A Plan-Based Analysis of Indirect Speech Acts 1 
C. Raymond Perrault 
Department of Computer Science 
University of Toronto 
Toronto, Canada M5S 1A7 
James F. Allen 
Department of Computer Science 
University of Rochester 
Rochester, New York 14627 
We propose an account of indirect forms of speech acts to request and inform based 
on the hypothesis that language users can recognize actions being performed by others, 
infer goals being sought, and cooperate in their achievement. This cooperative behaviour is 
independently motivated and may or may not be intended by speakers. If the hearer 
believes it is intended, he or she can recognize the speech act as indirect; otherwise it is 
interpreted directly. Heuristics are suggested to decide among the interpretations. 
1. Introduction 
Austin \[1962\] was one of the first to stress the 
distinction between the action(s) which a speaker per- 
forms by uttering a sentence (such as informing, re- 
questing, or convincing) and the truth conditions of 
propositions contained in the sentence. Actions have 
effects on the world, and may have preconditions 
which must obtain for them to be felicitously per- 
formed. For actions whose execution involves the use 
of language (or speech acts), the preconditions may 
include the speaker holding certain beliefs about the 
world, and having certain intentions or wants as to 
how it should change. 
As well as being important to the study of natural 
language semantics, speech acts are important to the 
designer of conversational natural language under- 
standing systems. Such systems should be able to 
recognize what actions the user is performing. Con- 
versely, if such a system is to acquire information or 
request assistance from its user, it should know how 
and when to ask questions and make requests. (See 
Bruce \[1975\] for an early attempt.) 
Cohen and Perrault \[1979\] (hereafter referred to as 
CP) argue for the distinction between a competence 
t This research was supported in part by the National Re- 
search Council of Canada under Operating Grant A9285. Thanks 
to Phil Cohen, Michael McCord, Corot Reason, and John Searle for 
their comments. We assume the usual responsibility for remaining 
inaccuracies, misunderstandings, and downright errors. 
theory of speech acts, which characterizes what utter- 
ances an ideal speaker can make in performing what 
speech acts, and a performance theory which also ac- 
counts for how a particular utterance is chosen in giv- 
en circumstances, or how it is recognized. We are 
only concerned here with a competence theory. 
In Perrault, Allen, and Cohen \[1978\] we suggested 
that it is useful to consider speech acts in the context 
of a planning system. A planning system consists of a 
class of parameterized procedures called operators, 
whose execution can modify the world. Each operator 
is labelled with formulas stating its preconditions and 
effects. A plan construction algorithm is a procedure 
which, given a description of some initial state of the 
world and a goal state to be achieved, constructs a 
plan, or sequence of operators, to achieve it. 
It is assumed there, and in all our subsequent work, 
that language users maintain a model of the world 
(their beliefs) and a set of goals (their wants). One 
person S's beliefs may include beliefs about another 
person A's beliefs and wants, including A's beliefs 
about S, etc. We do not concern ourselves with obli- 
gations, feelings, etc., which clearly can also be affect- 
ed by speech acts. 
CP discuss criteria for judging the correctness of 
the preconditions and effects of the operators corre- 
sponding to speech acts, and specifically those of the 
acts INFORM and REQUEST. However, the condi- 
tions on INFORM and REQUEST given in CP are at 
best necessary and certainly not sufficient. In particu- 
Copyright 1980 by the Association for Computational Linguistics. Permission to copy without fee all or part of this material is granted 
provided that the copies are not made for direct commercial advantage and the Journal reference and this copyright notice are included on 
the first page. To copy otherwise, or to republish, requires a fee and/or specific permission. 
0362-613X/80/030167-16501.00 
American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 167 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
lar they say nothing about the form of utterances used 
to perform the speech acts. Several syntactic devices 
can be used to indicate the speech act being per- 
formed: the most obvious are explicit performative 
verbs such as "I hereby request you to ...", and mood 
(indicative for assertions, imperative for requests to 
do, interrogative for requests to inform). But the 
mood of an utterance is well known to not completely 
specify its illocutionary force: 1.a-b can be requests to 
close the door, 1.c-e can be requests to tell the an- 
swer, and 1.f can be an assertion. 
(1.a) I want you to close the door. 
(1.b) Will you close the door? 
(1.c) Tell me the answer. 
(1.d) I want you to tell me the answer. 
(1.e) Do you know what the answer is? 
(1.f) Do you know that Jack is in town? 
Furthermore, all these utterances can also be intended 
literally in some contexts. For example, a parent leav- 
ing a child at the train station may ask 1.g expecting a 
yes/no answer as a confirmation. 
(1.g) Do you know when the train leaves? 
The object of this paper is to extend the work in 
CP to account for indirect use of mood, loosely called 
indirect speech acts. The solution proposed here is 
based on the following intuitively simple and inde- 
pendently motivated hypotheses: 
(1) Language users are rational agents engaged in 
goal seeking behaviour. Among these goals 
are the modification of the beliefs and goals 
of other agents. 
(2) Rational agents are frequently capable of 
identifying actions being performed by others 
and goals being sought. An essential part of 
helpful or cooperative behaviour is the adop- 
tion by one agent of a goal of another, fol- 
lowed by an attempt to achieve it. For exam- 
ple, for a store clerk to reply "How many do 
you want?" to a customer who has asked 
"Where are the steaks?", the clerk must have 
inferred that the customer wants steaks, then 
he must have decided to get them himself. 
This might have occurred even if the customer 
had intended to get the steaks him or herself. 
Cooperative behaviour must be accounted for 
independently of speech acts, for it often oc- 
curs without the use of language. 
(3) In order for a speaker to successfully perform 
a speech act, he must intend that the hearer 
recognize his intention to achieve the effects 
of the speech act, and must believe it is likely 
that the hearer will be able to do so. This is 
the foundation for the philosophical account 
of speech acts. 
(4) Language users know that others are capable 
of achieving goals, of recognizing actions, and 
of cooperative behaviour. Furthermore, they 
know that others know they know, etc. A 
speaker may intend not only that his actions 
be recognized but also that his goals be infer- 
red. 
(5) Thus a speaker can perform one speech act A 
by performing another speech act B if he in- 
tends that the hearer recognize not only that 
B was performed but also that through coop- 
erative behaviour by the hearer, intended by 
the speaker, the effects of A should be 
achieved. The speaker must also believe that 
it is likely that the hearer can recognize this 
intention. 
The process by which one agent can infer the plans 
of another is central to our account of speech acts. 
Schmidt et al \[1978\] and Genesereth \[1978\] present 
algorithms by which one agent can infer the goals of 
another, but assuming no interaction between the two. 
We describe the process in terms of a set of plausible 
plan inference rules directly related to the rules by 
which plans can be constructed. Let A and S be two 
agents and ACT an action. One example of a simple 
plan inference rule is: 
"If S believes that A wants to do ACT then 
it is plausible that S believes that A wants to 
achieve the effects of ACT." 
From simple rules like this can be derived more com- 
plex plan inference rules such as: 
"If S believes that A wants S to recognize 
A's intention to do ACT, then it is plausible 
that S believes that A wants S to recognize 
A's intention to achieve the effects of ACT." 
Notice that the complex rule is obtained by introduc- 
ing "S believes A wants" in the antecedent and conse- 
quent of the simple rule, and by interpreting "S recog- 
nizes A's intention" as "S comes to believe that A 
wants". Throughout the paper we identify "want" 
and "intend". 
We show that rules of the second type can account 
for S's recognition of many indirect speech acts by A, 
i.e. those in which S recognizes A's intention that S 
perform cooperative acts. 
To distinguish the use of, say, the indicative mood, 
in an assertion from its use in, say, an indirect request, 
the speech act operators REQUEST and INFORM of 
CP are reformulated and two further acts S.REQUEST 
and S.INFORM are added. These surface level acts 
are realized literally as indicative and imperative utter- 
ances. An S.REQUEST to INFORM is realized as a 
question. The surface level acts can be recognized 
immediately as parts of the higher level (or illocution- 
ary level) acts, to which the simple plan construction 
168 American Journal of Computational Linguistics, Volume 6, Number 3°4, July-December 1980 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
and inference rules can apply. Alternatively, the com- 
plex rules can be applied to the effects of the surface 
acts, and the intended performance of one of the illo- 
cutionary acts inferred later. 
For example, there are two ways an agent S could 
be led to tell A the secret after hearing A tell him 
"Can you tell me the secret?". Both start with S's 
recognition that A asked a yes/no question. In the 
first case, S assumes that A simply wanted to know 
whether S could tell the secret, then infers that A in 
fact wants to know the secret and, helpfully, decides 
to tell it. In the second case S recognizes that A 
intends S to infer that A wants to know the secret and 
that A intends S to tell A the secret, and thus that A 
has requested S to tell the secret. 
In general, several of the plan inference rules could 
apply at any time, and none of them guarantees a valid 
consequence. The application of the rules is con- 
trolled by a set of heuristics which rate the plausibility 
of the outcomes. 
Following a review of the relevant aspects of 
speech act theory in section 2, section 3 outlines our 
assumptions about beliefs, goals, actions, plans, and 
the plan inference process. Section 4 shows how the 
speech act definitions and the plan inference process 
can be used to relate literal to indirect meanings for 
REQUESTs and INFORMs. We show how utterances 
such as 1.h-l, and even 1.m can be used as requests to 
pass the salt, and what the origin of the several inter- 
pretations of 1.m is. 
(1.h) I want you to pass the salt. 
(1.i) Do you have the salt? 
(1.j) Is the salt near you? 
(1.k) I want the salt. 
(1.1) Can you pass the salt? 
(1.m) John asked me to ask you to pass the salt. 
Similarly we show how 1.n can be used to inform 
while 1.o cannot. Section 5 relates this work to the 
literature, while section 6 suggests further problems 
and draws some conclusions. 
(1,n) Do you know that the train is late? 
(1.o) Do you believe that the train is late? 
The speech act recognition process described here 
has been implemented as a computer program and 
tested by having it simulate an information clerk at a 
railway station. This domain is real, but sufficiently 
circumscribed so that interchanges between clerk and 
patrons are relatively short and are directed towards a 
limited set of goals. The program accepts as input 
simple English sentences, parses them using an ATN 
parser, and produces as output the speech act(s) it 
recognized and their associated propositional contents. 
It can handle all the examples discussed here. Details 
of the implementation can be found in Allen \[1979\]. 
2. Introduction to Speech Acts 
2.1. Basic Definitions 
Prior to Austin \[1962\], logicians considered the 
meaning of a sentence to be determined only by its 
truth value. However, Austin noted that some sen- 
tences cannot be classified as true or false; the utter- 
ance of one of these sentences constitutes the per- 
formance of an action, and hence he named them 
performatives. To quote Austin: "When I say, before 
the register or altar, etc., 'I do', I am not reporting on 
a marriage: I am indulging in it". 
Examples like this, and his inability to rigorously 
distinguish performative sentences from those which 
purportedly have truth value (which he called 
constatives) led Austin to the view that all utterances 
could be described as actions, or speech acts. He clas- 
sified speech acts into three classes, the locutionary, 
illocutionary, and perlocutionary acts. 
A locutionary act is an act of saying something: it 
is the act of uttering sequences of words drawn from 
the vocabulary of a given language and conforming to 
its grammar. 
An illocutionary act is one performed in making an 
utterance; "promise", "warn", "inform" and "request" 
are names of illocutionary acts. In general, any verb 
that can complete the sentence "I hereby <verb> you 
{that I to} ..." names an illocutionary act. An utter- 
ance has illocutionary force F if the speaker intends to 
perform the illocutionary act F by making that utter- 
ance. Verbs that name types of illocutionary acts are 
called performative verbs. From now on, we take 
speech acts to mean the illocutionary acts. 
Perlocutionary acts are performed by making the 
utterance. For example, S may scare A by warning A, 
or convince A of something by informing A of it. The 
success of a perlocutionary act is typically beyond the 
control of the speaker. For example, S cannot con- 
vince A of something against A's will, S can only pres- 
ent A with sufficient evidence so that A will decide to 
believe it. Perlocutionary acts may or may not be 
intentional. For instance, S may or may not intend to 
scare A by warning A. 
Searle \[1969\] suggests that illocutionary acts can be 
defined by providing, for each act, necessary and suf- 
ficient conditions for the successful performance of 
the act. Certain syntactic and semantic devices, such 
as mood and explicit performative verbs, are used to 
indicate iUocutionary force. 
One of the conditions included in Searle's account 
is that the speaker performs an illocutionary act only if 
he intends that the hearer recognize his intention to 
perform the act, and thereby recognize the illocution- 
ary force. This is important for it links Austin's work 
American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 169 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
on speech acts with the work of Grice on meaning, 
and is discussed in the next section. 
2.2. Communication and the Recognition of Intention 
Many philosophers have noted the relationship 
between communication (or speaker meaning) and the 
recognition of intention (Grice \[1957, 1968\], Strawson 
\[1964\], Searle \[1969\], Schiffer \[1972\].) Grice presents 
informally his notion of a speaker meaning something 
as follows: 
"'S meant something by x' is (roughly) 
equivalent to 'S intended the utterance of x 
to produce some effect in an audience by 
means of the recognition of this intention'" 
In other words, in order for S to communicate M 
by uttering x to A, S must get A to recognize that S 
intended to communicate M by uttering x. To use and 
example of Grice's, if I throw a coin out the window 
expecting a greedy person in my presence to run out 
and pick it up, I am not necessarily communicating to 
him that I want him to leave. For me to have success- 
fully communicated, he must at least have recognized 
that I intended him to leave. The same arguments 
hold when discussing illocutionary acts. For example, 
the only way S can request A to do ACT is to get A 
to recognize S's intention to request A to do ACT. 
2.3. The Indirect Speech Act Problem 
The relation between speech acts and the devices 
used to indicate them is complicated by the fact that 
performative verbs are seldom present and the same 
device can be used to perform many illocutionary acts. 
The interrogative mood, for example, can be used to 
request: "Can you pass the salt?" 
question: "Do you know the time?" 
inform: "Do you know that Sam got married?" 
warn: "Did you see the bear behind you?" 
promise: "Would I miss your party?" 
As many authors have pointed out, an utterance 
conveys its indirect illocutionary force by virtue of its 
literal one (Searle \[1975\], Morgan \[1977\], Morgan 
\[1978\]). "It's cold here" can function as a request to, 
say, close the window, in part because it's an assertion 
that the temperature is low. 
Most of the literature on the treatment of indirect 
speech acts within the theory of grammar stems from 
the work of Gordon and Lakoff \[1975\] (hereafter 
GL). They claim that direct and indirect instances of 
the same speech act have different "meanings", i.e. 
different logical forms, and they propose a set of 
"conversational postulates" by which literal forms 
"entail" indirect ones. The postulates for requests 
correspond to conditions that must obtain for a re- 
quest to be sincere. For A to sincerely request B to 
do ACT, the following sincerity conditions must hold: 
(1) A wants ACT. 
(2) B can do ACT. 
(3) B is willing to do ACT. 
(4) B will not do ACT in the absence 
of the request. 
They then propose that one can convey a request 
by asserting a speaker-based sincerity condition 
(condition 1), or querying a hearer-based sincerity 
condition (conditions 2-4). 
The postulates for indirect requests given in GL do 
not account for the readings of 2.3a and 2.3b as re- 
quests, and although more rules could be added (and 
some should be weakened) we believe this solution to 
be misguided. 
(2.3a) Is the salt near you? 
(2.3b) John asked me to ask you to pass the salt. 
GL's postulates directly relate the literal form of 
one speech act to the indirect form of another. Thus 
they do not predict why certain acts allow certain indi- 
rect forms. For example, the postulates do not ac- 
count for why 2.3c-d can be requests while 2.3e-f 
cannot. But 2.3e is infelicitous as a (literal) question 
since there is no context where one can acquire infor- 
mation by querying one's own mental state. Utterance 
2.3f is a reasonable question but even if the speaker 
found out the answer, it would not get him any closer 
to acquiring the salt (by having the hearer pass it). A 
theory of indirect speech acts should capture these 
facts; GL's does not (although they agree it should). 
(2.3c) I want the salt. 
(2.3d) Do you want to pass the salt? 
(2.3e) Do I want the salt? 
(2.3f) Does he want to pass the salt? 
Similarly, GL's postulates fail to explain the rela- 
tion between indirect forms of different speech acts. 
For example, 2.3g can be an assertion that P and 2.3h 
cannot, for the same reasons that 2.3i can be a request 
to do A and 2.3j cannot. 
(2.3g) I want you to know that P. 
(2.3h) Do I want you to know that P? 
(2.3i) I want you to A. 
(2.3j) Do I want you to A? 
The hearer's knowing that P obtains is an intended 
perlocutionary effect of an informing act, just as the 
hearer's doing an act A is an intended effect of a re- 
quest. A speaker can indirectly inform or request by 
informing the hearer that the speaker desires the per- 
locutionary effect of that act, and intending that the 
hearer recognize the speaker's intention that the perlo- 
cutionary effect should be achieved. 
This paper shows that what GL achieve with their 
postulates can be derived from the five hypotheses 
given in the Introduction. Our proposal here is a de- 
170 American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
velopment of Searle \[1975\]. It requires separating the 
surface form conditions completely from the defini- 
tions of the illocutionary acts and introducing an inter- 
mediary level, the surface acts. 
