Natural Categories for More Natural Generation
Cline, Ben E.
Nutter, J. Terry
MetadataShow full item record
Psychological research has shown that natural taxonomies contain a distinguished or basic level. Adult speakers use the names of these categories most frequently and can list a large number of attributes for them. They typically cannot list many attributes for superordinate categories and few list additional attributes for subordinate categories. Because natural taxonomies are important to human language, their use in natural language processing systems appears well founded. In the past, however, most AI systems have been implemented around uniform taxonomies in which there is no distinguished level. It has recently been demonstrated that natural taxonomies enhance language processing systems by allowing selection of appropriate category names and by providing the means to handle implicit focus. In previous research, we have argued that benefits from the use of natural categories can be realized in multi-sentential connected generation systems. We briefly summarize the psychological research on natural taxonomies that relates to natural language processing systems, the use of natural categorizations in current natural language processing systems, and the results of our previous research in which we show how natural categories can be used in multiple sentence generation systems to allow the selection of appropriate category names, to provide a mechanism to help determine salience, and to provide for the shallow modeling of audience expertise. We then describe additional benefits of natural categories in generation systems by demonstrating that natural categories provide a mechanism that aids selection of discourse schemes and increase the efficiency of inheritance.