Cognitive agents often acquire knowledge of complex phenomena by reading a book. For example, a naive cognitive agent may acquire knowledge of how air-conditioners work by reading a popular science book such as The Way Things Work by David Macaulay [1988]. In general, understanding the natural language description of a device, comprehending how the device works, and acquiring a device model, involves a complex interplay between language, comprehension, memory, problem solving and learning processes. In addition, these processes use many different kinds of knowledge including semantic knowledge of the domain, episodic knowledge from past experiences in the domain, and the information provided in the text.
But most computational models of text interpretation deal with language understanding in vacuum, in more or less complete isolation from other processes. Typically, they either propose a largely bottom-up process in which the interpretation is constructed from the text alone, or a largely top-down process in which a precompiled knowledge structure helps to generate expectations and provides a template for filling in specific details given in the text. In interpreting real texts, however, neither the text always provides sufficient information to enable the construction of a satisfactory interpretation nor does the reader always have a precompiled knowledge structure that matches the text. Our theory of language understanding for device comprehension and knowledge acquisition not only combines bottom-up and top-down strategies for language processing, but it also integrates the language process with memory, comprehension, problem-solving and learning processes.
In contrast to multi-strategy or multi-task theories, we call our theory multi-faculty because it unifies multiple cognitive faculties, not just multiple tasks or strategies within a specific cognitive faculty such as language. The multi-faculty theory is embodied in an operational, but still evolving, computer program called KA.
In [Pittges et. al. 1993], we described an early version of the KA system that unified language, memory and comprehension processes in the service of understanding a new design problem stated in English. We also showed how past problem-solving experiences retrieved from long-term memory enable the understanding of new problems. In [Peterson et. al. 1994], we described a new version of the KA system that not only integrated language, memory and comprehension processes but also unified them with problem solving . We also showed how problem solving helps to evaluate the output of the language, memory and comprehension processes. The above work grew out of our earlier theory of adaptive design in which new design problems are solved and new designs are constructed by adapting past design cases [Goel 1991a, 1991b].
In this article, we describe new work on the KA project that differs from and adds to earlier work in two aspects. Firstly, the input to KA now is not a description of a design problem, but an English language description of a device from the book ``The Way Things Work.'' Secondly, the new version of KA not only unifies language, memory, comprehension, and problem-solving processes but also integrates learning with them. This new work grows out of an evolving theory of adaptive modeling in which comprehension of the workings of a system is represented and organized in the form of a structure-behavior-function (SBF) model, and SBF model of a new device is constructed by adapting old models of familiar devices [Goel 1991b, 1996].
Since we already have described the process of language understanding in KA in earlier papers, we will not repeat it here; [Peterson, Mahesh and Goel 1994] provides a detailed account. Instead, we (i) describe our framing of the problem of device comprehension as an abduction task, (ii) present a high-level account of the knowledge and strategies KA uses for addressing this task, and (iii) discuss how KA acquires a SBF model of new devices from English language description.