A novel alignment model inspired on IBM Model 1

We present an extension to IBM Model 1 for training word-to-word lexicon probabilities. This model takes into account a given fixed segmentation of the source and target sentences in the estimation of the statistical dictionary. Our experimentation on the Europarl corpus shows that a statistical consistent improvement in the translation quality can be achieved by including our proposed model as a new information source in a log-linear combination of models. 1 Statistical Machine Translation The goal of Machine Translation is the translation of a text given in some source language into a target language. We are given a source language sentence f = f1 . . . fj . . . fJ which is to be translated into a target language sentence. Among all possible target language sentences, we will choose the sentence ê = e1 . . . ei . . . eI which maximises the posterior probability. Such statement is formalised in the Fundamental Equation of Machine Translation: ê = argmax e {Pr(e|f)} = argmax e {Pr(e) · Pr(f |e)} . (1) The argmax operation denotes the search problem, i.e. the generation of the output sentence in the target language. The decomposition in Eq. (1) allows an independent modelling of the target language model Pr(e) and the (inverse) translation model Pr(f |e)1, known as source-channel model [1]. This decomposition has a very intuitive interpretation: the translation model Pr(f |e) will capture the word relations between both input and output languages, whereas the language model Pr(e) will ensure that the output sentence is a well-formed sentence belonging to the target language. Many statistical translation models [2–5] try to model word-to-word correspondences between source and target words. Known as statistical alignment models, these models typically yield the following equation: