Speech Acts and Rationality 
Philip R. Cohen 
Artificial Intelligence Center 
SRI International 
and 
Center for the Study of Language and Information 
Stanford University 
Hector J. Levesque 
Department of Computer Science 
University of Toronto" 
1 Abstract 
This pallet derives the ha.sis of a theory, of communication from 
a formal theov,.' of rational interaction. The major result is a 
<h, mon~t fallen t hat. ilh,c,tionary acts need not I)e primitive, and 
.ee, I uot he reco~'nized..\s a t,'st case. we derive Searle's con- 
dit ions on reqt,est in~ from pri,ciples of ralionality coupled with 
a ~;ric~,an theory of iml~erativ,.s. The theory is shown to dis- 
tingui.~h insincere or nonserious imperatives from tr~le requests. 
\['~xlensions to indirect .~peech acts. and ramifications for natural 
language ~ystcms are also brieily discussed. 
2 Introduction 
'\]'he tlnifyin~ tilt'me of m,wh c-trent pragmatics antl discourse 
re~earrh is that the c.herence .f dialogue is to he folnnd in tile 
iuleraclinn of the cottver~alll'~' 1~61rI.I. Thal is, a speaker is re- 
garded a~s planning his ,lllcrance,~ re achieve his goals, which 
n,ay involve in{h..lwing a hean'r by the ,,se of comm,micative 
or "speech" acts. (-)u receiving an lltler~tnce realizing such an 
action, the hearer altempls Io infer the ~peaker's goal(s) anti to 
qndeffland how the 11llerat|rv fnrthcrs them. The hearer then 
adopts new goals (e.~.. to re-pond to a reqllest, to clarify the pre- 
vious ~peaker'~ lllll'r~ince or ~.:f,al) and plan~ his r~wn utterances 
to acl,ie:'e those. :\ cotl,cel'?~alion enslle~, I 
This view of language a.~ p.rposefid art ion has pervaded ('om- 
putational I,inzui-~ics re~carch, and ha.~ re~,lted in numerous 
protoCyl~e systems \[I, 2, 3..',. 9. 25, 27\]. llowever, the formal 
foundations underlying 01n... %v~l.ems haw" heen unspecified or 
.nder~peril'ied. In this ,.late ,~\[' affairs, one cannot characterize 
what a ~,y.',tem .~llould ih~ independently from what it does. 
This paper hl,gins to rectify this sit-ation by presenting a 
fl~rmalizalinn of rational interaction. ~pon which is erected tile 
he~itmin~'- r,f a theory of rein m~miralion attd ~peech acts. Inter- 
;wtion is d~.riv~,d fr~,m prmcil~h,.~ of rational action for indivi,h,al 
a~enas. ~.. well as lwinciph's -\[ helief and goal adoption among 
a~enls. The h~sis of a theory nf purposefi,l communication thus 
"F, ll~,w ,,I' th~ Canadian lr, sti~,~t~- f~)r A,'.wanc~d R-search. 
~This re~,.areh was mad- W,~sdde ;n part hy a gilt from ~he Systems Dew.l- 
opm~.n~ \["~.md:~ti,,n. and in part t,y suFport fr-m ti~e r)efens~ Advanced 
R~.se~rrh \['roje.rts .Ag,ncy un.h'r C,~n~ra.ct Nf~I)t)3D.8.I-K-0078 wilh the 
.N~v:~| \['~lec~ronic Systems C,,mm~nd. The views and om¢lusions eon- 
tain~'d in thls document ~re ~hos~" of the ~uthor~ and should not be inter- 
preted ;~ representa, tive of the. omci~.| policies, ~ither expre~ed or implied, 
oi" the Defense ~dvanced Research Projects Agency or the United States 
(Jovernment. Mu~h nf this rrsearrh was done when the second a.uthor 
wa~ employed at the Falrehild ('~m,r~ and Instrument Corp. 
emerges as a consequence of principles of action. 
2.1 Speech Act Theory 
Speech act theory was originally conceived a~s part of action the- 
ory. Many of Austin's \[.l\] insights about the nature of ~peech 
acts, felicity conditions, anti modes of lath,re apply equally well 
to non-communicative actions. Searle \[2G\] repeatedly mentions 
lhat many of the conditions he attributes to variol,s illocution- 
ary acls (such as requests anti qm,stions) apply more ~e:.,rally 
to non-communicative action. \]lowever, re~earcher~ have ~rad- 
ually lost ~ight of their roots. In recent work \[3~ I illoc,ltior,a~" 
acts are formalized, antl a logic is proposed, in which propertie~ 
of IA's (e.g., "preparatory conditions" and "mode~ of achieve- 
ment') are primitively stip.laled, rather than derived front more 
h~ic principles of action. We helieve this approach misses sig- 
nificant generalities. "\['hm paper ~hows how to derive properties 
of illocutionary acts from principh,s of rationality, .pdating the 
formalism of \[10J. 
Work in Artificial Intelligence provided the first forntal 
gro.nding of speech act theory in terms of plannin~ and plan 
rerog~nitmn, cldminalin~ in Perra.h and \lh.n'~ \[:2:~ I I I...ry of 
indirect speech acts. Xhwh ~,I" o~0r re~earch i~. in.~lfir~'d I,~ lhrir 
analyses, llowe~er, one major ingredien! ~I" their the.ry r:m be 
shown to he redundant in01 illocutionary acts. All do. in- 
ferential power nf the recolfnition of their dloc~itionary acts wa.s 
already available in other "operators'. Nevertheless, the natu- 
ral langlnage systems based on this approach \[I. ,-3\] always had 
to recognize which illocutionary act was performed in order to 
respond to a tnser's utterance. Since the illocutionary acts were 
unnecessary for achieving their ell'errs, so too wa.~ their re~'n~ni- 
tion. 
The stance that illocutionary arts are not primitive, and need 
not he re;og'nize(l, is a lih..ratmg one. ()nee taken, it l)ecomes 
apparent that many of the (lifl~cuhies in applying ~l),,ech act 
theory to discourse, or to computer systems, stem from taking 
these acts too seriously - i.e., too primitively. 
3 Form of the argument 
We show that illocutionary acts need not be primitive hy de- 
riving Searle's conditions on requesting from an independently- 
motivated theory of action. The realm of communicative action 
is entered following Grice \[13i -- by postulating a correlation 
between the ,ntterance of a sentence with a certain syntactic fea- 
ture (e.g., its dominant clause is an imperative) and a complex 
49 
propositional attitude expressing the speaker's goal. This atti- 
tude becomes true as a result of uttering a sentence with that 
feature. Because of certain general principles governing beliefs 
and goals, other causal consequences of the speaker's having the 
expressed goal can be derived. Such derivations will be "summa- 
rized" as lemmas of the form "If (conditions) are true, then any 
action making (antecedent) true also makes (consequent) true\] 
These lemmas will be used to characterize illocutionary acts. 
though they are not themselves acts. For example, the lemma 
called REQUEST will characterize a derivation that shows how 
a heater's knowing that the speaker has certain goals can cause 
the hearer to act. The conditions licensing that chain will be col- 
lected in the REQUEST lemma, and will be shown to subsume 
those stipulated by Searle \[261 as felicity conditions. However, 
they have been derived here from first principles, and without 
the need for a primitive action of requesting. 
The benefits of this approach become clearer as other illocu- 
tionary arts are derived. We have derived a characterization 
of the speech act of informing, and have used it in deriving 
the speech act of questioning. The latter derivation also allows 
us to disting~tish real questions from teacher/student questions, 
and rhetorical questions. However. for brevity, the discussion of 
the.,e speech acts has been omitted. 
