Learning Components of Computational Models from Texts

Abstract

The mental models of experts can be encoded in computational cognitive models that can support the functioning of intelligent agents. This paper compares human mental models to computational cognitive models, and explores the extent to which the latter can be acquired automatically from published sources via automatic learning by reading. It suggests that although model components can be automatically learned, published sources lack sufficient information for the compilation of fully specified models that can support sophisticated agent capabilities, such as physiological simulation and reasoning. Such models require hypotheses and educated guessing about unattested phenomena, which can be provided only by humans and are best recorded using knowledge engineering strategies. This work merges past work on cognitive modeling, agent simulation, learning by reading, and narrative structure, and draws examples from the domain of clinical medicine.