In this thesis, I describe a system I built that produces instantiated representations from descriptions embedded in natural language. For example, in the sentence 'The girl walked to the table', my system produces a description of movement along a path (the girl moves on a path to the table), instantiating a general purpose trajectory representation that models movement along a path. I demonstrate that descriptions found by my system enable the imagining of an entire inner world, transforming sentences into three-dimensional graphical descriptions of action. By building action descriptions from ordinary language, I illustrate the gains we can make by exploiting the connection between language and thought. I assert that a small set of simple representations should be able to provide powerful coverage of human expression through natural language. In particular, I examine the sorts of representations that are common in the Wall Street Journal from the Penn Treebank, providing a counterpoint for the many other sorts of analyses of the Penn Treebank in other work. Then, I tum to recognized experts in provoking our imaginations with words, using my system to examine the work of four great authors to uncover commonalities and differences in their styles from the perspective of the way they make representational choices in their work. Thesis Supervisor: Patrick Henry Winston Title: Ford Professor of Artificial Intelligence and Computer Science

There has been substantial recent interest in integrating knowledge based reasoning (KBR) and case-based reasoning (CBR) within a single system due to the potential synergisms that could result. Here we describe our recent work investigating the feasibility of a combined KBR-CBR application-independent system for interpreting multi-episode stories/narratives, illustrating it with an application in the domain of interpreting urban warfare stories. A genetic algorithm is used to derive weights for selection of the most relevant past cases. In this setting, we examine the relative value of using input features of a problem for case selection versus using features inferred via KBR, versus both. We find that using both types of features is best (compared to human selection), but that input features are most helpful and inferred features are of marginal value. This finding supports the idea that KBR and CBR provide complimentary rather than redundant information, and hence that their combination in a single system is likely to be useful.

There has been substantial recent interest in integrating knowledge based reasoning (KBR) and case-based reasoning (CBR) within a single system due to the potential synergisms that could result. Here we describe our recent work investigating the feasibility of a combined KBR-CBR application-independent system for interpreting multi-episode stories/narratives, illustrating it with an application in the domain of interpreting urban warfare stories. A genetic algorithm is used to derive weights for selection of the most relevant past cases. In this setting, we examine the relative value of using input features of a problem for case selection versus using features inferred via KBR, versus both. We find that using both types of features is best (compared to human selection), but that input features are most helpful and inferred features are of marginal value. This finding supports the idea that KBR and CBR provide complimentary rather than redundant information, and hence that their combination in a single system is likely to be useful.