American Journal of Comprttationd Lin~tticr 
Microfjchc $7 
A GOAL-ORIENTED MODEL OF N DALOGE 
Jms A. Moore 
James; A* hsvin 
Wittiam C. Maw 
USCflnf ormation Scicncc~ lnclitde 
4676 Admiralty Way 
Marina del Rey, Cdifornia 9029 1 
Thc research reported hereto was gupporlrd by the Porronnel md Tra,in,trtg P~tr:lrr 11 
Programo of thc Olfica of Navel R~crnch, Contrltl N00014-75-C- 07 IQ, undr~ ~P!@A 
Ordcr Number 2930 from tho Cybarml,~, Tcchnclogy Qff~~cc cf th. ~dv~ncrd P~i~rartbt 
Projecto Agency. 
A Modcl of Dialogue 
SUMMARY 
Within a view of ianguagc users as problem solvers, speakers arc seen as creating 
uttcrsnccs in pursuit of thcir own goals. Dialogue "works" because this activity tends to 
scrvc goals of both participants. ,k Hcncc, for the model of dialogue cornprchcnsion 
prcscnted here, rccognifion of these goals of the speaker is central to the comprehension 
4" 
of dialo~uc, 
Wc have found that dialogues arc conrposud of structured interactions rcpresanted 
by collections of knowledge which dcscri be the intcrrelatcd goals of the participants. 
We call thcsc knawlcd~c structures "Dialogue-games" (DGg). This paper describes DGs 
in gcncral, a particular one (the Hclping-DG) in some detail, how OGs are used by our 
Dialog~c-game Model (DGM), and the benefits of this model. 
A DG consists of three parts: the Parameters (the two roles fillcd by thc 
participants, and the topic), the Pnramcter Spccifications (a set of predicates on the 
Pararndcrs), and the Components (a sequcncc of goals hgld by the participants in the 
course of the dialogue). 
For cxample, in the Helping-DG, the Parameters are HELPER, HELPEE (the rolcs) 
and TASK (the topic). The Spccifications are: 1) The HELPEE wants to pertorm the 
TASK: 2) thc HELPEE wants to bc a6/e to do it but 3) the HELPEE is not able to. 4) The 
ILELPER wants to enable the HELPEE to do the TASK and 5) the HELPER is ab/e to 
provide this help. The Componcntr specify that 1) the HELPEE wants to establish a 
context by describing a collection of unexceptional events (a parlial performance of the 
TASK): 2) he also wants to dcscribc some sort of unresirable surprise: then 3) the. 
I-IELPER wants to cxplain the violation of expectation so that the HELPEE can avoid it a'nd 
get on with the TASK. 
The DGM makes use of DGs in fiue stages of processing: Nomination, Recognition, 
Instantiation, Conduct and Termination. 
Thc DGM models each participant's knowlcd,ge, goal and attention states. A 
mcchanisrn adds to the attention state, cmcepts "suggested" by those already in attention. 
Whcn a hearer sees himself or his partner as potentially filling a role in a DG(by fulfilling 
one or more demands of the DG's Specifications) then that DG is brought into attention 
(Nominated), 
DGs can bc nominated by weak evidence: Recognition is the step of verifying that 
these DGs arc plausibly consistent with the currcnt state of the model. Thosa which arc 
not are eliminated from attention. 
A Model of Dialogue 
DGs which survive Lhc Recognition stagp arc Instantiated by asserting (as 
assumptions) all the Specifications not ye1 rcprcscnicd as holding. For example, when a 
person says "Do you have a match?", instantiation, by the hcarer (of the-~ction-sck DG) 
dcrivcs asscrtionr; that thc speaker does not have a match and wants the hcarer to give 
him one, 
The Conduct of ihc DG is rnodclcd by tracking the pursuit and fulfillment of the 
participants' goals as rcprescntcd in the Components. 
Whcn tho DGM dctcpls that onc parlicipant no longer regards a Specification as 
holding, this crcatcs an cxpcctation of the Termination of this phase of the 
dialogue--there is no longer a possibility that it will serve both participants' goals, 
Thc appendix contains a detailed hand-simulation of the DGM assimilating a 
scgrncnt of a dialogue. 
A Modcl of Dialogue 
CONTENTS 
2. St~tcrrrcnt of thc Problam 
3. Post Rcscarch on Language Comprchcnsion 
4. Thc Shopc of Ihc Thcory 
5. Thc Dialogue-came Modcl 
5,l What's in a Game? 
5.1.1 Paramctcrs 
5.1.2 Paramctcr Specifications 
5.1.3 Components 
5.1.4 Bidding and Accepting 
5.2 The Helping-gamc, an Examplc 
5.3 Dialogue-gamcs in the Comprchcnsi~n cf Cialcd~e 
5.3.1 Proccssine Environment 
5.3.2 Nomination 
5.3.3 Rccogni tion 
53.4 Instantiation 
A Modol of Dialogue 
5.3.5 Condtrct 
5.3.6 Tcrmination 
5.4 The Dialogue-game Proccsscs 
5.4.1 Long-term Memory (LTM) 
5.4.2 Workspace (WS) 
5.4.3 Parser 
5.4.4 Protcus 
5.4.5 Match 
5.4.6 Deduce 
5.4.7 Dialogue-game Manager 
5.4.8 Pronoun Processes 
6. Deficiencies in Current Man-machine Communication 
Appendix -- Sirnulslion of the Uiaip~;ue-l;nmu Model 
A Modcl of Dialogue 
STATEMENT OF THE PROBLEM 
Thc broadest goal of our research has been to improve the sorry state of interactive 
man-machine communication, including its appearance of complexity, rigidity, lack of 
continuity and the difficulty it poses for many people to acquire useful levels of 
cornpctcnce. In our pursuit of this goal, we have adopted-the following two assumptions: 
Assumption L: When pcople communicate with machines, they do so by 
using their already well-dcvelopcd ability to communicate with other people. 
Assumption 2: The effectiveness of this communication is diminished by any 
adaptation required of the human. 
A scientific understanding of how people cvmmunicate is thus relevant to the design 
of man-machine communication schemes, but suchknowledge is seldom used in the design 
process. Since human communication skills have not been characterized at a level of 
clctsil appropriate for guiding design, interface designers have not been able to take into 
account some major determinants of their succcss. 
The opcrativc goal of our research was therefore to creatd a mode/ of human 
communicafion at an appropriate level of detail to benefit man-machine communication 
design. Any form of communication must be based on ,knowledge shared by the 
individuals engazed in that communication. However, the nature of this shared 
knowladgs and how is it uscd in the communicative process have not been well 
undcrstood. We have developed a working hypothesis which has deeply affected the 
r-csearch: 
Hypothesis: People know that certain kinds of goals may be pursued by 
communication, and they know which kinds of communication acts correspond 
to which goals. Tho use of this knowledge is essential to comprehending 
dialogue. 
In particular, a pcrson generates an utterance to advance one or more of his own goals. 
Thus, to assimilate a particular utterance, it is necessary to identify why the person said 
i't. 
Working with this hypothesis, we have conducted three related investigations: 
A Model of Dialogue. 
1. A study of naturally occurring language to discover regularities of usage 
and to determine what these regularities mean lo the users of the language. 
2. The representation of there regularities as knowledge structures and 
processes in a dialogue model. 
3. The establishment of standards by which the model's performance can 
be compared with that of humans on closely related tasks. 
We have adopted two additional, tactical constraints on the task: 
1. We have modeled only the receptive aspects of communication. 
2. Wc-have examined only dialquc communication, interaction in real-time, 
by exactly two people. These dialogues were conducted over a rcsttictcd 
medium so ihat there was no visual or intonational communication not captured 
in thc tr~nscript. 
A ~odcl of Dialogue 
PAST RESEARCH ON LANGUAGE COMPREHENSION 
Most of the research jnto language comprehension has focused on the 
comprchcnsion of single sentences or fragments of sentences. However some research 
has indicated the importance of the context created by surrounding sentences on the 
comprchcnsion of an individual sentence. One specific model for the form of this 
multi-scntcntial knowledge is the "story schema", organized within a story grammar 
(Rurnclhart, 1975). This model has been supportbd' by the results 6f story recalls 
(Rumcl hart, 1975: Thorndyke, 1977). Other similar kinds of theoretical constructs for 
organizing multiple srntcnces of stories have been proposed called: "frames" (Minsky, 
1975: Charniak, 1975), "scripts" (Schank Q Abelson, 1975), and "commonsense 
algorithms" (Ricgcr, 1975). 
To account for the conduct and comprchcnsion of dialogues, mu1 ti-scntcntial 
knowlcdgc units have also bccn proposed by linguists and sociolinguists to explain 
certain kinds of rcgul ari tics observed in naturally occurring dialogues. These 
rcgularitics have bccn called "rules" by Labov & Fanshel (19.74) and "sequences" by 
Sacks, Schagloff, & Jefferson (1974). 
Once these multi-scntential knowledge units are evoked, they serve as a basis for 
comprehending the~successive inputs. This is achieved by generating expectations and 
by providing a framework for integrating the comprehcnsion of an utterance with that of 
its prcdcccssors. Recently, we have propased (Leuin & Moore, 1976: 1977, Mann, 
Moorc Rr Lcvin, 1977) multi-scntential knowledge units that are specified primarily by 
the speaker's and hcarcr's goals. Thcsc goal-oriented units, which we call 
Dialogue-gsmcs[l], specify the kinds of language interactions in which people engage, 
rathcr than the spccific content of thcsc intcractions. Pcoplc use langua~c primarily to 
comrnunicatc with other pcoplc lo achieve their own goals. Thc Dialoguc-game 
mu1 ti-scntontial structures wcrc dcvcloped to represent this knowledge about language 
and how it can be uscd to achicve goals. 
-[I] Thc term "Oialoguc-game" was adopted by analogy from Wittgcnstcin's term 
'Yan~uagc game" (~itt~cnstcin, 19 8). a Howcvcr, Dialogue-games reprcs6ht knowlcdgo 
?= 
pcoplc hive dbout language as uscd to purGuo goals, rathcr than Wittgcnrtcin's mop 
~cncral. notion. Althoueh other "game-." arc similar, thc propcrtios of .Dia)ofiuo-,gornes 
arc only thooo dcscribod here. For example, thoy aro not nacoroarify cornpctitivc, 
consciouily pursued, or zcro-sum. 
A Model of Dialogue 
An important problem for rcscarchcrs of language comprehension is posed by 
scntences with which tho speaker performs what philosophers of language have called 
"indirect spccch acts" (Searle, 1969). The direct comprehension of these sentences 
fails to derive the main communicative~effcct. For example, declarative scntenccs can be 
used to seek information ("1 nced to know your Social Security number."): questions can be 
u-scd to convey information ("Did you know that John and Harriet got married?") or to 
request an action ("Could you pass the salt?''). These kinds of utterances, which have 
bccn extensively analyzed by philosophers of language (Austin, 1962: Searle, 196 9, 
1975: Grice, 1975), are not handled satisfactorily by any of the current theories of the 
dircct comprchcnsion of language. However, these indirect language usages ara 
widespread in naturally occurring language--even two-year-old children can 
comprehend indirect requests for action almost as well as dircct requests (Shatz, 1975). 
