Toward a large-scale formal theory of commonsense psychology for metacognition

Robust intelligent systems will require a capacity for metacognitive reasoning, where intelligent systems monitor and reflect on their own reasoning processes. A large-scale study of human strategic reasoning indicates that rich representational models of commonsense psychology are available to enable human metacognition. In this paper, we argue that large-scale formalizations of commonsense psychology enable metacognitive reasoning in intelligent systems. We describe our progress toward developing 30 integrated axiomatic theories of commonsense psychology, and discuss the central representational challenges that have arisen in this work to date. Commonsense Psychology and Metacognitive Reasoning Metacognition is defined in terms of commonsense psychological concepts. The term is meant to refer to the suite of human mental behaviors for monitoring and control, including the selection of goals, assessing progress toward those goals, adopting new strategies for achieving goals, and even abandoning a goal entirely. In providing examples of human metacognition in practice, situations where people employ reasoning strategies are often cited. For instance, metacognition is exhibited when a person sticks a note on their bathroom mirror so that they will remember to do something each morning, or when students assess the relative difficulty of exam questions in order to determine which ones to tackle first. Gordon’s large-scale analysis of human planning strategies (Gordon, 2004) revealed that all strategies have metacognitive aspects to them. In this work, 372 strategies that people use in their everyday lives to accomplish personal and professional goals were collected from ten different real-world planning domains (e.g. governance, warfare, and artistic performance). These strategies were then encoded as preformal definitions aimed at identifying the formal knowledge representations that will be required to enable the creation of cognitive models of human strategic reasoning. Of the 988 unique concepts that were required to author preformal definitions of these strategies, 635 were identified as pertaining to mental states and processes. Gordon (2002) organized these 635 concepts into 30 representational areas (e.g. Memory, Explanations, Expectations, Goal management, Plan adaptation, Execution control), a set which stands as the most comprehensive characterization of human commonsense psychological models available today. Figure 1 illustrates these 30 representational areas by clustering them around the central areas of Knowledge, Envisionment, Goals, Planning, and Execution. There is a growing interest in trying to create intelligent systems that are themselves metacognitive, in that they monitor and control their own reasoning processes to respond proactively to problems and perform better with less need for human intervention. There may be several different metacognitive reasoning approaches that would be successful across different types of intelligent systems, depending on their computational tasks. However, intelligent systems designed to cooperate adaptively with humans will need to utilize representational models that can be aligned with those of their users. To build cooperative, adaptive intelligent systems that engage in metacognitive reasoning, it will be necessary to develop large-scale inferential models of human commonsense psychology. This paper describes our efforts in authoring formal axiomatic theories of human commonsense psychology. First, we describe how we elaborated the representational requirements for strategic reasoning identified by Gordon (2004) by conducting a large-scale analysis of English phrases related to mental states and processes. Second, we discuss our progress in authoring axiomatic theories for the 30 representational areas and the lessons learned so far. We then conclude with directions for future work. Authoring Large-Scale Theories of Commonsense Psychology Our authoring approach was to identify clearly the representational requirements of large-scale theories of commonsense psychology using analytic and empirical techniques before beginning to encode this knowledge as formalisms in first-order predicate calculus. We began with an analysis of Gordon’s 30 representational areas of commonsense psychology, and the 635 concepts that were sorted among them. Because these concept lists contained only the terms necessary for the adequate definition of the real-world strategies that were analyzed as part of that study, some redundancy and gaps were evident. For example, the representational area of Managing Expectations (dealing with the events that people expect to happen in future states) listed the term Expectation violation, referring to the mental event of being surprised by something that occurs, but does not include in the list of eight terms a corresponding concept for Expectation confirmation, referring to the mental event of realizing that one’s expectations have been met. To elaborate these 635 concepts as well as reduce the redundancy that was apparent, we decided to use natural language as additional empirical evidence for the commonsense psychological concepts that were necessary to manipulate in formal theories. Beginning in Summer 2002 and ending in Fall 2004, we conducted a large-scale effort to identify every possible way to express in the English language all of the concepts related to the 30 representational areas. This work was completed by first identifying multiple ways to express each of the 635 concepts in the English language. These examples were then used as a launching point for large group discussions aimed at eliciting even more examples (typically dozens for each concept). The resulting sets of examples were then organized to determine cases where the existing concept set for a representational area lacked a concept that was expressible in language (a term needed to be added) and cases where language made no distinction between two existing concepts (two concepts needed to be combined into one). Computational linguistics graduate students then identified full sets of synonymous expressions for each of the examples, and authored finite-state transducers capable of automatically tagging expressions of commonsense psychology concepts in English text. Gordon et al. (2003) evaluated the quality of this approach to authoring finitestate transducers for four of the 30 representational areas, reporting an average precision performance of 95% and an average recall performance of 82%. Through this approach, a final set of 528 concepts was identified that describe the representational requirements of formal theories of commonsense psychology. With these requirements in hand, we then began the process of applying more traditional knowledge representation methods of formalization and axiomization.