DFKI-LT - Dissertation Series


Stefan Thater : Minimal Recursion Semantics as Dominance Constraints: Graph-Theoretic Foundation and Application to Grammar Engineering

ISBN: 978-3-933218-26-1
125 pages
price: € 15

order form

This thesis defines a translation from Minimal Recursion Semantics into dominance constraints, two relevant and closely related scope underspecification formalisms. Due to fundamental differences in the way the two formalisms interpret underspecified descriptions, the translation is restricted to a large class of underspecified descriptions that share certain structural properties.
On the one hand, the translation clarifies the precise relationship between the two formalisms: they are to a large extent essentially equivalent, but differ for certain types of underspecified descriptions. On the other hand, the translation is practically relevant as it allows to share resources and tools available for the two formalisms: existing broad coverage grammars computing MRS descriptions can be used to derive dominance constraints for a wide variety of natural language sentences, and the highly efficient algorithms based on dominance constraints can be used to process MRS descriptions.
The translation is based on graph-theoretical concepts. We will show that underspecified descriptions of both formalisms can be represented as dominance graphs, and identify the important subclass of (weak) nets. Nets have important structural properties, which in particular allow for a correct translation of MRS descriptions into dominance constraints. An interesting aspect arises from the restriction of the translation to nets.
We will show thatmost underspecified descriptions that the English Resource Grammar, a state-of-the-art broad coverage grammar, computes are actually nets, and provide evidence that non-nets are linguistically meaningless i.e., they do not encode the meaning of the underlying sentence correctly. We demonstrate that the concept of a net can be used in grammar verification to detect incorrect rules in the syntax-semantics interface.