Indirect speech acts can be handled within the framework. 
although, again, we cannot present the analyses here. Briefly, 
axioms similar to those of Perrauh and Allen {22\] can be sup- 
plied enabling one to reason that an agent has a goal that q, 
~iven that he also has a goal p. When the p's and q's are them- 
selves goals of the hearer (i.e.. the speaker is trying to get the 
hearer to do something), then we can derive a set of lemmas for 
i,,lirect requests. Many of these indirect request lemmas corre- 
spond to what have been called %herr-circuited" implicatures. 
which, it was suggested \[211 underlie the processing of utterances 
of the form "Can you do X?'. "Do you know y?", etc. l,emma 
formation and lemma application thus provide a familiar model 
of-herr-circuiting. Furthermore. this approach shows how one 
ran use general purpose reasoning in concert with convention- 
alized b~rms (e.g., how one can reason that "Can you reach the 
salt" is a request to pass the salt), a problem that has plagnwd 
most theories of speech acts. 
The plan for the paper is to construct a formalism based on 
a theory of action that is sufficient for characterizing a request. 
Most of the work is in the theory of action, as it should be. 
4 The Formalism 
To achieve these goals we need a carefl:lly worked out (though 
perhaps, incomplete) theory of rational action and interaction. 
"!'he theory wil~ be expressed in a logic whose mndet theory is 
ba.,ed (loosely) on a possible-worlds semantics. We shall propose 
a logic with four primary modal operators -- BELief, BMB, 
~,f)AL. and AFTER. W~th these, we shall characterize what 
agents need to know to perform actions that art, intended to 
achieve their ~oals. The .zgents do so with Ihe knowledge that 
other agents operate similarly. Thus, agents have beliefs about 
.'her'~ gcals, and they have goals to influence others' beliefs 
and goals. The integration of these operators follows that of 
Moore {20l, who analyzes how an agent's knowledge affects and 
is affected by his actions, by meshing a possible-worlds model 
of knowledge with a situation calculus model of action \[18\]. By 
adding GOAL, we can begin to talk about an agent's plans, 
which can include his plans to influence the beliefs and goals of 
others. 
Intuitively, a model for these operators includes courses of 
events (i.e., sequences of primitive acts) " that characterize what 
has happened. Courses of events (O.B.e.'s) are paths through a 
tree of possible future primitive acts, and after any primitive act 
has occurred, one can recover the course of events that led up 
to it. C.o.e.'s can also be related to one another via accessiblity 
relations that partake in the semantics of BEL and GOAL. Fur- 
ther details of this semantics must await our forthcoming paper \[17\]. 
As a general strategy, the formalism will be too strong. First, 
we have the usual consequential closure problems that plague 
possible-worlds models for belief. These, however, will be ac- 
cepted for the time being. Second, the formalism will describe 
agents as satisfying certain properties that might generally he 
true, but for which there might be exceptions. Perhaps a process 
of non-monotonic reasoning could smooth over the exceptions, 
but we will not attempt to specify such reasoning here. Instead, 
we assemble a set of basic principles and examine their conse- 
quences for speech act use. Third, we are willing to live with the 
difficulties of the situation calculus model of action - e.g., the 
lack of a way to capture tnse parallelism, and the frame prob- 
lem. Finally. the formalism should be regarded as a de,~eription 
or specification Bran agent, rather than one that any agent could 
or should use. 
Our approach will be to ground a theory of communication in 
a theory of rational interaction, itself supported by a theory, of 
rational action, which is finally grounded in mental states. Ac- 
cordingly, we first need to describe the_behavior of BEL, BMB. 
GOAL and AFTER. Then, these operators will be combined 
to describe how agents' goals and plans influence their actions. 
Then. we characterize how having beliefs about the beliefs and 
goals of othe~ can affect one's own beliefs and goals. Finally, 
we characterize a request. 
To be more spe~iflc, here are the primitives that will be used, 
with a minimal explanation. 
4,1 Primitives 
Assume p, q, ... are schema variables ranging over wffs, and 
a, b • • are schematic variables ranging over acts. Then the 
following are wlfs. 
4.1.1 tVffs 
~p 
{p v q} 
(AFTEI'~. a p} - p is true in all courses of events that obt,-,in from 
act a's happening';, (if a denotes a halting act). 
(DONI:'. a) - The event denoted by a has just happened. 
(AGTa x) - Agent xistheonly agent of act a 
a ~ b -- Art a I)r~cedes act b in the current course of events. 
3 z p ,~here p contains a free occurrence of variable z. 
x-~.y 
True. False 
(BEL x p) - p foUows from X'S beliefs. 
{~OAL x p) -- p fotlotps from x's goals. 
{BMB x y p} .- p/~llows from x's beliefs about what is mutually 
believed by x and y. 
:P'w chls paper, the only events that will be considered &re primitive acts. 
3Th&t is. p is true in ~.11 c.o e.'s resulting from concatenating the current 
c.o.e, with the c.o.e, denoted by a. 
50 
4.1.2 Action Formation 
If a, b, c, d range over sequences of primitive acts, and p is a 
wff. then the following are complex act descriptions: 
a:b -- sequential action 
a \[ b -- non-deterministic choice (a or b) action 
p? -- action of positively testing p. 
def (IF p a b) -- conditional action = (p?:a) 1 (~pT;b), as in dy- 
namic logic. 
(UNTIL p a) -- iterative action d*~ (~p:a)';~p? (again, as in 
dynamic logic). 
The recta-symbol "1-' will prefix formulas that are theorems, 
i.e.. that are derivable. Properties of the formal system that will 
be assumed to hold will be termed Propositions. Propositions 
will be both formulas that should always be valid, for our forth- 
coming ~emantics, and rules of inference that should be sound. 
No attempt to prove or validate these propositions here, but we 
do so in It 7\]. 
4.2 Properties of Acts 
We adop! ,In' ,Isual axioms characterizing how complex actions 
behave .mh'r AFTER, a.s treated in a dynamic logic (e.g., \[20\]) 
namely, 
Proposition t Propert*es o/complez aet~ --~ 
(AFTER 
(AFTER 
(AFTER 
AFTER atttl 
ties: 
Proposition 
Proposition 
Proposition 
Propositlon 
Proposition 
a:b p) --- (AFTER a (AFTER b p)). 
a\]b p) -= (AFTER a p) ^ (AFTER b p). 
p't q) -= p ^ q. 
DONE will have ~he following additional proper. 
2 V act (AFTER act (DONE x act)) 4 
$ Va \[{DONE (AFTER a p)?:a) ~ p\] 
4 \[lb. ~D,q then 
(DONE ~?:a) :~ (DONE ,')?;a) 
,5 p -= {DONE p?} 
6 (DONE \[(p 3 q) ^ p\]?} .~ (DONE q?) 
Our treatment of acts requires that we deal somehow with the 
"frame problem" \[18\]. That is, we must characterize not only 
what changes as a resuh of doing an action, but also what does 
not change. To approach this problem, the following notation 
will he convenient: 
Definition t (PRESERVES a p) d.f P ~ (AFTER a p) 
Of co.rse, all theorems are preserved. 
Temporal concepts are introduced will DONE (for past hap- 
penings) and <> (read "eventually'}. To say that p was true at 
~(,me point in the past, we use 3a (DONE p?:a). <> is to he 
regarded in the "branching time* sense \[I 1\], and will be defined 
more rigorously in !17\]. Essentially, OP is true iff for all infinite 
extensions of any course of events there is a finite prefix satis- 
fying p. OP and O~p are jointly satisfiable. Since OP starts 
"now ", the following property is also true, 
*(AFTER t (DONE t)), where t is term denoting a primitive act 
(or a sequence of primitive actsl, is ant always true since aft ;~t '~ay 
change the values of terms (e.g., an election changes the value of the term 
(PRESIDENT U.S.)) 
