Three Typed Pragmatics for Dialogue Structure Analysis 
Hitoshi IIDA*, Takayuki YAMAOKA* and Hidekazu ARITA** 
* A'I2R Interpreting Telephony Research Laboratories 
Sanpeidani, hmidani, Seika-cho,Souraku-gun, Kyoto 619-02, ,Japan 
\[{iida/yamaoka}%atr-la.atr.co.jp@uunet. UU.NET\] 
** Mitsubishi Electoric Corporation 
8-1-1 Tsukaguchi-honmaehi, 
Amagasaki, tlyougo 661, Japan 
\[arita%sys.crl.melco j unet@uuneL UU .NETI 
Abstract 
An experimental system for dialogue 
structure analysis based on a new type plan 
recognition model for spoken dialogues has been 
implemented. This model is realized by using 
four typed plans which are categorized into three 
kinds of universal pragmatics and a ldnd of task- 
dependent knowledge related to common action 
hierarchies. The experimental system is 
characterized by higher modularity and 
computational efficiency through defining a 
hierarchical usage order between these 
knowledges. The system can grasp a dialogue 
structure making it possible to solve problems 
related to spoken dialogue interpretation. 
1. INTRODUCTION 
An efficient, and smooth communication 
be'~ween humans is generally realized in spoken 
dialogues. This fact is mainly supported by 
various ellipsis exp:°essions concerning old information, the 
dialogue participants, zero- 
pronouns - especially in dapanese, substitutional 
verbs and so on. Ass result, each utterance is 
fragmental. A sequence of these utterances 
generally construct a whole dialogue step by 
step. \['ragmental utterance comprehension by a 
hearer can be achieved using knowledge of the 
dialogue sitamtion, context intbrmation, domain 
dependent knowledge, especially the donmin 
dependent action hierarchy\[Litman871, 
universal pragmatics concerning how to advance 
a dialogue, maintain dialogue cooperation 
between dialogue participants, etc., and 
language specific pragmatics \[ Levinson831. Such 
ellii;sis-resolution-is one of the main problems in 
discourse understanding. Several approaches 
were proposed and implemented in \[fiendrix781, 
\[Shimazu79\], \[Carbonel1831, etc. These 
researchers used various heuristic rules and did 
not make a clear distinction between domain 
~pecific knowledge and pragmatics. As a 
consequence, the user interface became 
inflexible because the system basically handled 
expected utterance patterns. 
*(1) :A discourse goal is characterized by a discourse 
expectation which dialogue participants mutually believe 
as an expectation. Therefore, a discourse goal is very vague 
because a expectation depends on various attentions or empathies and the participant's knowledge, 
*(2) :Allowing embedded turn-takings. *(3):A communicative act is basically defined as an 
abstract action, one that effects the bearer's thinking or deci,qiommaking and which can be described by a plan 
schema (cf. Cohen84). 
Recently, a plan-based dialogue 
understanding approach was developed using a 
kind of pragmatics metaplan: a 'discourse plan' 
incorporated with domain knowledge, 'domain 
plan' \[Litnmn871. This approach was based on 
Allen & Perrault's plan recognition strategies 
\[AllenS0\], \[PerraultS0\], \[Allen871. By way of 
contrast, a pragmatic-based approach was 
proposed in order to understand intersentential 
elliptical fragments \[Carberry891. She claimed 
Litnmn's strategies could not recognize a 
surprise ors doubt conveyed by an elliptical 
fragment, for example "$10,000 ":",. which is not a 
complete sentence including postulated speech 
acts 'request' or 'inform'. In addition, she also 
claimed that metaplans representing 
communicative goals should be dealt with 
adequately. She newly introduced 'discourse 
expectation' rules for grasping interaction 
between an information-seeker and an 
information-provider and 'discourse goal' rules 
for identifying an information-seeker's 
conversational or communicative goa\[*(l). 
These rules are heuristics for interpreting an 
elliptical fragment which explicitly indicates no 
linguistic clues to interpreting speech act.s. On 
the other hand, in order to make the general 
understanding mechanism clear, a surprise or a 
doubt fragment such as mentioned above, must 
be understood from recognizing an irregular 
meaning from the view of gaps between common 
sense (a certain prerequisite eondition) and the 
declaration. Furthermore, studying how to 
express a query, an answer, a confirmation, a 
surprise or a doubt in a context and a dialogue 
situation is necessary. 
In consideration of making a dialogue 
understanding model including such an 
understanding process, a method to handle 
pragmaties and utterance or dialogue must be 
studied. This l?aj)er shows three typed 
pragmaties use(t 1or cooperative dialogue 
development, as well as a dialogue structure 
