In this paper, we present an approach to statistical machine translation that combines the power of a discriminative model (for training a model for Machine Translation), and the standard beam-search based decoding technique (for the translation of an input sentence). A discriminative approach for learning lexical selection and reordering utilizes a large set of feature functions (thereby providing the power to incorporate greater contextual and linguistic information), which leads to an effective training of these models. This model is then used by the standard state-of-art Moses decoder (Koehn et al., 2007) for the translation of an input sentence. We conducted our experiments on Spanish-English language pair. We used maximum entropy model in our experiments. We show that the performance of our approach (using simple lexical features) is comparable to that of the state-of-art statistical MT system (Koehn et al., 2007). When additional syntactic features (POS tags in this paper) are used, there is a boost in the performance which is likely to improve when richer syntactic features are incorporated in the model.
For Full Paper : Click Here