Figure 2 illustrates the general functional architecture of KA. Here we only describe the processes linked by bold-faced arrows in the figure.
The long-term memory contains episodic knowledge of previously encountered devices. Each device case has an associated case-specific structure-behavior-function (SBF) model that explains how the device works [Goel 1991a, 1991b]. The SBF model of a device explicitly represents the structural elements and their configuration, the functions, and the internal behaviors of the device. Each behavior specifies a causal process in the device; the causal processes specify how the device structure results in its functions. In particular, they specify how the device functions are composed of the functions of the structural elements of the device. The SBF model for each device case is expressed in a common ontology that arises out of earlier work on device representations [Bylander and Chandrasekaran 1985; Sembugamoorthy and Chandrasekaran 1986; Chandrasekaran, Goel and Iwasaki 1993]. The SBF ontology defines the domain concepts and the relations between them, and constitutes the conceptual knowledge of the KA system.
The language process uses lexical and conceptual knowledge to generate cues for the memory process as well as preliminary interpretations for the comprehension process. Conceptual knowledge refers to knowledge of the domain concepts and the relations between them as characterized by the SBF ontology. The language process contains a large semantic network that takes the output of the parser as input and produces conceptual interpretations. The nodes and the links in the network are based on the SBF ontology of domain concepts and the relations between them. The spreading-activation mechanism in the network uses an early-commitment processing strategy with robust error-recovery to resolve word-sense ambiguities [Eiselt, 1987]. The mechanism resolves word-sense ambiguities by considering processing choices in parallel, selecting the alternative that is consistent with the current context, and deactivating but retaining the unchosen alternatives for as long as space and time resources permit. If some later context proves the initial decision to be incorrect, retained alternatives are reactivated without reaccessing the lexicon or reprocessing the text. [Peterson, Mahesh and Goel 1994] provides details of language processing in KA.
The memory process uses cues generated by language process as probes into the long-term memory. It accesses device cases and associated case-specific SBF device models and puts them into a working memory for use by the comprehension and problem-solving processes. The memory process also stores newly learned models in the long-term memory. The device cases are indexed by the functions of the stored devices; the SBF models are indexed by the cases. This indexing scheme is borrowed from our earlier work on adaptive design [Goel 1991a, 1991b].
The comprehension process constructs a SBF model for the new device by adapting the SBF device models accessed by the memory process. It uses generic (abstract, skeletal) modification plans for the task of adapting SBF models of known devices to construct a model of the new device. The selection of relevant modification plans is based on the functional and structural differences between the SBF model of the known device and the description of the new device. KA's method constructing the new SBF model is identical to that of adaptive modeling [Goel 1991b, 1996].
The learning process uses the SBF model of the new device to learn appropriate indices for storing the model in the long-term memory. The new indices depend both on the contents and organization of the memory and the functional and causal explanation provided by the SBF model. Again, KA's method of index learning is identical to that of adaptive modeling [Bhatta and Goel 1995]. (The problem-solving process in Figure 2 plays no direct role in this process of acquiring a SBF model of a new device from an English language description.)