This dissertation presents an approach to language understanding that treats all ambiguity resolution as a problem of reference resolution: grounding references to episodic memory. This model of language understanding is evaluated with an implementation of DMAP (Direct Memory Access Parsing) called REDMAP (Reference resolution in Episodic memory for DMAP). DMAP is a language understanding model that recognizes its input by mapping phrasal patterns to existing knowledge structures, updating memory with new information only as needed.;REDMAP works with a large logic based memory (evaluated with ResearchCyc 1.28 million assertions). It uses lexically driven rules to form candidate sets of assertions, and queries memory to ground references in those assertions to existing instances. Assertions from subsequent sentences are merged with running interpretations by identifying how new references are mapped to existing references. Mappings are evaluated by propagating remindings to existing instances to the new references. These instances are substituted into the new assertions, and memory is queried for their existence. If found these assertions support the reference mapping. Additionally, these queries will simultaneously ground any new unmapped references, if possible.;A corpus of simplified English texts describing people, places, and events that span multiple sentences and multiple texts was used to evaluate the accuracy and scalability of this approach. This dissertation provides strong support for two claims. Claim 1: A memory-based reference resolution algorithm (REDMAP) can provide broad coverage of and extending an existing large knowledge base by grounding to existing episodic memory as it parses and can use that memory to reduce ambiguity. Claim 2: The reading rate (mean time per sentence for a text) of the REDMAP algorithm is empirically independent the number of references in the text and the length of input (number of sentences). The evaluations also provide weaker support for an additional two claims. In contrast to supporting claim 1, the savings obtained by REDMAP are mitigated by the cost of additional interaction with memory; however the overhead is shown to be constant and minimal. Furthermore, REDMAP performs knowledge integration while resolving references, alleviating the need for this to be conducted as a separate step.
展开▼