analysis and understandinfi model using a plan 
recognition approach. 'lhe pragmatlcs are 
described by a 'dialogue plan', 'communication 
plan' and 'interaction plan'. 
2. THREE TYPED PRAGMATICS 
Three types of universal pragmaties (el. 
LevinsonS3) can be classified and described by 
the following plans: Interaction-Plan - a plan 
basically eharaeterized by a dialogue turn- 
taking*(2) which describes a sequence of 
communicative acts*(3), Communication-Plan - 
a plan which determines how to execute or 
370 
achieve an utterance goal or dialogue goals, and 
Dialogued:qan : a plan for establishing a 
dialogue construction, e.g. a cooperative 
dialogue*(4). 
For example, in order to achieve the goal of 
registering for the conference tim following 
sequer, ce must usually be performed (Domain- 
Plan}: obtain a registration form, fill out the 
form and return it to the secretariat. In such a 
telephone dialo}'ue, if something is needed to 
execute the gore action, a request to send it can 
be made, or it will be offered to you 
(Co:mmunication-t)lan). To complete the 
cooperative infbrnmtion-seeking, the hearer will 
resl)ond to the speaker's request*(5) (Interaction- 
Plan). Beiorc building a whole dialogue 
structure, the speaker should utter the opening 
section of the dialogue, especially on the 
telephone. Furthermore when the dialogue is 
finished, the speaker should wide up the 
dialogue (Dialogue-Plan). 
Each plan is described in terms of a schema 
formulation (plan-schema). 
A plan-schema has various slots to describe both 
an action's inner properties, e.g. HEADER and 
41 r~ ,l • • PI{LI)ICAIE&CASES, and relationsttlps 
between the action and prerequisite 
states/actions, effects, etc., e.g. PREREQUISITE, 
DECOMPOSITIONS, CONSTRAINTS, 
EFFECTS. 
A definite hierarchical order among these 
plans is available as follows; 
Interaction-Plan > Communication-Plan > Domain-Plan > Dialogue-Plan. 
3. DIALOGUE ANALYSIS 
There are several linguistic phenomena 
which are hard to interpret, such as ellipses, 
referring pronouns and substitutional 
expressions. Both information from the 
established context and expectations from the 
current dialogue situation are required to 
resolve such problems. In order to get these 
information a dialogue structure which indicates 
the goal hierarchy of utterances in a dialogue 
must be constructed. A dialogue analysis is 
required and it is necessary to determine ; 
(1) how to infer each goal of an utterance 
within a dialogue, 
(2) how to make clear the relationships 
between goals within the dialogue. 
For the first problem, a plan recognition 
inference method is adequate for identifying an 
utterance intention because the intention can be 
inferred by recognizing the speaker's plan by 
chaining communicative acts regarded as speech 
acts in a specific domain \[Allen80), \[Perrault80\], 
\[Litman 87\]. For the second problem, the 
*(4) :A dialogue global construction usually has an opening 
section and closing section. I lere, such a linguistic 
phenotmnon is regarded as language-universal. 
*(5) :On the other hand, in order to complete the cooperative 
interaction, when the speaker imparts information, the 
hearer will confirm what the speaker has said according to 
the speaker's belief in the hearer's intention. 
*(6) An active chart parser has been developed for 
Japanese dialogue analysis on a unification based grammar, 
which is based on IIPSG and JPSG \[Pollard88), \[Gunji87\], 
\[Kogure89\]. Furthermore, many discourse entities can be 
identified by using NP Identification Method \[Nogaito88\]. 
domain..specific, knowledge is related to the 
action and objects, especially the action 
hierarchy which is used to grasp the utterance ~ 
oal, and three kinds of pragmatics between 
unmns involved in a spoken dialogue are used 
to grasp the dialogue development. 
3A Communicative Acts 
First, in order to recognize the speaker's 
plan, it is necessary to recognize the turn-taking 
patterns. Communicative acts \[Cohen84\] are 
introduced . In Fig.l, an example of 
communicative acts in a cooperative task- 
oriented dialogue, e.g. 'queries and explanations 
regarding registration for an international 
conference' is shown. 
A communicative act in the demand class 
and a corresponding act in the response class 
make a turn-taking pair. This is recognized by 
the interaction plan. A communicative act is a 
decomposition element of an interaction plan. 
Denmnd Class Response Class 
Ask-value Inform-value 
"Whet, is the deadline? .... The deadline for the 
paper is,J une 3." 
Confirm-value . Affirmative 
"The deadline is Jtme 3, isn't it? .... Yes, it is." 
Negative 
"No, it isn't." 
Request-action Accept 
"Will you send me the lbrm? .... OK" 
Reject 
"I'm afraid I can't." 
Give..offer Accept-oiler 
