ATNS USED AS A PROCEDURAL DIALOG MODEL 
D. Metzing, University of Bielefeld, 
Germany 
Summary 
An attempt has been made to use an Aug- 
mented Transition Network IO as a 'proce- 
dural dialog model'. The development of 
such a model appears to be important in 
several respects: 
- as a device to represent and to use 
different dialog schemata proposed in 
empirical conversation analysis; 
- as a device to represent and to use 
models of verbal interaction; 
- as a device combining knowledge about 
dialog schemata and about verbal inter- 
action with knowledge about task- 
oriented and goal-directed dialogs. 
A standard ATN should be further de- 
veloped in order to account for the 
verbal interactions of task-oriented 
dialogs. 
Introduction 
4 Naturally occuring task-oriented dialogs 
are the joint product of the inter- 
actions of (at least) two participants 
who know how to cooperate, i. e. who 
know how to organize 
- their social interactions 
- their verbal interactions, and 
- their task-oriented interactions. 
The amount of such interactions which 
seem to be necessary in specific task- 
oriented dialogs may 
(i) depend on a number of factors 
given in advance such as: readi- 
ness to cooperate (a), preciseness 
of the task representation (b), 
amount of mutually shared task- 
specific knowledge (c), amount of 
knowledge about the other parti- 
cipant (d) and personal factors as 
for example competence and (self-) 
confidence (e); 
(ii) depend on procedures apt to modi- 
fy these factors (a - e) in an 
efficient and positive way; 
(iii) depend on procedures used for 
task resolution and result ex- 
planation. 
Participants of naturally occuring task- 
oriented dialogs are able to make use 
of these factors and procedures in a 
skillful and flexible way, but such 
properties are still lacking even in 
experimental dialog systems. 
In past natural language processing 
research considerable efforts have been 
made to process the structures under- 
lying sentences or texts. Procedures 
have been developed which build up 
deep structures of sentences or which 
determine macro-structures or event 
skripts ('frames') underlying texts. In 
the next two years special efforts will 
be made to process the structures under- 
lying task-oriented dialogs. 
Representation of Interactional 
Knowledge 
In coversation analysis, systematic 
accounts of the sequential organization 
of dialog interactions have been de- 
veloped, e. g. for turn taking, for 
opening sequences, for closing or re- 
pair sequences 7 or for different types 
of task-oriented verbal interaction 
as a whole (e. g. giving advice, direc- 
tions, explanations) 12 But these 
accounts have only Deen of a structural 
type, not of a procedural type. A for- 
mal representation has rarely been 
attempted 6 and an integration or inter- 
action of different knowledge sources 
is generally not considered. 
--487-- 
In the subsequent sections we will argue 
for a level of representation guiding 
the social interactional and the verbal 
interactional aspects of task-oriented 
dialogs. A personal belief or knowledge 
component will use information of this 
interactional level together with infor- 
mation of a task level as well as infor- 
mation of a sentence/text level. We will 
argue for a procedural representation of 
interactional knowledge and we think 
that the usefulness of ATNS I0'3 for such 
a representation should be examined in 
more detail. 
'Parsing Interactions' 
The approach presented here differs from 
other computational dialog models in the 
following way: 
- A dialog model is not based on an un- 
2 derlying dialog prototype specifying 
essentially task-oriented information. 
It is claimed that the social inter- 
actional and the verbal interactional 
aspects of task-oriented dialogs are 
important enough to be represented in 
a detailed way on a special level. 
- Dialog proporties are not only exa- 
mined by problem solving techniques I. 
Instead, extended parsing techniques 
are used in 'parsing interactions' and 
they are supposed to be helpful in de- 
termining the interactional structure 
underlying utterances. 
Let us further specify the kind of inter- 
actional knowledge which participants of 
certain types of task-oriented dialogs 
are supposed to have as well as ways to 
represent it. The participants will 
generally know how to manage the social, 
the verbal and the task-oriented inter- 
actions. They will generally know about 
several rather invariant, components of 
a certain type of task-oriented dialogs 
as well as of a normal sequence of these 
components. They know about the detailed 
(alternative) structures of each compo- 
nent, the choice of which may depend on 
factors as were mentioned above(a - e). 
They know 
- how to initiate a social contact/a ver- 
bal interaction and how to respond 
positively/negatively to this initia- 
tive; 
- how to continue/to maintain an inter- 
action, 
- how to signalize interest, competence 
or difficulties of) understanding, 
- how to organize turn-taking, 
- how to initiate the termination of a 
social contact/a verbal interaction 
and how to respond positively or nega- 
tively to it. 
Part of this knowledge may be described 
as sequences of social/verbal inter- 
actions, formally to be represented as 
connected (sub-) networks of an ATN. 
I. e. social/verbal interaction is seen 
as a process of path selection of (at 
least) two participants in a network of 
states and state transitions. These 
(sub-) networks should be set up on an 
empirical basis (recordings of naturally 
occuring task-oriented dialogs). 
An Example 
Some properties of the interactional 
knowledge mentioned above may be repre- 
sented in an ATN in a straightforward 
way, whereas the representation of 
others seem less straightforward. 
Constituents 
Let us assume that some types of infor- 
mation-giving dialogs may be assigned a 
-488 
network structure with the following 
constituents of social/verbal inter- 
action: 
vl 
T ~ 
< 
Z © 
H 
@ 
/9 
r..) 
o-5 
! 
1 
21 
Each (sub-)network represents sequences 
of interaction 
- on the level of a dialog type (a), 
- on the level of speech act sequences 
(b), or 
- on the level of turn-taking organi- 
zation (c). 
The arcs indicate (alternative) state 
transitions/(alternative) courses of 
interaction. They may be marked or un- 
marked ('JUMPs'). 
According to fig. I a dialog interaction 
may start: 
- with a contact initiating utterance 
("EXCUSE ME ..."), level a; 
- with a question introducer ("MAY I 
ASK YOU ...") level b; 
- with a task-specific question ("WHERE 
DO I FIND ..."), level b; 
- with a 'turn introducer' ("YES"I "...") , 
level c. 
Possible actions may be skipped 
(cf--~4--~arrow), repeated (loons;itera- 
tion) or some sequences nay ~e embedded 
into other sequences (cf. QUESTION/ 
ANSWER arc in the QUESTION/ANSWER sub- 
network; recursion). 
These subnetworks are connected to other 
networks with the same type of infor- 
mation or with other types of infor- 
mation: 
linguistic information in order to re- 
cognize/generate different forms (e. g. 
to initiate a contact, to introduce 
questions, topass turns) or direct/ 
indirect ways to ask a question ("9~ERE 
DO I FIND ..." vs "I DO NOT KNOW ...", 
"I SUPPOSE YOU KNOW ..."); task-oriented 
information in order to build up coherent 
sets of answers (on the basis of a 
task-model). 
Note that some utterances may serve 
--489-- 
several interactional roles ("MAY I 
ASK YOU ..." uttered at the beginning 
of a dialog initiates a contact and 
introduces a question.) Special tests 
on arcs will recognize this and the 
corresponding actions will build up an 
interactional structure according to such 
multiple roles of constituents. (A way 
to see them from different perspectives.) 
Interpretation of Arcs 
The interactiona± information represented 
in an ATN-subnet may be used to plan and 
to guide the recognition/generation of 
social/verbal interaction as part of a 
task-oriented dialog. The information 
represented on an arc should be used for 
both, for recognition as well as for 
planning/generation. The structure of the 
(sub-)networks is useful in the sense 
that 'normal' courses of interaction are 
explicity represented. So they are 
expected and indicate a kind of inter- 
actional coherence of a task-oriented 
dialog. But speakers may violate such 
normally respected dialog sequences and 
they can cope with this fact. It is 
therefore desirable to make a more 
flexible use of the information represen- 
ted in a network. E. g. it seems desirable 
to calculate a (not yet existing) tran- 
sition on the basis of task-specific 
cues and/or utterance cues (for example 
when a participant reopens an already 
closed subtask or when he repeats or 
reopens an already executed move or 
when he suddenly terminates an inter- 
action. 
Further developments 
It seems desirable to have an extended 
ATN parser in order to cope with unex- 
pected dialog sequences. In implementing 
aspects of social/verbal interactions 
one should examine carefully efforts 
made 
5,9 - to use an ATN in a more flexible way 
- to combine recognition and generation 
in an ATN 8, and 
- to build up several interacting ATNs". 