Our theory of indirection will however share with 
GL some problems brought up by Sadock \[1970\], 
Green \[1975\], and Brown \[1980\]. These are discussed 
further in section 4.5. 
3. Plans, Plan Construction, and Plan Inference. 
Our analysis of indirect REQUESTs and INFORMs 
relies on the inference by the hearer of some of the 
goals of the speaker and of some of the actions which 
the speaker is taking to achieve those goals. Section 
3.1 outlines the form of the models of the world which 
language users are assumed to have, in particular their 
beliefs about the world (and about other agents), and 
their goals. In section 3.2 we define actions and how 
they affect the belief model. The rules for plan con- 
struction and inference are considered in sections 3.3 
and 3.4. Because of space limitations, this section is 
very sketchy. More detail, motivation, and problems, 
are available in Allen \[1979\] and Allen and Perrault 
\[19801. 
3.1. Beliefs, Knowledge, and Goals 
3.1.1. The Belief Model 
We assume that every agent S has a set of beliefs 
about the world, which may include beliefs about oth- 
er agents' beliefs. Agents can hold false beliefs. As 
Quine \[1956\] pointed out, belief creates a context 
where substitution of coreferential expressions need 
not preserve truth-value. 
We add to a first-order language with equality the 
operator B, and B(A,P) (usually written BA(P)) is to 
be read "A believes that P", for any formula P. The 
B operator is assumed to satisfy the following axiom 
schemas (inspired by Hintikka \[1962\]), where P and Q 
are schema variables ranging over propositions, and A 
ranges over agents: 
(B.0) all theorems of First Order Predicate Calculus 
(B.1) BA(P ) -~ BA(BA(P)) 
(B.2) BA(P ) ^ BA(Q ) ~.~ BA(P ^ Q) 
(B.3) BA(P) v BA(Q) ~ BA(P v Q) 
(B.4) BA(~P) ~ ~BA(P) 
(B.5) (3x) BA(P(x)) =~ BA((~x)P(x)) 
(B.6) (BA(P =~ Q) A BA(P)) ~-~ BA(Q) 
The rules of inference are Modus Ponens and: 
If T is a theorem, then BA(T ) is a 
theorem, for every agent A. 
i.e. every agent believes every valid consequence of 
the logical axioms. 
The partial deduction system used in the implemen- 
tation of Allen \[1979\] is based on Cohen \[1978\]. The 
foundations for a more elaborate system can be found 
in Moore \[1979\]. 
3.1.2. Knowing 
The word "know" is used in at least three different 
senses in English. One may know that a proposition P 
is true, know whether a proposition P is true or know 
what the referent of a description is. 
We define "A knows that P", written KNOW(A,P), 
as P ^ BA(P). This is weaker than some definitions of 
"know" in the philosophical literature, where, among 
other things, "A knows that P" entails that A believes 
P for the "right reasons"; i.e. knowledge is true and 
justified belief (Ayer \[1956\], but see also Gettier 
\[1963\]). If S believes that A knows that P, S is com- 
mitted to believing that P is true. 
Unfortunately, the meaning of "A does not know 
that P" is not captured by ~(P a BA(P)), but by the 
weaker (P ^ ~BA(P)), i.e. 
~KNOW(A,P) -= P A ~BA(P ) 
In other words, if S believes A does not know P, then 
S must believe that P is true in addition to believing 
that A does not believe P is true. This problem is 
analogous to the wide/narrow scope distinction that 
Russell found in his account of definite descriptions 
(Russell \[1919\]). One solution to this problem is to 
consider KNOW as a "macro" whose expansion is 
sensitive to negation. Details may be found in Allen 
\[1979\]. 
A knows whether a proposition P is true if A 
KNOWs that P or A KNOWs that ~P. 
KNOWlF(A,P) -= KNOW(A,P) v KNOW(A,~p) 
Knowing what the referent of a description is re- 
quires quantification into belief. One of its arguments 
is a formula with exactly one free variable. 
KNOWREF(A,P(x)) 
(\]y) ((Vz) P(z) --- y= z) 
^ BA((Vz ) P(z) ~ y = z) 
A KNOWREF the departure time of TRAIN1 if 
TRAIN1 has a unique departure time y, and if A be- 
lieves that y is TRAINI's unique departure time. 
3.1.3. Wanting 
We let W(A,P) (usually written WA(P)) mean 
"agent A wants P to be true". P can be either a state 
or the execution of some action. In the latter case, if 
ACT is the name of an action, WA(ACT(b)) means 
"A wants b to do ACT". 
The logic of want is even more difficult than that 
of belief. It is necessary for us to accept the follow- 
ing: 
American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 171 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
(W.1) WA(P ) ~- BA(WA(P)). 
(W.2) WA(P ^ Q) ~.~ WA(P ) A WA(Q). 
The most interesting interactions between the belief 
and want operators come from the models that agents 
have of each other's abilities to act and to recognize 
the actions of others. This will be further discussed in 
the following section. 
3.2. Actions and Plans 
Actions model ways of changing the world. As 
with the operators in STRIPS (Fikes and Nilsson 
\[1971\]), the actions can be grouped into families rep- 
resented by action schemas, which can be viewed as 
parameterized procedure definitions. An action sche- 
ma consists of a name, a set of parameters with con- 
straints and a set of labelled formulas in the following 
classes: 
Effects: Conditions that become true after the 
execution of the procedure. 
Body: a set of partially ordered goal states that 
must be achieved in the course of executing the 
procedure. In the examples given here, there will 
never be more than one goal state in a body. 
Preconditions: Conditions necessary to the suc- 
cessful execution of the procedure. We distin- 
guish for voluntary actions a want precondition: 
the agent must want to perform the action, i.e. he 
must want the other preconditions to obtain, and 
the effects to become true through the achieve- 
ment of the body. 
The constraints on the parameters consist of type 
specifications, and necessary parameter interdependen- 
ties. Each action has at least one parameter, namely, 
the agent or instigator of the action. In the blocks 
world, for example, the action of putting one block on 
top of another could be defined as: 
PUTON(a,bl,b2) 
constraints: AGENT(a) A BLOCK(bl) A BLOCK(b2) 
precondition: CLEAR(bl) A CLEAR(b2) 
A Wa(PUTON(a,bl,b2)) 
effect: ON(bl,b2) 
The preconditions, effects and body provide infor- 
mation to the plan construction and inference process- 
es so that they can reason about the applicability and 
effect of performing the action in a given context. 
Finally, the body of the action specifies what steps 
must be achieved in the course of the execution of the 
action. Primitive actions have no bodies; their execu- 
tion is specified by a non-examinable procedure. 
All agents are assumed to believe that actions 
achieve their effects and require their preconditions. 
We need the following axioms: 
For all agents a and b, and for all actions 
ACT, if PRE is the precondition of ACT and 
EFF its effect then: 
(ACT.l) BA(ACT(b) =~ PRE). 
(ACT.2) BA(ACT(b) ~ EFF). 
Every predicate and modal operator in these axioms, 
and throughout the paper, should be indexed by a 
state or time. The resulting logic would be, according- 
ly, more complex. The issue is raised again in sect. 6. 
3.3. Plan Construction 
A plan to transform a world W\[0\] (represented by 
a formula) into a world W\[n\] is a sequence of actions 
A1 ..... An such that the preconditions of Ai are true 
in W\[i-1\], and Ai transforms world W\[i-1\] into W\[i\]. 
An agent can achieve a goal by constructing and 
then executing a plan which transforms the current 
state of the world into one in which the goal obtains. 
This can be done by finding an operator which, if 
executed in some world, would achieve the goal. If its 
preconditions are satisfied in the initial world, the plan 
is complete. Otherwise, the planning process attempts 
to achieve the preconditions. This simple view of plan 
construction as a "backward chaining" process can be 
refined by assuming different levels of "detail" in the 
representation of the world and of the operators. This 
view (as developed in Sacerdoti \[1973, 1975\], for 
example) allows plans constructed at one level of de- 
tail to be expanded to a lower level through the bodies 
of their constituent acts. 
As noted earlier, the agent of an action must be- 
lieve that its precondition is true to believe that his 
executing the action will succeed. For agent A to plan 
that agent S should perform action ACT, A must 
achieve that S should believe that the precondition of 
ACT holds, and S's beliefs should not be inconsistent 
with A's, i.e. it must be true that BA(KNOW(S,P)), 
where P is the precondition of ACT. 
We assume that an agent cannot do an action with- 
out wanting to do that action. Thus a precondition of 
every action ACT by an agent A is that 
WA(ACT(A)). 
We are concerned with the model that agents have 
of each other's plan construction and inference proc- 
ess, and consider these two processes as consisting of 
chains of plausible inferences operating on goals and 
observed actions. The processes are specified in two 
parts: first as schemas of rules which conjecture that 
certain states or actions can be added to a plan being 
constructed. The plausibility of the plans containing 
the result of the inferences is then evaluated by rating 
heuristics. Thus the plan construction and inference 
rules are not to be interpreted as valid logical rules of 
inference. 
172 American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
The first three plan construction (PC) rules are: 2 
(PC.EA) \[Effect-action rule\] For any agent A, if 
Y is an effect of action X, then if A wants Y to 
hold, it is plausible that A will want action X to 
be done. 
(PC.AP) \[Action-precondition rule\] For any agent 
A, if X is a precondition of action Y, and if A 
wants Y to be done, then it is plausible that S will 
want X to hold. P$, 
(PC.AB) \[Action-body rule\] For any agent A, if 
A wants an action Y to be done, and if X is a 
part of the body of Y then it is plausible that S 
will want X to be done. 
If X and Y are systematically replaced by one of 
the pairs in Figure 1, then rules PC.EA, PC.AP, and 
PC.AB can all be written as 
WA(Y) =c=> WA(X) 
with =c=> indicating that the rule is a construction 
rule. 
We also need a rule based on KNOWIF: 
(PC.KI) \[KNOWlF rule\] For any agent A, if A 
wants P to be true, then it is plausible that A 
should want to know whether P is true. 
WA(P) =c=> WA(KNOWIF(A,P)) 
3.4. Plan Inference 
For every plan construction rule 
WA(Y) =c=> WA(X), 
and every agent S, there is a corresponding plan 
inference (PI) rule which is written 
BsWA(X ) =i= > BsWA(Y ). 
The following rules correspond to the effect-action, 
action-precondition, and action-body rules of the pre- 
vious section: 
(PI.AE) \[Action-effect rule\] For all agents S and 
A, if Y is an effect of action X and if S believes 
that A wants X to be done, then it is plausible 
that S believes that A wants Y to obtain. 
(PI.PA) \[Precondition-action rule\] For all agents 
S and A, if X is a precondition of action Y and if 
S believes A wants X to obtain, then it is plausi- 
ble that S believes that A wants Y to be done. 
(PI.BA) \[Body-action rule\] For all agents S and 
A, if X is part of the body of Y and if S believes 
that A wants X done, then it is plausible that S 
believes that A wants Y done. 
There are two inverses to the KNOWIF rule: if A 
wants to know whether P is true, then A may want P 
to be true, or A may want P to be false. 
2 Throughout the rest of the paper agent A will usually de- 
note the constructor/executor of plans, and S (or System) the 
recognizer of plans (usually constructed by A). 
X 
ACT 
preconditions of ACT 
body of ACT 
effects of ACT 
ACT 
ACT 
Figure !. Arguments for PC/PI rules. 
(PI.KP) \[Know positive\] For all agents S and A, 
BsWA(KNOWIF(A,P)) =i=> BsWA(P) 
(PI.KN) \[Know negative\] For all agents S and A, 
BsWA(KNOWIF(A,P)) =i=> BsWA(~P) 
PI.W is the special case of the precondition-action 
rule where the precondition is the want precondition: 
(PI.W) \[Want rule\] For all agents S, A, and C 
and for all actions ACT whose agent is C, it is 
plausible that 
BsWA(Wc(ACT)) =i=> BsWA(ACT ) 
3.4.1. The Plan Inference Process 
The plan inference rules generate formulas which 
the recognizing agent believes are possible. A sepa- 
rate mechanism is used to evaluate their plausibility. 
An agent S attempting to infer the plans of another 
agent A starts with an observed action of A and a 
(possibly empty) set of goals or expectations which S 
believes A may be trying to achieve. S attempts to 
construct a plan involving the action and preferably 
also including some of the expectations. 
Plan inference is a search through a space of partial 
plans each consisting of two parts. One part is con- 
structed using the plan inference rules from the ob- 
served action (and called the alternative); the other is 
constructed using the plan construction rules from an 
expected goal (and called the expectation). 
The partial plans are manipulated by a set of tasks 
which decide what rules are to be applied, what 
"merges" between alternatives and expectations 
should be attempted, and when the process terminates. 
The partial plans and their associated tasks are rated 
by a set of heuristics, and the most highly rated task is 
executed first. 
3.4.2. Rating Heuristics 
The rating of a partial plan reflects how likely it is 
to be part of the "correct" plan, i.e. the plan the 
speaker is executing. If several incompatible inferenc- 
es can be made from one point in the alternative, then 
its rating is divided among them. The heuristics de- 
scribed in this section are based on domain independ- 
ent relations between actions, their bodies, precondi- 
tions, and effects. The need for more domain depend- 
ent measures is discussed later. 
American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 173 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
The heuristics are described here only in terms of 
increasing or decreasing ratings of partial plans. 
Decrease the rating of a partial plan in which 
the preconditions of executing actions are 
currently false. 
Decrease the rating of a partial plan contain- 
ing a pending action ACT by an agent A if A 
is not able to do ACT. 3 
Decrease the rating of a partial plan in which 
the effects of a pending act already obtain or 
are not wanted by the planner. 4 
Other heuristics depending on how well the utter- 
ance fits with the expectations are not immediately 
relevant to understanding indirect speech acts and will 
not be discussed here. One further heuristic is added 
in section 4.3. 
In general several rating heuristics are applicable to 
an partial plan. Their effects on the rating of the par- 
tial plan are cumulative. 
3.4.3. Extending the Inference Rules 
A hearer S identifies the illocutionary force of an 
utterance by recognizing that the speaker A has cer- 
tain intentions, namely that S should recognize some 
intention P of A's. This can be represented by a for- 
mula of the form BsWA(BsWA(P)). To do the recog- 
nition, the simple plan construction and inference rules 
of sections 3.3 and 3.4 must be extended so that they 
can operate on these nested formulas. This can be 
done by assuming that every agent is aware that other 
agents construct and infer plans in the same way he 
can. In fact, both the simple inference and construc- 
tion rules are necessary to derive the extended infer- 
enee rules. 
The extended rules are specified by "meta-rules" 
which show how to construct new PC/PI rules from 
old ones. The first extended construction rule (EC. 1) 
is: A can achieve that S recognizes that A wants the 
effect of ACT by achieving that S recognizes that A 
wants ACT to be done, assuming that S would infer 
that the effects of ACT are also desired. The same 
rule applies if we replace "wants the effect of ACT" 
and "wants ACT to be done" by any pair of Y and X, 
as given in Figure 1. We assume all these sul~stitu- 
3 This definition is the same as Cohen's CANDO relation. 
Being able to do an action means that the action's preconditions are 
either presently true, achieved within the existing plan, or can be 
achieved by a "relatively simple plan", which we take to be a single 
action whose preconditions are presently true or achieved in the 
existing plan. 
4 We have avoided the problem here of planning to do a task 
that requires one to deny a subgoal temporarily so that some action 
can execute, and then needing to reaehieve that (presently true) 
goal. 
tions are possible in rules EC.1 - EC.3 and EI.1 - 
EI.3. 
(EC.1) If BsWA(X) =i=> BsWA(Y) is a PI rule, then 
WA(BsWA(Y)) =c=> WA(BsWA(X)) is a PC rule. 
Similarly we can generate the corresponding PI rule: 
(EI.1) If BsWA(X) =i=> BsWA(Y) is a PI rule, then 
BsWA(BsWA(X)) =i=> BsWA(BsWA(Y)) is a PI 
rule. 
EI. 1 allows prefixing BsW A to plan inference rules. 
Plan construction rules can also be embedded: if A 
wants S to want to do ACT, then A should be able to 
achieve this by achieving that S wants the effect of 
ACT, and by relying on S to plan ACT. In other 
words: 
(EC.2) If Ws(Y) =c=> Ws(X) is a PC rule, then 
WA(Ws(X)) =c=> WA(Ws(Y)) is a PC rule. 
Correspondingly, 
(EI.2) If Ws(Y) =e=> Ws(X) is a PC rule, then 
BsWA(Ws(Y)) =i=> BsWA(Ws(X)) is a PI rule. 
Finally, any agent A can plan for S to recognize 
A's intention that S plan, and for S to be able to rec- 
ognize this intention in A. For example, A can plan 
for S to recognize A's intention that S want to close 
the door by planning for S to recognize A's intention 
that S watlt the door closed. These rules are obtained 
by using EI.2 as the PI rule which is "extended" by 
EC. 1 and El. 1. 
(EC.3) If Ws(Y) =c=> Ws(X) is a PC rule, then 
WABs(WAWs(X)) =c=> WABs(WAWs(Y)) is a PC 
rule. 
