COMMUNICATING WITH MULTIPLE AGENTS* 
Elizabeth A. Hinkelman and Stephen P. Spackman 
Deutsches l'~orschungszentrum f/it Kfinstliche lntelligenz 
Sl;uhlsatzenhausweg 3, 66123 Gerlmmy 
hinkelman@dtki.uni-sb.de, stephen@_acm.org 
Abstract 
Previous dialogue systems have focussed on dia. 
logues betwe(:n two agents. Many ~q)plications, 
however, require conversations between several 
l)articipants. This paper extends speech act deft- 
nitions to handle multi-agent conversations, based 
on a model of multi-agent belief attribution with 
some unique properties. Our approach has |,lie ad- 
vantage of capturing a lnnnlmr of interesting phe- 
nomena in a straightforward way. 
Motivation 
The rise of spoken language NI, P applications has le.d 
to increased inte.rest in such issues as real time pro-- 
cessing and on-line error recovery. But dialogue is an 
inherently online process; this manifests in such linguis- 
tic phenomena as turntaking \[Sacks et al., 1974\], repair 
\[Schegtoff et el., 1977\], and content grounding \[Chu:k 
and Schaefer, 1989\]. Grounding is the phenonlenon 
that establishes shared beliefs, lbr which simply mak- 
ing or hearing an utterance does )tot suffice, It makes 
hearers into hill participants who actively signal suc- 
cess or failure of communication, as in this excltange: 
Steph: |;hat's friday at seven then. 
Lgnn : at, seven. 
Our long term goal is to show how, Ji'om the per- 
spective of a pattie|peril, one plans and acts in an envi- 
ronment with other communicating individuals, even 
when those other individuals are not perfectly reli- 
able, and ewm when the groups involw~.d may be large 
enough that it is impractical to model all participants. 
For examl)le, consider this familiar exchange, from the 
point of view of someone who remembers that |;tie next 
grou I) meeting is on tuesday: 
Jan : so we should dro I) the ram. cancel tit('. 
meeting on thursday. 
Les tuesday 
All,; tuesday 
Lou yeah. 
Jan \[yes, right. 
Ilere both our sub jeer, and another participant of_ 
fer a correction, which is confirmed by l,ou and by the 
original speaker. Other participants may I)e pre.sent. 
In this paper, we focus on the elfects of comnm- 
nicative actions on the particil)ant's model of the situ- 
at, ion. In contrast with previous diMogue work, we are 
*The wm'k underlying this paper was suppovlx,.d by a re- 
si!al'(:h gratl(;, \])'KZ I'FW 9002 0, fl'om t, he Gel'tn&n thmdes- 
nfinisterium ffir 1,'orschung mid Technologie to the 1)FK1 
project DISCO. 
interested ill czuses whe.re there are more than three 
agents, in a group of ten or more, it is hard to imagine. 
how a participant can track the beliefs and disbeliet~s 
of others accurately; it, may not even be practical to 
track who they all are. 
The advantages of analysing natural language utter- 
elites ~us coilimnnicative actioi,s are by now well unde> 
stood; they serve to slmnnarise conversations for long- 
term storage \[Schupeta, t993\], as a basis for generation 
\[Moore and Paris, 1989\], in top--down prediction of ut- 
terance flmction and structure \[Alexanderssou el el., 
1!)94\], and most importantly, to provide a represen- 
tation of natural language utterances that is uniform 
with that used in general facilities for planning and 
action \[Allen, 1{)83\]. 
We tollow \[Traum and llinkehnan, 1992\] in regard- 
ing speech acts as fully joint actions between conver- 
sational participants. Not only are joint speech acts 
co..operatively undertake.n, but they have at least nora-. 
|nelly.joint etDcts: if they complete but still fail to re- 
sult in shared goals or shared beliefs, this should be 
attributable to politeness, dishonesty (of. \[Perrault, 
1991)\]), or other social funelions. 
This perslmctiw; on speech act processing forces us 
to deal with issues of jointly hehl goals and beliefiq at a 
w.'ry b~usic level. These matters are by now quite well- 
studied, but analytic solutions from logical tirst prim 
cipies tend to be relatiw;ly complex, yielding neither 
perspicuous notations nor plausible computational or 
cognitive models. In short, normative analyses are not 
necessarily descriptive, ones. 
