging the number of edit operations by the length of the edit path. The ex- perimentation in this work was carried out using the NED. 2.1 Median String Given a set E = e1, . . . , en, . . . , eN of translation hypotheses from N MT systems, let Σ be the vo- cabulary in the target language and Σ∗ be the free monoid over that vocabulary (E ⊆ Σ∗). The me- dian string of the set E (noted as M(E)) can be formally defined as: M(E) = argmin e′∈Σ∗ N � n=1 �wn · …

Content truncated to 500 words