One theory proposcd to account for these indirect uses of language is based on the 
concept of "convcrsotional postulates" (Grice, 1975: Gordon Q Lakoff, 197 1). If the 
dircct comprchcnsion of an utterance is implausible, then the indirect meaning is derived 
using these postulates. Clark & Lucy (1 975) formalized and tested this model, and found 
that people's rasponse times tend to support a three-stage model (deriving the literal 
mcaning, check its plausibility and, if implausible, dcriving the "intended" meaning" from 
convcrsational rules). 
In general, this approach to indired speech acts is infc~ence-bascd, depending on 
the application of conversational rules to infer the indirect meaning from the dircct 
mcaning and the context. A different approach has been proposcd by Lsbov~fiFans~cl 
(1 974) and by Levin & Moore (1976: 1977). Multi-sentential knowledge, organizing a 
scgmcnt of language interaction, can form the basis for deriving the indikect effect of 
uttcr~ncc within the segment. For example, a multi-sentential structure for an 
information-seeking interaction can sypply the appropriate context for interpreting the 
subscqucnt utterances to s~ck and t-hen supply information. The infcrcncc-bascd 
approach rcquircs one set of convcrs~tional rulc-, for information requests, a dif fcrcnt 9ct 
of rulcs for answers to these rcquc:ts, and a way to tic thcnc two rulc sets together. The 
Dialogue-game model postulates a single knowlcd~e struclurc for this kind of interaction, 
with coopcrating proccssc; for: (1) rccognizinp; when this kind of interaction is proposcd, 
(2) using this knowlcdgc to comprchcnd uttcranccn within its scope, and (3) identifying 
when the interaction is to be terminated, 
A Modcl of Dialogue 
THE SHAPE OF THE THEORY 
Our thcory of human language use has bccn strongly influenced by work in human 
problem solving (Ncwcll XI Simon, 1972) in which the bchavior of a human is modeled as 
an information. processing system, having goals to pursue and selecting actions which 
tend to schicvc thcsc goals. Wc view humans as engaging in linguistic bchavior in order 
to advance the state of certain of thcir eoals. Thcy dccide to use language, they sclcct 
(or accept) thc other participant for a dialogue, they choose the details of linguistic 
cxprcssion -- all with the expectation that some of their desired state specifications can 
thcrcby be rcalizcd, 
In this thcory of lancuagc, a participant in a linguistic exchange views the other as 
an indcpcndcnt information-processing system, with separate knowledge, goals, abilities 
and acccss lo the world. A spcsker has a range of potcntial changes he can cffcct in his 
li~tcncr, a corresponding collection of linguistic actions which may result in each such 
chance, and some notion of the conscqucnccs of performing each of these. The spcokcr 
may view the hcarcr as a resource for information, a potential actor, or as an object to bc 
moldcd into sorrrc dcsircd state. 
A dialogue involves two speakers, who altcrnatc as hearers. In choosing to initiate 
or conlinuc tho cxchany,~, a participant attcmpts to satisfy his own goals: in intcrprcting 
on uttcrancc of his partner, each participant attcmpts to find the way in which that 
utterance serves the goals of his partner. Thus a dialoguo continues because the 
participants continue to scc it as furthering thcir own goals. Likewise, when the dialoguc 
no lon~or serves the goals of one of the participants, it is redirected to new goals or 
tcrminatcd. 
this rrlcchanism of joint interaction, via cxchange of uttcranccs, in pursuit of dcsircd 
t.itcs, is uscful for ochiovihg ccrtain relatcd pairs +of participanls' ~oals (c.~., 
Itv~rni rlr./tcact~inc, buyinc/sc\ling, gctting hcip/~iving hclp, ...). Many of thcsc paired sets 
of I correspond to hichly structured collections of knowlcdgc, shorcd by thc 
rncrnbcrr, of thc langunpc community. Thcsc ~ollcctions specify such things as: 1) what 
chnractcri:tics an individual must havc to cngagc in a dialogue of this sort, 2) how this 
dialocuc is initiated, pursued and tcrminatcd, 3) what ranee .of infarmation can bo 
comrnunicotcd imp1 icitl y, and 4) undcr what Circumstances tho dialoguo will "succeed" 
(scrvc tho function for which it was initiated) and how this &ill bo cxhibitcd in the 
participants7 bck~uvior. 
VJo h~~o allr:mptc:d to rrproscnt those collr:.ctiong_of knowlcdgc and tho wily in 
wt~ich thcy arc u:cd to facilit~to tho cornprchcnsion of a diala~ua, in tha Oinloe~~o t;nrno 
Mod(-? I. 
A Model of Dialogue 
THE DIALOGUE-GAME MODEL 
This section describes our Dialogue-game Model at its current state of 
dcvclopmcnt. It starts with a brdef overview of dialogue and how it is structured, then 
describes the dominant knowledge structures which guide the model, and finally 
dcscribcs a set of processes which apply these knowledge structures to text to 
comprehend it 
Within the mb.dcl., each participant in a dialogue is simply pursuing his own goals of 
thc moment. The two participants interact smoothly because the conventions of 
communication coordinate their goals and give them continuihg reasons to speak and 
listen. These goals have a number of attributes which are not necessarily consequences 
of cithcr human activity in general, or communication in particula'r; but which are 
nonetheless characteristic of human communication in the form of dialogue: 
1. Goals are cooperatively esta5lished. Bidding and acceptance 
activities serve to intfoduce goals. 
2. Goa/s.aremufua//yknown. Eachpartyassumesorcomestp 
know goals of the othcr, and each interprets the entire dialogue relative to 
currently known goals. 
3. Goalsareconfieu~edbyconvenfion. Setsofgoalsforusein 
dialogue (and othcr lwguage use as well) are tacitly'known and employed by 
all competent spe;l&rs ofthe language. 
4. Goa/s are bilateral. Each dialogue participant assumes goals 
complementary to those of his partner. 
5. Gas/ssreubiguilous. Ahoarerviewsthcspe~kerasalways 
having goals hc is pursuing by speaking. Furthermore, the hearer recognizes 
and uses thcsc goals as part of his understanding of the utterance. 
An ~ninlerrupted dialogue goes through three phases: 
establishing goals, 
pcir sui ng eobl s, 
dccommitting from goals. 
A Modcl of Dialogue 
Typically this sequcncc is repeated several times over the coursc of a few rninutcs. 
We havc crcotcd knowlcdse structurcs to rcprescnt these convcntions,, and 
proccsscs to apply the conventions to actual dialo~ucs to comprehend them; Since the 
knowledsc structures dominatc all of the activity, they are described first. The 
assimilation of an uttcranco in the dialogue is rcprcscntcd in this model by a sequence of 
modifica\ions of a "Work~pacc"[2] which rcprcscnfs the attention or awareness af the 
listening party. Tho modificN~tions arc roughly cyclic: 
1. A ncw item of text f is brought into attention through the 
"Par scr."[-21 
2. Interpretive conscqucnces~of T are developed in the Workspace by 
a variety of proccsscs. 
3. An exprcssian E appears in thc Wor'kspace which specifics the 
relation between i and the imputed goals of the spcaker of T. 
This final cxprcssion is of coursc a formal expression in the knowledge 
representation of the modcl. E rcprcsents the proposition (held by the hcarer) that in 
uttering T, the spcaker was performing an act in pursuit of G, a-spbaker's goal known to 
thc hcarer. Sucrcssful comprchcnsion is cquatcd with relating tcxt to salisf action of 
spcakcr's goals. 
To makc an explicit account of dialoguc in this way, wc now describc the knowledge 
structures that rcprcscnt those c~~nvcgtions which supply tho goals for the participants to 
pursue. In particular, wc will anewcr thc following thrco questions: 
1. What is thc knowlcd~o wo arc rcprescnting within tho dofinition of o 
pclrlicular Dialogue-gamc7 
2, How is this knowledge? used to modcl tho roccptive acts of dialogue 
participant8 
-----e--*-e-----e--'-m-dew- 
[2]'Thc F3ar.;cr and thc Workspzlco oro port% of tho procarg model and aro der,~r~t~i!d in a 
later sceti,~n. 
k Model of Dialogue 
3. What*sort of processes docs it take to support this model? 
A Dialogue-game consists of thrcc? parts: a set af Parameters, a collection of 
Spco/~c,~ll'ons that apply to these Paramctcrs throughout the conduct of the game, 
and a partially ordcrcd set of Components characterizing the dynamic aspects of the 
came. For the bslancc of this section, we will elaborate on these three parts and 
cxcmplify these with an cxalliple of the Helping-game. 
Dialo~ue-games capture a certain collection of inforhation, common across many 
dialv~ucs. However, the individual participants involved and the content subject df the 
dialoguc may yary freely over dialogues described by the same Dialogue-game. To 
represent this, each Dialogue-game has a set of Parameters which assume specific values 
for each particular dialogue. 
Thc dialogue types wc have represented so far as Dialogue-games have each 
required only thrcc Parameters: the two participants involved (called "Roles"), and the 
subjcct of the dialogue (called "Topic"). 
Pararndcr Spccifications 
Onc of the major aspects distinguishing various types of d'ialogucs is the set of goals 
hcld by the participants. Another such aspect is the set of kno'wledgc states of the 
participants. We have found that each type of dialogue has a char$cteristic set of eaal 
and knowledge states of the participants, vis-a-vis each other and the subject. Within 
the formalism of the Dialogue-game, these are called the Parameter Spccifications, and 
are rcprpscnted by a collection of predicates on the Parameters. 
These Spccifications are known to the participants of the dialogue, and the 
requirement that they be satisfied during the conduct of a game is used by thp participants 
to signal what Dialogue-games they wish to conduct, to recognize what game is being bid, 
to dccide how to respond to a bid, to conduct the game once the bid is accepted, and to 
terminate the garno when appropriate. These Spccifications also provide the means with 
which to explain the implicit, but clearly succcssful, communication which accompanics 
any natural dialog~c. Examples and discussions of these Specifications will accompany 
the following description of !he Helping-game. 
Componcnts 
While the Paramctsr Spccificstions represent those aspects of a dialogue type that 
rcmain constant throughout the course of a dialogue of that type, we have also found that 
certain aspects change in systematic ways. Thcse are reprcsented in Dialogue-games a5 
Components. In the Dialogue-games we have developed so far, the Components are 
rcprcscntcd 9s a set of participants' subgoals, partially ordered in time. 
Bidding and Accepting 
Eiddinp, and Acccptancc arc entry operations which people use to cntcr 
Dialo~uc-games. Bidding 
1. identifies thc game, 
2. indicates the bidder s interest in pursuing t'hc game, 
3. idcn tifies the Psramctcr configuration intcnded. 
Bidding is performed many diffcrcnt ways, often very bricfly. It is typically the 
source of a great deal of implicit comrnunicotion, since a bricf bid can cornmunicatc all of 
the Pararncters and thcir Specifications forjhe Dialogue-game being bid. 
Acceptance is one of tho typical responses to a Bid, and leads to purwit of thc game. 
