A Proposal for Incorporating Analogically Learned Constructions in a Feature Based Parsing Framework

Structure mapping theory is a computational model of analogy that has recently been used to learn FrameNet constructions from a small corpus of annotated text. This paper proposes an approach that uses constructions learned in this way to bootstrap the performance of an existing natural language understanding system with a more traditional feature-based chart parser, EA NLU. We examine the benefits of analogically learned constructions as well as the challenges involved in applying these generalizations to novel text.