Donald Bren School of Information and Computer Sciences

Solving the mysteries of human language understanding inevitably requires an answer to the question of how the language understander resolves ambiguity, for human language is certainly ambiguous. But ambiguity leads to choices between possible explanations, and choice opens the door for mistakes. Unless we are willing to believe that the human language understander always makes the correct choice, any explanation of ambiguity resolution must be considered incomplete if it does not also account for recovery from an incorrect decision.

This dissertation describes a new approach to lexical ambiguity resolution during sentence understanding which is implemented in a program called ATLAST. Many computational models of natural language understanding have dealt with lexical ambiguity resolution, but ATLAST is one of the few models to address the associated problem of error recovery. ATLAST's ability to recover from an incorrect lexical inference decision stems from its ability to retain unchosen word meanings for a period of time after it selects the apparently context-appropriate meaning of an ambiguous word. The short-term retention of possible lexical inferences permits ATLAST to recover from incorrect decisions without backtracking and reprocessing text, and without keeping a record of possible choices indefinitely.

The principle of retention provides a solution to the problem of error recovery which is compatible with current psycholinguistic theories of lexical disambiguation. Furthermore, the existence of some form of retention in lexical disambiguation is supported by the results of experiments with human subjects. This dissertation includes a discussion of these results and speculation on how the principle of retention might be extended to account for recovery from erroneous higher-level inference decisions.