Acccptsncc cxhibi t:. 
1. acknov~lcdy,rncnt that 3 bid t~ss hccn rnndc, 
2. rscognitisn of thc particul;~r OiaIo~;uc-garrlr? ;~nd 13aramc:tcr~ liid, 
3. 
ar,re.crn~nt to pursuc the l;;~trio, 
4. 
assumption of thc A~ccptor'; rolfc in It)(: Ui;~lol;~~t: 1:;jrtlc:. 
Acccptoncc is o1tc.n implicit, c peci.~Il y in I c1Iii!i'vtbly inforrn;ll dial oj;lrrl. c:an bc 
indi~atcd by stat(:rncnk of clgr(:(:nir:r~t or ;~pprov;ll, or hy br:l;inn~rl~ to ptrr;iJcl thc 1;hrnc. 
(i, sttcrnpts to satisfy it-ic goals). A to clcc(~ptar~rc inrlt~do rcjr:cting, 
nc~otiatinr, and i~norinc. 
Ciddin~ and acccptnricc appear to ljn pi~rt' of 1;arnc r:c~lry for ;dl of thc 
Dial oguc- games of ord~nary ad~~lt dislu;.uc. Ti~y are .il:n i nv'olvcd in f;amc trtrrni n:~Iian. 
In thc ca:r of t~rmindion, thrce altcrnat~vc.: ;rrfa po:<iblc: ~rllrlrruptiun :~nd 5pont:lncous 
torminotian by ei thcr 1;oal x~lisfclcIior\ or uncotl(llI~on:~I i;o:~l f:~iI~~rcr. 
A Model of Dialogue 
Once a eamc has bccn bid and acccptcd, the two participants each pursue the 
subgoals spccificd for their role by the Components of this game. Thcse subgoals are 
mutually complcn~cntary, each set facilitating the other. Furthermore, by the time the 
tcrrninati on stage has been rcachcd, pursuit of the Component-specified subgoals will 
have assurcd satisfaction of the higher, initial goals of the participants, for which the 
Came was initiated in thc first place. 
In this scction, wc c.xhibit a specific Diolocuc-game: the Helping - gn/ne. This 
game is prcscntcd in an informal rcprcsentation, in order to emphasize the informational 
content, rather than the representational power of our formalism. Later in this report we 
will prcscnt thc formal analocue of this same game. In what follows, the bold face 
indicates the information contained in tho representation of this particular Dialogue-game: 
the tcxt in regular type is explanatory commentary. 
The (annotated) Helping-game. 
--------1---11---------------- 
Pnmn~clcrs: HELPEE, HELPER, and TASK. 
The HELPEE wants help from the HELPEE. The TASK is 
some sort of a problem, otherwise unspecified. 
Paran~eler Specificat ions: 
HELPEE: wants to perform TASK. 
HELPFE: wants lo be able lo perform TASK. 
HELPEE: not ab/e lo perform TASK. 
HELPEE: permitted tii perfirm TASK. 
MELPFF: a person. 
A Modcl of Dialogue 
Thcse Spccifications not only constrain who would qualify 
as filling the rolc of HELPEE, but also provide reliable information 
about the HELPEE, given that this individual is believed to be 
engaged in the Helping-game. This prohibits someone from 
asking for help on a problem he did not want solved. Similarly, if 
one rcc~ivcs what he judges to be a sincere request for help to do 
some task, the helper normally as:umes that the requester has the 
necessary authority to do the task, if only he knew how. 
HELPER: wants to help HELPEE perform TASK. 
HELPER: able to provide help. 
HELPER: a person. 
So, in ordcr to be a HELPER, an individual must be willing and 
able to provide the needed assistance. Since this Dialogue-game 
rcprcscnts shared knowledge, the HELPER knows these 
Spccifications, and therefore will not bid the Helping-game to 
someone who is not likely to meet them. And similarly, no one 
who fails to meet these Specifications (and knows he fails) will 
accept a bid for the Helping-game with himself as HELPER. 
Components of the Helping - game: 
Thcre are three components: the first two constitute the 
"Diagnosis" phase to communicate what the problem is. 
. HELPEE wants HEfPFR to know about a sef of unexcepfiona/, acfuuii/ 
events. 
'The HELPEE sets up a context by describing a situation 
where everything, so far, is going well. Since the HELPEE 
assumes that the TASK is understood by the HELPER, he also 
assumes that the HELPER shares his expectations for r~bsequent 
activity. 
A Modcl of Dialogue 
2. HELPEEwc.mts HELPERfoknowabouf: 
// a set of exceptional events whkh occurred 
or 
2) a set of expected, unexcepflbnal events which did not occur. 
This pattern of a Helping-game is sufficiently well known to 
the participants, that the HELPEE almost never needs to actually 
ask a question at this point. By simply exhibiting a failure of 
expectation, the HELPEE has communicated that this acts as a 
block to his successfully pursuing the TASK. The HELPER is 
expected to explain why the failure occurred and how HELPEE can- 
avoid it or otherwise continue in the TASK. 
The third componcnt specifies the 'Treatment" phase where 
the HELPER communicates an explanation for the perceived 
failure. 
3 HELPER wants HELPEE fo know about an action which will avoid the 
undesired event or cause the desired one. 
The context description enables the HELPEE to identify a 
collection of activities which he understands, and in which the 
HELPEE is attempting to participate. The 
violation-of-expectation description points out just where the 
HELPEE's image of the activities differs from the correct image. 
It 
is from this area of difference that the HELPER selects an action 
for the HELPEE. 
A Model of Dialogue 
#ia/o,mue - games in the Con?pre/rension of Dialogue 
In this section we describe the five stages of dialogue assimilation and detail the 
involvement of Dialogue-games with ~ach stage: 
1) nomination, 
2) recognition, 
3) instantiation, 
4) conduct, 
5) termination. 
Proccssi ng Environment 
Our description of the model should be viewed as representing the changing 
coy,nitive state of one of the participants, throunhout the course of the dialogue. That is, 
two models are involved, one for each participant. Since the same processing 
occurs for both, we will describe only one. 
Thc Dialogue-Game Modcl consists of a Long-Term Memory (LTM), a Workspacc 
(WS), and a set of proccsscs that modify the contents of WS, contingent upon the contents 
of LTM and WS. LTM conbins a rcprescntation of the knowledge that the partigular 
di ologuc participant bl ines to the dialogue bcfo~e it starts. This includcs knowlcdgc 
about the world, relevant objects, processes, concepts, the cognitive statc of his partner 
in dialogue, rules of inference and evidence, as well as linguistic knowlcdp;e (words and 
thcir semantic rcprcscntation, case frames for verbs and predicates and the multi-turn 
language s'trbctures, the Dialogue-games). 
WS is the volatile short-term rncmory of thc modcl, containing all the partial and 
temporary rcsults of processing. The contcnte of WS at any momcnt rcprc.;cnt thc 
madel's state of comprchcnsion and focus at that point. Tho processes arc autonomous 
specialists, opcrl~tiny: indcpcndcntly and in parallel, to modify thqentitics in WS (callcd 
"activations"). Thcsc proccsscs arc also influcnccd by the contents of WS, as well a.; by 
thc knowlcdgc in LTM. Thus, WS is the place in which thcso concurrcnt~ly operating 
proccsscs interact with each othcr. This anarchistic control structure rcscrnblcs that 
the HEARSAY system (Erman, Fennel, Lesser & Rcddy, 1973) 
A Modcl of Dialogue 
Nomination 
When dialo~uc participants propose a new type of interaction, they do not 
consistently use any single word or phrase to introducc the interaction. Thus we cannot 
dcterminc which Dialogue-games represent the dialoguc type through a simple 
invocation by namc or any othcr pre-known collection of words or phrases. Instcad the 
diolo~uc type is cornmunicatcd by attempts to establish various entities as the values of 
the Psi actcrs of ihe dcsircd Dialogue-game. Thus, an utterance which is 
cornprchcndcd as associating an entity (a-person or a concept) with a Parameter of a 
Di aloguc-game suggests that Dialogue-game as a possi bilily for initiation. 
The Dialogue-Game Modcl has two ways in which these nominations of ncw 
Dialo~uc-games occur. One of the processes of the modcl is a "spreading activation" 
proccss call& Protcus (Lcvin, 1976). Protcus gcncratcs new activations in WS on thc 
basic of cclations in LTM, from concepts (nodes in the semantic network) that are already 
in WS. Protcus brings into focus concepts somehow related to those already thcrc. A 
collection of concepts in WS leads to focusing on some aspect of a particular 
Dialocuc-game, in this sense "nominating" it as a possible new Dialoeue-game. 
MATCH and DEDUCE are two of thc modcl s processes which operatc in .conjunction 
to ccncratc ncw activations from existing ones, means of finding and ap,prp@rulc-like 
transformations, Thcy operate through partial match and plausible i nfcrencetcchniques, 
and if thcy activate Pardrnetcrs, thcn the Dialogue-gsrnc that contains those ~ararndtcrs 
bccomcs nomina,tcd as.a candidate Dialogue-game. Match and Deduce operate to,gether 
as a kind of production system (Newell, 1973). 
For cx;lmplc, from the input utterance: 
"I tried to send a message to <person> at <computer-site3 and it~didn't go." 
the following two scqucnccs of associations and inferences result: 
(la) I trjcd toX. 
(25) 1 wpntcd to X. 
(3a) 1'want to X. 
(4a) HELPEE wants to do TASK. 
( I b) It didn't go. 
(2b) What I tried to do didn't work. 
(3b) X didn't work. 
(4b) I can't X. 
(58) 1 don't know ha to X. 
(6b) HEL.PEE\doc$n7t know how to do TASK. 
A Model of Dialogue 
(Where: I = HELPEE and X = do TASK = send a message to <person> at <computer-site>.) 
At this point, (45) and (6b), since they are both Parameter Specifications for the 
Helping-game, cause the model to focus on this Dialogue-game, in effect nominating it as 
an organizing structure for the dialogue being initiated. 
Thc proccsscs described so far are reasonably unselective and may activate a 
number of possible Dialogue-eamcs, some of which may be mutually incompatible or 
othcrwisc inappropriate. The Dialogue-garnc Manager investigates each of the 
nomi natcd Dial.oguc-games, verifying infcrcnccs based on the Parameter Specifications, 
and eliminating< those Dialogue-gamcs for which one or more Specifications are 
contraclictcd. 
A second rncchanism (part of Protcus) identifies those activations which are 
incornpati blc and scts about accumulating evidence in support of a decision to accept one 
and dclctc the rest from the WS. 
Fdr cxarnplc, suppose the question 
"How do I get RUNOFF to work?" 
lcads to the nomination of two games: 
Info-scck-game (pcrson asking question wants to know answer) 
and 
Info-probe-game (pcrson asking qucdi~n wants to kn~w if other knows answcr) 
Thcso Iwo Dialocuc-~nrncs have a lot in common hut differ in ono crucial aspect,: In the 
Info-scck-gamc, thc qucctioncr docs not know the answcr to thc question, whilc in thc 
Info-probc-game hc doc:. Thcsc two prcdicatcs arc rcprcscntcd in the Parameter 
Spc.cifications of t hc two Di al.ogue-games, and upon thcir joint nomination are discovarcd 
to bc contradictory. Prolcus rcprcsent: this di~covery with a structure which ha5 the 
cffcct of climinatinp, thc conflicting Dialogue-came with the least supporting evidence. 