References 

I. Appelt, D. E. "Problem solving Applied 
1980 to Lanauage Generation", 
Proc. of the 18 th Meeting 
of ACL, 59 - 63. 

2. Bobrow, D. G.,and the PARC understander 
1977 group, "GUS - I, a Frame 
Driven Dialog System" 
Artificial Intelligence 
8:2. 

3. Goldstein, I. P., M. L. Miller PAZATN: 
1976 A Linguistic Approach 
to Automatic Analysis 
of Elementary Pro- 
gramming Protocols, 
MIT AI MEMO 388. 

4. Grosz, B. "Discourse Knowledge", 
1978 in: D. Walker (ed.) Un- 
derstanding Spoken Lan- 
guage, 229 - 344 

5. Kwasny, S., N. Sondheimer "Un- 
1979 grammaticality and Extra- 
Grammaticality in Natural 
Language Understanding 
Systems", 
Proc. of the 17 th Meeting 
cf ACL, 19 - 24. 

6. Metzinq, D. Implementation of Inter- 
1979 actional Processes 
Memo University of Biele- 
feld. 

7. Sacks, U., E. Schegloff, G. Jefferson 
"A Simplest Systematics 
for Conversation 
for the Organization 
of Turn-Taking for 
Conversation", 
Language 50, 696-735. 

8. Shapiro, S. C. "Generalized Augmented 
1979 Transition Network 
Grammars for Generation 
from Semantic Networks", 
Proc. of the 17 th 
Meeting of ACL, 25 - 29. 

9. Weischedel, R. M., J. E. Black "If the 
1980 Parser Fails" 
Proc. of the 18 th 
Meeting of ACL, 
95 - 96. 

10. Woods, W..A. "Transition Network 
1970 Grammars for Natural 
Language Analysis~, 
CACM 13, 591 - 606. 

11. - Research in Natural 
1978 Language Understanding, 
BBN-Report No. 3963 

12. Wunderlich, D. "Ein Sequenzmuster 
fur Ratschl~ge: Ana- 
lyse eines Beispiels", 
to appear'in: D. Metzing 
(ed) Dialogmuster und 
Dialogprozesse (Hamburg). 
