Explorations in story understanding using contextually accessed semantic networks
Robert L. Young · 1982
This thesis explores the problems of narrative understanding, and of representing the knowledge required for this task. Several considerations lead to an extended semantic network formalism in which access is strictly controlled. This representational formalism is built on three principles, each shown to have important implications for the task of narrative understanding. (1) Controlled visibility, achieved by manipulating only the visibility of arcs, is a very useful representational device. (2) One may very profitably apply controlled visibility to the structured inheritance problem. (3) The notion of a method for representing and manipulating implicit information, is an extremely valuable representational addition. Within this representational framework, the goal is to describe how successive sentences of a narrative are understood. is the term used to describe this overall process of merging a new sentence with the preceding portion of the narrative. Textual coalescence involves two major activities. (1) Input sentences are enriched, using the knowledge base, with context-independent information. (2) Rules of textual ordering are used to generate possible coalescence centers from an existing text representation (and the knowledge base). Coalescence of new input with previous text occurs when a node in the enriched input is identified with one of the coalescence centers. The relating old and new will typically include implicit links, and is represented using virtual information. This path describes the (possibly quite complex) relationship which should be made fully explicit in coalescing the new input with the preceding text. Other processes, such as referent identification, are subordinated to the more important process just described. The operations needed for coalescence are easily described and reasonable to perform because the representational system provides the appropriate tools. No standard deduction system is used within the model.