Such support might be, for cxamplc, cither the knowledge that the speaker is th~ hearcr's 
tcachcr or that hc is a novice prograrnrncr (which would icnd support f6r tho choicc of the 
Info-prabc-errme or Info-seek-garnc, respectively). 
A Model of, Dialogue 21 
Through these proccsscs, the number of candidate Dialogue-games is reduced until 
those remaining are rompatible with each other and with the knowledge currently in WS 
and in LTM. 
Instantiation 
Oncc a proposed Dialogue-game has successfully survived the filtering proce.Bses 
describe-d above, it is thcn instantiated by the Dialogue-game Manager. Those 
Parameter Specifications not prcviou:ly known (represented in the WS) are established 
as newly infcrred knowledge about the Parameters. A large part of the implicit 
communication between dialogue participants is modeled through instantiation. 
To illustrate this, suppose that the following come to be represented in WS (i.e., 
known) in the course of assimilating an utterance: 
SPEAKER does not know how to do a TASK. 
SPEA~ER wants to know how to do that TASK. 
SPEAKER wants t-o do the TASK* 
Thcso are adequate to nominate the Helping-game. In the process of instantiating this 
Dialogue-game, the following predicates are added to WS: 
SPEAKER believes HEARER knows how to do TASK, 
SPEAKER believes HEARER is able to tell him how to do TASK. 
SPEAKER believes HEARER is willing to tell him how to do TASK.. 
SPEAKER wants HEARER to tell hirn.how to do TASK. 
SPEAKER expects HEARER to t'cll'him how to do TASK. 
Thc model predicts that jhcsc predicates will bo implicitly communicated by an 
utterance which r;uccecds in instantiating thc Helping-game. This corresponds to a 
dialogue in which "&I can't gct this thing to work" io taken to eommuhicate thot'thc speaker 
wnnts to "get this thing to work" (even,though, on the surface, it is only a simple 
declarative of the speaker9$ abi'lity). 
Conduct 
Oncc a Dialogue-game is instantiated, tho Dialogue-game Managcr ic guided by thc 
Corhponcnts in cornprohcnding tho rect of the dialogue. Thars Component$ ore gook for 
A Model of Dialogue 
the dialogue participants. For the speaker, thesegoals guide what he is next to say: for 
thc hcarcr, these provide expectations for the functions to be served by the speaker s 
subscqucnt utterances. 
Thcse "tactical" goals are central to our theory of language: an utterance is not 
dccmcd ta be comprehended until some direct consequence of it is seen as serving a goal 
imputed to the spcakcr Furtherfiore, although the goals of the ~omponcnis arc active 
only within the conduct ef a particular game, their pursuit leads to the satisfaction of the 
goals described in the Parameter §pccifications, which were held by the participants 
prior to the evocation cf ihe Dialogue-game. 
In the case of the Helping-game, the goals in tho "diagnostic" phase arc that thc 
tIELPEE dcscribc a scquoncc of related, uncxceptional cvcnte leading up to a failure of his 
cxpectotions. Thcse goals model the state cyf th6 HELPER as he assimilates this initial 
part of thc dialogue, both in that he knows how tho HE~PEE is attempting lo dcscribc his 
problcrn, and also that thc HELPER knows whcn this phase is past, and thc time has come 
(Ihc "trcatrncnt" phase) for him to provide the help which has been implicitly rcquesteel. 
The processes described above perform thc identification and pursuit of 
Dialo~ue-games. How, then, arc DGs terminated? Thc Parameter Specifications 
rcprescnt thosc a:pects of dialogues that arc constant over that particular type of 
dialogue. The Oinloguc-Game Modcl pushas this a step further in specifying that the 
Dislo~uc-,game continues only sr long ar the Parameter Specifications continuc to hold. 
Whcnevcr any predicate in the Specificelion ceases to hold, then tho model prcdicls tho 
i rnpcr~ding tcrrnin;jtion of this Dio1oguc;garnc. 
For r_.xnrnplct, if the IIIIPEF' na longer wants to pcrfarm thn TASK (r:~lhrlr by 
nccornplid--~ir~f: it or by nb:rndonirl~ that goal), hv 'indicl~trr.i fhia with nn irttcsriinct. wlrich 
hd; for tr:rrnin:il~on. The: lit,lp~ng game ihr:n tnrrnln:\tc~; this corrc;pr~nrl; lo lhq 
:imultsrir.ou~ tcrrr~instion of thc holp~ng interact~on. If the HELPER bccomc; unwilling lo 
give hclp, or discovtrrs that hc 1s unelble, thc:n Ihr! Eirlptng-game also terminiltcs. Again, 
we haddo one simplc rule ,lh?it corcr; h di.dcr:;!y of casos--a rule for tormination that 
captures tho variety uf ways that the dialogue? wa havo ~tudied end. 
A Model of Dialogue 
The Dialogue - ,name Processes 
In this section we describe the major process elements of the Dialogue-Game 
Model. All the major parts and their connectivity are shown in Fieure 1. Thcse parts 
(two rncmorics and six Proccsscs) will each be described separately. Thc appendix 
contains an extensive, detailed trace of the model as it analyzes (via hand simulation) a 
naturally occurring dialogue fragment. Finally, we will summarize our experience with 
the model to date. 
Long-term Mcrnory (LTM) 
The Long-Term Memory is the rnodcl's representation of a participant's knowledge 
of the external world. It contains the initial knowledge states of the participants: the 
grammatical case frames, the semantic structures for word-senses, tho knowledge of the 
s~bjcct,matter of the dialopu~, the various ways in which dialoeues are structured, ctc. 
LTMis a semantic network, containing a set of nodes (also called concepts) and the 
relations that hold between them at the Iowost ievel. This information is stored in the 
form of triples: 
<node- 1 relation node-2> 
Wc have this machinery encoded and working--a 611 complement of read and write 
primitives for this representation. However, it has proven awkward for us to specify 
knowlcdge at this level, so we have implemented further machinery (named SIM) to 
tran.olatc n-ary predicates into these triples. Thuq far a predicate, P, having arguments 
A 1, A2, and A3, SIM can be given the input: 
PI: (Alpha P Beta Gamma) 
[mca~ing that P1 is defined to be an instance of P (the predicate always goes iR swd 
position) with arguments Alpha for Al, Beta for A2 and Gamma-?or A3.1 The resulting 
triples we created: 
<P1 PRED P> 
<P1 A1 ALPHA, 
<Pl A2 BETA> 
<PI A3 GAMMA> 
Let's examine a more concrete example; suppose we want to include in the LTM 
that: 
A Model of Dialogue 
Dialogue 
text 
Figure 1. The Dialogue-game Model 
A Modcl of Dialogue 
.Mary hit Johnwith arock. 
The predicate "HIT" ha's two mandatory arguments (subject, object) and an optional one 
(instrument).   he SIM representation of thi~~assertion (which we shall name Ql) is 
Ql:(MARY HIT JOHN ROCK) 
which translates into the foll~ing triples: 
<Q1 PRED HIT> 
(01 SUBJ MARY> 
41 OBJ JOHN> 
<Q1 INST 
ROCK, 
Workspace (WS) 
The Workspace is the model's representation for that information which the 
participant is activcly using, This memory corresponds roughly ta a model of the 
participant's focus of attention. 
While the L'TM is static during the operation of the model (we are not attempting to 
simulate learning), the WS is extremely volatile, with elements (activations) coming into 
and out of focus c~tinuously. All incoming sensations (i.e., utterances) appear in the WS, 
as do all augmentations of the participant's knowledge and goal statc. The 
representational format of the WS is the same as in LTM. Each node in the WS isa token 
(copy) of some node in LTM. Whenever some process determines that the model's 
attention (WS) should include a token of a specific node (C) from LTM, a new node (A) is 
created by copying C and this new node is added to the WS. A is referred to as an 
artlr,ntlnnn(ranrlihr,aiw- 
<A IAQ C> 
This rcprescntatim providcs the associative links between an object in attention, and the 
body of knowledge assbciated with it, but not yet broucH into attention. 
A Modcl of Dialogue 
This module produces activations representing each successive utterance to be 
processed. These rcprcscntations are generated from the surface string using a 
standard ATN Grammar similar to those developed by Woods (19701 and Norman, 
Rumelhart, fir the LNR Research Group (1975). We use a case grammar represeniation, 
with each utterance spccificd as a main predicate with a set of parameters. Bccausc this 
module is a conventional parser whose implementation is well understood, we hove so far- 
produced hand parses of the input utterances, following an ATN grammar. 
Protcus 
This is a sprcodine activation mechanism, which modifies thc activation of conccpts 
spccificd as rclatcd in LTM whenever a givcn conc~pt bccomcs active. This mcchanism 
provides a way to intcgratc top-down and bottom-up processing within a uniform 
framework (Lcvin, 1976). The Dial ogue-Game iilodel uses Protcus to activate a 
conccpt, given tha a number of closcly relatcd conccpts (Componcnts, fcaturcs, 
instances, etc.) arc active. 
I9 
Protcus opcratcs on all currwt activations to modify their salience", a numbcr 
associated with each activation that generally represents the importance or rclcvancc of 
the conccpt. Two kinds of influence relations can exist bctwccn conccpts: cxcitc or 
inhibit. If an excite relati~n exists, then Protcus increases the salience af the 
activation of that concept in proportion to the saliencc of the influencing conccpt. The 
higher the salience of an activation, the larger its influence on directly related conccpts. 
If an inhibit relation is spccificd. then Process decreases the salience of the activation 
of the neighboring conccpt. 
Match 
TI- is Proccxi idkntifics conccpts in LTM that arc congruent to cxistinc activationr. 
The Diologuc-Game Modcl conthins a numbcr of cquivalencc-like relations, which Mal'ch 
uses to idcntify a conccpt in LTM as rcprcscntin~ thc same thing as an activation of somo 
~ccrningly different concept. Once this equivalent conccpt is found, it is activated. 
Dcpcnding on how this conccpt io dcfincd in LTM, its activation may havo cffects. on othcr 
processes (for cxamplo, if thc canccpt is part of a rulc, Dcducc may bo invoked). 
Match can be vicwcd as an attcrnpt to find an activation (A) in WS and a Concc!pt (C) 
in LTM which correspond, accordintta some set of critcr The basic tactic i., to ottrrrnpt 
A Model of Dialogue 
to find a form of cquivalencc relationship between A and C, without delving into their 
structure at all. 