Proposition 7 t- p 30P 
Also, we have the following rule of inference: 
Proposition 8 I/I- a ~ fl then O(a v p) ~ O(3 v p) 
4.3 The Attitudes 
Neither BEL, BMB. nor GOAL characterize what an agent 
actively believes, mutually believes (with someone else), or has 
as a goal, but rather what is imph'cit in his beliefs, mutual be- 
liefs, and goals, s That is, these operators characterize what 
the world would be like if the agent's beliefs and mutlml beliefs 
were true, and if his goals were made true. Importantly. we 
do not inch,de an operator for wanting, since desire~ m,ed not 
he consistent. We ass.me that once an agent has sorted o~lt 
his possibly inconsistent desires in deciding what he wishes to 
achieve, the worhls he will he striving for are consisteal. ~'on- 
versely recognition of an agent's plans n,'ed not, com, ider that 
agent's possibly inconsistent desires. F,zrthermore. there is al~o 
no explicit operator for intending. If an agent intends to bring 
about p, the agent is usually regarded as also being able to bring 
about p. By using GOAL, we will be able to reason about the 
end state the agent is aiming at separately from our reasoning 
about Iris ability to achieve that state. 
For simplicity, we assume the usual Hintikka axiom schemata 
for BEL \[I,SI, and we introduce KNOW by definition: 
Definition 2 (KNOW x p) ~f p ^ (BEL x p) 
4.3.1 Mutual Belief 
Human communication depends crucially on what is mutually 
believed \[I, 6, 7, 9, 22, 23, 2.1\]. We do not use the standard 
definitions, but employ (nMB y x p), which stands for y's belief 
that it is mutually believed between y and x that p. (BMB y 
x p} is true iff (BEL y \[p A (BMD x y p)\]). ~ BMB has the 
following properties: 
Proposition 9 (BMB y x pAq) =- (BMB y x p) A 
(DMB y x q) 
Proposition 10 (BMB y x pDq) 3 
((BMD y x p) 3 (BMB y x q)) 
Proposition 11 1/I-,~ 3 # then 
~-(BMB y x ~) :3 (BMB y x J) 
Also, we characterize mutual knowledge as: 
Definition 3 (MK x y p)d.=f P ^ (BMB x y p) ^ 
(BMD y x p)r 
5For an exploration of the issues involved in explicit vs. implicit belief, see ilel. 
SNotice that (BMB y x p) $ (BMB x y p). 
~This definition is not entirely correct, but is adequate for present 
purposes. 
51 
4.3.2 Goals 
For GOAL, we have the following properties: 
Proposition 12 {GOAL x {GOAL x p)) ~ (GOAL x p) 
If an agent thinks he has a goal, then he does. 
Proposition 13 {BEL x {GOAL x p}} - {GOAL x p} 
Proposition 14 {GOAL x p} ^ {GOAL x p~q) 
{GOAL x q)8 
The following two derived rules are also useful: 
Proposition 15 If i" o D ~ then 
~'(GOAL x a) D (GOAL x ~) 
Proposition t0 Ilk- a A ;1 D "7 then 
I-{BMB y x (GOAL x ~)) ^ (BMB y x {GOAL x ~)} :~ 
(BMB y x {GOAL x "~)) 
More properties of GOAL follow. 
4.4 Attitudes and Rational Action 
Next. we must characterize how beliefs, goals, and actions are 
related. "the interaction of BEL anti AFTER will be patterned 
after Moore's analysis \['20l. In particular, we have: 
Proposition IT v x. act (AGT a x) D 
(AFTER act (KNOW x (DONE act))) 
Agents know what they have done. Moreover, they think certain 
effects of their own actions are achieved: 
Proposition 18 (BEL x {RESULT x a p)) 3 
(RESULT x a (BEL x p)). tvhere 
def Definition 4 (RESULT x a p) = (AFTER a p) ^ 
(AGT a x) 
The major addition we have made is GOAL. which interacts 
tightly with the other operators. 
We will say a rational agent only adopts goals that are achiev- 
able, and accepts as "desirable" those states of the world that 
are inevitable. To characterize inevitabiJities, we have 
Definition 5 (ALWAYS p) 4.~ Va (AFTER a p) 
This says that no matter what happens, p is true. Clearly, we 
want 
Proposition 19 lf~-r~ then ~- (BEL x (ALWAYS ,~)) 
That is, theorems are believed to be always true. 
Another property we want is that no sequence of primitive 
acts is forever ruled out from happening. 
Proposition 20 ~" Va (ACT a) ~ ~(ALWAYS ~(DONE a)), 
where (ACT a) ~f ~(AFTER a --(DONE a)) 
One important variant of ALWAYS is (ALWAYS x p) (rel- 
ative to an agent), which indicates that no matter what that 
aqent does, p is true. The definition of this version is: 
d~f Definition 6 (ALWAYS x p) = Va {RESULT x a p) 
A u:~eful instance of ALWAYS Is (ALWAYS pDq) ill which no 
matter what happens, p still implies q. We can now distinguish 
between p :~ q's being logically valid, its being true in all courses 
of events, and its merely being true after some event happens. 
SNotice that it pDq is true (or even believed} but (GOAL x pDq) is not 
true, we should not reach this conclusion since some act could make it 
laise. 
4.4.1 Goals and Inevitabilities 
What an agent believes to be inevitable is a goal (he accepts 
what he cannot change). 
Proposition 21 (BEL x {ALWAYS p)) ~ (GOAL x p) 
and conversely (almost), agents do not adopt goals that they 
believe to be impossible to achieve -- 
Proposition 22 No futility -- (GOAL x p) 
~(BEL x (ALWAYS ~p)) 
This gives the following useful lemma: 
Lemma I Inevitable Consequences 
(GOAL x p) A (BEL x (ALWAYS p~q )) D (GOAL x q) 
Proof: By Proposition 21, if an agent believes pDq is always 
true, he has it as a goal. Hence by Proposition 14, q follows 
from his goals, 
This lemma states that if one's goal is ac.o.e, in which p holds, 
and if one thinks that no matter what happens, pDq, then one's 
goal is a c.o.e, in which q holds. Two aspects of this property 
are crucially important to its plausibility. First, one must keep 
in mind the "follows from* interpretation of our propositional 
attitudes. Second, the key aspect of the connection between 
p and q is that no one can achieve p without achieving q. If 
someone could do so, then q need not be true in a c.o.e, that 
satisfies the agent's goals. 
Now, we have the following as a lemma that will be used in 
the speech act derivations: 
Lemma 2 Shared Recoqnition 
(BMB y x {GOAL x p)} A 
(BMB y x (BEL x (ALWAYS p~q))) 3 
(BMB y x (GOAL x q)) 
The proof is a straightforward application of Lemma I and 
Propositions 9 and 10. 
4.4.2 Persistent goals 
In this formalism, we are attempting to capture a number of 
properties of what might be called "intention" without postu- 
lating a primitive concept for "intend". Instead, we will combine 
acts, beqiefs, goals, and a notion of commitment built out of more 
primitive notions. 
To capture ,me grade of commitment than an agent might 
have towards his goals, we define a persistent goal. P-GOAL, 
to be one that the agent will not give up until he thinks it has 
been an:(stied, or until he thinks he cannot achieve it. 
Now, in order to state constraints on c.o.e.'s we define: 
d*f Definition T (PREREQ x p q) = 
Vc (RESULT x ¢ q) ~ 3 a (a ~ c) A (RESULT x a p} 
This definition states that p is a prerequisite for x's achieving q 
if all ways for x to bring about q result in a course of events in 
which p has been true. Now, we are ready for persistent goals: 
52 
dlt Definition 8 (P-GOAL x p) = 
(GOAL x p) ^ 
\[PREREQ x ((BEL x p) v 
{BEL x (ALWAYS x ~p))) 
~(GOAL x p)l 
Persistent goals are ones the agent will replan to achieve if his 
earlier attempts to achieve it fail to do so. Our definition does 
not say that an agent must give up his goal when he thinks it is 
satisfied, since goals of maintenance are allowed. All this says is 
that somewhere along the way to giving up the persistent goal, 
the agent had to think it was true (or belie~,e it was impossible 
for him to achieve). 
Though an agent may be persistent, he may be foolishly so 
beca,se he ha.~ no competence to achieve his goals. We charac- 
terize competence below. 
