AI Magazine Volume 8 Number 1 (1987) (© AAAI) RESEARCH IN PROGRESS Mallory Selfridge, Donald J. Dickerson, and Stanley F. Biggs Cognitive Expert Systems and Machine Learning: Artificial Intelligence Research at the University of Connecticut Research at the Artificial Intelligence Laboratory of the Uni- judge performance of corporations and learn about them by versity of Connecticut is currently focused on a number of reading real-world sources, and (3) fundamental research in projects addressing both fundamental and applied aspects of computer models of cognitive development. next-generation expert systems and machine learning. We CMACS: Learning Causal Models of Physical believe that these next-generation expert systems will have Mechanisms by Understanding Real-World Natural to be based on cognitive models of expert human reasoning Language Explanations and learning in order to perform with the ability of a human expert. Consequently, we term such next-generation expert The causal model acquisition system (CMACS) (Daniel1 systems cognitive expert systems. 1985; Klimczak 1986; Selfridge, Daniell, and Simmons Cognitive expert systems should display three charac- 1985) addresses how an expert system can learn causal teristics. First, because expert human reasoning and learning models of physical mechanisms by understanding real-world rely in part on qualitative causal models and large-scale natural language explanations of these mechanisms. Follow- event-based memory structures, cognitive expert systems ing research conducted by deKleer and Brown (1983; 1984)) should rely on similar knowledge. Second, because human CMACS represents physical mechanisms as combinations of experts are skilled at acquiring knowledge, often through natural language interaction, cognitive expert systems should learn through real-world natural language interac- tion. Third, cognitive expert systems should not merely Abstract In order for next-generation expert systems to learn from the input provided but should be aware of the demonstratethe performance, robustness,flexibility, and learning information that has not been provided and ask the user for it. ability of human experts, they will have to be based on cognitive (In the extreme case, interacting with such a system consists models of expert human reasoning and learning We call such next- entirely of answering its questions.) Thus, we seek to build generation systems cognitive expert systems. Research at the Artificial Intelligence Laboratory at the University of Connecticut cognitive expert systems that model expert human reason- is directed toward understanding the principles underlying ing, learn from real-world natural language interaction, and cognitive expert systems and developing computer programs ask questions about what is not understood. embodying those principles The Causal Model Acquisition To achieve these goals, our research falls into three ma- System (CMACS) learns causal models of physical mechanisms by jor areas: (1) diagnostic expert systems that learn causal understanding real-world natural language explanations of those models for physical mechanisms by understanding real- mechanisms. The Going Concern Expert (GCX) uses business and world natural language explanations, (2) expert systems that environmental knowledge to assesswhether a company will remain in business for at least the following year The Business Information System (BIS) acquires business and environmental Mallory Selflidge is an assistantprofessor in the Departmentof Computer knowledge from in-depth reading of real-world news stories. These Science and Engineering, Donald .I Dickerson is a professor in the systems are based on theories of expert human reasoning and Depal tment of Psychology,and StanleyF. Biggs is an associatePI ofessorin learning, and thus represent steps toward next-generation cognitive the Department of Accounting at the University of Connecticut, Storrs, expert systems. Connecticut06268 SPRING 1987 75 components whose behavior it already knows about. This ing abilities to include a much wider variety of explanations; component knowledge consists of semantic component and (3) exploring CMACS’s possible role in an intelligent frames and associated knowledge of component behavior, tutoring system. which are similar to deKleer and Brown’s confluences. In order to learn a causal model of a new physical mech- anism from an explanation, CMACS translates each explan- GCX: Using Event-Based Memory Structures in an atory statement into a representation of its meaning. If the Expert System to Make Going-Concern Judgments meaning expresses a direct causal connection, CMACS The going-concern expert (GCX) (Biggs and Selfridge 1985, builds the appropriate piece of causal model. If the statement 1986a, 1986b; Selfridge, Biggs, and Krupka 1986) is being expresses causality between two components that cannot di- developed to address whether a company is a going concern, rectly be connected to each other, CMACS infers the inter- that is, whether it will remain in business for at least the mediate components. If the causality expressed by the state- following year. The question is important because it is a fun- ment is ambiguous, CMACS uses its knowledge of plausible damental part of an auditor’s certification of a company’s component interconnections to infer the correct causality financial statements. If, for example, auditors determine that and build the correct piece of causal model. If the statement a company is not a going concern, then generally accepted refers to a subcomponent of another component, CMACS accounting principles are not strictly applicable. Interviews merges the subcomponent information with the other com- with auditors have revealed that not only do they use knowl- ponent in the model. After processing the entire explanation, edge of financial performance in making going-concern CMACS examines the causal model it has built and searches judgments, but they also have extensive knowledge of busi- for components that have inputs or outputs which are not ness and environmental factors which represent the underly- connected to other components. Such incompletely con- ing causes of financial performance. Thus, building GCX nected components represent areas of ignorance for has involved developing both a financial-reasoning capabil- CMACS; it generates natural language questions about these ity and a capability to reason causally from business and en- areas and builds additional pieces of the causal model from vironmental factors. The going-concern judgment is an ap- the user’s answers. propriate research topic not only because it is an important After acquiring a causal model for a mechanism, part of auditing but also because it is representative of a large CMACS can use the model to answer questions and generate set of similar problems, both within auditing and within the explanations about the mechanism and to understand mecha- larger context of understanding organizational behavior. nism behavior and diagnose mechanism failures. CMACS GCX begins with financial data from a company’s an- answers questions and gives explanations by expressing ele- nual reports for a period of time and with business and envi- ments of the model through a natural language generator ronmental knowledge covering this same period. The busi- (Cullingford et al. 1982). It understands mechanism behav- ness and environmental knowledge consists of a network of ior and diagnoses failures by (1) reasoning backward business and environmental events linked by temporal and through the model from external behavior (for example, causal relationships. In addition, this knowledge is orga- gauge behavior) to hypothesize possible control changes or nized and indexed through a set of higher-level knowledge component failures that might cause the behavior, (2) per- structures, similar to Schank’s memory organization packets forming qualitative simulation of each hypothesized control (MOPS) (Schank 1982), that capture business and environ- change or mechanism failure to generate predictions about mental goal structures and generalizations. Finally, causal subsequent gauge behavior, and (3) comparing predicted links from business and environmental knowledge to a finan- gauge behavior with actual behavior to diagnose the control cial model of the company are maintained to enable GCX to change or component failure. understand the financial effects of business and environmen- CMACS currently knows 45 different components and tal events. has a vocabulary of 175 words. It successfully learns causal GCX uses a set of financial reasoning rules to calculate a models of an air conditioner and a gas turbine by reading number of financial measures and evaluate these measures to explanations taken from repair and maintenance manuals. make judgments about the company’s financial perfor- After learning, CMACS can answer questions and generate mance. It then accessesits knowledge of the company’s busi- explanations about each mechanism and can understand ness and environment and uses it in a series of processing mechanism behavior and diagnose mechanism failures steps to refine its judgments. First, it understands the under- within the limits of its qualitative reasoning abilities. Our lying reasons for the company’s financial performance. It current research with CMACS involves improving these might understand, for example, that a company’s recent op- abilities by (1) developing techniques for qualitative reason- erating loss was caused by