Learning Operator Semantics by Analogy

This paper proposes a cognitive model for human procedural skill acquisition based on problem sol\ing in problem spaces and the use or’ an&)g! for buildmg the reprcscnration of operator semantics. Protocol data of computer-naive subjects learning the EhlACS text editor suaests that they use their knowledge of typewriting to decide Hhich commands to use in performing editing tasks. We propose a formal method of analysis that compares operators in two problem spaces (based on posrcondirion similarit)) and generates misconceptions (based on preand postcondition differences). Comparing these predicted misconceptions bith error data and \.erbal comments in problem sohing episodes validates this analysis. The Phenomena and the Question Analysis of several experimental protocols of computer-naive people learning the EMACS text editor suggest that they were reasoning by analogy from their knowledge of typewriting. The context of text editing spontaneously evokes the analogy to typewriting, because of the similarity of the keyboards, the similarity of the computer screen to the typed page, and the similarity of the tasks in editing and typing. The use of the typewriter analogy was also prompted by the teachers, with remarks such as: “Yeah, go ahead. It works just like a regular typewriter.” The learners’ verbal data suggest that they were indeed taking this advice. For example: The task is to move to the beginning of a word. The teacher says: “Hitting the space bar still doesn’t move you ahead one space.” Learner: “That’s part of my problem, because on a regular typewriter, you can just zip right in there and do your thing.” Such verbal comments suggest that the learners were engaged in problem solving and were using the analogy to the typewriter to figure out what editing operations were appropriate. As scientists, we should be skeptical about taking these verbal statements at face value. The question is whether this apparent use of the typewriter analogy actually plays a significanl role in the learning and performance, Our strategy to examine this question is: Given a general model of analogical learning consistent with the protocols, we develop a specific analysis of the misconceptions that should arise in trying to import knowledge from the typewriter domain to the editor domain. We then test whether these predicted misconceptions are supported by the learners’ performance data. * Beginning September 1983, address will be: Sarah A. Douglas; Department of Computer Science; University of Oregon; Eugene, of? 97403. Thomas P. Moran Xerox Palo Alto Research Center 3333 Coyote Hill Road