Quality that stands out: comparing Opentrad and Google for the Spanish-Galician pair


News:Quality that stands out: comparing Opentrad and Google for the Spanish-Galician pair

Two of the main trends in current machine translation are those represented by structural transfer-driven machine translation systems (with linguistic information) and those represented by statistical machine translation systems (without linguistic information). When the distance between the languages is short, as in the case of the Romance languages (Spanish, Galician, Portuguese or Catalan, among others), where the focus of attention is concentrated more on linguistic questions than on those of structure, it is obvious, as we will be showing, that structural transfer-driven machine translation systems offer results that far exceed those achieved by the statistical systems.


The table below shows a comparison between the statistical machine translation system of Google and the structural transfer-driven system used by Opentrad, for the Spanish-Galician pair:



machine translation

translation direction




es > gl




es > gl




gl > es




gl > es




As we can see, the Google machine translation system (which uses the statistical model, and therefore without linguistic knowledge) achieves results that are clearly poorer than those of the machine translation system used by Opentrad, according to the NIST and BLEU evaluation methods, which compare texts translated by a human with those translated by a machine translation system. The BLEU measurement gives a score of between 0 and 1, which indicates the similarity between the text translated by the machine translation system and the one used as a reference (translated by a human); the value closest to 1 represents greater similarity. The NIST measurement is an evolution that has arisen out of the BLEU method. It takes into consideration more parameters to measure the accuracy of a translation; the higher the score, the more accurate the translation will be.


*Results obtained by means of measurements carried out with mteval-v11b.pl script on the basis of the translation of a 500-line document and with a translation produced by a human as the reference.

  • imaxin|software
  • Eleka