The input to the task of comprehending how a fire extinguisher works in KA is constituted of three elements: the above text, the annotations on the accompanying diagram, and a symbolic representation of the diagram. The symbolic representation of the diagram constitutes a structural model which specifies only the structural elements and the topology of their connections in the fire extinguisher.
But what characterizes acceptable output of the task? We view the task of comprehending how a device works from a natural language description of the device as an instance of the very general abduction task. The abduction task takes a given set of data as input and gives a ``best'' explanation for the data as output [Josephson and Josephson 1994]. But now the question becomes what characterizes a best explanation?
Figure 2: The KA Architecture.
The explanation of a device must not only specify the structural elements and the functions of the device, but it must also specify how the structure results in the functions. That is, it must specify how the device structure gives rise to the causal processes that result in the device functions. Thus we characterize a device explanation as a functional and causal model of the internal workings of the device. A best explanation of a device must satisfy three properties. First, the explanation must account for as much of the input as possible - ideally, it would cover the whole input. Second, the explanation must be consistent with the input. That is, no element of the explanation can be inconsistent with any element in the input. Third, the explanation must be internally consistent. That is, no two elements of the explanation can be mutually inconsistent. In sum, in KA an acceptable output of the task of comprehending how a device works from a natural language description of the device is a functional and causal model of the working of the device that accounts for as much of the description as possible, is consistent with the entire description, and also is internally consistent.