Direct memory access parsing

Lexical ambiguity has been a pervasive problem for computer models of natural language understanding, or "parsing," since the original work in machine translation from the fifties. I argue that this problem arises as a natural consequence of organizing these models around a lexically-indexed data base of parsing knowledge. What is required instead is a model of natural language understanding organized around the system's memory. The Direct Memory Access Parsing system organizes parsing knowledge around concepts in memory. A general process of memory search integrates models of top-down prediction and bottom-up reference; predictions control which linguistic knowledge is brought to bear, and the input controls which concepts become active. The result is a system in which memory controls the parsing process. This dissertation describes the algorithms and approaches used in constructing the Direct Memory Access Parser.