Aside from relatiw~ly involved calculations, there 
are sew:ral sources of dilliculty: 
• when nrultiple participants are involved, the nnm- 
I)er of 'belief spaces ) (patterns of modal embedding) 
tends to blow ul) rapidly; 
• when the actual state of atl;~irs regarding the extent 
of others' knowledge is unknown (a~s is the ease \['or 
an online converse(renal participant) the nmnber of 
cases to be considered can become large; 
,, when dealing with large organisations, some form of 
aggregate modelling I)ecomes an absolute necessity. 
Consider, for instanee, (;lie case in which you believe 
that the governnieut knows your iltcome, frol-n lash year. 
What you believe is not that each individnal govern- 
men| mnl)loyee knows it, bnt that anyone from tile tax 
department, wl,ose 1)usiness it is to know, and who 
actually wants to, will. Thus we would typically as- 
~';ilnie that an employee of the tax department who, 
in a professional capacity, lnakes contact, with your 
accountant, would actually have this information to 
hand. We want to abstract away from the commuui- 
1191 
cation channels that make this possible while retaining 
the availability of the conclusion; and we would ideally 
like to do so in a manner consistent with the needs of 
online dialogue processing. 
In the next section we describe a method of repre- 
senting the information necessary for processing mul- 
tiparticipant speech acts, one which treats groups as 
agents in their own right, and does so in such a way 
that speech acts can operate on them naturally and 
directly. 
Corporate Agents and Attitude 
Propagation 
The basis of our model is the corporate agent. These 
'agents' represent groups, but, like individual agents, 
they may have beliefs (which we write BagentP ) and 
goals (GagentP) ascribed to them directly. Thus, they 
can be thought of as intermediate in status between 
sets of component agents and prototypical group mem- 
bers. They differ from simple sets in three striking 
ways: first, they are intensional structures that may be 
distinct even when co-extensive (as, for example, when 
the members of the marketing department form a vol- 
leyball team on tuesday nights); second, attitudes are 
ascribed to them directly, and potentially at variance 
with any attitudes we might a.scribe to their members, 
or other subagents (see the discussion section); and 
third, that (other than perhaps in the case of a 'real', 
singleton agent) subsethood is not a sufficient condi- 
tion for subagency--some intramural volleyball teams 
are clearly, as social entities, agents of the company, 
and others rather the reverse. 
While not in a position to make detailed psycho- 
logical claims, we believe that structures of this kind 
are compatible with what we know about the linguistic 
and cognitive handling of aggregates in other contexts. 
These corporate agents will be used to represent 
both long-term social groups and transient groups of 
conversational participants. 
(sfe~veJa~y~S also a subagent ~ 17 eryonc 
D_J =t U' U U 
Figure 1: Agents 
In the remainder of this paper we illustrate rela- 
tionships between agents with diagrams such as that in 
figure 1. Here the playing-card shapes represent agents 
and the heavy lines connecting them the subagent rela- 
tion (_): the agents include the system itself (sec.Jan), 
the system's boss (Jan), Jan's office, their coworker 
Les, their common corporate employer (WidgetCorp), 
and another 'random' person, Steph, who does not be- 
long to WidgetCorp. Later we will represent attitudes 
ascribed to agents by small stlapes within the playing 
cards and their propagation by thin curved arrows be- 
tween them. 
Note that since we are discussing the system's own 
representation of the world, the double-bordered play- 
ing card really represents the system's self-model and 
the whole diagram the system; but we shall not be- 
labour the point here, assuming that information pass- 
es freely between the two. 
The model we use to compute the transfer of attitudes 
between agents is approximate, partly to simplify com- 
putation and partly because it is in any case unusual 
to have enough empirical data to compute an exact so- 
lution. The same model (with parametric variation in 
the domain knowledge) is applied to all the attitudes 
the system ascribes to agents; in our present imple- 
mentation, these may be either beliefs or goals) Un- 
like representations based on conventional logics of be- 
lief, it does not introduce nested contexts unless they 
are explicitly represented in the content of utterances 
(as would be the case with a sentence like "But Lynn 
thinks we should have the meeting anyway."), though 
extended reasoning processes may compatibly do so. 
