Kathrin Spreyer Department for Computational Linguistics Heidelberg University http://www.cl.uni-heidelberg.de/~spreyer/ In this talk we present a very simple method for training data-driven dependency parsers on partial (i.e., fragmented) trees. We will demonstrate how the method can be formulated within the two major paradigms for dependency parsing: transition-based parsing and graph-based parsing. We investigate the usefulness of the approach in a setting where the training data were created automatically by means of annotation projection across a parallel corpus. Projection of dependency trees typically yields incomplete target language structures, and only a small fraction of the projected dependencies form complete trees. The performance of our "fragment parsers" is compared to that of conventional "tree-based" parsers which are trained only on this restricted subset of complete projections. Our results show that the fragment parsing approach yields parsers that (i) perform at least as well as their tree-based counterparts, and (ii) tend to outperform the tree-based baseline on long-distance dependencies.