Talk by Carlos Gómez Rodríguez

Dynamic oracles and non-monotonicity for unrestricted non-projective dependency parsing

by Carlos Gómez Rodríguez (http://www.grupolys.org/~cgomezr/)

(Joint work with Daniel Fernández-González)

A dynamic oracle is a training enhancement for greedy transition-based parsers that alleviates error propagation by training on non-optimal sequences of transitions, to increase the robustness of the model. In this talk, I will present the first dynamic oracle for a dependency parser supporting unrestricted non-projectivity, which works in linear time for the Covington parser and achieves significant improvements in accuracy over the standard oracle. Then, I will show further accuracy improvements by enriching the parser with non-monotonicity, in such a way that it can fix previous mistakes by deleting already-built dependency arcs. The non-monotonic version of the parser can be trained by several variants of the dynamic oracle, based on tight approximations of the loss.