This paper presents an unsupervised approach to learning translation span alignments from parallel data that improves syntactic rule extraction by deleting spurious word alignment links and adding new valuable links based on bilingual translation span correspondences.