A. Bryan Loyall and Joseph Bates
We are studying how to create believable agents that perform actions and use natural language in interactive, animated, real-time worlds. We have extended Hap, our behaviorbased architecture for believable non-linguistic agents, to better support natural language text generation. These extensions allow us to tightly integrate generation with other aspects of the agent, including action, perception, inference and emotion. We describe our approach, and show how it leads to agents with properties we believe important for believability, such as: using language and action together to accomplish communication goals; using perception to help make linguistic choices; varying generated text according to emotional state; and issuing the text in real-time with pauses, restarts and other breakdowns visible. Besides being useful in constructing believable agents, we feel these extensions may interest researchers seeking to generate language in other action architectures.