Only if this fails arc their respective substructures examined. In ths 
sccond case, the same match which was attempted. at the top level is tried bctwccn 
corresponding subparts of A and C. Match proceeds in five steps: 
1. Is it alrcady known that A is an activation of C? If so, the match ferminates 
with a positive conclusion, 
2. Is there any other activation (A7j, and/or conccpt (C') such that A"is known 
to bc a view of A, C 1s known ta bc a kind of C', and A' is known (by step 1) to 
bc an actlvation of C'? The relations (i.. is a view of ...) gnd (... isba kind of ...) 
rcprescnt stored relations between pairs of activations 1 and' concepts, 
rcspcctivcly. One concept "is a kind of" another conccpt rep,~~~,lts, a 
s~pcrclass inclusion, trua for all timcand cdntexts. '(Whdever else he might 
be, John is a kindof huma'n being:) On the other hand, one activation may be "a 
view of" another only under certain circumstances--a conditional, or tactical 
relationship. Undcr diffcrent.conditions, if is appropriate to view John as a 
Husband, Father, Child; Hcl p-seeker, Advice-giver, e tc, 
3. A list of matched pairs of activations and concepts reprcscnt 
corrcspondcnccs found el scwhmc, with which match must be consi stcnt. 
(N.B.: this Match, as we will see later, may be in service of anothcr Match 
galled' on siructuros containing the current A and C.) If thc par [A,C] is a 
matchcd pair, then these tyo have been previously found to match,.so we may 
hcrc concl'udo the same thing and Match exits, 
4. On the other hand, if there is either an X or a Y such that [A,X] (or [Y,C]) is 
a matchcd pair, then replace this match with an attempt to match C and X (or A 
and Y). 
5. Finally, if the matchahas ncither succeeded nor failed by this pqint, !hcn 
Match is called recursi~~ly on all corresponding ~ubparts of A ahd C, 
pail~~isc. That is, c.~., if A and C have only thrcc subpartsip common (soy, 
SU3J, 013J and PRED) thn Match((SUBJ of A),(SUBJ of C)), Match((OBJDof 
A),(OBJ of C)) and ~atch((~~€~ of A),(P~EO of C)) arc attempted. Only if all of 
thcsc subordinate matches succeed is the top-level Match said to succccd. 
Clearly, for structures of significant complexity, Match may eventually call itc,clf 
rccursivcly, to an arbitrory depth. 
However, since each subordinate call is on a strictly 
smaller unit, this process must coitvcrge. 
A Madcl of Dialogue 
Our experience has shown us that this type of mechanism plus a collection of 
rewrite rules enable us to eventually map a wide variety of input parsing structirrcs to 
pre-stored, abstract knowledge structures, in a way that a significant aspect of their 
intended meaning has been assimilated in the process. 
Dcducc 
This opcratcs to carry out a rule when that rule has become active. Rules are of 
the form (Condition)->(Action), and Dcduce scnscs thc activity of a rule and applies the 
rule by activating the concept for the action. Whatcver corresponocnces were evolved 
in the coursc.of cccating the activation of the condition (left) half of ihe rule are carried 
over into thc activation of the action (right) half. The combination of Match and Dcduce 
a em. gives us the capability of a production syct 
Thc operation of Dcducc is relatively simple. It is called oniy when d rule is active 
in the WS. Dcducc attempts to match the left half of this rule with some other activation 
in the WS. (This has ty-pically already been done by match.) Assuming this is 
accomplished, Dcduce crestet an ac,tivalion of the right half of the rule, substi luting in the 
activation for all subparts for which thcke are correspondences with the icft half. 
Once a Dialogue-game has been activated (by Protcu.,) as possibly the 
comrnunrcation form being bid for a dialogue, the Dialogue-game Manager uses it to guide 
thc assimilation of successive utterances of the dialogue, through four stages: 
1. establish the Parameter values and verify that no Specification is 
contradictcd, 
2. cstabli ;h olhcrwi sc unsupported Specifications as assumptions, 
3. cstrrblish the Components as goal$ of thc participanfs, 
4. dctcct the circumstancc~ which indicate that the Dialogue-game is. 
terminating and represent thc conscquenccs of this. 
Thc first two of thesc phaces hsppcn in parallel. 
Whcn the Manager accesses 
each of thc Para~nctcrs, they arc found either to have activations in thc WS or not. 
If 
they do, the cgrrespondcnccs bclwccn activation and Psrsmetcr are established in the 
WS. This correspond: lo a::ignint a value to thc Paramcicr for this particular evocation 
of thc Dialogue-game. Any Parameter that has no activation ic put on a list which is 
A Model of Dialogue 
periodically chcckcd in the hope that later activity by the Manager will lead to the 
creation of appropriate activations. 
For each of the Specifitations, a check is made to determine if it already has an 
activation in WS. (In most cases, the activation of some of these Specifications will have 
led to the activity of the Dialogue-game itself.) The Specifications having activations need 
no further attention. 
For all remaining Spcc,ifications, activations are created substituting for the 
Parametcrs as determined above. At this stage, the Dialogue-game Manager calls 
Protcus to determine the stability of thcso new activations. Any new activation which 
contradicts existing activations will have its level of activity sharply reduced by Proteus. 
If this happens, the Dialogue-game Manager concludes that some of the necessary 
preconditions for the game db not hold (are in conflict with current understanding) and 
that this particular game should be abandoned. Otherwise, the new activaticns stand as 
new knowledge, following from the hypothesis that the chosen game is appropriate, 
The Dialogue-game has now been successfully entered: the Manager sets up the 
third phase, creating activations of the Dialogue-game's Components, with appropriate 
substitutions. (By this time, any unresolved Parameters may well have -activations, 
permitting their resolution.) This sets up all of the game-specific knowledge and goals for 
both participants. 
Finally, the Manager detects that one of the Specifications no longer appears to 
hold. This signals the impending termination of the Dialogue-game. In fact, the 
utterance whikh contai'ns this information is a bid to terminate. At this point, if the 
partici pants7 initial goals are satisfied (thus contradicting the Specification which calls for 
the prcsence of those goals) the interaction ends "successfully". Otherwise, the 
Dialogue-game is terminated for some other reason (e.g., one participant's unwillingness 
or inability to continue) and would generally be regarded as a "failure". These 
consequences are infcrred by the Manager and added to the WS. When a Dialogue-game 
has terminated, its salience goes to zero and it is removed from the WS. 
Pronoun Proccsscs 
The Dialogue-Gamo Model contains a set of Pronoun Processes, including an 
I-Process, a You-Process, and an It-Process. Each of those is invoked whenever the 
/' 
associated surface word appears in an input utterance,, and operates to identify some 
preexisting activationthat can be seen as a view of the same object. 
# Model of Dialogue 
Each of these Processes search the curreot context, as represented by the current 
sct of actirations in the WS, using tho katures specified there ro identify a set of possible 
co-rcfctcntial expressions. When there is more than one possibility, the one with a 
hi ghcr salience is sclectcd. 
A Model of Dialogue 
DEFICIENCIES IN CURRENT MAN-MACHINE COMMUNICATION 
With the understanding we new have of the multi-sentential aspects of human 
communication, it is easy to see why man-machine cornrnun'ietion appears so alien, highly 
restrictive, uncomprehending and awkward. This is because major regulation and 
interpret4 f ion structures are missing. 
In Table 1, we compare human dialogue and typical man-machine com.munication 
with respect to some of these features, The table designates a "sender" ancfa "receiver" 
which should be identified with the person and the computer, respeclively, in the 
man-machine communication case. 
ASPECTS OF NATURAL COMMUNICATION 
ADDRESSED BY DIALOGUE-GAME THEORY * 
SENDER'S GOALS KNOWN TO RECIPIENT 
PARTICIPANTS CAN DECLARE THEIR GOALS 
GOALS PERSIST OVER SEVEFiAL MESSAGES 
GOALS IDENTIFIED WITH EACH MESSAGE 
COMMUNICATION PLANS USED 
IMPLICIT COMMUNICATION TAKES PLACE 
HUMAN MAN- 
DIALOGUE MACHiNE 
YES NO 
YES NO 
YES NO 
YES NQ 
YES LITTLE 
YES LITTLE 
Table 1: fi @omparison of man-rnan and man-machine communlcalion 
Conventional man-machino communication froqucntly givec tho user o sense that 
the computcr is operating "out of contcwt", since he must continually respccify what ir 
rclcvant to the ongoing dtaloguc. In human communication it ic the shared awarcncss of 
each other's goal structures which pcrrnitr ficm to retain and fosuo.on what ie relevant. 
Man-machine communication sccms aimless and undirected bacauro no analogous body of 
knowlcdgc is being used tc facili tatc and interpret the communicotioh. 
A Modcl of Dialogue 32 
The ideal interface, arid the sort toward which this research is direded, would be 
continuously askihg itself: "Why did he say that?". From answers to this, the interface 
would infcr just what the human was expecting as a response. This would con~titute a 
major slcp toward the enabling tho intcrface to servo the actual (rather than the poorly 
cxprcsscd) needs of the user. Finally, such an intorface would require much lco~ 
adaptation on thc parf of the user, and so, by our original hypotheois, would significantly 
enhance the cffcctivencss of the man-machine partnership. 
A Model of Dialogue 
This paper has described a research effort into the modeling of human dialogue. 
The purpose of this research has been to uncover and describe in process models, 
reculari tics that occur in dialogue. It is hopcd that the enhqnced understanding of human 
communication which rcsul ts, will facilitate the development of more natural (and thus 
more effective) man-machine interfatcs. 
Thc principal regularity w'e have discovered is a collection of knowledge and goal 
structures, called Dialoeue-games, which seem to be crucial in understanding the 
structure of naturally-occurring dialogues. According to the theory we have proposed, 
one or more of these Dialogue-games serve as the major organizing influence on every 
human didlogue. 
Each Dialogue-game specifies what knowledge each person must have to ehgage in 
such a dialogue, and what goals of the participants might be served by that interchange. 
A Dialogue-game also spccifies, as a sequence of "tactical" goals, the manner in which the 
dialogue is conducted. 
The Diatogua-game Model is a collection of cooperative processes which 
continuausly updated a representation of each participant's attention state in a 
Workspaco. The model recognizes when a particular Dialogue-game is being bid, 
accepted, pursued and terminated, and represents these states appropriately in the 
Workspace. A particular Dialogue-game, the Helping-game, was described in some 
dctajl. A simulation of the evocation and use of the Helping-game on a segment of natural 
dialogue is contained in the Appendix. 
Our experience so far with the Dialogue-game Model has reinforced our 
hypothcses that an understanding of the goal-serving aspects of dialogue is a powerfdl . 
tool in understanding the individual di~alogues. 
A Model of Dialogue 
A Modcl df Dialogue 
APPENDIX -- SIMULATION OF THE DIALOGUE-GAMES MODEL 
Example of the.Oia[ogue Modcl in Action 
In this appendix we describe an extensive simulation of the arrent state of the 
Dialogue-game Model. We make use of a particular version of the Helping-game and alsc 
explore another structure, an Execution Scene, which describes the customary events 
surrounding the successful execution of a particular program (Runoff). 
We start by describing this more detailed version of the Helping-game, introducing 
names for the various aspects, to be used later. 
Next we show a short, naturally 
occurring dialogue between a computer operator and a user. 'Then we describe the 
operation of the Dialoeuc-garnc'Model as if assimilates this dialogue, up to the point at 
which it concludes that thc Helping-game is an appropriate structure throu~h which to 
understand the subr;cqucnt utterances. 
