Integrating Natural Language Subtasks with Bayesian Belief Networks

The development of automatic natural language understanding systems remains an elusive goal. Given the highly ambiguous nature of the syntax and semantics of natural language, it is often impossible to develop rule--based approaches to understanding even very limited domains of text. The difficulty in specifying rules and their exceptions has led to the rise of probabilistic approaches where models of natural language are learned from large corpora of text. These models usually serve as simple classifiers for particular subtasks such as word sense disambiguation or discourse segmentation. While successful in these limited roles, it is unclear that multiple classifiers can be combined to create comprehensive natural language understanding systems. Instead, we believe that recent advances in modeling and reasoning with uncertain information offer an appropriate framework for building such systems. We are developing and evaluating new algorithms that learn Bayesian belief networks from la..