Dialogue Model and Associative Processing
--- A Basic Consideration on Generating Cooperative Answers through Associative Mechanism with Dialogue Example Base ---

Tadahiro KITAHASHI, Noboru BABAGUCHI, and Atsuko TAKANO

The Institute of Scientific and Industrial Research, Osaka University
8-1 Mihogaoka, Ibaraki 567 Japan

Generation of cooperative answers is one of the important issues for realization of natural language interface system. Last year we reported an overall model of recognizing diversified patterns of Question \& Answer pairs based on presuppositions underlying questions, which have been regarded to be essential in many studies on cooperative answers. This year we have been pursuing an application of the model to generation of cooperative answers.
In a similar way of formalizing diversified patterns of Q-A pairs, we formalize the process of generating cooperative answers based on the following two points. One is speaker's misunderstanding of the hearer's situation that is detected from discrepancy or contradiction in the presuppositions under questions. In such cases, the replies may point out the misunderstandings or may search answers satisfying the intentions of them. The other is consideration about some background aiming at helping the speaker understand the situation clearly. In the process of formalizing generation of answers, we have realized that there is a pattern of answers which is unable to be generated by an algorithmic procedure unlike other answer patterns. We name the case as 'suggestions for speaker's creating answer',
One of the pragmatic presupposes at speaker's inquiry to a hearer is that the hearer has enough knowledge for answering the question. (In this paper, it is assumed that a speaker asks a question and a hearer replies.) The above case arises from the contradiction of the pragmatic presupposition. The 'suggestions for speaker's creating answer' should be relevant to supporting the process of speaker's inference for creating an answer to his own question. The following utterance R1 is an example of the 'suggestive knowledge for speaker's creating an answer' to the question Q1.

Q1: Kujira ha honyu-rui desuka. (Is a whale mammal?)
R1: Kodomo ha umimasuyone. (It breeds babies, doesn't it?)

R2 is another example of 'suggestions for speaker's creating an answer' to Q2. Here the speaker may be able to infer a reply to his own question Q2 by referring the hearer's utterance R2, who also has no idea about the answer.

Q2: Sanzenin no sanpai kyaku ha hotondo jyosei desuka.
(Are most worshipers of Sanzenin women?)
R2: Jyakkouin no sanpai kyaku ha hotondo jyosei desu.
(Most worshipers of Jyakkouin are women.)

As before, it is considered to be difficult to figure out an algorithmic procedure for generating such kind of answers. Then we have taken an approach to preparing an association mechanism based on an example-base of dialogues for replying questions at this kind of situation.
We propose a model of generating answers based on the association mechanism with an example-base of dialogues. We are examining the flexibility of the method by the experiment of retrieving data similar to key cases from the database of old temples in Japan. Then, we are going to consider about applying the mechanism to generation of 'suggestions for speaker's creating an answer'.
Our method firstly evaluates the viewpoint for a given question and retrieves a Q-A pair whose question is conceptually similar to a given question from the example-base of dialogues based on the viewpoint. Then it modifies the answer example utterance of the retrieved Q-A pair to generate an answer appropriate to the given question in an assumed situation.
In this method, semantic structures of utterances are represented by an extended case structure. They are composed of predicates, attributes of predicates and case elements. Attributes of predicates represent modal information or pragmatic information of the predicates. We consider fourteen types of information as the values of the attributes such as State, Possible, Cause, Guess and Past. We are currently sorting out the type of cases as follow;Agent, Object, Tense, Location, Tool, Degree, Cause and Comparison. The following expressions are examples of the semantic structure of question R3 and question R4.

R3:Hiroshima made donokuraide ikemasuka.
$<$(iku possible),((Location,("genzaichi",Hiroshima)),(Degree,HOW))$>$
R4:Hiko-ki to Shinkansen ha dochira ga toku desuka.
$<$(tokudearu state),((Object,(WHICH,Hiko-ki,Shinkansen)))$>$

Here, context processing problems such as anaphoric reference and elliptical case elements are supposed to have been resolved. The example-base of dialogues consists of Q-A pairs, where utterances are represented in the form of the semantic structures. Although the example-base has a flat structure now, it should be systematized in future for efficient reference.
The similarities between given questions and example questions in the example-base of dialogues are evaluated by the conceptual similarities between their predicates, their attributes and their case elements each by each. In order to define the measure of the similarity between predicates, we are constructing the hierarchical tree of predicates. The measure of two predicates is defined using the distance between them on the tree. The similarities between each two cases elements are also evaluated using the distance between them on the hierarchical tree of nouns.
We are now building the dialogue example base at the information desk of sightseeing. We are going to test the feasibility of our method by applying it to many sample questions, in the situation where hearer does not have enough knowledge for answering the question.