Evaluating Top-Down versus Bottom-Up Knowledge Acquisition Strategies for use in Rule Based Models of Emotional Common Sense

1 Abstract Three ways to build rule based models of emotional common sense reasoning are described: a top-down approach, inspired by a popular theory in the eld, a bottom-up or demo driven approach, and a hybrid approach. A means of evaluating each of the models is described. One aspect of the evaluation involves comparing a given model's performance against an empirically derived benchmark of human behavior. The second aspect involves measuring computational and knowledge engineer resources required by a given approach in order to ascertain hypotheses made about the virtues of each approach. 2 Introduction Knowledge based systems can be constructed to exhibit functionalities that are useful for emotionally intelligent agents. As R. Pfeifer (1987) has indicated, the emotions are closely linked with common sense reasoning. In such systems the same bits of knowledge can be used to model the deductive processes, produce explanations for inferences the system makes, or abduce explanations for others' emotions. Furthermore, understanding emotional phenomena is an essential part of common sense reasoning. Unfortunately, knowledge acquisition is an extensive, labor intensive process (Rich & Knight, 1991). Given the tight coupling between common sense and the emotions, it is advantageous to use a system that contains large amount of common sense background knowledge. One such system is Cyc. Cyc contains a common sense knowledge base and an inference engine. The encoding such of everyday knowledge remains one of the central aims of the project (Lenat & Guha, 1988). The Cyc knowledge base (KB) contains a large number of common sense facts stated in a special logical language known as CycL. The English representation of such common sense facts include a vast range of statements including; mammals have fur. birds fly (except penguins). if it is daytime, it is light outside. Cyc is not the only such system in which to conduct a project such as this. The claims I make below are meant to be generic to any endeavor that involves a knowledge based approach to emotion. In what follows I propose a means of conducting a knowledge acquisition eeort that compares diierent approaches to emotional common sense knowledge acquisition. There are three beneets of such work: advancement of the state of the art in systems with emotional common sense, increased knowledge about the value of diierent approaches towards emotional knowledge acquisition, and theories of emotion with greater precision, clarity, and consistency.