Answering Comparison Questions in SHAKEN: A Progress Report

Abstract

An important class of questions for knowledge based systems concern comparisons, such as "How is X like Y?" and "How are X and Y different?" This paper describes how we have used a cognitive simulation of analogical processing to answer such questions, to support domain experts in entering new knowledge. We outline techniques for case construction and summarization of comparison results that have been developed and refined based on an independent formative evaluation. In addition to these techniques, we discuss the role of the comparison system in SHAKEN, the larger system in which they are embedded, and our plans for further improvements.

Achieving Intelligence Using Prototypes, Composition, and Analogy

“…The approach to answering a similarities and differences question in AURA follows the conventional model of analogical reasoning (Nicholson and Forbus 2002): case construction, candidate inference computation and summarization. A case for a concept contains its taxonomic information, slot values, and constraints.…”

Large-Scale Analogical Reasoning

“…and What are the differences between A and B? The structure of the computation is the same as that used in (Nicholson and Forbus 2002): case construction, candidate inference computation and summarization. We explain our implementation of these steps and discuss how they differ from previous work.…”

“…For every assertion in the KB, it is possible to query the justification system to check if it was directly asserted or derived, and if it was derived, from which concept it was derived. If the only justification for it is the concept C, we assume that the assertion is local to C. In previous work, the problem of what to include in a case was solved by relying on either an arbitrary depth bound on inference or based on what the domain expert chose to keep visible while saving the concept (Nicholson and Forbus 2002). Neither of these approaches was found to be effective.…”

“…An ideal evaluation would have compared the outputs from the current system with the outputs from (Nicholson and Forbus 2002). We loaded the current KB into the prior system, but since the KB is much bigger it crashes that system.…”

“…In a previous effort to use analogical reasoning for comparison questions, three problems were found (Nicholson and Forbus 2002): (a) the system answered with too much information, and the salient similarity or difference was not highlighted; (b) analogical inference found some incorrect differences; and (c) some expected similarities were missing. As an illustration of these problems consider the comparison between a Eukaryotic-Cell and a Prokaryotic-Cell 1 .…”