It is clear that sophisticated dynamic three-dimensional models of faces, such as that shown by Pelechaud at IJCAI93, have the power to delight audiences and convey expressive information[Pelachaud, Viaud, & Yahia1993]. Nonetheless, it may also be argued that much emotional content can be delivered in schematic format as well. Cartoon faces (such as those in Calvin and Hobbs, for example, convey much about the current appraisals of the cartoon character, and schematic faces have been used in clinical settings to help children identify their emotions. Are such faces, which are much less computationally expensive to manipulate, able to convey emotions in a consistent way?
In our own work we use a set of approximately seventy schematic faces, covering up to three intensities in each of the twenty-four emotion categories [Elliott, Yang, & Nerheim-Wolfe1993]. Sixty of these have been included in a morphing module so that faces gradually break into a smile, decay from rage back to a default state, and so forth. The module runs in real time, allowing run-time control over face size, rate of morph, and rudimentary mouth movement (for when the agent is speaking). The system thus allows for over 3000 different morphs, a range not possible with 3D representation. The morphs run on a 66 Mhz IBM PC (with sound and speech cards) concurrently with midi playback, text-to-speech, speech recognition, and the background emotion simulation.
Assuming that either representation can be effective, the question still arises about the effectiveness of the emotion representation on which the dynamic face depends. Our approach is that, at present, low-level personality and emotion representations are too complex to simulate complex social interaction, and that content theories of personality and emotion embedded in individual domains are too simplistic. Hence the middle ground, using an architectural approach (e.g., at the level of schematic face representation) consistent across all aspects of the system.