Unifying Representation and Generalization: Understanding Hierarchically Structured Objects

Wasserman, Kenneth

Hierarchies are pervasive. They are used to organize and describe many artificial and natural phenomena. In general, humans are very good at understanding them. It therefore seems reasonable to give computers the same ability if they are to be "intelligent". The integration of representation and generalization is necessary in order to understand hierarchically structured objects. In this thesis we address the issues involved and present a scheme, MERGE, designed to be used in computer systems that understand and automatically classify instances of hierarchies in a given domain. The MERGE scheme uses a form of dynamic generalization-based memory in order to achieve this integration. Representations of individual hierarchies are stored in terms of how they vary from previously created generalized concepts. Memory is continually reorganized as new data becomes available to a MERGE-based system so that it accurately reflects the known information. The overall effect of this scheme is that representations of individual hierarchies are enhanced by the use of information in the knowledge base. These representations are in turn used to enhance the knowledge base by permitting more and better generalizations to be made. We have developed two MERGE-based computer systems that intelligently understand hierarchies. CORPORATE-RESEARCHER is a program that learns about upper-level corporate management hierarchies when it is fed representations of corporate charts. RESEARCHER is a larger, natural language processing program that reads and understands patent abstracts about physical objects. Both programs serve as intelligent information systems that automatically classify representations of instance hierarchies.



More About This Work

Academic Units
Computer Science
Department of Computer Science, Columbia University
Columbia University Computer Science Technical Reports, CUCS-177-85
Published Here
November 1, 2011