4.4.3 Competence 
I'e.ple are ~omet imes experls in certain fiehts, as well as in their 
own bodily movements. For example, a competent electrician 
will form correct plans to achieve world states in which "elec- 
trical" .-tares of affairs obtain. Most aduhs are competent in 
achievimz worhl states in which their teeth are brushed, etc. 
We will say an agent is COMPETENT with respect to p if, 
whenever he thinks p will tnJe after some action happens, he is 
correct: 
def Definition 9 (COMPETENT x p} = 
Va (BEL x (AFTER x p)) 2) (AFTER a p} 
One property of competence we will want is: 
Proposition 23 Vx. a (AGT x a) 
(ALWAYS (COMPETENT x (DONE x a))), where 
Definltlon I0 (DONE x a) a---'f (DONE a) .'~ (AGT a x) 
That is. any person is always competent to do the acts of 
which he is the agent. ~ Of course, he is not always competent 
to achieve any particular effect. 
Finally. ~iven all these properties we are ready to describe 
rational agents. 
4.5 Rational Agents 
i~elow are properties of ideally rational agents who adopt per- 
~i.~tent gnals. 
First. a~ents are carefuh they do not knowingly and deliber- 
ately make their persistent goals impossible for them achieve. 
Proposition 24 (DONE x act) 2) {DONE x p?;act), where 
p %'J (P-GOAL x q) ~ ~(DEL x (AFTER act 
(ALWAYS x ~p))) v 
~(COAL x (DONE x act)) l0 
in other words, no deliberately shooting onessetf in the foot. 
Now, agents are cautious in adopting" persistent goats, since 
they must eventually come to some decision about their feasi- 
bility. We require an agent to either come up with a "plan ~ to 
Sl}ecause of Proposition 2. all Proposition 23 says is that if a competent 
agen,, believes his own primitiw act halts, it will. 
~nNotice *hat tt is eruciad that p be true in ~he sane world in which the 
agent does act, hence the use ,if "p?;aet*. 
achieve them -- a belief of some act (or act sequence) that it 
achieves the persistent goal -- or to believe he cannot bring the 
goal about. That is, agents do not adopt persistent goals they 
could never give up. The next Proposition will characterize this 
property of P-GOAL. 
But, even with a correct plan and a persistent goal. there 
is still the possibility that the competent agent never executes 
the plan in the right circumstances -- some other agent has 
changed the circumstances, thereby making the plan incorrect. 
\[f the agent is competent, then if he formulates another plan. it 
will be correct for the new circumstances. But again, the world 
could change out from under him. Now, just as with operating 
systems, we want to say that the world is "fair" - the agent will 
eventually get a chance to execl,te his plans. This property is 
also characterized in the following Proposition: 
Proposition 25 fa,r EzecuHon -- The agent u,dl prentually 
form a plan and ezeeute *t. believing it achieves his persistent 
goal in e,rcumstanees he believes to be appropriate for its sucees.~. 
V x (P-GOAL x q) 2) 
0\[3 act' (DONE x p?;act')\] v 
\[BEL x (ALWAYS x ~ql\[}, 
where p 4=*¢ (nEL x (RESULT x act' q)) 
We now give a crucial theorem: 
Theorem I Consequences of a pers,stent goal -- If .~omeone 
has a pers*stent goal of bringing about p, and brmgm 9 ~l~ut p is 
usffhin his area of competence, then eventually either p becomes 
true or he wall believe there is nothing that can be done to achiet, e 
P 
¥ y (P.GOAL y p) A (ALWAYS (COMPETENT y p)) D 
(> (p v (BEL y (ALWAYS y ~p})) 
Proof sketch: 
Since the agent has a persistent goal. he eventually will either 
find and execute a plan. or will believe there is nothing he can 
do to achieve the goal. Since he is competent with respect to p, 
the plans he forms will be correct. Since his plan act' is correct, 
and since any other plans he forms for bringing about p are also 
correct, and since the world is "fair', eventually either the agt,nt 
executes his correct plan, making p true, or the agent comes to 
believe he cannot achieve p. A more rigorous proof can be found 
in the Appendix. 
This theorem is a major cornerstone of the formalism, telling 
us when we can conclude  p, given a plan and a ~oal. and is 
used throughout the speech act analyses. \[f an agent who is not 
COMPETENT with respect to p adopts p a.s a persistent goal, 
we cannot conclude that eventually either p will be true (or the 
agent will think he cannot bring it about), since the agent could 
forever create incorrect plans. \[f the goal is not persistent, we 
also cannot conclude OP since the agent could give it up without 
achieving it. 
The use of ~ opens the formalism to McDermott's "Little 
Nell* paradox \[19l. tt In our context, the problem arises as 
follows: First, since an agent has a persistent goal to achieve p, 
~lLittle Nell is tied to the railroad tracks, and will be muhed by the neXt 
train. Dudley Doright is planning to save her. McDermott claims that, 
according to various A\[ theories of planning, he never will, even though 
he always knows just what to do. 
53 
and we assume here he is always competent with respect to p, 
~p is true. But, when p is of the form Oq (eg., <>(SAVED 
LITTLE-NELL)), <><>q is true, so <>q is true ~ well. Let us 
assume the agent knows all this. Hence, by the definition of 
P-GOAL, one might expect the agent to give up his persistent 
goal that <>q, since it is already satisfied! 
On the other hand, it would appear that Proposition 25 is 
sufficient to prevent the agent from giving up his goal too soon, 
since it states that the agent with a persistent goal must act on 
it, and, moreover, the definition of P-GOAL does not require the 
agent to give up his goal immediately. For persistent goals to 
achieve <>q. within someone's scope of competence, one might 
think the agent need "only" maintain <>q as a goal, and then 
the other properties of rationality force the agent to perform a 
primitive act. 
Unfortunately, the properties given so far do not yet rule out 
Little Nell's being mashed, and for two reasons. First, NIL 
denotes a primitive act -- the empty sequence, llence, doing it 
would satisfy Proposition 25, but the agent never does anything 
substantive. Second, doing anything that does not affect q also 
satisfies Proposition 25, since after doing the unrelated act, <>q 
is still true. We need to say that the agent eventually acts on q! 
To do so, we have the following property: 
Proposition 26 (P-GOAL y Oq) 3 
O\[(P-GOAL y q) v 
(rtgL y (ALWAYS y ~q))\], 
That is. eventually the agent will have the persistent goal that 
q, and by Proposif ion 25. will act on it. If he eventually comes to 
believe he cannot bring about q, he eventually comes to believe 
he cannot bring about eventually q as well, allowing him to give 
up his persistent goal that eventually q. 
4.6 Rational Interaction 
This ends our discussion of single agents. We now need to char- 
acterize rational interaction sufficiently to handle a simple re- 
qt,?st. First, we ,.liscuss cooperative agents, and then the effects 
of uttering sentences. 
4.6.1 Properties of Cooperative Agents 
We describe agents as sincere, helpful, and more knowledgeable 
than others about the t~lth of some ~tate of affairs. Essentially, 
O.,,~e concepts capture (quite ~iml)li,qic) constraints on influegc- 
ing ~omeone clse's beliefs and goals, and on adopting the beliefs 
and goal~ of someone else ~ one'~ own. More refined versions 
are certainly desirable. Ultimately. we expect such properties of 
cooperative agents, a.s embedded in a theory of rational inter- 
action, to provide a formal description of the kinds of conver- 
sational behavior ~rice \[1-t\[ describes with his "conversational 
m;Lxims". 
First, we will say an agcnt i~ SINCERE with respect to p if 
whenever his goal is to get someone else to belietpe p, his goal is 
in fact to get that person to knom p. 
dec Definition tl (SINCERE x p) = 
(GOAL x (laEL y p)) D (GOAL x (KNOW y p)) 
An agent is HELPFUL to another if he adopts as his own 
persistent goal another agent's goal that he eventually do some- 
thing (provided that potential goal does not conflict with his 
own I. 
