Semi-Supervised Block ITG Models for Word Alignment

Labeled training data for the word alignment task, in the form of word-aligned sentence pairs, is hard to come by for many language-pairs. Hence, it is natural to draw upon semi-supervised learning methods (Fraser and Marcu, 2006). We introduce a semisupervised learning method for word alignment using conditional entropy regularization (Grandvalet and Bengio, 2005) on top of a BITG-based discriminative model. Our preliminary experiments show improvement in the alignment quality compared to a strong supervised model (Haghighi et al., 2009). Let L = {〈xi,yi〉}1 be a set of labeled examples where xi is an input and yi is its output label, and U = {xj} U 1 be a set of unlabeled examples. The goal of semi-supervised learning is to take into account both labeled and unlabeled data in finding a good mapping from input to output. In the word alignment problem, the label y is the word alignment for the sentence pair in x.