Our work on KA builds on many lines of research in cognitive science including natural language understanding, device comprehension, knowledge acquisition, mental models and model-based reasoning, case-based reasoning and learning, and abductive explanations. Due to limitations of space, however, here we only outline its relationship to earlier work that lies at the intersection of language understanding, device comprehension, and acquisition of device models.
Lebowitz's  RESEARCHER program read natural-language texts in the form of patent abstracts, specifically disk drive patents, and updated its long-term memory with generalizations made from these texts. Its knowledge representation scheme was oriented toward device objects and their structural relationships, which was a departure from most natural language understanding systems of that time which had typically focused on intentional actors and events. The output of the processing was a generalized representation in the form of a structural model of the disk drive which specified its components and the topological relationships among them. The system stored this structural model in its long-term memory and later used this knowledge to aid in the top-down understanding of additional patent texts. However, RESEARCHER's emphasis on components and structural relationships left it unable to build functional and causal models of the mechanisms described. In other words, the system effectively knew how a disk drive was constructed, but it did not know how it worked. In sharp contrast, KA takes a structural model of the new device as part of its input.
Dyer, Hodges, and Flowers  describe EDCA, a conceptual analyzer which serves as a natural language front-end for EDISON, a naive design problem solver. EDCA uses knowledge of the function of physical devices to produce an episodic description of a device's behavior as described by an input text. This episodic description can then be used to generate a new device model to be integrated into long-term memory. The result is a much more comprehensive understanding of the device's functionality than was possible with RESEARCHER, but EDCA's analysis of the device description is not fully integrated with the processes for generating new device models and incorporating them into memory. EDCA, in other words, is but a front end to EDISON.
As Selfridge  notes, separating the process of analyzing the language input from constructing and incorporating the new model is misguided --- the process of understanding a device description is the process of constructing a functional and causal model of that device. This is the approach that we have followed in our work on KA. We believe that this approach enables KA to correct the shortcomings of both RESEARCHER and EDCA.