Turn-Taking in Discourse and Its Application To The Design of Intelligent Agents

We are interested in studying turn-taking in dialogue, in particular human-computer discourse. Humans and computers can communicate in any number of ways, such as through an operating system command line, a direct manipulation interface, or even natural language. Turn-taking is a basic fact of life that interacting agents must deal with: who should say what, and when? In some situations, turn-taking is a simple matter of following a predefined script, such as reciting wedding vows, or following the World Wide Web HTTP protocol. However, in much human-human dialogue, a more flexible and dynamic form of turn-taking is needed since it is impossible to always predict how a conversation will progress. In natural language dialogue systems, it is often assumed that explicit turn-ending signals will be given by the conversants. For example, such systems typically require that you type a sentence and then press , which is a signal to the computer that it is its turn to talk. However, explicit turn-ending signals are not practical in other domains. For example, in a graphical user interface, the user points and clicks with a mouse, while the computer responds by opening windows, moving objects, etc. In this case, the user does not give explicit signals as to when she has finished "speaking". As in human-human dialogue, it is necessary to carefully observe the other conversant and watch for relevant points to take a turn as they arise. To do this effectively, a general theory of turntaking is needed to determine why a turn should be taken, when it should be taken, and how turn-taking goals interact with an agent’s other goals. With respect to agents in general, we believe that since agents may engage in interaction that can be treated as a kind of conversation, knowing how and why to communicate with other agents is a necessary ingredient for success. In particular, agents meant to interact with a human must be able to deal with people in a natural and flexible manner. Below we outline a general framework for turn-taking that could be applied to any dialogue situation involving intelligent interacting agents.