In this simplified model, the propagation of atti- 
tudes is performed lazily, as determined by their se- 
mantic relevance to the varions agents involved. Ide- 
ally, this judgment would derive from social world- 
knowledge and information about the purposes of 
groups; in our current implementation it is approxi- 
mated using a simple classification of corporate agents, 
participant roles and message topics into the domain 
ontology. Delays and chance are not modelled; all rel- 
evant attitudes are presumed to propagate between 
agents, subject to the following constraints: 
idgetCorp 
"'"/\] ~ i'"'"'"'">'x., impossible- 
/ / /.. ",, 7L27 "°" 
~_~'sec'Jan UJan OSteph 
Figure 2: Common context constraint 
1Since our nlodel is not analytic we do not want or need 
a notion of 'knowledge': the system lacks direct access to 
empirical truth and to social consensus, and does not have 
the cognitive sophistication to validate argmnents against 
stone independent notion of rationality. In short, none of 
the usual theories of truth can in principle be made to 
apply. 
1192 
l. Attitudes propagate only between superageut and 
subagent (or vice-versa). This stipulation anlounts 
to saying that comnmnication only occurs between 
agents in the presence of some common social con- 
text. Of course, new corporations can be introduced 
when new social groups form. 
Thus in ligure 2, beliefs ascribed to WidgetCorp 
can propagate to the subagents, Jan and Jan's elec- 
tronic secretary; but they do not reach Steph, who 
is not a subagent of WidgetCorp. ~ We use the con- 
vention that attitudes are drawn filled in if they 
are known by direct evidence, hollow otherwise; and 
that dotted structures are 'negated'---they do not 
arise ms drawn because they violate some constraint 
under discussion. 
~i filet'Jan 
'''' conflicting prior J ~blocked- belief 
~____~sec'Jau W Jan 
Figure 3: Propagation as default 
2. Attitude propagation is only a default (the particu- 
lar choice of default logic need not concern us here). 
If there is direct evidence for ascribing a contrary 
attitude to an agent, propagation from an exter- 
nal som:ce is inhibited. This property is crucial to 
modelling dishonesty, negotiation, compromise, and 
atypical group members in general. 
Such blocking is illustrated in figure 3. In this 
ease our model of Jan fails to inherit a goal from 
otfice.Jan because it conflicts with another goal (the 
square box) for the ascription of which to Jan we 
have prior independent evidence. 
3. The system may never assume that its own attitudes 
antomatically propagate upwards to a snperagent. 
The ut)ward attitude propagation path models tt,e 
effect of external agents independently attending to 
their various communicative goals, but the system 
nmst still plan--and execute- its own actions. 
Thus, in figure 4 the system--see.Jan--is pro- 
hibited from simply assuming that its own beliet~ are 
shared by its employer, though those of fellow em- 
ployees would be propagated when otherwise consis- 
tent. (Some hmnans seem to relax this constraint.) 
2In order to place some limit on the promiscuity of atti- 
tude propagation, it seems best to insist that indirect trans- 
fer must occur through a siligle agent that is a transitive 
Sllper- or Sill)- agellt of both terniinal agents, Thus, even 
if Jan and Steph both belonged to some peer of Widget- 
Corp with a similar seniantic domain, propagation would 
still be not permitted along the resulting N-shaped path. 
Gonlmon membership in Everyone will not transmit beliefs 
eithe.r, because its relewuice tilter is maximally restrictive. 
/ ffice.Jan //,// \ 
requires explicit action 2 
U'" 
Figure 4: The exceptional natnre of self ~ 
office.Jan 
Figure 5: The need for speech 
4. Nonce corporations, introduced dynamically to rep- 
resent the temporary grouping of participants in 
an active conversation, never inherit attitudes from 
their subagents, but must acquire them <as the ef- 
fects of observable actions. The idea here is that 
while participating in (or directly observing) a con- 
versation, the system is in a position to observe the 
construction of the public record of that conversation 
\[Lewis, 1983\] directly, and this record consists exact- 
ly to the attitudes we wish to ascribe to tile conver- 
sational group itself. In conversation even a new 'un- 
spoken understanding' should be based on inference 
fi'om observed communication, and not just the sys- 
tem's private beliefs about other participants' views. 
