Dynamic computation of spatial reference frames in narrative understanding
Albert Hanyong Yuhan · 1991
For a given natural language directional reference, an orientational system with a determined origin and directionalities in which the reference should be interpreted is called the spatial reference frame. Correct recognition of spatial reference frames is critical in understanding spatial information expressed in natural language narratives. This dissertation presents a model solution to the spatial reference frame problem. The solution is based both on extensive grammatical analysis of input sentences' story context by keeping track of the story's contextual goals and the deictic centers. This research demonstrates the robustness of the proposed model by empirically testing the performace of an AI system that understands a short narrative story. This system (called CASSIE), based on SNePS for its inference capability and on ATN for its natural language sentence parsing and generation, goes through with each input sentence four processing phases: the initial interpretation, the extended interpretation, the immediate inference, and the extended inference. Equipped with powerful narrative understanding capability, CASSIE resolves spatial reference frame problems utilizing seventeen proposed resolution rules derived from three general strategies for narrative processing. The specific objective of this dissertation, namely, the resolution of the spatial reference frame problem is embedded in the global goal of understanding natural language narratives with their full contextual coherence maintained.