"Shall I send you the tbrm?" "Yes, please" 
Reject.offer 
"No, thank you." 
Eig.1 At) Example oi' Commurticative Acts 
3.2 Inference Mechanism 
'An utterance meaning is represented by 
illocutionary speech act types and propositional 
contents obtained from a tlead-Driven 
unification-based active chart parser*(6) for 
Japanese dialogues. The parsing result is 
described by a feature structure and tim system 
input is modified into a communicative act with 
propositional contents. These consist of a certain 
predicate, an 'utterance type', and some 
variables, in particular, 'speaker', 'hearer', and 
'topic'. 'Topic' is, on a surface level, an NP 
marked with the Japanese special particle, 'wa', 
or the compound particle 'nituite'. 
The plan recognizer (1) assumes a goal. (2) If 
a particular goal cannot be found, then stop else 
goto next (3). (3) Infer a chaining path from an 
input to the goal. If success, stop. Else return to 
the first process (1) in order to try to find the next 
candidate. The chaining process between plans 
generally finds a candidate plan from the 
current state (IIEADER) to an action list 
represented in I)ECOMPOSITtON. Ilowever if 
this fails, the chaining will be continued in 
accordance with PREREQUISITE and EFFECT. 
In order to manage the current 
understanding state, the system uses two slacks. 
UN1)ERSTANDING-IAST stores completed 
plans as the current understanding state, and 
GOAL-IAST maintains incomplete plans 
regarded as possibilities and expectations for 
-2- 371 
future goals . An overview of a dialogue 
structure construction process is shown in Fig.2. 
4. AN EXPERIMENTAL SYSTEM 
This analysis model is realized by using lout" 
typed plans which are categorized into Ulree 
types of universal pragmatics and a type o£ task- 
dependent knowledge related to common action 
hierarchies. The system has been implemented 
in Symbolics Common Lisp. A dialogue structure 
is represented by both completed plans and 
incomplete plans stored in the two stacks. 
Therefore, the system can understand dialogue 
meanings and can offer a dialogue structure 
using the contents of both stacks. Four model 
dialogues regarding an international conference 
registration taken from slightly modified inter- 
keyboard dialogues in Japanese have been 
applied. For example, the system can understand 
a Japanese substitutional expression, e.g. "O- 
isog-i kudasai" (literal translation : 'Hurry 
please') which shows no agent, no object and no 
verb, because the current topic which is focused 
on an action in the domain plan is known in the 
system and the omitted verb (e.g. 'Return-Form') 
can be identified under the scope dominated by 
the topic. 
5. CONCLUSION 
Litman & Allen introduced a set of discourse 
plans, each one corresponding to a particular 
way that an utterance can relate to a discourse 
topic. They distinguish discourse plans from a 
set of domain plans. The dialogue structure 
analysis model basically follows the above idea 
and uses new three typed pragmatics: 
interaction plan, communication plan and 
dialogue plan. By introducing these plans, the 
mech~mism for constructing a dialogue structure 
becomes - clear because of the way a surface 
utterance is connected with both pragmatics and 
the domain-specific knowledge, and by reducing 
the search space using a hierarchical order of 
applying knowledge, computational efficiency is 
improved. 
Acknowledgement 
The authors would like to thank The 
President Dr. Akira Kurematsu and our other 
colleagues at ATR Interpreting Telephony 
Research Laboratories for their encouragement 
and thought-provoking discussions. 
Loi a_~_931o u e st ruct u r_Lg\] 
: D~a~~ Previous utterances 
-- } t~ o-~R'T; MAKE ,RESi~,~ (Interpretation) / IDECOMPOSITION; \[ GET-FOaM J /" spl: I'd like to make a registration for the conference, 
FILL-FOR~a Spl: What should I do7 , 
sph Please send me 
DIRECTION __ G,T-EO~M ,p~.k.rC ..... .°,,,O, ..... ,y~ t.,Oo.T.A~,o.l,,.,,, ,P' 1 ~. ,.r oon  
I TopIc form 
IOemai.n P!l.a n I J.~ADE.: SE.O-EO~ J 
PREREQUISITE: KNOW (ADORESS&NAME) IEFFECT: HAVE -A-FORM J 
T---- 
(Predlctlon$) If_communication Plan i 
IL---~EADER: iNTRODUCE-DOMAiN-PLAN \[interaction Planl -~ 
....... • /i//l//////z/// r" .... "I \[OECOMPOSlI"ION: \[ REQUES\]-ACTION*UNI "--~HEAD£R: REQUEST-ACXION-UNIT\] ~ 
.... L.IFILL'FO~M!.~ (Predkeon$) \[ \[ WILL-OO.ACTION,UNIlJ \]|DEEOMPOSl\]ION; REOUEST.ACTION ~--~ REQUEST-ACTION ~'~ 
: ,- ...... , ~- ...... . l ActEd,, --------Jl ~/,,//,-////(~z. ((./..~ 
I ACHIEVE.KNOW | .... I RETURN-FORM I .... r-----'* : ....................... : (Predictions) J... \[ACCEPT I "': Predl~tln9 candidate : 
; {sp2: All ¢lghl. : 
........................ . 
Fig.2 An Overview of a Dialogue Structure Construction Process 
379, 
3 - 

