Corrective Dependency Parsing

We present a discriminative model for correcting errors in automatically generated dependency trees. We show that by focusing on “structurally local” errors, we can improve the overall quality of the dependency structure. Defining the task by way of a locality constraint allows us to search over a large set of alternate dependency trees simply by making small perturbations to individual dependency edges. This technique requires no additional data for training as it uses the original training data and parser to generate a set of parses from which the training examples are generated. We present experimental results on a Czech corpus using four different parsers, both projective and non-projective, showing the robustness of the technique.