Once this hypothesis for the form of thedialogue has been chosen, we continue the 
simulation to examine how Jhc model dccidcs that a particular Execution Scene is 
appropriate for assimilating the content of the dialogue. 'Next, we see how this choice of 
occnes cnhances the set of goals imputed to the speaker, thus facilitating the 
cornprehcnsion of what he is saying. Finally, we summarize our experience with the 
Dialogue-game Model so far. 
A Dctai/ed Structure for the help in,^ - game 
What fol4ows is the substance of the communication structure we have namcd the 
Hclping-game. In the interests of clarity af presentation, the formal structureo of the 
definition have been expressed in prose. However, the elements of the following 
description correspond one-to-one to those in the actual Helping-game used in ?the 
simulation. 
HELPING-GAME 
Paramctcrs: 
Thc parameters are two roles (HELPER and HELPEE) and a topic (TASK/HG). 
A Modcl of Dialogue 
Parameter specifications: 
The HELPER and HELPEE are each a kind of person. 
H1 = A goal of the HELPEE is that he perform TASK/HG. 
H2 = It is not true that HELPEE is able to perform this TASKIHG. 
H5 = The HELPEE wants to be ablc to pcrform the TASKIHG. 
(being able to perform ?he task is a subgoal of 
performing the task) 
H6 = The  ELP PER is able t.o enable the HELPEE to .perform the TASK/HG. 
H8 .I Thc HELPER is willing(= is ablc to want to ...) to enable the 
HELPEE to perform the TASKIHG. 
H10 = The HELPEE is permitted to perform the TASK/HG. 
HI 1 = Thc HELPEE wants the HELPER to enable him to perform the TASKIHG. 
(bcing enabled to perform the task.is6a subgoal of 
performing the task) 
Game components: 
HGX 1 = The HELPEE knows of a particular execuiion scene, XS/HE. 
[note: an execution scene is a flowchart-like description 
of thc use af a particular process; more details below] 
HGX2 = The HELPEEknowe that his perceiving the terminal state of XS/HE 
would satisfy his wanting to perform TASKIHG. 
HGXPC= (Thus) The HELPEE wants to perceive XS/HE in this terminal 
state, 
(this perception is a subgoal of performing the TASKIHG) 
ACTION/GUOD = an ACTION of XS/HE which was realized in the past. 
HGX3 = The HELPEE knows he has perceived this ACTION/GOOD. 
HGX4 = T~C"HELPEE knows he had expected to perteive it. 
HGX5 = Thc HELPEE knows hc wants to perceive this ACTION/GOOD. 
(pcrcciviny: the ACTION/GOOD is a sub~oal of perceiving tho 
[dcsirc.dl tcrmi nal date of the XS/HE) 
kCTIOIJ/BAD = an ACTION of XS/t1E which was not rcalized in tho past. 
HGX6 = Thc HELPEE knows that hc did not perccivc ACTIONfBAD. 
HGX7 = The tiELPEE knows that he had cxpcctcd to pcrceivo it. 
HGX8 = The HELPEE want.; to pcrccive ACTIOId/BAD. 
(pcrcciving the ACTIONIBAD is a subgoal to perceiving the 
terrnirral state of XS/HE.) 
HGX9 T Thc HELPEE wants to describe what happclncd which was both 
cxpcctcd and wanted, thc ACTIOl,l[s]/G000. 
(dcscribinc thcse A~TIOFJ[~]/G~~D is a subgoat of having 
thc HELPER enablo the HELPEE to pcrform tho TASK/HG.) 
HGXlQ= Thc HELPEE wants to dcscribo what dtd not hsppcn that he 
A Model of Dialogue 
expected, and wanted, the ACTION[s]/BAD. 
(describing these ACTION[s]/BAD is a subgoal of having 
the HELPER enable the HELPEE to perform the TASK/HG.) 
The Dialogue to be Modeled 
What follows is a transcr.ipt of a naturally occurring dialogue between a computer 
operator (identified as "0') and a user ("L'3 who has "linked" to the operator, in an 
attempt to solve a problem. 
Thcre has been virtually no "cleanup" of this transcript, except to remove 
extraneous typing hat had appeared on the operator's console listing as a result of the 
op&ratine system printing routine status messages. The choice of words, and even 
spcliinc, are exactly as typcd by the participants. (We have segmented the text by 
interposing carriage-returns as we deemed appropriate.) 
Dialogue OC 11 7 
LINK FROM rL1, TTY 42 
I. : How ,do 1 get runoff to work, 
I kcep xeqtn it 
but it just grabs my input file 
and then says done 
but gives me no output? 
GA 
0: The output comes out on the line printer 
L : Throw it away 
but can I get it to go to a file? 
GA 
0: Confirm your commands with a comma 
and you'll be queried for files, dc. 
GA 
A Model of Dialogue 
L: Thanx mucho 
BREAK 
The subsequent simulation is of the model .processing the first five segments, the 
entire first utterance. Each utterance is ingested one at a time, by the Parser, and the 
assimilation proceeds until a quiescent state is reached (much more detail, below) 
whereupon the next segment is parsed and input for processing. 
The identification of thc hclping-gamc 
How docs thc model know to evoke the Helping-game? To exhibit answers to thi~ 
and subscqucnt questions, we lead tho reader through a simulation of the model as it 
proccsscs the beginning of dialogue OC117. We indulge in fhe samc use of prorie for 
formalism as aboQe, again with the same assuranaes of correspondcnccs with tho actual 
sirnulati on. 
Thc simulation proceeds in cycles: in each cycle, we exhibit the operation of a 
sinzlc processor, performing one iteration of its function. We do not address hcrc the 
is:uc; of how the model would select w.hich processor to call next. In fact, our dcsign 
calls for these processors to be mgxirnally autonomous and parallel in their operation, 
operating whcncvcr circumstances are ripe for their function and dormant otherwise. 
Thc format of this sirnuistion is as follow;: Thc cycle number is first, in the form: 
:cy,mcnt nurnbcr9--cycle number in this scgrncnt,. Next is tho name of the praccwor 
operating in this cyclc. Aftcr that is EI description of the nature of the pracossiny. donc 
duPlnb that cyclc. Finally, tharo is a list of tha rcsults for this cycle, that is, ;dl tho 
irnportljnt changw in WS, 
iily tho drxcription is at a vcry d~tjiled level. But after a whiln, tha 
operations bccotl~c cxtrernr~ly rcpctitive so tho dcstription bccomas less d~~tailcd, 
focusing only on thc unique aspects of thc current opcrstion. In this cxmple. each 
processor is called at lcastonco in thc processing of each scgment: Match, Dcduco and 
Protcu; bear thc major burden, having scvoral invocations each pa? segment. 
Cycle 1-1 -- Parse. 
The parser reads one utterance/segment of input and translates it into the formalism 
for activations in the workspace. No claim 'is made that this translation retains all the 
content of the original text, only that it is adequately faithful to the level of detail we are 
simulating. 
Results: Case/9 (= (0 perceives that L asks (how do I get Runoff working?))) is activated. 
Cycle 1-2 -- I-processor 
Certain words (e,g. pronouns, determiners) are taken to be signals that a reference 
is being made to conccpts introduced elsewhere. Sne presence of a concept in the 
workspoce corresponding to one of these words lcads to the calling of the. 
process-specialist which attempts to resolve the implied reference. Thus, the presence 
of "I" in the text leads to the calling of the I-process, whose sole function is to determine 
the referent of the .'I" and modify the stored concept to reflect this. This process judges 
that if L is asking a question which contains "I" as its subject, then this constitutes 
adequate evidence to hypothesize that "I" is being used to refer to L. 
Results: 0 perceives that L asks (how does L get Runoff working?) 
Cycle 1-3 -- Match 
Match is always on the lookout for pairs of nodes, one in the WS and the other in the 
LTM, such that the activation (node in WS) matches the concept (node in LTM). This is 
taken to be evidence that the activation is also to be tiken as an activation of the matched 
concept. It should be understood that we areaexamining only some of the succewful 
matches which occurred, 
Starting in this cycle, we see a pattern which recurs regularly, and which accounts 
fcr a significant piece of the action, as the model assimilates the dialogue. 
Match 
dcterrnincs that a particular activation matches the left half (condition side, if part, etc.) of 
a production-like rule srorcd in LTM. This successful match leads to the identification of 
the corrcspondcnces between the aspects of the activation and those of the left half of 
the rule, ae well as creating an activation of the rule itself. The activation of a rule leads 
to calling the Deduce processor in thenext cycle, which applies the activated rub to the 
node in the WS responsible for the rule's activation. This application of a rule (which 
also results in thc removal of the rule's activation from the WS) creates a new activation 
structure in the WS. 
A Modcl of Dialogue 
In othcr words, the introduction of a piece of knowledge suggests that a certain 
transformation (e.g.,'"Whcncver you know X, you can conclude Y.") is appropriate. This 
transformation is applied to the stimurus knowledge to generste a conclusion: a new piece 
of knowledge. 
In this particular case, the above result structure is found to match the left half of 
Rule0 = If Opcrceives a proposition, 
then 0 knows that proposition. 
with thc correspondenccs 
Case/l (= (L asks(How do I get Runoff working?))) is activated. 
corresponds to the proposition. 
(This rulc ~cprcscnts the approximation that what is perceived is accepted at face value.) 
Sincc Casc/g is now scon to be an activation of the Left-half of RuleO, an activation 
for the rule itself is created in the WS, 
Results: Case19 is an activation of Left half of RuJeO. 
Casa/j.corresponds to the proposition in RuleO. 
An activation of Rule0 is entered into WS. 
Cycle 1-4 -- Deduce 
Sincc a rule is active in WS, Ocduce is called in an attempt to apply thc rule. The 
Match has guar'antced that the necessary correspondences exists betwoen the left half of 
the rule and the n'odc which is its activation. To apply the rOle, Deduce creates an 
activation of the righthalf, with the corresponding sub-parts substituted. 
Rcsul ts: RO- 1 = 0 knows Casc/l 
Activation of RuleO deleted from WS. 
Cycle 1-5 -- Match 
Match finds that RO-1 matches the left half of: 
Rule1 = If 0 knows (L asks about a proposition), 
then 0 knows (L does not know about that proposition). 
A Madcl of Dialogue 
Rcsults: RO-1 is an activation of the loft half of Rulel. 
Case/l corrcsponds to (L asks about a proposition) 
Case12 = (How does L get Runoff working) corresponds to the 
proposition. 
An activation of Rule1 is created in the WS. 
Cycle 1-6 -- Deduce 
Dcduce applies Rule 1 to RO- 1, substituting according to the discovered 
corrcspondcnccs. 
Rcsults: R1-1 (= 0 knows (L does not know Case/2), is activated.) 
Activation of Rule 1 deleted from MS. 
Cycle 1-7 -- Match 
Match Rl-1 with left half of 
Rule3 = If 0 knows that a person does not know how to perform a 
task, 
thcn 0 knows that that person is not able to perform 
the task. 
Rcsults: R1-1 is an activation of the left half of Rule3, 
L corresponds to the person mentioned. 
Get corccoponds to Perform. 