Definition 12 (HELPFUL x y) a,¢= 
'Ca (BEL x (GOAL y (}(DONE y a))) ^ 
~(GOAL x ~(DONE x a)) D 
(P-GOAL x (DONE x a)) 
Agent x thinks agent y is more EXPERT about the true of p 
than x if he always adopts x's beliefs about p as his own. 
def Definition 13 (EXPERT y x p) : 
(BEL x (BEL y p)) :3 (BEL x p) 
4.0.2 Uttering Sentences with Certain aFeatures" 
Finally, we need to describe the effects of uttering sentences with 
certain "features" \[141, such an mood. In particular, we need 
to characterize the results of uttering imperative, interrogative, 
and declarative sentences t: Our descriptions of these effects 
will be similar to Grices's \[131 and to Perrauh and Allen's {22\] 
%urface speech acts'. Many times, these sentence forms are not 
used literally to perform the corresponding speech acts (requests, 
questions, and assertions). 
The following is used to characterize uttering an imperative: 
Proposition 27 Imperatives: 
V x y (MK x y (ATTEND y x}) 3 
(RESULT x \[IN4PER x y "do y act" 1 
(laMB y x 
(GOAL x 
(BEL y 
(GOAL x 
(P-GOAL y (DONE y act) ))))))) 
The ac: !IMPER speaker hearer 'p\] stands for "make p t r~w" 
Proposition 27 states that if it is mutually known that y is at- 
tending to x, is then tile result of uttering an imperative to y 
to make it the case that y has done action act is that y thinks 
it is mutitally believed that the speaker*s goal is that y should 
think his goal is foe y to form the persistent goal of doing act. 
We also need to a~sert that IMPER preserves sincerity about 
the speak,'r's coals and helpfulness. These restrictions c,~uld be 
loosened, but maintaining them is simpler. 
Proposition 28 {PRESERVES \[IMPER x y "do y act'\] 
(BMB y x (SINCERE y (GOAL y p)))) 
Proposition 29 (PRESERVES \[IMPER x y "rio y ;Jet'\] 
(HELPFUL y xt) 
All t ',ricean "feature'-based theories of communication need 
to acco,mt for cases in which a speaker uses an utterance with a 
feat'tre, but does not have the attitudes (e.g.. beliefs, and goals) 
'2llowever, #e can only present the analysis of imperatives here. 
tall it is not mutually known that y is attending, for example, if the speaker 
i~ not speaking to an ~udience, then we do not say what the result of 
uttering an imperative is. 
54 
usually attributed to someone uttering sentences with that fea- 
ture. Thus, the attribution of the attitudes needs to be context- 
dependent. Specifically, proposition 28 needs to be weak enough 
to prevent nonserious utterances such as "go jump in the lake ~ 
from being automatically interpreted as requests even though 
the utterance is an imperative. On the other hand, the formula 
must be strong enough that requests are derivable. 
5 Deriving a Simple Request 
In making a request, the speaker is trying to get the hearer to do 
an act. We will show how the speaker's uttering an imperative 
to do the act leads to its eventually being done. What we need 
to prove is this: 
Theorem 2 Result o\[ an Imperative -- 
(DONE \[(MK x y (ATTEND y x)) ^ 
(BMB y x 
(SINCERE x 
(GOAL x 
(P-GOAL y (DONE y act)))))^ 
(HELPFUL y x)l?; 
lIMPER x y "do y act'\]) :3 
O(DONE y act) 
We will give the major steps of the proof in Fi~lre I, and 
point In their justifications. The full-fled~'ed proofs are h'ft to 
Ihe ,,nergetic reader. All formula.s preceded by a * are supposed 
t,, be Irue just prior to performing the IMPER, are preserved by 
il. an,I thus are implicitly conjoined to formulas 2 - 9. By their 
placement in the proof, we indicate where they are necessary for 
making t he deductions. 
E~entially. the proof proceeds as follows: 
If it is mutually known that y is attending to x. and y thinks it 
i~ mutually believed Ihat Ihe e-conditions hohl. then x's ,lltering 
an imlwrative to y to do some action results in formula (2). Since 
h i~ mutually believed x is sincere about his goals, then (:~) it is 
miltually believed his goal tndy is that y form a persistent goal 
to ,Io the act. Since everyone is always competent to do acts of 
which they are the agent. (.1) it is mutltally believed that the act 
will eventually be done, or y will think it is forever impossible 
to do. But since no halting act is forever impossible to do, it 
is (.3) mutually believed that x's goal is that y eventually do it. 
Ih, nee, 16) y thinks x's ~oa\] is that y eventually do the act. Now, 
~ince y is helpfillly disposed towards x, and has no objections Io 
doing the act. 17) y takes it on as a persistent goal. Since he 
is alwa.w competent about doing his own arts, 18) eventually it 
~ill I,.,Ione or he will think it impossible to do. Again. since it 
is n(,I f~)rever impossible. (3) he v, ill eventually do it. 
W,. have shown how the p,.rforming of an imperative to do 
an act leads to the act's evemually being done. We wish to 
create a number of lemmas from this proof (and others like it) 
to characterize iilocutionary acts. 
8 Plans and Summaries 
6. t Plans 
A plan for agent "x" to achieve some goal "q" is an action term 
~a" and two sequences of wits ".no', ~Pl".... "pt," and "q0", 
"qz", ... ~qk" where "qk" is ~q" and satisfying 
I. I- (BEL x (poApt A ...Ap~} 
(RESULTxa qoaptA...APk ))) 
2. h (BEL x (ALWAYS (p~a Ch-t) D q,))) i=l,e....k 
In other words, given a state where "x" believes the "pi ~, he 
will believe that if he does ~a" then "q0" will hold and moreover. 
given that the act preserves pi, and he believes his making "qi-i ~ 
true in the presence ofpi will also make "qi* tale. Consequently, 
a plan is a special kind of proof that 
I- (BEL x ((Po^.-. A Pk) ~ (RESULT x a q))) 
and therefore, since 
(BEL x p) D (BEL x (BEL x p)) 
and 
(BEL x (p ~ q)) D ((BEL x p) D (BEL x q)). are axioms of 
belief, a plan is a proof that 
h (BEL x (p. A ...^ p~)) ~ (BEL x (RESULT x a 'l)) 
Among tile corollaries to a plan are 
}- (BEL x ( (Po a ... ^ p,) ~ (RESULT x a q,))) i=\[ .... k 
and 
}- (BEL x ( (p," a ... a Pi) ~ (ALWAYS q~-i D qi))) 
i: 1 .... k \]=l" .... k 
There are two main points to be made about the~e corollaries. 
First of all, since they are theorems, the implications can be 
taken to be believed by the agent "x" in every, state. In this 
sense, these wits express general methods believed to achieve 
certain effects provided the assumptions are satisfied. The sec- 
ond point is that these corollaries are in precisely the form that 
is required in a plan and therefore can be used as justification for 
a step in a filture plan in much the same way a lemma becomes 
a single step in the proof of a theorem. 
6.2 Summaries 
We therefore propose a notation for describing many ~t,.p~ of 
a plan as a single summarizing operator. A 3ummary consists 
of a name, a list of free variables, a distingafished free variable 
called the agent of the summary (who will always be list,,d tirst), 
an Effect which is a wff, a optional Body which is either an 
action or a wff and finally, an optional Gate which is a wff. The 
understanding here is that summaries are associated with agent 
and for an agent "x" to have summary "u". then there are three 
cases depending on the body of "u': 
I. If the Bodyof "u" is a wff, then 
(BEL x (ALWAYS (Gate ^ Bod~) ~ {Gate ^ Effect})) Is 
2. If the Body of "u" is an action term, then 
I- (BEL x (Gate ~ (RESULT agent Bod~ (Gate A Effect}))) 
:60f course, many actions change the truth of their preconditions. H~ndllng 
such actions and preconditions i$ straightforward. 
