Two-tier model of semantic memory for text comprehension

How can the background knowledge associated with our everyday concepts be represented in a machine and how can we obtain and encode such information? This thesis presents a new architecture for semantic memory that provides a framework for addressing the "background-knowledge" problem and discusses the implications of this architecture for a model of text comprehension. Semantic memory consists of two tiers: a relational tier that represents the underlying structure of our cognitive world expressed as a set of dependency relationships between concepts, and an analog semantic feature (ASF) tier that represents the common or shared knowledge about the concepts in the relational tier, expressed as a set of statistical associations. I present an information theoretic approach to automatically acquiring and encoding this knowledge from on-line text corpora. In this approach the background knowledge common to a community is encoded using a finite vocabulary of ASFs. The ASFs used are based on the category structure of a thesaurus. The two levels of semantic memory support two complementary views of comprehension. One view, the "fine-grain" view, captures the many details of interaction between context and world knowledge as time-trajectories through concept space. This view permits a deeper understanding of a text. A second view, the "coarse-grain" view, captures in the form of a weighted semantic graph called an interpretation graph, a set of explicit semantic relationships that can be used to reason about the understanding of a text, which includes the ability to summarize the text and extract what is important. This view corresponds to a shallow understanding of the text. Several computational techniques are presented for comparing at two levels--the relational and the ASF--the underlying similarity of two passages. The techniques developed are embodied in two computer programs--LeMICON, a structured connectionist implementation, and SSS, a symbolic implementation--designed to explore the system's comprehension of 16 short texts from the stock market domain. The thesis describes an architecture and a mode of processing in which memory is dynamic, exhibits hysteresis effects, and emphasizes what is new about the effect of a given input on the knowledge represented there.