System Combination Using Joint, Binarised Feature Vectors
Christian Federmann · International Conference on Computational Linguistics · 2012
We describe a method for system combination based on joint, binarised feature vectors. Our method can be used to combine several black-box source systems. We first define a total order on given translation output which can be used to partition an n-best list of translations into a set of pairwise system comparisons. Using this data, we explain how an SVM-based classification model can be trained and how this classifier can be applied to combine translation output on the sentence level. We describe our experiments for the ML4HMT-12 shared task and conclude by giving a summary of our findings and by discussing future extensions and experiments using the proposed approach.