55 
I. 
2. 
Given 
P27, P3, P4, I 
3. Pll, P12, 2 
(DONE \[(MK x y (ATTEND y x)) A 
(.conditions)\]?; 
lIMPER x y "do y act*\]) 
(BMB y x (GOAL x (BEL y (GOAL x 
(P-GOAL y (DONE y act)))))) A 
*(BMB y x (SINCERE x 
(GOAL x (P-GOAL y (DONE y act))))) 
(BMB y x (GOAL x (P-GOAL y (DONE y act)))) ^ 
*(BMB y x (ALWAYS 
(COMPETENT y (DONE y act)))} 
(BMB y x (GOAL x O\[(DONE y act) v 4. 
(BEL y (ALWAYS ~(DONE y aet)))l)) ^ TX, Plf, 3 
,(BMB y x ~(ALWAYS ~(DONE y act))) 
5. (BMB y x (GOAL x O(DONE y act))) A P160 P20, P8, 4 
6. (BEL y x (GOAL x O(DONE y act))) ^ Def. BMB 
(HELPFUL y x) 
T. (P-GOAL y x (DONE y act)) ^ Def. of HELPFUL, MP 
• (ALWAYS (COMPETENT y (DONE y act))) 
8. <>\[(DONE y act) v (BEL y (ALWAYS ~(DONE y act)})l ^ T1 
• ~(ALWAYS ~(DONE y act)) 
9. <>(DONE y act} P20, P8 
Q.E.D. 
Figure 1: Proof of Theorem 2 -- An imperative to do an act result~ in its eventually bein 9 done. 14 
One thing worth noting about summaries is that normally the 
wiTs used above 
~" (BEL x (Ga:e D ...)) 
will follow from the more general wff 
I- (;ate D ... 
llowever, this need not be the ca,~e and different agents could 
have different summaries (even with the same name). Saying 
that an agent has a summary is no more than a convenient 
way of saying that the agent always believes an implication of a 
certain kind. 
7 Summarization of a Request 
The following is a summary named REQUEST that captures 
steps 2 through steps 5 of the proof of Theorem 2. 
\[REQUEST x y act\]: 
Gate: it) (BMB y x (SINCERE x (GOAL x 
(P-GOAL y (DONE y act))))) ^ 
(2) (BMB y x (ALWAYS 
(COMPETENT y (DONE y act)))) 
(3) (BMB y x ~(ALWAYS ~(DONE y act))) 
Bo~i~. (BMB y x 
(GOAL x 
(BEL y 
(GOAL x (P-GOAL y {DONE y act))))}) 
Effect: (BMB y x (GOAL x O(DONE y act))) 
This summary allows us to conclude that any action preserv- 
ing the Gate and making the Bod!/true makes the Effect true. 
Conditions (2) and (3) are theorems and hence are always pre- 
served. Condition (1) was preserved by assumption. 
Searle's conditions for requesting are captured by the above. 
Specifically, his "propositional content" condition, which states 
that one requests a future act, is present as the Effect because 
of Theorem 2. Searle's first "preparatory" condition -- that the 
hearer be able to do the requested act, and that the speaker 
think so is satisfied by condition (2). Searle's second prepara* 
tory condition -- that it not be obvious that the hearer was 
going to do the act anyway -- is captured by our conditions on 
persistence, which state when an agent can give up a persistent 
goal, that is not one of maintenance, when it has been satisfied. 
Grice's "recognition of intent* condition \[12, 13\] is satisfied 
since the endpoint in the chain (step 9) is a goal. Hence, the 
speaker's goal is to get the hearer to do the act hy means, in 
part, of the (mutual) recognition that the speaker's goal is to 
get the hearer to do it. Thus, according to Grice, the speaker 
has meant,,,, that the hearer should do the act. Searle's revised 
Gricean condition, that the hearer should "understand" the lit- 
eral meaning of the utterance, and what illocutionary act the 
utterance "counts as* are also satisfied, provided the summary 
is mutually known, le 
T. 1 Nonserious Requests 
Two questions now arise. First, is this not overly complicated? 
The answer, perhaps surprisingly, is "No'. By applying this 
REQUEST theorem, we can prove that the utterance of an im- 
perative in the circumstances specified by the Gate results in 
the Effect, which is as simple a propositional attitude as anyone 
would propose for the effect of uttering an imperative -- namely 
that it is mutually believed that the speaker's goal is that the 
hearer eventually do the act. The Bod V need never be considered 
16~'he further elaboration of this point that it deserves is outside the ~cope 
. ot this paper. 
56 
unless one of the gating conditions fails. 
Then, if the Body is rarely needed, when is the "extra" em- 
bedding (GOAL speaker (BEL hearer ...)} attitude of use? 
The answer is that these embeddings are essential to preventing 
nonserious or insincere imperatives from being interpreted un- 
conditionally as requests. In demonstrating this, we will show 
how Searle's "Sincerity ~ condition is captured by our SINCERE 
predicate. 
The formula (SINCERE speaker p) is false when the speaker 
does something to get the hearer to believe he, the speaker, has 
the goal of the bearer's believing p, when he in fact does not have 
the goal of the heater's knowing that p Let us see see how this 
would he applied for "Go jump in the lake', uttered idiomati- 
cally. Notice that it could be uttered and meant as a request, 
and we should be able to capture the distinction between serious 
and nonserious uses. In the case of uttering this imperative, the 
content of SINCERE. p p =((:OAL speaker (P-GOAL hearer 
(DONE hearer/JUMP-INTO Laker\]))). 
Assume that it is mutually known/believed that the lake is 
frigidly cold (any other conditions leading to -,.{GOAL x p) 
would do as well. e.g., that the hearer is wearing his best suit, 
or that there is no lake around). So, by a reasonable axiom of 
goal formation, no one has goals to achieve states of affairs that 
are objectionable (assume what is "objectionable" involves a 
weighing of alternatives). ~o, it is mutually known/believed that 
~(GOAL speaker (DONE hearer \[JUMP-INTO Laket\])), and 
so the speaker does not believe he has such a goal. l'l The 
consequent to the implication defining SINCERE is false, and 
because tile result of tile imperative is a mutual belief that the 
speaker's goal is that the hearer think he has the goal of the 
bearer's jumping into the lake, the antecedent of the implica- 
tion is true. Hence, the speaker is insincere or not serious, and 
a request interpretation is blocked, is 
In the case of there not being a lake around, the speaker's goal 
cannot be that the hearer form the persistent goal of jumping 
in some non-existent lake. since by the 3/0 Futility property, the 
hearer will not adopt a goal if it is unachievable, and hence the 
speaker will not form his g~al to achieve the unachievable state of 
affairs (that the hearer adopt a goal he cannot achieve). }tence, 
since all this is mutually believed, using the same argument, the 
speaker must be insincere. 
8 Nonspecific requests 
The ability conditions for requests are particularly simple, since 
as long as the hearer knows what action the speaker is referring 
to. he can always do it. He cannot, however, always bring about 
some goal world. An important variation of requesting is one in 
which the speaker does not specify the act to be performed; he 
merely expresses his goal that some p be made true. This will be 
captured by the action lIMPER y 'p\] for ~make p true*. Here, 
tTThe speaker's expressed goat is that the hearer form t persistent gold 
to jump in the lake. But. by the /neeitails Coassqasaees lemma, given 
that a c.o.e, satisfying the speaker's goal also hu the heater's eventually 
jumping in (since the hearer knows what to do), the speaker's goal is also 
• c.o.e, in which the hearer eventually jumps in. In the same way, the 
speaker's goal would also be that the hearer eventually gets wet. 
