Symbolic preference using simple scoring

Abstract

Despite the popularity of stochastic parsers, symbolic parsing still has some advantages, but is not practical without an effective mechanism for selecting among alternative analyses. This paper describes the symbolic preference system of a hybrid parser that combines a shallow parser with an overlay parser that builds on the chunks. The hybrid currently equals or exceeds most stochastic parsers in speed and is approaching them in accuracy. The preference system is novel in using a simple, three-valued scoring method (-1, 0, or +1) for assigning preferences to constituents viewed in the context of their containing constituents. The approach addresses problems associated with earlier preference systems, and has considerably facilitated development. It is ultimately based on viewing preference scoring as an engineering mechanism, and only indirectly related to cognitive principles or corpus-based frequencies.