Tho state of Runoff working corresponds to the task. 
An activation of Rule3 is'created in the WS. 
Cycle 1-8 -- Dcduce 
Deduce applies Rule3 to R1-1. 
Results: R3-1 (= 0 knows that R3-11= (L is not able to perform 
(getting Runoff working)) is activated). 
Activation of Rule 3 deleted from WS. 
Cycle 1-9 -- Match 
A Model of Dialogue 
Match R3- 11 with H2 = Helpee is not able to perform the task. 
Results: R3- 11 is an activation of H2, 
(~ctting Runoff working) corrcsponds to the task. 
L corrcsponds to tho Hclpeo 
Cycle 1-10 -- Match 
Match RO- 1 with left 112 of: 
Rulc2 = if 0 knows (L asks about a proposition), 
then Q knows (L wants to know about that proposition). 
Rcsults: RO-1 is an activcition of the left half of Rule2. 
Case11 corrcsponds to (L asks ...), in Rule 2. 
Casc/2 corrcsponds to the proposition. 
An activation of Rule 2 is created in the WS. 
Cycle I- 1 1 -- Deduce 
Dcduce applies Rule2 to RO-1. 
RcsulEs: R2-1 (= 0 knows (L wants to know about Case(2) is activated). 
Activation of Rule 2 dcletcd from WS. 
Cycle 1- 12 -- Match 
Match R2-1 with lcft half of 
Rule4 5 If 0 knows (a person wants fo know how 
to perform a task), 
then 0 knows (that person wants to perform that task). 
Results: R2- 1 is an activation if the lcft half of Rule4. 
L corrcsponds to the person. 
(getting Runoff to work) corresponds to the task. 
An activaiion of Eule 4 is created in the WS. 
Cycle 1- 13 -- Deduce 
A Model of Dialogue 
Deduce applies Rule4 to R2- 1, 
Results: R4- 1 (a O knows (L wants to perform (getting Runoff working)) is activated), 
Activation of Rule 4 deleted from WS. 
Cycle 1- 14 -- Match 
Match R4-11 with H1 = Hclpoo wants to perform a tesk. 
Rcsults: R4- X 1 is an activation of H2, 
L corrcsponds to tho Hclpcc. 
(Getting Runoff working) corresponds to the task. 
Cycle 1-15 -- Match 
Match RO- 1 with left half of 
RuleVa = If 0 knows (a person says 
(hc cxccutcs a process with an instrument)), 
then 0 knows (I'hat person is saying 
(he pcrforms (the a-xecution of the procoss) 
with thc instrument). 
Rc~ults: RO-1 is on activation of tha left half of RuloVa, 
L corrcspondo to thc person. 
(getting Runoff working) corresponds to (... executes a process ...) 
How car respond^ t,o tho instrument (i.e., the means). 
An activation of Rule Va is croatod in the WS. 
Cycle 1- 16 -- Deduce 
Deduce applies AuleVa to RO- 1. 
Rcsults: RVa-1 (= 0 knows ( 1 asks (how do I perform (getting.Runaf# working)?)) is 
activated).. 
Activation of Rule Va deleted frorii WS. 
Cycle 1- 17 -- Match 
A Model of Dtalogua 
Match RVe- 3 with Left,hatf of 
RutaZa = If 0 knows (apcrson asks how to perform a ta$k), 
then 0 knows (that porson wants O to onable him 
to perform that task). 
RosuIf S: RVa- 1 is an activation of tho left half of Ru~B~B, 
L corrcsponds to that person. 
(L gcttiny, Runoff to work) corscspondr to the task. 
An activatian of Rule 2a is created in the WSc 
Dcducci applies Rule2a to RVa- 1 
Rcsulto: R2- 1 (= 0 knows (L wants 0 to enable him (L) to get Runoff working) is activated). 
Activation of Rule 2a dcleted from WS. 
Cycle 1- 19 --- Match 
Match R2a- 1 with H11 = Hclpce wants Helper to enable him toto a task, 
Rceults: 0 corrcspofdis to H'elpcr. 
L corrcsponds to Hclpce. 
(L getting Runoff to work) corrcsponds to tho task. 
Cycle 1-20 -- Protcus 
H1, H2 & HI1 provide .Protcus with enough evidence to create an activation of the 
Helping-Garnc. 
Rcsul ts: An activation of the Helping-game is created in the WS. 
Cycle 1-21 -- Dialogue-game Manager 
Thc-presence of an activation of a Dialogue-game in the WS leads to the calling of 
thc processor spccializcd in this.catcgory of knowledge. The Dialogue-game Manager 
(OGM) makes use of a set of correspondences that have already been established by the 
matches which led to the activations of HI, H2, and H11: 
A Modol of Dialogue 
Previous Rosults: L corresponds to Hclpee 
0 corresponds to Helper 
Case13 (= (Runoff working)) corresponds to the task. 
Once ar) activation of a eamc has led to the calling of the DGM, the Manager accesses 
thc cntira collection of information about tho game from the LTM representation bf it. 
Tho itcms of knowledge in tho game, with tho particular parameters of this situation 
substituted apprcrpriatcly, fall into one of three categories: 
1. Alrcady known to hearor (e.g. HI, H2 & H11). Items in this category are 
simply icnorod, since it serves no purpose to re-assert them. 
2. Contradict knowlcdgc already held by the hearer (e.g., if 0 already k,ncw, 
for sure, that t knew all about Runoff). If any item falls into this category, the 
hypothesis that this game is active is simply abandoned as inaccurate. 
3. Mtms ncithcr previously known or contradicted (the majority of the 
contcnt of the typical case). In this case, tho DGM creates activations of 
thcsc items to represent the collection of implitit knowledge that follows from 
a recognition of the proposed game. 
Rcsul ts: Activations are created for all of the following: 
H5 = L wants to be able to get (Runoff workin$) himself. 
(being able to get (Runoff working) is a subgoal 
to performing (Runoff working).) 
H6 = 0 is able to enable L to gct (Runoff working). 
H8 = O is able to want to enablc [i.e. is willing to enable] 
L to gct (Runoff working). 
101.C is permitted to get (Runoff working). 
Thc game also contains a collection of knowledge having to do with the conduct of 
thc game, rather than what the parllcipantsneed to successfully evoke it. These items of 
knowledte and goals are also established as activations by thaDGM at this time: 
Resutts: Activations are created or all of the following: 
HGx1 = L knows of an execution scene (XS/HE). 
HGx2 = L knows that if he perceives a particular 
terminal state of this scene, this will 
satisfy his wanting to perform the task. 
HGX2C= (Thus) L wants to perceive this terminal state 
A Modcl of Dialogue 
of XS/HE. 
An ACTION/GOODis mACTION within the specification of 
XS/HE whi'ch occurred in thc past. 
HGX3 = L knows that he has pcrceivc tho ACTIQN/GOOD. 
HGX4 = L knows ho expected to perceive it. 
HGXS = L wantcd to pcrccivc it. 
An ACTlON/BAD is an ACTION within the specification of 
XS/HE which has not occurred in thc past, 
HGX6 z L knows he has not perceived the ACTION/BAD. 
HGX7 = L knows hc expcctcd to perceive it, 
HGXB = L knows hc wanted to pcrccivc it. 
(perceiving the ACTION/BAD is a subgoal to perceiving 
the dcsircd tcrminal state of XS/HE.) 
HGX9 = L wants to describe the ACTION[s]/GOOD [to 01. 
(this describing is a subgoal to (0 enables L to 
pcrform the task) 
HGXlO= L wants to describe the ACTION[s]/BAD [to 01. 
(this describing is a subgoal to (0 enables L to 
perform the task) 
Proccsscs, procedures, ccremonics, and the like, may have an associated execution 
scenc, which is in effect an abstract description of a complete performance of the object 
dcscribcd. The cxccuiion scene rcscmblcs a flowchart, with the boxes being actions of 
onc of thc active agents i'nvolvcd. 
In this case, the execution scenc is for Runoff, a program which reads a file 
specified by the uscr, formats the contents of the file, and outputs this formated material 
onto cithcr the line printer or another file. The execution scene of Runoff, as stored in 
our model, is similar to figure A-1. 
A Model of Dialogue 
START 
I 
XSA4 a Uscr initiates Runoff 
I 
1 
XSA-2 = Runoff requests a file name. 
I 
\I 
XSA-3 = User types a file name. 
I 
I 
XSA-4 = Runoff requosts a confirnlation. 
I 
I 
[one of thc fol louing two paths is taken:] 
I I 
I I 
1 = user types coninla. 
XSA-21 = user types carriage return. 
1 I 
I I 
XSA-12 = Runoff reads (grabs) 
XSA-22 = Runoff reads (grabs) 
input f i le. 
input file. 
I I 
1 I 
XSA-13 R~~nof f rec~c~csts output XSA-23 - ~unof f produces (gives) 
f i le nanle. 
output on l irle printer. 
I I 
I I 
XSA-14 - User types output f i l e nanle. XSA-24 = Runoff tgpes DONE. 
I I 
I F 1.N I SH 
XSA-15 .= Runoff' produces (gives) 
output' on output fi le. 
I 
I 
XSA-16 = Runoff- types DONE. 
1 
FI I\J 1 SH 
Figure A-1. XS/RO. THE RUNOFF EXECUTION SCENE. 
A Model of Dialogue 
Cycle 1-22 -- Proteus 
As a result of the numerous rcfarcncar to Runoff and XS/HE, tho activationo for 
thcsc two conccpts are "highly active". ~onsc~ucntl~, when Protcus io called, tho 
eonccpt XSjRO (the execution sccno of tho Runoff proccos) bocarnoo active and, duo to its 
similarity to XSfHE, is taken to be equivalent to it. Sinca XS/RO is more dctailad (contein~ 
more information) than XS/HE, XS/RO is used in place if XS/HE in all of the expressions 
introduced in Cycle 1-21. 
Something wo pnsscd ovcr in thc earlier examples was tha issuo of vyhcn tho modcl 
is willine to stop processing a given piece of tcxl and eo on la the nexf onc. It scorns 
inappropriate lo demand that tho rncdol wring all possiblo information end deductions out 
of each utterance. Yet there must bo soms demands mada on tho assimilation. An 
altcrnatc form of tho question is: what ncedr of his own does the hearer see the incoming 
text as potentially satisfyinc? We have taken the position that a hearer (tentatively) 
understands an utterance, when he successfully views it as serving some goal imputed to 
the spcskcr. That is, to a first approximation, the hearer has assimilated an utterance if 
hc fisures out why thc spcakcr said it. 
Thc modcl has already established (HCX9 and HGX10, above) that L wants to 
dcscribc (implicitly, to 0) certain action; in XS/RO bat L expected to perceive, and in 
sonw csscs, did, Thus, in thc following uttcrsnces, we see the modol matching the 
parsed input structure with one of thcsc two goals, thus it is sccn as bcing in service of a 
goal of thc spcakcr, and need bc examined no further (for tho time being). 
In thc subscqucnl example, we use two ncw rules: RS (Satisfaction) and RQ 
(Quicsccncc). RS dctcrrnincs when an uttcranco is sccn to satisfy a speaker's goal and 
RQ, rcscts to this dcfectcd satisfaction by marking the utterance quicsccnt. 