The fact that we still permit conversational 
groups to inherit from superagents allows us to place 
a discussion within a social context that supplies 
shared background a,ssumptions. The fact that we 
permit their subagents t,o inherit from them models 
the actual adoption of information from the public 
record by individual participants, including the sys- 
tem itself, without additional mechanism. 
Figure 5 depicts this situation: noncel, the coll- 
versational grouping, represents a shared social con- 
strnct distinct from our understanding of Jan's pri- 
vate views. This allows us to deal gracefully with 
the situation in which we, see.Jan, catch (or perhaps 
even conspire with) Jan in telling a lie. 
The most important property of this model of at- 
titude ascription is that the only belief spaces it im 
troduces are those that are independently motivated 
1193 
by identified social groupings or the records of the ac~ 
tual conversations in which the system participates. 
This reduces the chance that the system will become 
mired in irrelevant structural detail, and specifically 
avoids the 'spy novel' style of belief space nesting that 
is characteristic of cla,ssical normative models. Attri- 
bution by default inference allows an individual to be 
represented as a member of several different groups 
holding conflicting beliefs, and inheriting only those 
beliefs consistent with those represented as being held 
privately. 
The results are thus substantially different from 
those obtail, ed in classical logics \[Allen, 1983; Kraus 
and Lehmann, 1988; Appelt, 1985; Cohen and 
Levesqne, 1990\]. They differ from other path-based 
algorithms \[Ballim and Wilks, 1991\] in the provision 
of semantic relevance conditions and in addressing the 
need for shared attitudes by ascribing them directly to 
groups, rather than by maintaining complex accounts 
of which agents believe what. This allows us to de- 
scribe and process conversational mechanics withont 
recourse to nested (x believes that y believes that...) 
belief spaces, though snch structures may remain nec- 
essary for other, less routine feats of cognition. 
In the next section we show how our model of at- 
titude ascription can be used to implement multipar- 
ticipant speech act processing. 
Multiparticipant Speech Acts 
As in \[Traum and Hinkelman, 1992\], we assume that 
a core speech act is ultimately realised by a sequence 
of utterances that ernbody the grounding process. The 
model requires thai; the definitions of the speech acts 
themselves abstract away from grounding and provide 
high level actions that can be integrated with non- 
linguistic domain actions in planning. Using our multi- 
agent attitude attribution mechanism, we can simplify 
matters fllrther, defining speech acts as joint actions 
whose effects apply directly to the conversational group 
being modelled. 
Consider the generalised action operator repre- 
senting one simple core speech act: 
Informal) 
conditions : BbPA b K aAlivea 
effects : Bap 
This Inform is a true joint action. Agent a is the 
nonce corporation representing all the participants tak- 
en jointly (the live predicate requires that this nonce 
correspond to an a ongoing conversation). Though the 
singleton subagent b is the source of the information, 
the action has its effect directly on our model of the 
group. From that point propagation downwards to the 
individual participants is a function of the attitude as- 
cription model, and is subject to the constraints giv- 
en above. (The system effectively assumes that corre- 
sponding updates actually take place in the minds of 
the conversational participants.) 
The correctness of this formulation relies on two 
facts. The first is that the grounding structure realis- 
3Our current implementation actually deals with (mail 
rather than live speech, and must cope with mul(.iple acl;ive 
dialogues. 
ing the core speech act operator ensures the content is 
successfully conveyed. The second is that if a speech 
act that has an efl~ct on the conversational gronp is 
flflly realised and properly grounded, then any hearer 
who dissents from those effects must explicitly act to 
cancel them. That is, acknowledgement of receipt of 
a message establishes a presumption of assent to the 
content. Note, however, that when the speech act re- 
mains unchallenged this means only that the conversa- 
tional participants will let it stand as part of the public 
record; it does not mean that they are tndy persuaded 
of its content, and the rules we have given only pre- 
dict that they adopt it if there is no evidence to the 
contrary. 
St, ecessful requests have effects on the goals rather 
than the beliefs of the group. It is crucial that both 
communicative and noncommnnicative are introduced. 
