Improving Word Alignment Based on Extended Inversion Transduction Grammar

We propose a fusion of Inversion Transduction Grammar model with IBM-style notation of fertility to improve wordaligning performance. In our approach, binary context-free grammar rules on the source language, accompanied with orientation preferences on the target, and fertilities of words are leveraged to construct a syntax-based statistical translation model. Our model, inherently possessing the characteristic of ITG restrictions and allowing for many consecutive words aligned to one and vise versa, outperforms original ITG model and GIZA++ not only in alignment error rate (23% and 14% error reduction) but in consistent phrase error rate (13% and 9% error reduction) as well. Better performance in these two evaluation metrics will lead to better phrase-based machine translation with great possibility.