Constraint-based RMRS Construction from Shallow Grammars

Semantic formalisms such as MRS [Copestake et al., 2003], UDRT [Reyle, 1993], or CLLS [Egg et al., 2001] provide elegant solutions for the treatment of semantic ambiguities in terms of underspecification. These semantic formalisms are not constrained to specific syntactic frameworks. MRS, for example, is mainly used with HPSG grammars, but may also be built from LFG grammars using co-description as in the Bergen NorGram project. Also, HPSG grammars have been used for construction of alternative semantic representations, e.g. using the “glue” approach in [Asudeh and Crouch, 2001] for construction of intensional logic representations, or UDRT in [Crouch et al., 2001]. Underspecification in these semantic formalisms is naturally concerned with the underspecification of semantic ambiguities – most prominently scope. In recent work, [Copestake, 2003] has investigated a novel aspect of underspecification in the design of semantic formalisms, which is more concerned with the representation of partial semantic information, as it might be obtained from shallow, i.e. incomplete syntactic analysis. The main rationale for this type of underspecification is to ensure monotonicity, and thus upwards compatibility of the output of shallow parsing with semantic representations obtained from full syntactic parsing. Thus, Copestake’s design of RMRS – Robust Minimal Recursion Semantics – provides an important contribution to a novel line of research towards the integration of shallow and deep NLP. While previous accounts in [Crysmann et al., 2002, Daum et al., 2003, Frank et al., 2003a] focus on shallow-deep integration at the syntactic or sub-syntactic level, Copestake aims at integration of shallow and deep NLP components at the level of semantic representations. In our paper we will present a constraint-based syntax-semantics interface for construction of RMRS representations from shallow grammars. We design a modular interface that can be flexibly adapted to existing shallow grammars of various types – ranging from chunk grammars to context-free grammars. We investigate the possibilities and challenges for principle-based RMRS construction from shallow syntactic structures, and present an implementation in a typed feature structure formalism. We briefly review the RMRS formalism designed by [Copestake, 2003], and present an architecture for a principle-based syntax-semantics interface for RMRS construction from shallow grammars. We argue for a unification-based approach for RMRS construction, to account for (underspecified) argument binding in languages with case-marking as opposed to structural argument identification properties. The architecture we propose is especially designed to support flexible adaptation to different types of shallow to intermediate-level syntactic grammars that may serve as a basis for RMRS construction. The major challenge for principle-based semantics construction from shallow grammars is the flat and sometimes noncompositional nature of the structures they typically produce. We present RMRS semantics construction principles that can be applied to flat syntactic structures with various degrees of partiality. Finally, we compare our work to alternative accounts for construction of deep representations from shallow grammars.