The first goal below is noncommunicative and repre- 
sents simply that the requested action be performed. 
Note that although the requested action's (possibly 
corporate) agent participates in the dialogue, there is 
no restriction that it not include the requester. Writ- 
ing Bifap for Bop V Ba~p and O for eventually, we 
have 
Requestae 
conditions : agent e E a A live a 
effe.cts : (~a<~ e A GaBifa@ e 
The second goal in the effects is a communicative 
one; the group acquires the goal of finding out whether 
the requested action will be performed. The conse- 
quence of this is that the requester gets an indication 
of whether the request was successful: even lmder the 
assunrption of co-operativity, goal conflicts and plan 
constraints sometimes lead to the rejection of a suc- 
cessfully communicated request. 
In tile next section we describe how OOSMA, our 
implemented calendar management system, processes 
all actual exchange. 
Processing an N-Way Speech Act 
Speech acts like the above can now fignre in the plan- 
and inference- based algorithms of communicative in- 
telligent agents. Since dialogue may include unpredict- 
ed events, such agents must be able to react to chang- 
ing circumstances rather than relying completely on 
advance planning. As each incoming speech act ar- 
rives, the agent updates its beliefs and goals; these be- 
liefs and goals are the basis for subsequent action. This 
is not only appropriate for the interface between task 
and dialogue, but absolutely crucial for the grounding 
process. 
A typical application task for Noway speech acts 
in a multiagent environment is appointment schedul- 
ing, with dialogue systems serving as personal appoint- 
ment secretaries to human agents. Our implemented 
system, COSMA, operates in this domain. We model 
a human/secretarial pair as a kind of corporate agent 
in which beliefs about appointments propagate up and 
down from both members, and in which goals about 
appointments propagate fl'om the hnman up to the pair 
and from there down to the secretary. 
When this example begins, the dialogue system 
1194 
(see.Jan) has the role of a personal aplmintment sec- 
retary to a human agent (Jan), forming the hu- 
man/secretarial corporation ofliee..hm. Jan sends 
sec.aaa email text; referring to a pre-existing appoint- 
Inent: 
Jau: Cancel tit(', meeting with the 
hardware group. \[--',sec.a an\] 
'\]_'he COSMA system interprets this input by first 
constructing a nonce corporation for the new dialogue, 
nonce1, with 
dan, see.Jan E nonce/ E ofl~ce.aan 
Q =GOcancel~e~.ja.meeting2 ~ office.Jan 
i =G Blrn~*l~canc°l~'l=" meeling2 
- :~. Transmission by core al~eech act i~ L _~ 
ec.Jatl Jan 
Figure 6: Making a request 
All members of a nonce corporation inherit belief~ 
and communicative goals from it;. The interpretation 
of the first utterance as a speech act is: 
Request{Jan, sec.Jan} Cancel sec.danmeeting2 
This interpretation is checked for consistency with 
context according to the method of \[llinkelmau, 1992\], 
and forms an acceptable reading. Its effects on the 
group are asserted (Q, @ in figure 6): 
G norn:e l ~ ca heel see.dan meeting2 
G nonce l B if nonce l 0 cancelsec.3a n meeting2 
- _~officc..lau ~ =(y ~'cancol 8ec,l~n lnteetlllg 2 
I=GB r Ocancel . meethlg2 
A =Ocancol ~eaanltlemhng2 
/ / 
g, 
Figure 7: t/.esponding 
When it has finished processing all inputs, the sys- 
tem exalnines its own goals in order to determine what 
actions, if any, it will perform. It finds no immediate 
priwtte goals, but there are two that it inherits. Be- 
cause it is a participant; in the ongoing discussion with 
Jan it inherits the nom'e's communicative goal gif... 
(@ in figure 7). It also inherits the goal to ensure 
that the cancellation actually does happen (@)4 (A 
less compliant agent thau the current COSMA wouhl 
not acquire non-communicative goals directly from the 
nonce, but would obtain the cancellation goal indirectly 
through office.Jan. The implementation could be w.'ry 
similar, because the indirect inheritance path can be 
compiled out when the nonce is initially constructed.) 
