Analogy in Learning by Reading

Natural language understanding is an important problem in artificial intelligence, and a challenging one. Analogy is a powerful tool that has been applied to a number of important AI tasks. This thesis describes three contributions that integrate language understanding and analogical reasoning. The first contribution is to demonstrate that analogical retrieval of previously stored sentence choice cases can be used for semantic disambiguation. This includes both word sense disambiguation and other forms of semantic disambiguation. This allows a system to avoid introducing error while reading. The second contribution is to show that by exploiting the connectivity properties of the semantic interpretation of a text, a cognitive system can extract subsets of that interpretation that are useful for analogical reasoning tasks. The third contribution is an ontology of analogical dialogue acts that can be used to improve understanding of text passages that contain explicit analogies. This allows a system to make use of textual analogies, rather than having them be a source of noise. All three contributions are supported by experiments that demonstrate their utility in improving understanding. The work described in this dissertation is integrated into a larger cognitive architecture.