Adapting Stochastic LFG Input for Semantics

LFG c(onstituent)-structure and f(unctional)-structure analyses provide the detailed syntactic structures necessary for subsequent semantic analysis. The f-structure encodes grammatical functions as well as semantically relevant features like tense and number. The c-structure, in conjunction with the phi-mapping, provides the information on linear precedence necessary for semantic scope and anaphora resolution. In this paper, we present a system in which a stochastic LFG-like grammar of English provides the input to the semantics processing. The LFG-like grammar uses stochastic methods to create a c-structure and a proto f-structure. A set of ordered rewrite rules augments and reconfigures the proto f-structure to add more information to the stochastic output, thereby creating true LFG f-structures with all of the features that the semantics requires. Evaluation of the resulting derived f-structures and of the semantic representations based on them indicates that the stochastic LFGlike grammar can be used to produce input to the semantics. These grammars provide the advantages of LFG structures, e.g. the explicit encoding of grammatical functions, in conjunction with the advantages of stochastic systems, e.g. providing connected parses in the face of less-than-ideal input.