The dialogue system thus retrieves the following 
goals: 
Gsec.Jan ~ canc elsec.Janmeeting2 
Gsec.,lanBifnoncel 0 cancelsec.danmeeting2 
These goals become input for the planning process. 
The first goal can be achieved in the current context 
by ltrst opening the appointment file, then perform- 
ing a stored subplan \['or cancelling appointments that 
includes modifying the database entry and notifying 
the participants. Our reactive algorithm allows com- 
nlunicative phms to be freely embedded within domain 
actiotls, and vice versa. 
ltaving found this sequence of actions, the system 
now knows that the ~> ... part holds. It is therefore able 
to plan to Informnoncel(O ...), satisfying the second 
goal (@). The output for the second goal is: 
hfformnoncel (O ca ned see.dan meeting2) 
see.,lan: Ok, I'll cancel it. \[--~-3all\] 
WidgctCorp 
Figure 8: Informing the group 
Finally, it must conlph.'te the execution of its plan 
to satisfy the first goal by updating the appointment 
file and notifying MI participants. The notification step 
involves constructing a suitable conversational I:tOltce, 
this time a descendant of WidgetCorp itself (in spo 
ken dialogue this requires, aside front setting up the 
necessary internal data structures, meeting the ad- 
dressees and greeting them; when communicating via 
email the analogous requirement is just composing a 
suit~ble mail message header). Then, as show in figure 
8, the system initiates a further Inform action of its 
OWll', 
Informnonce4 (-,7) meeting2) 
whk:h (:an be verbalised as follows: 
see.dan: The meeting of Monday, Feb. 13 
at 3 I'M will not take place. \[~sec.aan, 
Jan, gou, Le G Lee\] 
4Nol.e that if the system were asked, it could now inflw 
that, Jan also has these goals, but tohat this is not part of 
the speech act interpretation algorithm itself. 
1195 
Discussion 
An important property of the corporate agent model 
presented in this paper is its scaling behaviour. Al- 
though the number of 'agents' in a nontrivial world 
model may be large, we only introduce belief spaces 
corresponding to 'actual' objects about which the sys- 
tem has knowledge. In particular, the corporate agents 
that are used correspond to either durable social gronp- 
ings or records of actual conversations. Individual 
speech act definitions, though they acconnt for all the 
agents in the dialogue, need make reference to at most 
two agents. 
In contrast with normative models, our speech act 
processing model at no point requires that individual 
addressees be modelled. Of course, dialogue is typical- 
ly motivated by the desire to modify the addressees' 
mental states, but our system is free to make these 
updates on demand. Thus, so long as constructing de- 
tailed partner models is not independently necessary, 
the effort required to plan and respond to speech acts 
remains almost constant as the number of conversa- 
tional participants grows. 
We have thus achieved the extension of speech 
acts to multiagent environments, a step beyond other 
speech act based models\[Dols and van der Sloot, 1992; 
Meyer, 1992; Bunt, 1989; Traum and Ilinkehnan, 
1992\]. In the process, we have reduced the complexity 
of the task/dialogue interface. 
WidgetCorp 
~~ ~'- I nonce5 
Figure 9: WidgetCorp expands 
An interesting consequence of not needing to nlodel all 
members of a conversational group is that it becomes 
unnecessary to identify them. While in some circum- 
stances this may be an advantage, it does leave the 
door open to an interesting glitch: without an indepen- 
dent check, the system's model of who it is addressing 
may turn out to be inaccurate. A related thing can 
happen when the system plans on the basis of attitude 
propagation: it can perform an action that 'ought' to 
result in a given agent's coming to hold some view 
through social processes, but since social channels are 
quite imperfect, the message never gets through. Hu- 
man agents are at times more cautious, and may mod- 
el delays in the grapevine, but this 'lost sheep' phe- 
nomenon occurs sufficiently ofl;en in real life to make 
the utterance "Oh, I'm sorry, I guess you weren't at 
the meeting." sound very familiar. 
Generalisatiou remains a hard problem, of course. 
Our system has no special advantages when faced with 
a question like "Are conservatives porcupine-lovers?" 
Vague questions about large groups require extensive 
search or some independent theory of generalisation, 
and seem to be difficult even for hunians. 
