Model of Dialog<P>
-- Some Roles of General Dialog Manager --
<P>
Riichiro MIZOGUCHI and Yoichi YAMASHITA
<P>
The Institute of Scientific and Industrial Research, Osaka University<P>
8-1, Mihogaoka, Ibaraki, Osaka, 567 Japan<P>
e-mail: miz@ei.sanken.osaka-u.ac.jp
<P>\baselineskip 5.5mm
	There are increased expectations for realizing dialog between machines 
and humans through spoken language.  We must solve a number of problems 
before such a natural dialog is realized.  Some problems are inherent to 
dialog or speech interaction and independent of the problem solver.  They 
should be dealt with in the interface system with the dialog management, and 
not in the internal function of the individual problem solver.  From such a 
viewpoint, we are developing an general speech I/O interface independent of 
the dialog task.  In this year, we discussed two roles of the dialog manager 
based on our former studies on the speech interface.  One is the sentence 
generation from concept representation concerning the dialog context and the 
other is topic identification for spoken dialog understanding.  

	In our frame work of speech output based on the concept-to-speech 
conversion system, the problem solver determines what to say using concept 
representation.  Then, the general speech interface modifies the concept 
representation according to the dialog context and converts them into 
synthetic speech.  The dialog manager in the interface has roles of 
determining the surface sentence of the output message taking account of 
the dialog context.  Therefore we investigated the difference between the 
sentence expressions generated with and without the dialog context.  An 
experiment of sentence generation from concept representations was carried 
out by 3 and 4 subjects for two dialog scenarios, respectively.  Differences 
between two sets of the sentence were categorized into 11 factors, such as 
insertion of conjunctions or other words, ellipses, and so on.  Another 
experiment was carried out in order to investigate preference of the 
differences, that is factors of the sentence expression.  Many kinds of 
factors got high scores of preference, over 80\%.  Such preferred factors 
should be incorporated into sentence generation.  We propose ideas of 
generating sentences with preferred factors using the dialog context modeled 
in terms of the topic transition model and information transmission acts.  

	We have been developing a method of predicting the next utterance by 
using the dialog model.  Our basic idea of the utterance prediction is 
arrangement of utterance patterns based on utterance pairs and instantiation 
of the patterns according to predicted topics.  The identification/prediction 
is one of the most important issues in our utterance prediction mechanism.  
We discuss a method of predicting the topic for ambiguous utterance 
information, given as bunsetsu lattice.  The method predicts the topic of 
the utterance using top-down and bottom-up information, which are topic 
transition model, named TPN, and correlation of the word category and the 
topic.  The former is extracted from dialog corpus by hand.  The latter is 
automatically trained for each topic by using topic-labeled dialog corpus.  
The rank of the correct topic among 21 topic candidates is 3.7 and 2.6 for 
the bunsetsu lattice and the correct sentence, respectively, only with 
bottom-up information.  Furthermore, by using a topic transition model, 
these score are improved to 3.0 and 2.1, respectively.
<P>
Keywords: speech interface, dialog model, dialog context, sentence generation, 
topics, utterance prediction<P>