KA is a computational theory of a complex cognitive phenomenon. From the viewpoint of cognitive science, one of the major advantages of building complex and elaborate, yet detailed and precise, computational theories such as KA is the identification of interesting interactions among the different processes. At the start of the KA project, we enumerated a set of ten high-level hypotheses about these interactions [Goel and Eiselt 1991]: (i) understanding natural language descriptions of physical devices enables acquisition of device models, (ii) situating language processing in problem solving identifies the meaning of the ``meaning'' of a device description, (iii) past cases and case-specific models, that originally provided the knowledge structures for addressing a class of design problems, also provide the knowledge structures for language processing, (iv) the SBF language for representing device models, originally developed to address design problems, provides the conceptual knowledge needed for text interpretation, (v) the model-based scheme for indexing the stored cases and case-specific models in long-term memory, again originally developed to address design problems, is appropriate for supporting language processing, (vi) the language process generates adequate cues for probing the long-term memory, (vii) the memory process retrieves relevant cases and associated models from the long-term memory into the working memory, (viii) the retrieved case-specific models act as expectation generators, (ix) the model-based expectations guide the language process, and (x) the language process generates adequate cues for guiding the comprehension process in adapting the retrieved models to construct a model for the new device.
Now at the end of this project, we can confidently assert that the KA theory helps to greatly refine these hypotheses, to make them more precise and explicit. We conclude this article with a brief discussion of how the KA theory has helped to refine the last of the ten hypotheses above because this initially surprised us. We found that language processing provides only limited guidance to the comprehension process in adapting the SBF model of a known device (e.g., the spray can) to construct a model of the new device (e.g., the fire extinguisher). The products of the language process do indicate some of the many differences between the two devices. But most of the important differences come from the structural models of the two devices. Also, the text does enable limited verification of the modified model to insure that the new model is consistent with the text. But we were initially surprised to find that language processing does not clearly indicate the precise content and form of the new device model. There are two apparent explanations for this. First, the device descriptions in Macaulay's The Way Things Work are coarse-grained while our SBF models, which need to support multiple reasoning processes, are fine-grained. This might be resulting in a mismatch between the text and the model so that text can provide only limited help in adapting the model. Second, the diagram that accompanies the textual description of a device is given to KA in the form of a symbolically represented structural model of the device. This might be resulting in some loss of information. Perhaps more importantly, this may imply that the visual process, and not the language process, might be especially important for model adaptation and construction.