Improving Coverage and Runtime Complexity for Exact Inference in Non-Projective Transition-Based Dependency Parsers

We generalize Cohen, Gómez-Rodríguez, and Satta's (2011) parser to a family of non-projective transition-based dependency parsers allowing polynomial-time exact inference. This includes novel parsers with better coverage than Cohen et al. (2011), and even a variant that reduces time complexity to O(n^6), improving over the known bounds in exact inference for non-projective transition-based parsing. We hope that this piece of theoretical work inspires design of novel transition systems with better coverage and better run-time guarantees. Code available at https://github.com/tzshi/nonproj-dp-variants-naacl2018