Some requirements and approaches for natural language in a believable agent

Abstract

The goal of creating resource-bounded believable agents that use language and action raises interesting issues for the various components of the agent. In this paper we explore these issues, focusing particularly on their relation to natural language understanding in such agents. The specific issues we address are: responsiveness and interruptability; pursuing multiple independent goals concurrently; designing and managing groups of goals to execute concurrently for specific effects (for example, pointing and talking or understanding such gesture and language combinations); understanding incomplete or ungrammatical language or behaviors; the effects of failure in a larger agent; and consistency between the components of the agent. For each of these issues we argue why it is important for our goals, describe how it relates to natural language, describe our approaches to it, and describe the experience we are drawing on in making our conclusions.