I*11owever, we do not say what else might be derivable. The speaker's true 
goals may have more to do with the manner of his action (e.g., tone of 
voice), than with the content. All we have done is demoasnurata formally 
how • hearer could determine the utterance is not to be talteo ~r, face 
value. 
in planning this act, the speaker need only believe the hearer 
thinks it is mutually believed that it is always the case that the 
hearer will eventually find a plan to bring about p. Ahhough we 
cannot present the proof that performing an \[IMPER x y "p\] 
will make Op true, the following is the illocutionary summary 
of that proof: \[NONSPECIFIC-REQUEST x y p\]: 
Gate: (BMB y x (SINCERE x (GOAL x (BEL y 
(GOAL x (P-GOAL y p)))))) A 
(BMB y x (ALWAYS (COMPETENT y p))) 
(BMB y x (ALWAYS 
~-7 act' (DONE y q?;act'), 
where q ~( (BEL y (RESULT y act' p)))) 
Body:. (IJMB y x 
(GOAL x 
(BEL y 
(GOAL x (P-GOAL y p))))) 
Effect: (nMB y x 
(GOAL x OPt) 
Since the speaker only asks the hearer to make p true. the 
ability conditions are that the hearer think it is mutually be- 
lieved that it is always true that eventually there will be some 
act such that the hearer believes of it that it achieves p (or he 
will believe it is impossible for him to achieve). The speaker 
need not know what act the hearer might choose. 
9 On summarization 
Just as mathematicians have the leeway to decide which proofs 
are useful enough to be named a.s lemmas or theorems, so too 
does the language user. linguist, computer system, and speech 
act theoretician have great leeway in deciding which summaries 
to name and form. Grounds for making such decisions range 
from the existence of ilfocutionary verbs in a particular lan- 
guage, to efficiency. However. summaries are flexible -- they 
allow for different languages and different agents to carve up 
the same plans differently. ,o Furthermore, a summary formed 
for efficiency may not correspond to a verb in the language. 
Philosophical considerations may enter into how much of a 
plan to summarize for an illocutionary verb. For example, most 
illocutionary acts are considered successful when the speaker has 
communicated his intentions, not when the intended effect has 
taken hold, This acgues for labelling as Effects of summaries in- 
tended to capture illocutionary acts only formulas that are of the 
form (BMI3 hearer speaker (GOAL speaker p)), rather than 
those of the form (BMB hearer speaker p) or (BEL hearer p), 
where p is not a GOAL-dominated formula. Finally, summaries 
may be formed as conversations progress. 
The same ability to capture varying amounts of a chain of 
inference will allow us to deal with muhi-utterance or muhi- 
agent acts, such as, betting, complying, answering, etc., in which 
there either needs to be more than one act (a successful bet 
r.quires an offer and an acceptance), or one act is defined to 
require the presence of another (complying makes sense only 
in the presence of a previous directive). For example, where 
REQUEST captured the chain of inference from step 2 to step 
5, one called COMPLY could start at 5 and stop at step 9. 
tSRemember, summaries are actually beliefs of agents, and those beliefs 
need oct be shared. 
57 
Thus, the notion of characterizing illocutionary acts as lemma- 
like summaries, i.e., as chains of inference subject to certain 
conditions, buys us the ability to encapsulate distant inferences 
at "one-shot'. 
9.1 Ramifications for Computational Models of 
Language Use 
The use of these summaries provides a way to prove that various 
short-cuts that a system might take in deriving a speaker's goals 
are correct. Furthermore, the ability to index summaries by 
their Bodies or from the utterance types that could lead to their 
application (e.g., for utterances of the form "(.',an you do <X> ~) 
allows for fast retrieval of a lemma tlmt is likely to result in goal 
recognition. By an appropriate organization of summaries \[5\], 
a system can attempt to apply the most comprehensive sum- 
maries first, and if inapplicable, can fall back on less compre- 
hensive ones, eventuMly relying on first principles of reasoning 
about actions. Thus. the apparent difficulty of reasoning about 
speaker-intent can be tamed for tile "short-circuhed ~ cases, but 
more general-purpose reasoning can deployed when necessary. 
IIowever. the conil)lexities of rea.~oning about others' beliefs and 
goals remains. 
10 Extensions: Indirection 
Indireciion will be modeh'd ill tills framework a.s tile derivation of 
propositions (lUlling with the speaker's goals that are not stated 
as such by tile initial propositional attitude. For example, if we 
can conchlde from IBMB y x (GOAL x (GOAL y Nil that 
(BMB y x (GOAL x (GOAL y 0 q))), where pdoes not entail 
q, then. "loosely', we will say an indirect request has been made 
by x. 
(;iven the properties of O. (GOAL x p) D (GOAL x <C>P) is 
a dworcm. (GOAL x p) an(l ((;()At, x -li) ar~" mutually un- 
~ati~\[ial)le, hilt (COAL x OP) and (GOAL x O~p) are jointly 
~ali~liahh'. \["(}r examllh ", ((;OAL BILL OHAVE BILL HAM- 
MERI))) and (GOAL BILL <~(HAVE JOHN HAMMERI)) 
could both be part of a description of Bill's plan for John to get 
a hammer and give it to him. Such a plan could be triggered 
by Bill's merely saying "C, et tile ilammer" in the right circum- 
stances, such as when Bill is on a ladder plainly holding a nail. 
:0 A subsequent paper will demonstrate the conditions under 
which such reasoning is ~ound. 
I1 Concluding Remarks 
rhi~ i)alier tia.~ demonstrated tilat all illocutionary acts ne,'d 
ant t),' primitive. At least some can be derived from more basic 
priuciph.s of rational lotion, and an account of tile propositional 
attitudes affected by the uttering of sentences wittl decl.'u-ative, 
interrogative, and imperative moods. This account satisfies a 
number of criteria for a good theory of illocutionary acts. 
* Most elements of :he theory are independently motivated. 
The ~heory of rational action is motivated independently 
from any notions of communication. Similarly, the proper- 
ties of cooperative agents are also independent of commu- 
nication. 
l°Notice thllt molt theoritqt Ot Ipeech gta would treat the above utterance 
u Bed I I direct request. We do not. 
The characterization of the result of uttering sentences with 
certain syntactic moods is justified by the results we derive 
for illocutionary acts. as well as the results we cannot de- 
rive (e.g.. we cannot derive a request under conditions of 
insincerity ). 
Summaries need not correspond to illocutionary verbs in a 
language. Different languages could capture different parts 
of the same chain of reasoning, and an agent might have 
formed a summary for purposes of efficiency, but that sum- 
mary need not correspond to any other agent's summary. 
The rules of combination of illocutionary acts (character- 
izing, for example, how mnltiple assertions could consti- 
tute the performance of a request) are now reduced to nlles 
for combining propositional contents and attitudes. Thus, 
multi-utterance illocutionary acts can be handled by accu- 
mulating the speaker's goals expressed in multiple titter- 
antes, to allow an illocutionary theorem to be applied. 
Multi-act utterances are also a natural outgrowth of l|liS ap- 
proach. There is no rea.~on why one cannot apply mulliple 
illocutionary sunlniaries tO tile res0ill of utlt, ring a S¢'lllen¢¢'. 
Those sllmmaries, however, need not ¢'orre~pond Io illoc0f 
tionary verbs. 
The theory is naturally extensible to indirection (to lie ar- 
gued for hi another paper), to other illoc.tio.ary act, such 
u questions, commands, informs, a~sertions, and to tile act 
of referring \[gl. 
Finally. allllougti illocutionary act rerog'nition may h,, ~lricily 
unntwcssary, given the complexily of o01r proofs, it is likely to 
he loser011. I'\]~s,.nliallv. s01etl rec~l~nilhm would ;lillOlill~ to lh(. 
application (if ill,lc01tl*lnary Sllnlllllries llleort'nl.~ Io di.~cover the 
speaker'~ I~(ml(s L 
12 Acknowledgements 
We wo.ld like to thank Tom Blenko, Ih.rb (:lark, Michul 
(,eorg,.lr, David I~r~el, Bob Moore, (;(,off .NUli|ierg', Fernan(|o 
\[)ereira. flay Penault, .":,tan Rosenschein, Ivall ~ag, and ,~loshe 
Vacdi for valuable dise.ssions. 