(EI.3) If Ws(Y) =e=> Ws(X) is a PC rule, then 
BsWA(BsWA(Ws(Y))) =i=> BsWA(BsWA(Ws(X))) 
is a PI rule. 
Our "toolkit" is now sufficiently full to allow us to 
consider some speech acts and their recognition. 
4. Plan Inference and Indirect Speech Acts 
4.1. Speech Acts 
The definitions of the speech acts REQUEST and 
INFORM used in this paper are slightly different from 
the ones in Cohen and Perrault \[1979\] in that they 
rely on the existence of speech act bodies to account 
for indirect forms. Plans including speech acts are 
now thought of as having two levels, the illocutionary 
level and the surface level. Acts at the illocutionary 
level model the intentions motivating an utterance 
independently of the syntactic forms used to indicate 
those intentions. Acts at the surface level are realized 
by utterances having specific illocutionary force indi- 
cators. 
174 American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
'The first illocutionary level act is one by which a 
speaker informs a hearer that some proposition is true. 
INFORM(speaker, hearer, P) 
prec: KNOW(speaker,P) ^ 
W(speaker,INFORM(speaker,hearer,P)) 
effect: KNOW(hearer,P) 
body: B(hearer,W(speaker,KNOW(hearer,P))) 
For A to sincerely inform S that P is true, A must 
believe A knows that P is true and want to inform S 
that P (the preconditions), and must intend to get S to 
know that P is true (the effect), which is done by con- 
structing a plan that will achieve S's recognition of this 
intention (i.e. that Bs(WA(KNOW(S,P))~). A then 
must depend on S to bring about the effect: S must 
decide to believe what A said. This is made explicit by 
introducing an admittedly simplistic DECIDE TO 
BELIEVE act: 
DECIDE TO BELIEVE(agent, other, P) 
pree: B(agent,W(other,KNOW(agent,P))) 
effect: KNOW(agent,P) 
Thus A can INFORM S of P by achieving 
BsWA(KNOW(S,P)) followed by DECIDE TO 
BELIEVE(S,A,P). 
In many cases, agents reason about INFORM acts 
to be performed (by others or by themselves) where 
the information for the propositional content is not 
known at the time of plan construction. For example, 
A may plan for S to inform A whether P is true. A 
cannot plan for S to perform INFORM(S,A,P) since 
this assumes the truth of P. We get around this diffi- 
culty by defining INFORMIF, another view of the 
INFORM act. 
INFORMIF(speaker, hearer, P) 
prec: KNOWlF(speaker,P) A 
W(speaker,INFORMIF(speaker,hearer,P)) 
effect: KNOWIF(hearer,P) 
body: B(hearer,W(speaker,KNOWIF(hearer,P))) 
Similarly, it must be possible for A to plan for S to 
tell A the referent of a description, without A knowing 
the referent. This is the role of the INFORMREF act. 
INFORMREF(speaker, hearer, D(x)) 
pree: KNOWREF(speaker,D(x)) ^ 
W(speaker,INFORMREF(speaker, 
hearer,D(x))) 
effect: KNOWREF(hearer,D(x)) 
body: B(hearer,W(speaker,KNOWREF( 
hearer,D(x)))) 
Request is defined as: 
REQUEST(speaker, hearer, action) 
constraint: hearer is agent of action 
prec: W(speaker,action(hearer)) 
effect: W(hearer,aetion(hearer)) 
body: B(hearer,W(speaker,action(hearer))) 
The intention of a request is to get the hearer to 
want to do the action, and this is accomplished by 
getting the hearer to believe that the speaker wants 
the hearer to do the action and then depending on the 
hearer to decide to do it. 5 To explicitly represent this 
decision process, a CAUSE TO WANT act defined 
along the lines of the DECIDE TO BELIEVE act 
above is necessary. 
CAUSE TO WANT(agent, other, P) 
prec: B(other,B(agent,W(agent,P))) 
effect: W(other,P) 
As examples of the use of speech acts, "Tell me 
whether the train is here" and "Is the train here?", 
intended literally, are both REQUESTs by A that S 
INFORMIF the train is here. "When does the train 
arrive?", intended literally, is a REQUEST by A that 
H INFORMREF of the departure time of the train. 
Finally we define the two surface level acts: 
S.INFORM produces indicative mood utterances, and 
S.REQUEST produces imperative utterances, or inter- 
rogative utterances, if the requested act is an IN- 
FORM. These acts have no preconditions, and serve 
solely to signal the immediate intention of the speaker, 
the starting point for all the hearer's inferencing. 
S.INFORM(speaker, hearer, P) 
effect: B(hearer,W(speaker,KNOW(hearer,P))) 
S.REQUEST(speaker, hearer, action) 
effect: B (hearer,W(speaker,action(hearer))) 
The effects of S.INFORM match the body of the IN- 
FORM act, reflecting the fact that it is a standard way 
of executing an INFORM. It is important, however, 
that S.INFORM is only one way of executing an IN- 
FORM. The same relationship holds between the 
S.REQUEST and REQUEST actions. 
4.2. Recognizing IIIocutionary Force 
Given the speech act definitions of section 4.1, we 
say that A performed an illocutionary act IA by uttering 
x to S if A intends that S should recognize (and be 
able to recognize) that 
(1) x is an instance of a surface act SA, and 
(2) A intended S to infer (using the PI rules and 
associated heuristics) from A having performed 
SA that A wants to achieve the effects of IA. 
This definition allows more than one illoeutionary 
act to be performed by a single surface act. In this 
section we show how the hearer of an utterance can 
recognize the speaker's intention(s) indicated by a 
speech act, especially when these intentions are com- 
municated indirectly. 
5 See Cohen and Perrault \[1979\] for a discussion of why 
Searle's preparatory conditions "Speaker believes Hearer can do the 
action" need not bc part of the preconditions on REQUEST. 
American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 175 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
All inferencing by S of A's plans starts from S's 
recognition that A intended to perform one of the 
surface acts, and that A in fact wanted to do the act. 
All inference chains will be shown as starting from a 
formula of the form BsWA(A do the surface act). The 
object of the inferencing is to find what illocutionary 
level act(s) A intended to perform. The action-effect 
rule applied to the starting formula yields one of the 
form BsWA(BsWA(P)), i.e. S believes that A wants S 
to recognize A's intention that P. The inferencing 
process searches for plausible formulas of the form 
BsWA(IA(A)) where IA is an illocutionary level act. 
(1) S.REQUEST(A,S,PASS(S,A,SALT)) 
PI.AE (2) BsWA(PASS(S,A,SALT)) 
PI.BA (3) REQUEST(A,S,PASS(S,A,SALT)) 
Example 1. "Pass the salt," 
Example 1 shows a direct request to pass the salt, 
where the surface request maps directly into the in- 
tended request interpretation. 6 The actions relevant to 
the examples given here are: 
PASS(agent, beneficiary, object) 
prec: HAVE(agent, object) 
effect: HAVE(beneficiary, object) 
REACH(agent, object) 
prec: NEAR(agent, object) 
effect: HAVE(agent, object) 
Let us also assume that S presently has the salt, i.e. 
HAVE(S,SALT) is true, and mutually believed by S 
and A. 
The rating heuristics for the complex rules El. 1 to 
EI.3 are the same as for the PI rules but each heuristic 
may be applicable several times at different levels. 
For example, consider the frequently recurring infer- 
ence chain: 
(1) BsWA(ACT(S)) 
PI.BA (2) REQUEST(A,S,ACT(S)) 
PI.AE (3) Ws(ACT(S)) 
PI.PA (4) ACT(S) 
PI.AE (5) effects of ACT(S) 
It shows the line of inference from the point where S 
recognizes that A requested S to do ACT (at step (2)) 
to the point where the effects of the requested action 
are inferred as part of A's plan. Of interest here is the 
evaluation of the plausibility of step (3). Two heuris- 
tics are applicable. The proposition "Ws(ACT(S))" is 
6 To improve readability of inference chains in the examples, 
we drop the prefix BsW A from all propositions. The formula on 
line (n) follows from the one on line (n-l) by the rule at the begin- 
ning of line (n). Applications of EI.1 will be labelled "rule"/EI.l, 
where "rule" is a PI rule embedded by EI.1. Similarly, applications 
of EI.2 and EI.3 will be labelled "rule"/EI.2 and "rule"/EI.3, 
where "rule" is a PC rule name. 
evaluated with respect to what S believes A believes. 
(Remember that BsW A should appear as a prefix to all 
propositions in inference chains.) If BsBAWs(ACT(S)) 
is true, the request interpretation is considered unlike- 
ly, by the effect-based heuristic. In addition, the pre- 
conditions of ACT(S) are considered with respect to 
what S believes A believes S believes. This step will 
only be reasonable if S can do the action, by a 
precondition-based heuristic. 