Related to this is the Nixon diamond anomaly 
faced by many default inference systems. In our case, 
when we find propagation paths that will support the 
ascription of contradictory attitudes to a single agent, 
how should we choose between them? It turns out that 
selecting whichever result we first discover we would 
like to use is a surprisingly good solution. Such 'ar- 
bitrary' judgments tend to facilitate the conversation 
by using the inference mechanism not to seek a reli- 
able proof but to find the most convenient supportable 
argument, regardless of actual truth. 
Perhaps the most interesting deviation of our model 
from the behaviour of systems founded on mutual be- 
lief is the social fiction anomaly: one can fairly e~i- 
ly reach a state in which an attitude is ascribed to a 
corporation which is held by none of its members. In- 
credibly, this also corresponds to a familiar situation in 
everyday life. Three examples should serve to illustrate 
the point. In the first place, consider this exchange, in 
which Jan asks Les to compile a tedious report: 
Jan: I'll need that on my desk by friday. 
Les: friday, no problem. 
Such a dialogue may occur even when Jan will not 
have time to look at the report until the middle of the 
following week, and Les knows that the work cannot 
possibly be completed before the weekend. '~ 
Secondly, we propose the example of a couple who 
are jointly but not severally on a diet. When together, 
neither partner ever takes dessert, and this policy is 
verbally reinforced. Yet either of' them will happily join 
you in eating a large slice of strawberry cheesecake, if 
they are apart. 
Finally, imagine that you are a lone bicyclist ap- 
proaching a Very Large Hill. You might now say to 
yourself--a conversational nonce of one--"It's not far 
to the top!" Processing this speech act results in an- 
other unsupported belief. You can now try to be con- 
vinced. 
In light of all of tile above anomalies, it begins to ap- 
pear that human agents may be struggling with limi- 
tations similar to those of our own model. 
It may still be objected that onr model falls short 
in failing to support the detailed 'spy novel' reason- 
ing used in conventional logics of belief, keeping track 
of whether Loll believes that Lee believes that p, and 
whether or not common beliefs are truly mutual. Our 
response is threefold: 
* Reflective l)rol)lem solving is always an option, but in 
humans appears to be an independent process. Re- 
sponsiveness demands make it unsuitable for manda- 
tory online use m dialogue processing, though it may 
be important to use models (like ours) with which it 
can be integrated simply. 
* Conversational mechanisms have evolved to cope 
with the cognitive shortcomings of humans, qb the 
SThe authors disagree as to whether t, his par~,iculm' pat- 
tern is more likely to arise througla malice or optinlism. 
7196 
extent that the pertBrmance errors of a dialogue 
agent mirror huinan failings, co-operatiw.' recov- 
ery performance wil;h hlltrla\[l communicative tools 
should be enhanced. 
• Even given access to an ideal normative dialogue 
model, a fltll system would benetit li'om running a 
less precise and lnore descriptive model in parallel. 
This would a~ssist in isolating those parts of a com- 
municative plait where confllsioll on tim part of other 
agents is predictable. 
Corporate agents are an alternative to normative 
logics of belief which capture a mind)or of interesting 
social and commtmicative phenonmna straightforward- 
ly. With their help, we can refornmlate core speech act 
delinitions cleanly and seahd)ly tbr the case of nrany 
agents. The level of planning abstraction that results 
seems well-suited to the needs of intelligent commu- 
nicative agents operating in an environment that in-. 
(:htdes I/laity lnlnlall agents. 
References 
\[Alexandersson ctal., 1994\] Jan A lexandersson, l';lis- 
almth Maier, and Norbert l{eithinger. A robust 
and eflicient three-layered dialog component for a 
speech-to-speech translation system, subm. (Jon- 
ference on Applied Natural Language Processing, 
Stuttgart, 1994. 
\[Allen, 1983\] James Allen. ll.eeognizing intentions 
from natural langm~ge utterances. In Michael Brady 
and l{obert C. Berwiek, editors, Computalioual 
Models of Discourse. MIT Press, 1983. 
\[Appelt, 198'5\] D. E. Appelt. Planning English S'en- 
tences. Cambridge University l'ress, Can,bridge, 
1985. 
