Mildly Non-Projective Dependency Grammar

Syntactic representations based on word-to-word dependencies have a long tradition in descriptive linguistics, and receive considerable interest in many computational applications. However, dependency syntax has remained somewhat of an island from a formal point of view, which hampers the exchange of resources and computational methods with other syntactic traditions. In this article, we present new results on the formal relationship between dependency syntax and phrase-structure syntax. To do so, we introduce a formalism in which abstract properties of a dependency grammar such as generative capacity and parsing complexity can be controlled by constraints on the non-projectivity of the analyses permitted by the grammar. This gives rise to an infinite hierarchy of mildly context-sensitive dependency grammars. The class of grammars at the lowest point of this hierarchy is restricted to projective analyses and equivalent to contextfree phrase-structure grammar. The class of grammars at the second-lowest point is equivalent to tree-adjoining grammar. In contrast to projective structures, the dependency analyses permitted by this class have a close to perfect coverage on practically relevant treebank data.