Generating Dialogue

Naoyuki OKADA

Department of Artificial Intelligence, Kyushu Institute of Technology
Iizuka 820, Japan

The three items on the discourse structure of dialogue in "invitation" are discussed: dialogue environment, invitation structure, and dialogue process.
First, the environment of dialogue is shown. An intelligent agent simulating the protagonist of an Aesop fable was constructed previously. The agent (k-system) has desires, makes plans to satisfy them, recognizes its surroundings, takes action to execute the plans, gets emotional and utters those processes in natural language. Based on this, another agent (p-system) is constructed: first, a copy of k-system was made. Then high-level concepts are kept as they are, but low-level concepts or instances are changed to special one to the p-system. The parameters of physiological and emotional states are also changed. Thus, the p-system can be completed to have a conversation with the p-system.
The dialogue is as follows: the k-system tries to go to a mansion, find a pond in the garden, and drinks water, because he is thirsty very much. On the way to the mansion, he meets the p-system, and exchanges greetings. He knows the mansion is very dangerous now, and hesitates to proceed with the original plan further. The p-system reads his mind, and to invite him to change the plan and come to the river side at a bridge together with the p-system. Although the river side is a little bit far, he finally agrees to the p-system.
Next, the discourse structure of invitation is represented by networks. More than thirty examples of invitation dialogues are investigated in detail, and a representation scheme is proposed. It consists of two kinds of mental models: inviting and invited sides. Each model has two layers of networks: rough and detailed. The rough model of inviting side has two components: the goal explanation part and the persuasion/negotiation part. Depending on the opponent's response to the goal explanation, the persuasion/negotiation part chooses several strategies: to make opponent's another plan give up, to compromise with each other on time or location in executing the one's plan, to show the merits/demerits of one's/opponent's plan, and so on. On the other hand, the invited model has corresponding components to the inviting ones: the explanation comprehension part and the assertion/negotiation part.
The p-system, which tries to invite the k-system to do something, contains both the inviting model to achieve an invitation and the invited model to read the opponent mind. The k-system contains both models, too, but their roles of "main and sub-" interchange each other.
Third, the process of an invitation dialogue is analyzed according to the networks. Although the representation roughly indicates the transition of mental states of each side, it is a kind of "default value" of various transitions. The models are designed to well-manipulate such a transition. In particular, the complicated transition of the inviting model of the p-system is shown in detail as the dialogue proceeds. Thus, the models can be expected to be robust in various dialogues.
Finally, our summary is given: the discourse structure of invitation is made clear. Not only the inviting model but the invited one are represented by two-layer networks. The networks can interpret various invitations through flexible transitions.

Keywords: dialogue generation, discourse structure, mental process, invitation