To make more explicit the distinction between in- 
ferences in BsW A and inferences in BsWABsWA, let 
us consider two inference chains that demonstrate two 
interpretations of the utterance "Do you know the 
secret?". Lines 1-3 of Example 2 show the chain 
which leads S to believe that A asked a (literal) 
yes/no question; lines 1-6 of Example 3 show the 
interpretation as a request to S to inform A of the 
secret. Notice that in both interpretations S may be 
led to believe that A wants to know the secret. In the 
literal case, S infers A's goal from the literal interpre- 
tation, and may tell the secret simply by being helpful 
(lines 4-9). In the indirect case, S recognizes A's 
intention that S inform A of the secret (lines 1-6). 
Telling the secret is then conforming to A's intentions 
(lines 7-9). 
There is in fact a third interpretation of this sen- 
tence. If A and S both know that A already knows 
the secret, then the utterance could be intended as 
"If you don't know the secret, I will tell it to you." 
This requires recognizing a conditional action and is 
beyond our present abilities. 
4.3. The Level of Embedding Heuristic 
Two sets of PI rules are applicable to formulas of 
the form BsWABsWA(P): the simple rules PI.1 to PI.6 
operating "within" the prefix BsW A, and the rules 
generated by EI.1 and EI.3 which allow the simple 
rules to apply within the prefix BsWABsW A. To re- 
flect the underlying assumption in our model that in- 
tention will always be attributed if possible, the infer- 
ences at the most deeply nested level should be prefer- 
red. 
Of course, if the inferences at the nested level lead 
to unlikely plans, the inferences at the "shallow" lev- 
els may be applied. In particular, if there are multiple 
mutually exclusive inferences at the nested level, then 
the "shallow" inferences will be preferred. This re- 
flects the fact that the nested inferences model what 
the speaker intends the hearer to infer. If there are 
many inferences possible at the nested level, the 
speaker would not be able to ensure that the hearer 
would perform the correct (i.e., the intended) one. 
176 American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
PI.AE 
PI.BA 
PI.AE 
PI.W 
PI.AE 
PI.KP 
PI.PA 
PI.AE 
(1) S.REQUEST(A,S,INFORMIF(S,A,KNOWREF(S,SECRET(x)))) 
(2) BsWA(INFORMIF(S,A,KNOWREF(S,SECRET(x)))) 
(3) REQUEST(A,S,INFORMIF(S,A,KNOWREF(S,SECRET(x)))) 
(4) Ws(INFORMIF(S,A,KNOWREF(S,SECRET(x)))) 
(5) INFORMIF(S,A,KNOWREF(S,SECRET(x))) 
(6) KNOWIF(A,KNOWREF(S,SECRET(x))) 
(7) KNOWREF(S,SECRET(x)) 
(8) INFORMREF(S,A,SECRET(x)) 
(9) KNOWREF(A,SECRET(x)) 
Example 2. "Do you know the secret?" (yes/no question) 
PI.AE 
PI.AE/EI.1 
PI.KP/EI. 1 
PI.PA/EI.1 
PI.BA 
PI.AE 
PI.W 
PI.AE 
(\]) S.REQUEST(A,S,INFORMIF(S,A,KNOWREF(S,SECRET(x)))) 
(2) BsWA(INFORMIF(S,A,KNOWREF(S,SECRET(x)))) 
(3) BsWA(KNOWlF (A,KNOWREF(S,SECRET(x)))) 
(4) BsWA(KNOWREF(S,SECRET(x)) ) 
(5) BsWA(INFORMREF(S,A,SECRET(x)) ) 
(6) REQUEST(A,S,INFORMREF(S,A,SECRET(x))) 
(7) Ws(INFORMREF(S,A,SECRET(x)) ) 
(8) INFORMREF(S,A,SECRET(x)) 
(9) KNOWREF(A,SECRET(x)) 
Example 3. "Do you know the secret?" (indirect request) 
PI.AE 
PI.PA/EI.1 
PI.AE/EI.1 
PI.W/EI.1 
PI.BA 
(1) S.INFORM(A,S,WA(PASS(S,A,SALT)) 
(2) BsWA(Bs(WA(PASS(S,A,SALT)))) 
(3) BsWA(CAUSE TO WANT(A,S,PASS(S,A,SALT)) ) 
(4) BsWA(Ws(PASS(S,A,SALT)) 
(5) BsWA(PASS(S,A,SALT)) 
(6) REQUEST(A,S,PASS (S,A,SALT)) 
Example 4. "I want you to pass the salt." 
4.4. More Indirect Requests 
Example 4 shows the interpretation of "I want you 
to pass the salt" as a request. Taking the utterance 
literally, S infers that A wants him to know that A 
wants him to pass the salt. This yields proposition (2) 
which leads through the next three inferences to the 
intention that would be recognized from a request act, 
i.e. that A wants S to pass the salt (5). Notice that an 
application of the body-action rule to step (2) yields: 
INFORM(A, S, WA(PASS(S, A, SALT))), 
for, in fact, the speaker may be performing both 
speech acts. The level of inferencing heuristic favours 
the indirect form. 
The key step in Example 5 is the application of the 
know-positive rule from line (3) to line (4). Since, 
given the context, S assumes that A knows whether S 
has the salt, the literal interpretation (from (2)) would 
not produce a reasonable goal for A. This supports 
the nested know-positive inference, and attributes 
further intention to the speaker (4). Once this is 
done, it is easy to infer that A wants S to pass him the 
salt (5), hence the request interpretation. 
"Can you pass the salt?" and "Do you want to pass 
the salt?" are treated similarly, for they inquire about 
the preconditions on PASS(S, A, SALT). 
Example 6 begins like Example 5, leading to the 
inference that A wants S to be able to reach the salt 
(4). 7 Since being able to reach the salt is a precondi- 
tion to reaching the salt (5), which then enables pass- 
ing the salt (6), S can infer that he is being requested 
to pass the salt. "Is the salt near you?" can be treated 
in the same way, as being near the salt is a precondi- 
tion on reaching the salt. 
7 Let CANDO(S,ACT) be true if S believes the preconditions 
of ACT are true. 
American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 177 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
PI.AE 
PI.AE/EI. 1 
PI.KP/EI. 1 
PI.PA/EI. 1 
PI.BA 
(1) S.REQUEST(A,S,INFORMIF(S,A,HAVE(S,SALT))) 
(2) BsWA(INFORMIF(S,A,HAVE(S,SALT))) 
(3) BsWA(KNOWlF(A,HAVE(S,SALT))) 
(4) BsWA(HAVE(S,SALT)) 
(5) BsWA(PASS(S,A,SALT)) 
(6) REQUEST(A,S,PASS(S,A,SALT)) 
Example 5. "Do you have the salt?" 
PI.AE 
PI.AE/EI. 1 
PI.KP/EI.1 
PI.PA/EI. 1 
PI.AE/EI. 1 
PI.PA/EI.1 
PI.BA 
(1) S.REQUEST(A,S,INFORMIF(S,A,CANDO(S,REACH(S,SALT))) 
(2) BsWA(INFORMIF(S,A,CANDO(S,REACH(S,SALT))) 
(3) BsWA(KNOWIF(A,CANDO(S,REACH(S,SALT)))) 
(4) BsWA(CANDO(S,REACH(S,SALT))) 
(5) BsWA(REACH(S,SALT)) 
(6) BsWA(HAVE(S,SALT)) 
(7) BsWA(PASS(S,A,SALT)) 
(8) REQUEST(A,S,PASS(S,A,SALT)) 
Example 6. "Can you reach the salt?" 
PI.AE 
PI.PA/EI.1 
PI.AE/EI. 1 
PI.AE/EI.3 
PI.W/EI. 1 
PI.BA 
(1) S.INFORM(A,S,WA(HAVE(A,SALT)) 
(2) BsWA(Bs(WA(HAVE(A,SALT)))) 
(3) BsWA(CAUSE TO WANT(A,S,HAVE(S,A,SALT))) 
(4) BsWA(Ws(HAVE(A,SALT))) 
(5) BsWA(Ws(PASS(S,A,SALT))) 
(6) BsWA(PASS(S,A,SALT)) 
(7) REQUEST(A,S,PASS (S,A,SALT)) 
Example 7. "I want the salt." (= "I want to have the salt.") 
Example 7 includes in the step from (3) to (4), an 
application, through EI.3, of the effect-action rule. A 
informs S of A's goal of having the salt (2) and then 
depends on S's planning on that goal to infer the 
PASS action. Because the action is the "obvious" way 
of achieving the goal, S believes that A intended him 
to infer it. 
Since questions are treated as requests to inform, 
most of them are handled in a similar manner to the 
requests above. 4.4a-h can all be understood as ques- 
tions about the departure time of some train. 
(4.4a) When does the train leave? 
