This paper proposes how to convert projective dependency structures into flat phrase structures with language-independent syntactic categories, and use a head finder algorithm to convert these phrase structures back into dependency structures. The head finder algorithm is implemented by a maximum entropy approach with constraint information. The converted phrase structures can be parsed using a hierarchical coarse-to-fine method with latent variables. Experimental results show that the approach finds 98.8% heads of all phrases, and our algorithm achieves state-of-the-art dependency parsing performance in English Treebank.