REFERENCES 

\[Allen 801 Allen, J. F. and PerraulL, C. R. : Analyzing 
Intention in Utterances, 
Artificial Intelligence, Vol.15, pp143-178(1980L 

\[Allen87\] Allen, J. F. : Natural Language 
Understanding, The BeujaminlCummings Publishing Co. 
(1987). 

\[Carberry 891 Carberry, S. : A Pragmatics-Based Approach 
to Ellipsis Resolution, 
Computational Linguistics, Vol.15, No.2, pp75-96(1989). 

\[Carbonel183\]Carbonell, J. : Discourse Pragmatics and 
Ellipsis Resolution in Task-Oriented Natural Language 
Interfaces, 21st Annual Meeting of the ACL,pp164-168(1983). 

\[Cohen 84\] Cohen, P. R. : The Pragmatics or Referring and 
Modality of Communication, Computational Linguistics, 
Vol.10, No.3, pp97-146(1984L 

\[ttendrix 781 Hendrix, G. G. et al. : Developing a Natural 
Language Interface tx) Complex Data, ACM Trans, Vol.3, No.2, pp 105-147(i 978). 

\[Gunji87\] Gunji, T, : Japanese Phrase Structure 
Grammar, Dordrecht, D. Reidel (1987). 

\[lida 891 lida, II. eL al. : An Experimental Spoken Natural 
Dialogue Translation System Using a Lexicon-I)riven 
Grammar, European Conference on Speech Communications 
and Technology (1989). 

\[Kogure891 Kogure, K. : Parsing Japanese Spoken 
Sentences Based eta HPSG, 
the hit. Workshop on Parsing Technologies (1989). 

\[Levinson 83\] Levinson, S.' C. : Pragmatics, Cambridge 
University Press (1983). 

\[l,itman871 Litman, D: J. and Allen, J. F. : A Plata 
Recognition Model for Subdialogues in Conversations, 
Cognitive Science, Vol.i 1, pp163-200 (1987J. 

\[Nogaito881 Nogaito, I. and lida, ti. : Noun l'hrase 
Identification iu Dialogue and its Application, 2nd 
International Conf. on Theoretical and Methodological Issues in Machine Translation ofNatural I,anguage (1988L 

\[Perrault 801 Perrault, ~. R. and Allen, J. F. : A Plan-Based 
Analysis of Indirect Speecia Acts, Computational Linguistics, 
Vol.6, pp166-182(1980). 

\[Polard891 Pollard, C. and Sag, I. : lnfornaation-Based 
Syntax and Semantics - Vol.l Fundamentals, CSLI Lecture 
Notes, No.13 (I988). 

\[Shimazu79\] Shimazu, A. and lida, If. : Experimental 
Japanese l,anguage Question Answering System MSSS78, 
6th International Joint Conf. on AI, pp803-808 (1979). 