(4.4b) I want you to tell me when ... 
(4.4c) I want to know when ... 
(4.4d) Tell me when ... 
(4.4e) Can you tell me when ... 
(4.4f) Do you know when ... 
(4.4g) Do you want to tell me when ... 
(4.4h) Will you tell me when ... 
4.5. An Example of an Indirect INFORM 
An interesting example of an indirect INFORM is 
4.5a for it is very similar to 4.5b-c which both seem to 
only be requests. The interpretation of 4.5a as an 
indirect INFORM follows from the fact that inference 
chains which would make it a REQUEST are all inhib- 
ited by the heuristics. 
(4.5a) Do you know that the RAPIDO is late? 
(4.5b) Do you believe that the RAPIDO is late? 
(4.5c) Do you know whether the RAPIDO is late? 
In Example 8, the possible body-action inference 
from (2) to 
REQUEST(A,S,INFORMIF(S,A,KNOW(S,P))) 
is downgraded because the embedded inference to (3) 
is possible. The interesting case is the embedded 
know-negative inference which is also possible from 
(3). It implies that BsWA(~KNOW(S,P)), or equiva- 
lently 
(4.5d) BsWA(P ^ 2,Bs(P) ) 
178 American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
PI.AE 
PI.AE/EI. 1 
PI.KP/EI.1 
..PI.BA 
(1) S.REQUEST(A,S,INFORMIF(S,A,KNOW(S,P))) 
(2) BsWA(INFORMIF(S,A,KNOW(S,P))) 
(3) BsWA(KNOWIF(A,KNOW(S,P))) 
(4) BsWA(KNOW(S,P)) 
(5) INFORM(A,S,P) 
Example 8. "Do you know that P?" 
But such a goal is highly unlikely. A is attempting to 
achieve the goal ~Bs(P) by having S recognize that A 
wants P to be true! As a result, no speech act inter- 
pretation is possible from this step. For instance, the 
bodies of the acts INFORM(A, S, P) and INFORM(A, 
S, ~P) are BsWA(P A Bs(P)) , and BsWA(~P A 
Bs(~P)), respectively. Both of these are contradicted 
by part of 4.5d. Thus the know-negative possibility 
can be eliminated. This allows the know-positive in- 
ference to be recognized as intended, and hence leads 
to the indirect interpretation as an INFORM(A, S, P). 
4.5b has only a literal interpretation since both the 
know-positive and know-negative rules are applicable 
at the nested level; without a reason to favour either, 
the literal 
REQUEST(A,S,INFORMIF(S,A,Bs(P))) 
is preferred. The interpretations of 4.5c are similar to 
those of Examples 2 and 3. 
4.6. Using Knowledge of Deduction 
All the examples of indirect speech acts so far have 
been explained in terms of rules PI.1-PI.6, and com- 
plex inference rules derived from them. In this sec- 
tion, we give one more example relying on somewhat 
more specific rules. A full investigation of how many 
such specific rules are necessary to account for com- 
mon forms of indirect REQUESTs and INFORMs 
remains to be done. 
This example shows how a completely non-standard 
form can be intended indirectly. Suppose that A tells 
S 
(4.6a) "John asked me to ask you to leave" 
This has at least three possible interpretations: 
(4.6b) A is asking S to leave, and giving a reason. 
(4.6c) A wants to simply report the fact to S 
that John did the action of asking S to leave. 
(4.6d) A wants to inform S that John wants 
him to leave. 
Interpretations c and d can hold even if S decides 
that A actually does want him to leave. However, in 
these cases, he would not say that A intended to com- 
municate the intent that he leave, i.e. he would not say 
the utterance was a REQUEST. 
Both interpretations rely on axioms ACT.1 and 
ACT.2 (of section 3.2) which state that if some agent 
A believes that agent S executed some action ACT, 
then A may believe that the preconditions of ACT 
obtained before, and the effects of ACT obtained 
after, the execution of ACT. 
They also require a new PC/PI rule: if A wants S 
to believe some proposition P, then A may get S to 
believe some proposition Q, as long as A believes that 
S believes that Q implies P. 
(PC.I) WA(Bs(P)) =c=> WA(Bs(Q)) , 
if BABs(Q ~ P). 
(PI.I) BsWA(Bs(Q)) =i=> BsWA(Bs(P)) , 
if BsBABs(Q => P). 
In Example 9, S recognizes that A asked him to 
leave. The interpretation depends on S concluding 
that John performed his REQUEST successfully 
(through PI.I and ACT.2), and hence that A wants to 
request S to leave. It is then an easy step to infer that 
A wants S to leave, which leads to the request inter- 
pretation. Interpretation (c), a simple report of some 
previous action, follows from (2) by PI.BA. 
In Example 10, S recognizes that A intended to tell 
him that John wants him to leave. This depends on 
the fact that S concludes that John wanted to perform 
the REQUEST that A reported. Most of the needed 
inferences call for the use of EI.1 to embed simple 
inference rules twice. Note that an INFORM act 
could have been inferred at each of the four previous 
steps; for example, from (5) the body inference would 
produce 
INFORM(A, S, Wj(REQUEST(A, S, LEAVE(S))). 
But the inferences at the "BsWABsWj" level were so 
direct that they were continued. 
5. Gordon and Lakoff Revisited 
The examples of the previous section show how our 
plan inference rules account for the indirect interpreta- 
tions of the requests which GL's postulates were de- 
signed for, as well as several others. Our approach 
differs from GL's in that an utterance may carry both 
a literal and an indirect interpretation, and of course in 
that its inference rules are language independent. 
American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 179 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
PI.AE 
PI.I/EI.1 
(PI.AE/EI. 1)/EI. 1 
(PI.W/EI. 1)/EI. 1 
(PI.BA/EI. 1) 
PI.AE/EI.1 
PI.W/E1.1 
PI.BA 
(1) S.INFORM(A,S,REQUEST(J,A,REQUEST(A,S,LEAVE(S)))) 
(2) BsWA(Bs(REQUEST(J,A,REQUEST(A,S,LEAVE(S))))) 
(3) BsWA(BsWA(REQUEST(A,S,LEAVE(S)))) 
(4) BsWA(BsWA(Ws(LEAVE(S)))) 
(5) BsWA(BsWA(LEAVE(S))) 
(6) BsWA(REQUEST(LEAVE(S))) 
(7) BsWA(Ws(LEAVE(S)) 
(8) BsWA(LEAVE(S)) 
(9) REQUEST(A,S,LEAVE(S)) 
Example 9. "John asked me to ask you to leave." (Interpretation b) 
PI.AE 
PI.I/EI.1 
(PI.AE/EI. 1)/EI. 1 
(PI.W/EI. 1)/EI. 1 
(PI.AE/EI. 1)/EI. 1 
(PI.W/EI. 1)/EI. 1 
(PI.BA/EI.1)/EI. 1 
(1) S.INFORM(A,S,REQUEST(J,A,REQUEST(A,S,LEAVE(S)))) 
(2) BsWA(Bs(REQUEST(J,A,REQUEST(A,S,LEAVE(S))))) 
(3) BsWA(BsWj(REQUEST(J,A,REQUEST(A,S,LEAVE(S)))) 
(4) BsWA(BsWj(WA(REQUEST(A,S,LEAVE(S)))) 
(5) BsWA(BsWj(REQUEST(A,S,LEAVE(S)))) 
(6) BsWA(BsWj(Ws(LEAVE(S)))) 
(7) BsWA(BsWj(LEAVE(S)) 
(8) INFORM(A,S,Wj(LEAVE(S)) 
Example 10. "John asked me to ask you to leave." (Interpretation d) 
However, in some ways both solutions are too strong. 
Consider, for example, the following: 
(5.a) Can you reach the salt? 
(5.b) Are you able to reach the salt? 
(5.c) I hereby ask you to tell me whether 
you are able to reach the salt. 
Although 5.a-c are all literally questions about the 
hearer's ability, only 5.a normally conveys a request. 
Sadock \[1974\] suggests that forms such as 5.a dif- 
fer from 5.b in that the former is an idiom which is 
directly a request while 5.b is primarily a yes/no ques- 
tion. However, as Brown \[1980\] points out, this fails 
to account for responses to 5.a which follow from its 
literal form. One can answer "Yes" to 5.a and then 
go on to pass the salt. 
Brown proposes what she calls "frozen ISA forms" 
which directly relate surface form and indirect illocu- 
tionary force, bypassing the literal force. Frozen 
forms differ from normal rules mapping illocutionary 
forces to illocutionary forces in that they point to the 
relevant normal rule which provides the information 
necessary to the generation of responses to the surface 
forms. 
