CLUSTER: An Approach to Contextual Language Understanding

Understanding natural language in context requires the existence of a model of the surrounding world. Such a model must include representations for the objects and activities present in the language being processed, in the surrounding physical environment, and in relevant past experiences of the understander. The CLUSTER theory of contextual language understanding addresses this issue. CLUSTER has two main components: a theory of modeling the context of an ongoing conversation, and a theory of language analysis. The first component of CLUSTER theory concerns the model of the world which an understanding system has: the objects and processes in the world that are to be represented, and the relation of these to the understanding system itself. The model emphasizes those elements that are more significant in the current situation, while ignoring information that is relevant. This thesis characterizes a mechanism for maintaining a model of the world, the Context Modeler, and the model of the world maintained by it, called the Context Model. The Context Modeler constructs the Context Model during the course of interacting with another language user. The second component of CLUSTER theory concerns language analysis, the production of a representation of the meaning of a given sentence. Natural language analysis provides the basis for constructing the Context Model. The Context Model, in turn, is a major resource used by the language analyzer. This thesis shows the centrality of the role played by a model of the context to a system's ability to understand natural language input. In addition to characterizing these mechanisms, this thesis also describes particular implementations of CLUSTER. The implementation of CLUSTER's language analysis component is named PHRAN, for PHRasal ANanalyzer. The implementation of CLUSTER's context modeling component is referred to as The Context Modeler. Sponsored in part by Naval Electronic Systems Command and ARPA order no. 3041.