A Post-processing Approach to Statistical Word Alignment Reflecting Alignment Tendency between Part-of-speeches

Statistical word alignment often suffers from data sparseness. Part-of-speeches are often incorporated in NLP tasks to reduce data sparseness. In this paper, we attempt to mitigate such problem by reflecting alignment tendency between part-of-speeches to statistical word alignment. Because our approach does not rely on any language-dependent knowledge, it is very simple and purely statistic to be applied to any language pairs. End-to-end evaluation shows that the proposed method can improve not only the quality of statistical word alignment but the performance of statistical machine translation.