Feature-Based ITG for Unsupervised Word Alignment

Inversion transduction grammar (ITG) (1) is an effective constraint to word alignment search space. However, the traditional unsupervised ITG word alignment model is incapable of utilizing rich features. In this paper, we propose a novel feature-based unsupervised ITG word alignment model. With the help of rich features and regularization, a compact grammar is learned. Experiments on both word alignment and end-to-end statistical machine translation (SMT) task show that our model achieves better performance than the traditional ITG model with the AER of word alignment improved by 3 points and the BLEU score of SMT improved by 0.8 points.