Automatic Generation of Appropriate Greeting Sentences using Association System

When we humans start a conversation, we are greeting at first. If computer and robot are greeting like us, they can communicate smoothly with us because the next subject comes easily after greeting. That is to say, greeting conversation plays an important part to smooth communications in speaking. In this report, we describe a method of increase the number of appropriate greeting sentences for conversation and selecting sentences based on the situation by machine. Many of conversation system tend to use templates. Lots of chatter bots (Eliza, A.L.I.C.E., Ractor, Verbot, Julia etc.) have been developed. For example, Eliza(Weizenbaum, J 1965) which is one of the well-known system acts for counselling by a personification therapist agent. Eliza does not evaluate an answer of a partner for the reply. It memorizes only a part of the content that the partner spoke in the past and replies by using the word. It is prepared for several kinds patterns about the topic. Like these, as for the natural language processing, task processing type conversation (e.g. automatic systems for tourist information and reservations) becomes the mainstream. However, even under the limited situation, it is known that it is difficult to make a knowledge base of all response case. Moreover, a method using only the prepared template makes monotonous reply and a reply except sentences made by a designer don't appear. So, to make various sentences automatically by machine is important, more than the method to select sentences designer prepared. We herein propose an intelligent greeting processing by which a machine generates various reply sentences automatically by obtaining information about the surrounding state and then generating the best conversation response based on the situation. All sentences are extended automatically from a small quantity of model sentences by using the concept base which is kind of natural-language ontology/concept networks. Simply mechanical extension of conversation sentences makes many improper sentences. So, the proposed method uses language statistics information to delete the improper sentences. In addition, for greetings conversation, we suggest "status space" expressing a certain situation. This is a model to give a weight to sentences automatically by taking the consequence at two