Discontinuity and Non-Projectivity: Using Mildly Context-Sensitive Formalisms for Data-Driven Parsing

We present a parser for probabilistic Linear Context-Free Rewriting Systems and use it for constituency and dependency treebank parsing. The choice of LCFRS, a formalism with an extended domain of locality, enables us to model discontinuous constituents and non-projective dependencies in a straightforward way. The parsing results show that, firstly, our parser is efficient enough to be used for data-driven parsing and, secondly, its result quality for constituency parsing is comparable to the output quality of other state-of-the-art results, all while yielding structures that display dis-continuous dependencies.