\[Ballim and Wilks, 1991\] At~al Ballim and Yorick 
Wilks. Arti\]icial Believers: The Aseriplion of He- 
lief. Lawrence I';rlbaum Assoeiates~ 1991. 
\[Bunt, 1989\] II. C. lhmt. Information dialogues as 
communicative action in relation to t)artner mod- 
elling and inforntation l)roeessing. In M,M Taylor, 
1". Neel, and 1). G. llouwhuis, editors, 7'he Struetm'e 
of Multimodal Dialogue. Elsevier Science Publishers 
I/.V., :1989. 
\[Clark and Schaefer, 1989\] flerbert \[1. ('lark and l';d 
ward F. Schaefer. Contrilmting to discourse. (,'oqui- 
live Science, 13:259 294, 1.98!/. 
\[Cohen and Lew'.sque, 11990\] Phillip R. Cohen aud 
Hector J. Levesque. I'ersistence, intention, and eoln- 
mitment. In P. R. Cohen, ,/. Morgan, and M. E. 
Pollack, editors, Intentions in Communication. MIT 
Press, 1990. 
\[l)ols and van der Sloot, 1992\] I,'. J. I1. l)ols and 
K. van der Sloot. Modelling mutual elfi~et,; in beliet: 
based interactive systems. In W. J. Black, G. Sabah, 
and T. J. Wachtel, editors, Abduetion, Beliefs and 
(;ontext: Proeeedings of the seeotM 1'2,S'111117 ' PLUS' 
workshop in computational pra.qmatie.s, 1992. 
\[llinkehnan, 1992\] Elizabe.th A. llinkehnan, intona- 
tion and the Request/(,~uestion l)istinetiou. In Pro- 
ceedings of the h~ternational Co,fi~reuce on ,~,'poken 
Language Processing, P, anlf, Canada, 19!12. 
\[Kraus and l,ehmann, :1988\] Sarit graus and DauM 
l,ehmann. Knowledge, belief and time. In "\['hcore& 
ical (Jomputer ,5'cieuce, vohune 58, pages 155-174, 
1988. 
\[Lewis, 198:t\] I). Lewis. Scorekeeping in a Language 
Game, volume 1, chapter. 13. Oxford University 
Press, Oxford, 1983. l)hilosophical Papers. 
\[Meyer, 1992\] lt.alph Meyer. Towards a conceptual 
model of belief and intention in dialogue situations. 
In W. J. Black, G. Sabah, and T. J. Wachtel, editors, 
Abduction, Beliefs and Context: PTveeedings of the 
second E,qPRIT PLU,q workshop in computational 
pragmatics, 1992. 
\[Moore and Paris, 1989\] J.D. Moore and C.L. Paris. 
Planning text for advisory dialogues. In Proe. Conf 
of the 27th Ammal Meeting of the ACL, pages 203- 
211, Varmouver, l !/8!). 
\[Perrault, 1!)90\] C. II. l'erranlt. An al)plication of de- 
fault logic to speech act theory. In P. \[{. Cohen, 
J. Morgan, ~tnd M. E. Pollaek, editors, Intentions in 
Communication. MIT Press, 1990. 
\[Sacks et al., :1974\] 11. Sacks, E.A. Schegloff, and (I.A. 
Jef\[hrson. A simplest systematies for the organi- 
zation of tin:n-taking in conversation. Language, 
50:696 7a5, 1974. 
\[Schegloff el al., 1977\] E. A. Schegloff, G..letferson, 
and I I. Sacks. The preference for self correction in 
the organization of repair in conversation. Language , 
5:t:36l ,'t82, 1977. 
\[Schnpeta, :199:{\] Achinl W. Sclmpeta. Cosma: F, in 
verteilter terminplauer als fidlstudie der verteilten 
ld. In Jucrgen Mueller and l)onald ,qteinel:, ed- 
itors, lk'oopericrende A.qenten, Saarbriieken, Get:- 
many, 1!19:t. 
\[Tranln and llinkehnan, 19!12\] I)avid 11.. Traum and 
Elizabeth A. Ilinkehnan. Conversation acts in task- 
oriented spokeu dialogue. Computational Intelli- 
gence, 8(:11:,575 .599, 1992. Special Issue on Non- 
literal language. 
1197 