(Opcrationolly, this means that in tho next cycle, thc Parser is called to input the next 
scgmcnt af text.) 
Wc resume the example at the point where the first segment has been marked 
quicsccnt, and the Parser is called. 
Results: Casc9a = 0 pcrceivcs that L declares (I executed it). 
Cycle 2-2 -- I-processor 
A Model of Dialogue 
Rcsults: Case9a .I 0 perceives that L declares (L executed it). 
Cycle 2-3 v- It-processor 
Tha case frame aasociatcd with the concept "execute" epecifieo that the object 
canccpt is to bo a procces. The it-processor determines this and examines thc WS to 
scc if'it contains any active concept which is a process. In this case, therc is only ono: 
Runoff. Since this caso is so clcerly unambicuous, this simplo-minded reqolution scheme 
is adcqu~lc to tho tmk. (We hsva.outlinos for moro ambitious resolution ticherncs, but 
tho dialoeucs we havc examined havo not yot rcquirod them.) 
Rcsults: Case9a = 0 pcrceivcs that L declares (L executed Runoff). 
Cycles 2-4 & 2-5 -- Match and Dcducc 
As in cycles 1-3 and 1-4, Rule0 is used to transform "perceive" into "know". 
Results: RO- la = 0 knows that L declares (L executed Runoff). 
Cycle 2-6 -- Match 
Two.itcms in tho WS arc matched to thc two parts of the left half of RS: 
RS = If a pcrson knows a proposition 
and 
hc knows that a secbnd person wants that proposition, 
then the first person knows that the realization of the 
proposition satisfies the second person's desire for it. 
Results: RO- la = (0 knows (L declares...)) corresponds to 
(a pcrson knbws a proposition) 
0 corresponds to the first pcrson. 
(i declarer; ...) corresponds to the proposition, 
0 knows'~i~9 = (L want (L describe action/good)) 
corresponds to 
(he knows-thc second person wants that proposition). 
L corresponds to the secondpersori. 
(L describe action/go.od) corresponds to 
the proposition. 
A Model of Dialogue 
(L dcclares (L exccutcd Runoff)) corresponds to 
(L dcscri be acti on/good) 
dcclarc corrcsponds to describe 
(L executcd Runoff) corrcsponds to ((User initiate Runoff) past) 
thus, (L oxccutcd Runoff) corresponds to action/good 
An activation of RulcS is croatcd in the W$, 
Cycle 2- 7 -- Deduce 
Doduco applies RS to R0-la and HGX9. Activation of Rule S deleted from WS 
Rcsults: RS- la (= 0 knows ((L dcclarcs ...) satisfies (1 wants (L describe ...))) is activated). 
Cycle 2-8 -- Match 
Match RS- la with left half of RQ. 
RQ = If a pcrson knows ((person2 utteks something) satisfies 
(pcrson2 wants something else)) 
then thc first pcrson knows that ho comprehends 
(pcrson2 uttering something) as constituting the 
somcthing clsc that pcreon2 wanted. 
Results: RS-la corrcsponds to the icft half of RQ. 
0 corrcsponds to the first person. 
(pcrson2 utters sorncthmg) corresponds to 
(L dcclarcs (1 exccutcd Runoff)) 
L corresponds to pcrson 2 
(L cxccutcd Runoff) corrcsponds to sorncthing. 
(person2 wants something clsc) corresponds to 
(L wants (L dcscri be ...)) 
(L describe action/good) corresponds to something else. 
An activgtion of RQ is created in the WS. 
AR activation of RQ is createdinthe WS. 
Cycle 2-9 -- Dcduce 
Dcduce applies RQ to RS- la. 
A Model of Dialogue 
Results: RQ- la = 0 knows (0 comprehends 
(L declare (t execute Runoff)) 
as constituting 
(L describe aition/~ood)) 
Activation of Ru1e.Q deleted from WS. 
Cyclcs 3-1 to 3-8 
This set of cycles are exactly parallel to the preceding set. 
The structure 
implantcd into WS by the Parser is 
Case/Sb (= 0 pcrccivcs (L declares (it grabbed filelmine))) 
The It-processor translates "it" to "Runoff". Rule0 is used by Match and Deduce to 
rcplace "pcrccivc" with "know". Match and Deduce then apply RS and RQ, to determine 
that CasefSb is compr'ehendcd as constituting another instance of (L describes 
action/good) [XSA-12 or XSA-22, Runoff reads (grabs) input file] 
Cyclcs 4- 1 to 4-8 
Similarly, tho Parser-produced structure: 
CaseJ9c (= itsaid done) 
is also found to be comprchendod as constituting an instance of (L describes actionlgood) 
[XSA- 16 or XSA;24, Runoff types DONE]. 
Cyclcs 5- 1 to 5- 10 
A ncarly identical sequence of cycles applies to the next Paiser-input: 
Case/Sd (= 0 perceive L declare (It did not produce output),) 
cxccpt an additional MatchfDeduce cycle is needed to apply Rp: 
Rp = If a persondeclares that something didn't happen, 
then he is declaring he did-not perceive it happen. 
In this case, however, we determine that Case/Sd is comprehended as constituting 
an instance of (L wants (L describe action/bad)) [XSA-15 = Runoff produces output on 
output file -- or -- XSA-23 = Runoff produces output on line printer]. 
A Model of Dialogue 54 
Whqt wo havc nccn, then, is thc setting upaf the expectations that the speaker will 
(i.0. wants to) dcscribc soma thincs that went right, and some that didn't. The prcaoncc 
of thew oxpcctcltions ha6 cnsblcd the aseimil~tion of tho last four uttcranccs, leading to 
thc modclys awnrcncss that for L, stcps XSA-I, XSA-I2 or -22, end XSA-16 or -24 all 
procecdcd as cxpcctcd, but that L didn't pcrceivo Runoff producing any output. 
Mcch~nisms outsidc the scapc of this oxamplo dctcrmina that XSA-15 (Runoff pr&duccs 
output on output file) was pcrccivable to L (had it occurrcd), but that XSA-23 (Runoff 
produccs output on tho line printcr) was not. This leads to the concl~sion that XSA-23 
probably was what had occurrcd, and thug lo the subsequant explanation from 0. 

REFERENCES 
Austin, J. L. How fo do things with words. Cambridge, MA: Harvard Universify Press, 1962. 
Bales, R. F. /nferactive process ana/ysis, Cambridge, MA: Addison-Wesley, 1952. 
Charniak, E. Organization and inference in a frame-like system of common sense knowledge. In -R. Schank & B. 1. Nash-Webber (Eds.), Theoretics/ issues in natura/ /anguafe processing. Cambridge, MA: Bolt, Beranek and hewman, Inc., 1975. 
Clark, H. H., & Lucy, P. Understanding what is meant from what is said: A study in conversationally convcyed requests. Journa/ of Verbs/ Learning and Vcrba/ Behavior, 1975,/4,56-72. 
Erman, L. D, Fcnncll, R. D., Lesser, V. Re,& Rcddy, D. R. System organizations for s ccch undc r st andi ng. Proceedings of the Third /nternationa/ Joint Conference on Artificial /nfel/i~encc. Palo Alto, CA: Stanford University, 1973. 
Gordon, D., & Lakoff, G. Conversational postulates. Papers from the Seventh Regionid Meeting, Chicago Linguistic Society, 1971. 
Grice, H. P. Loeic and conversation. In P. Cole R, J. L. Vargan (Eds.), Synt@x and semantics. New York: Academic Press, 1975. 
Labov, W., 81 Fanshel, 0. Therapeutic discourse. Psychotherapy as conversation. Draftcopy, 1974. 
Lcvin, Jo A. Protcus: An activation framework for cognitive process models. Unpublished doctoral discertation, San Diego, CA: Univ. of Calif, San Oiego, 1976. 
Lcvin. J. A*, & Moore, J. A. Dialogue-games: A process model of natural language i ntcr ac t i on. Proceedings of the A/SB Summer Conference. Edinburgh, Scotland, July 1976.
Lovin, Jo A*, f5 Moore, J. A. Dialogue Gamcs: Meta-communication Structures for Natural Language Interaction (ISIIRR-77-53), Marina del Roy, CA: Information Scienccs Institute, 1977. Alsoinproso,CognitiveScienc&s, 1977,//4/. 
Mann, W. C. Why things are so bad for the computer-naive user (ISI/RR-75-32). Marina del Rey, CA: Information Sciences Institute, 1975. 
Mann, W. C., Moore, J. A*, & Levin, J. A. A comprehension model for human d1.a 1 ogue. Proceedings of the Fifth /nternaliona/ JoinY Conference on Artificia//nte//igence. Carnbridge,MA:MIT, 1977,forthcorning. 
Minsky, M. A framework for representing knowledge. In P. H. Winston (Ed.), The ps~chologyofcomputervision. NewYork:McGraw-Hill, 1975. 
Newell, A. Production systems: Models of control structures. In W. G. Chase (Ed.), Visual intormalion processing. New Y or k: Academic Press, 1 973. 
Ncweli, A*, & Simon, H. A. Human problem so/vJ,?g. Englewood Cliffs, NJ: Prentice-Hall, 1972. 
Norman, D. A., Rumelhart, D E. & the LNR Research Group. Exp/oralions in cognition. San Francisco: W. H. Freeman, 1975. 
Riegcr, C. The commonsense algorithm as a basis for computer models of human memory, inference, belief and contextual language comprehension. In R. Schank & E. L. Nash-Webber (Eds.), Theoretica/ issues in natura/ /anguage processing. Cambridge, MA: Bolt, Beranek and Newman, Inc., 1975. 
Rurnclhart, D. E. Notes on a schema for stories. In 0. G. Bobrow & A. Collins (Ed s.), Representation and understanding: Studies in cognitive science. New York: Academic Press, 1975. 
Sacks, H., Schegloff, E. A., & Jefferson, G. A simplest systematics for the organization of turn-taking for conversation. Language, 1974,50, 696-735. 
Schank, R. C., & Abelson, R. P. Scripts, plans and knowledge. Paper presented at the Fourth International Joint Conferehcc on Artificial Intelligence, Tbilisi, USSR, August 1975. 
Searle, J. R. Speechacfs:Anessayint/lephi/osophyo//ang13'~ge. Cambridge, England: Cambridge University Prcss, 1969. 
Searle, 3. R. Indirect speech acts. In P. Cole & J. L. Morgan (Eds.), Syntax and semantics. New York: Academic Prcss, 1975. 
Shatz, M. How young children respond to language: Procedures for anewering. Papers and Reports on Child language Development, Palo Alto, CA: Stanford Univeroi ty, 1975. 
Thorndyke, P. W. Cognitive structures in comprehension and memory of narrative discourse. Co,~nitivePtycho/ogy, 1977,9,77-110. 
Wittgcn~tcin, L. Philosophical inve~figations (3rd ed.). New Y ~rk: Macmillan, 1958. 
Woods, W. A. Transition network grammars for natural language analysis. Communicnf ions of the ACM, 1970, /3; 5 9 1,- 6 0 6. 