13 References 
1. AIh'n..!. F. A llhin-lla-'~ed atll)roa~'h I(i Sll,,0.ch act rrc.~nh.ion. 
"r,.,ctinic:ll I~.,.port 17.1. Di'p;lrtnit'!it of ('ornpill.('r ~cil'nce. 
llilivei~ity ()f 'r,)roiito, January. ll.)\]'¢.). 
2. Allen..I. \[:., Frith. A. M.. <\[" l,il,nan. I)..I. ARt ;(iT: The 
Rochester dialogue system. Proceedings of the .Vat,..,d 
Conference on Artificial Intelligence, Phtsh,r~h, I),'nn~yl- 
vanla, 1982. ¢I,-70. 
3. Appelt, D. Planning Natural Language Utterances to S(itisfy 
Multiple Goals. Ph.D. Th.. Stanford University, Stanford, 
California, December 1981. 
4. Austin, J. L.//ol. to do thinfs ~ith wo,da. Oxford University 
Press, London, 1962. 
58 
5. Bracbman. R., Bobrow, R., Cohen, P., Klovatad, J., Wel>- 
bet, B. L., & Woods, W. A. Research in natural language 
understanding. Technical Report 4274, Bolt Beranek and 
Newman Inc., August, 1979. 
6. Bruce, B. C., & Newman, D. Interacting plans. CognRiee 
Science ~, 3, 1978, pp. 195-233. 
T. Clark. H. H., & Marshall, C. Definite reference and mutual 
knowledge. In Elements of Discourse Understanding, Aca- 
demic Press, Joshi, A. K., Sag, !. A., & Webber, B., Eds., 
New York. 1981. 
8. Cohen, P. R. The Pragmatics of Referring and the Modality 
of Communication. Computational LinquMtics lO, 2, 198,1, 
pp. 97-1.16. 
9. Cohen, P. R. On Knowin 9 what to Say: Plannin 9 Speech 
Acts. Ph.D. Th., University of Toronto. Toronto, January 
1978. Technical Report No. 118, Dept. of Computer Sci- 
ence. 
lO, ( :ohen, P. R.. & Levesque, II. J. Speech Acts and the Rerog- 
nition of Shared Plans. Proe. ,\[ the Third Iliennial Con- 
ference, Canadian Society for (~omputa!ional Studies of In- 
telligence, Victoria. B. (;., May. 1980, 263-271. 
II. Emerson, E. A., and Ilalpern. J. Y. "Sometimes" and "Not 
Never" Revisited: On Branching versus \[.inear Time. ACM 
Sympa.~ium on Prinr~ple.~ of t)rt~jrammin9 Lanquaqes, 1983. 
12. (;rice. l{. I'..\|caning. Phdo,,ophiral Ret,ietp 66, 1957, pp. 
377-3,g8. 
13. Grice. II. 1'. Utterer'.~ Meaning anti Intentions. t'hilo.~ophi- 
cal Reriew 63, 2. 1969, pp. 1.17-177. 
14. Grice, t|. P. Logic and conversation. In t:ole., P. and Mar 
gan, J. \[,., Eds.,Syntaz and Semantics: Speech Acts , Ace. 
demic Press, New York,1975. 
15. Halpern, J. Y., and Moses. Y. O. A tluide to the Modal 
Logics of Knowledge anti Belief. Pr~. a/the Ninth Inter. 
national Joint (;on\[erenre on .4rtl\]ir:al intelligence, J.J( :AI, 
Los Angeles, ('alif.. Augnst, 1985. 
Levesque, tlector, J. A logic of implicit and explicit belief. 
Proceedings of the National t,'ofl/erence a/ the American As- 
~ciation for Artificial Intelligence, ~ustin, Texan, 198.1. 
esque, H. J., & Cohen, P. R. A Simplified I,ogic of In- 
~eraction. in preparation 
18. McCarthy. J.. A: Ilayes. P..1. ~ome Philo~.phical \['rnhlems 
from the :-;tandpoint of ..\rtifi¢ial l.h'lli~,ehce, In 3t¢~rhl,e 
intelh'fence .i American El.~evier, B. Mehzer & D. Michh'. 
Eds., New York. 1~;9. 
I9. McDermott, D. A temporal logic for reasoning about pro- 
cesses and plans. Cognitive Science ~, 2, 1982, pp. 101-|55. 
20. Moore, R. C. Reasoning about Knowledge and Action. 
Technical Note 191, Artificial Intelligence Center, SR! In- 
ternational, October, 1980. 
21. Morgan, J. L. Two types o\[ convention in indirect speech 
acts. In Syntaz and Semantics, Volume 9: Pragmaties, 
Academic Press, P. Cole. Ed., New York, 1978, 261-280. 
22. Perrault, C. R.. & Allen, J. F. A Plan-Based Analysis of 
Indirect Speech Acts. American Journal of Computational 
LinguiaticJ 6, 3, 1980, pp. 167-182. 
25. Perrauit, C. R., & Cohen, P. R. It's for your own good: 
A note on inaccurate reference. In Elements of Discourse 
Understandin9, Cambridge University Press, Joshi, A., Sag, 
i., & Webber, B., Eds., Cambridge, Mass., 1981. 
24. Schiffer, S. Mctmin 9. Oxford University Press, London. 
1972. 
25. Schmidt, D. F., Sridharaa, N.S., & Goodson, J. L. The 
plan recognition problem: An intersection of artificial intel- 
ligence ant| psychology. Artificial Intelligence 10, 1979. pp. 
45-83. 
26. Searle, J. R. Speech acts: ..In essay in the philosophy of 
language. (?ambridge University Pre~s, (:ambridge, 1969. 
27. Sidner, (',. 1,., Bates, M., lgobrow. R. ,1., Brachman, R. J., 
Cohen, P. R.. Israel, D. J., Wehber. B. l,., & Woods, W. A. 
Research in knowledge representation for natural language 
undecstaz,liog. Annual R,'p~)rt .1785, Bolt, Beranek and 
Newman Inc., November, 1981. 
28. Vanderveken. I). A Model-Theoretic Semantics \['or illocu- 
tionary Force. Logique et ,4n,dy.~e ~'6, 10::-I0.l, 19~q'l, pp. 
3,~9-39.~. 
59 
13 Appendix 
Proof of Theorem l: 
First, we need a lemma: 
Lemma $ Va (DONE x \[(BEL x (AFTER & p)) ^ (COMPETENT x p)\[?;a) :) p 
Proo/: 
L 
o 
3, 
4. 
5. 
Q.E.D. 
Va {DONE x \[(BEL x {AFTER & p)} A (COMPETENT x p)l?;a) Ass 
\[BEL x (AFTER x p)) A (COMPETENT x p) D {AFTER a p) Def. of 
COMPETENT, MP 
Ya (DONE x {AFTER & p)?;a} 2, P4 
p 3, P3 
Ya {DONE x \[(BEL x (AFTER a p)} A {COMPETENT x pJl?;a) D p Impl. lntr. 
Theorem I.. Vy (P-COAL y p) A (ALWAYS (COMPETENT y p)) D O(p v {BEL y {ALWAYS ~p))) 
Proo~ 
I, 
2. 
3. 
4. 
Q.E.D. 
{P.GOAL y (DONE y act}} A {ALWAYS {COMPETENT y {DONE y actJ}} 
O{3a {DONE y \[(BEL y (AFTER a p}}l?:a} v {BEL y (ALWAYS ~p}}} 
O{P v {BEL y (ALWAYS ~p}}} 
\[P-GOAL y (DONE y act}) ^ (ALWAYS (COMPETENT y {DONE y act}J} :) 
O(P v {BEL y (ALWAYS ~p))) 
ASS. 
I, P25, MP 
L3, P8, :2 
Impl. Intr., 3 
60 
