Uncertain Belief Reasoning and Metaphor-Based Reasoning in an Implemented Context-Based System
More specific web link: http://www.cs.bham.ac.uk/~jab/ATT-Meta/
This talk will outline the use of contexts for (a) reasoning about beliefs and (b) metaphor-based reasoning in ATT-Meta, an implemented AI system written in Quintus Prolog. The system is aimed mainly at natural language applications.
In the talk there will be an emphasis on (b) and on the interaction between (a) and (b). Task (a) is accomplished by a combination of simulative reasoning---using a context to encapsulate the target agent's alleged beliefs and reasoning---and non-simulative meta-reasoning. Arbitrary nesting of simulative contexts is supported.
Task (b) is accomplished by using a ``source-based pretence'' context to encapsulate reasoning within the terms of the source domain of the metaphorical view at hand. For example, for a sentence like ``The ideas were widely separated in John's mind'' there would be a context for reasoning about John's mind and ideas as if they really were a physical region and physical objects, respectively. We rely on the system having very general "conversion rules" (a type of context-bridging rule) for each specific metaphorical view it knows about, such as the the view of MIND AS PHYSICAL SPACE. The reasoning within pretence contexts is intended to link the often ridiculous literal meanings of metaphorical sentences with propositions that can be mapped by the conversion rules to target-domain propositions outside the pretence contexts. Pretence contexts can be nested, allowing a treatment of chained metaphor.
Most of the work on aspect (b) of ATT-Meta has been on metaphors for mental states. Thus, there is a tight interaction between (a) and (b): metaphorical input can defeat defaults about belief reasoning that would normally hold sway. Another type of interaction is that pretence contexts for (b) and simulative contexts for (a) can be nested within each other, allowing source domains involving agents to be properly handled and allowing other agents' metaphorical reasoning to be reasoned about.
A rather unusual feature of the system--whether as implemented belief reasoner or implemented metaphorical reasoner---is that the reasoning is (qualitatively) uncertain, resting mainly on default rules, and having a qualitative conflict-resolution mechanism based on a specificity heuristic. The system copes with the major complications that uncertainty gives rise to for aspects (a) and (b), notably the danger of an explosive increase in the number of hypotheses that need to be attended to, and the need to do conflict resolution across context boundaries as well as within individual contexts.
Although the system is implemented and rests on logic representations, it has not yet been fully formalized in the usual logical sense. Because of the central involvement of uncertainty, and because the provisions for uncertainty are still subject to change, complete formalization at the present stage would be difficult and perhaps inappropriate.
ACKNOWLEDGMENTThe project was supported in part by grants IRI-9101354 and CDA-8914670 from the National Science Foundation.
Slideshttp://www.cs.bham.ac.uk/~jab/VWS/title.html (Wed May 5 08:21:34 PDT 1999)
1 John A. Barnden
School of Computer Science
The University of Birmingham
Top of this page