The speaker of 5.b or 5.c may in fact want the 
hearer to reach the salt, as does the speaker of 5.a, but 
he does not want his intention to be recognized by the 
hearer. Thus it appears that from the hearer's point of 
view the chain of inferences at the intended level 
should get turned off, soon after the recognition of the 
literal act. It seems that in this case (Example 6 of 
section 4.4) the plausibility of the inferences after step 
3 should be strongly decreased. Unfortunately it is 
not obvious that this can be done without making ,the 
rating heuristics sensitive to syntax. 
The indirect interpretation can also be downgraded 
in the presence of stronger expectations. If a speaker 
entered a room full of aspiring candidates for employ- 
ment and said: "I want to know how many people 
here can write a sort/merge program" and then turn- 
ing to each individually asked "Can you write a 
sort/merge?" the question would not be intended as a 
request to write a program, and would not be recog- 
nized as such by a PI algorithm which rated highly an 
illocutionary act which fits well in an expectation. 
In several of the earlier examples of questions in- 
tended as indirect requests, the literal interpretation is 
blocked because it leads to acts whose effects were 
true before the utterance. The literal interpretation of 
5.d gets blocked because the reminding gets done as 
part of the understanding of the literal act. Thus only 
an indirect interpretation is possible. 
(5.d) May I remind you to take out the garbage? 
180 American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 
Sadock \[1970\] points out that some co-occurrence 
rules depend on conveyed rather than literal illocution- 
ary force. The morpheme please can occur initially 
only in sentences which convey a request. 
(5.e) Please, can you close the window? 
(5.f) Please, it's cold in here. 
(5.g) *Please, do you know that 
James is in town? 
But it can occur in final position only in sentences 
which both convey a request and are literally requests: 
(5.h) Can you close the window, please? 
(5.i) *It's cold in here, please. 
(5.j) *Do you know that John is in town, please? 
These remain problematic for Brown and for us. 
6. Conclusion 
We have given evidence in this paper for an ac- 
count of indirect speech acts based on rationality (plan 
construction), imputing rationality to others (plan 
inference), surface speech act definitions relating form 
to "literal" intentions, and illocutionary acts allowing a 
variety of realizing forms for the same intentions. 
The reader may object that we are suggesting a 
complex solution to what appears to be a simple prob- 
lem. It is important to distinguish here the general 
explanation of indirect speech acts (which is presented 
here partly through an algorithm) from the implemen- 
tation of such an algorithm in a practical natural lan- 
guage understanding system. We claim that the ele- 
ments necessary for a theoretically satisfying account 
of indirect speech acts are independently motivated. It 
is almost certain that a computationally efficient solu- 
tion to the indirect speech act problem would short-cut 
many of the inference chains suggested here, although 
we doubt that all searching can be eliminated in the 
case of the less standard forms such as 4.6a. The 
implementation in Brachman et al \[1980\] does just 
that. However, the more fundamental account is nec- 
essary to evaluate the correctness of the implementa- 
tions. 
Many problems remain. Other syntactic forms that 
have significance with respect to illocutionary force 
determination should be considered. For example, tag 
questions such as 
"John is coming to the party tonight, isn't he?" 
have not been analysed here (but see Brown \[1980\]). 
Furthermore, no "why" or "how" questions have been 
examined. 
Besides the incorporation of more syntactic infor- 
mation, another critical area that needs work concerns 
the control of inferencing. To allow the use of spe- 
cialized inferences, a capability that is obviously re- 
quired by the general theory, much research needs to 
be done outlining methods of selecting and restricting 
such inferences. 
This paper has concentrated on recognition. Allen 
\[1979\] shows how the construction algorithms would 
have to be modified to allow the generation of surface 
acts, including indirect forms. McDonald \[1980\] dis- 
cusses the planning of low-level syntactic form. 
According to the definition of INFORM of section 
4.1, any utterance that causes S to infer that A has a 
plan to achieve KNOW(S,P) by achieving 
BsWA(KNOW(S,P)) is considered by S to be an IN- 
FORM. Strawson \[1964\] argues that one level of 
recognition of intention is not sufficient for the defini- 
tion of a speech act. Schiffer \[1972\] gives a series of 
counterexamples to show that no finite number of 
conditions of the form BsWA(BsWA(...(KNOW(S,P))) 
is sufficient either. The solution he proposes is that 
the recognition of intention must be mutually believed 
between the speaker and the hearer. Cohen and Lev- 
esque \[1980\] and Allen \[forthcoming\] show how the 
speech act definitions given here can be extended in 
this direction. 
We have only considered acts to request and in- 
form because many of their interesting properties can 
be based on belief and want. At least primitive ac- 
counts of the logics of these propositional attitudes are 
available. Clearly there is room for much work here. 
Extending the analysis to other speech acts, such as 
promises, will require a study of other underlying log- 
ics such as that of obligation. 
There also remain many problems with the formali- 
zation of actions. We believe this work shows that the 
concepts of preconditions, effects, and action bodies 
are fruitful in discussing plan recognition. The opera- 
tor definitions for speech acts used here are intended 
to facilitate the statement of the plan construction and 
inference rules. However, their expressive power is 
insufficient to handle complex actions involving se- 
quencing, conditionals, disjunctions, iterations, paral- 
lelism, discontinuity, and afortiori requests and prom- 
ises to do such acts. They are also inadequate, as 
Moore \[1979\] points out, to express what the agent of 
an action knows (and does not know) after the success 
or failure of an act. Moore's logic of action includes 
sequencing, conditionals, and iterations, and is being 
applied to speech acts by Appelt \[1980\]. Much re- 
mains to be done to extend it to parallel and disconti- 
nuous actions typical of multiple agent situations. 
These difficulties notwithstanding, we hope that we 
have helped show that the interaction of logic, philoso- 
phy of language, linguistics and artificial intelligence is 
productive and that the whole will shed light on each 
of the parts. 
American Journal of Computational Linguistics, Volume 6, Number 3-4, July-December 1980 181 
C. Raymond Perrault and James F. Allen A Plan-Based Analysis of Indirect Speech Acts 

References 
Allen, J.F. A Plan-Based Approach to Speech Act Recognition. Ph.D. 
Thesis, Dept. of Computer Science, University of Toronto, 
1979. 
Allen, J.F. Recognizing Intentions from Natural Language Utter- 
ances, forthcoming. 
Allen, J.F. and Perrault, C.R., Analyzing Intention in Utterances, 
Artificial Intelligence, 15, 3, pp. 143-178, 1980. 
Appelt, D.E., Problem Solving Applied to Language Generation, 
Prec. of 18th Annual Meeting of Assoc. for Comp. Ling., pp. 
59-63, 1980. 
Austin, J.L. How To Do Things With Words. New York: Oxford 
University Press, 1962. 
Ayer, A.J. The Problem of Knowledge. London: Penguin Books, 
1956. 
Brachman, R.J., Bobrow, R.J., Cohen, P.R., Klovstad, J.W., Web- 
ber, B.L., and Woods, W.A., Research in Natural Language 
Understanding, Report 4274, Bolt, Beranek, and Newman, 
Cambridge, Mass., 1980. 
Brown, G.P. Characterizing Indirect Speech Acts :A Computational 
Approach, Am. J. of Comp. Ling., this issue, pp. 150-166, 1980. 
Bruce, B.C. Belief Systems and Language Understanding, Technical 
Report 2973, Bolt Beranek and Newman, Cambridge Mass, 
1975. 
Cohen, P.R. On Knowing What to Say: Planning Speech Acts, Techni- 
cal Report no. 118, Dept. of Computer Science, University of 
Toronto January, 1978. 
Cohen, P.R., and Levesque, H.J., Speech Acts and the Recognition 
of Shared Plans, in Prec. of 3rd Con f. of Can. Soc. for Comp. 
Studies of Intelligence, Victoria, B.C., pp. 263-271, 1980. 
Cohen, P.R. and Perrault, C.R. Elements of a Plan Based Theory of 
Speech Acts, Cognitive Science, 3, 3, pp. 177-212, 1979. 
Cole, P. and Morgan, J.L. Syntax and Semantics, Vol 3: Speech Acts. 
New York, Academic Press, 1975. 
Fikes, R.E. and Nilsson, N.J. STRIPS: A New Approach to the 
Application of Theorem Proving to Problem Solving. Artificial 
Intelligence 2, pp. 189-205, 1971. 
Genesereth, M.R. Automated Consultation for Complex Computer 
Systems, Ph.D. Thesis, Harvard University, 1978. 
Gettier, E.L. Is Justified True Belief Knowledge?. Analysis 23, pp. 
121-123, 1963. Reprinted in M.D. Roth and L. Galis (eds.) 
Knowing, New York: Random House, 1970. 
Gordon, D. and Lakoff, G. Conversational Postulates. in Cole and 
Morgan (eds), 1975. 
Green, G. M. How to Get People to Do Things with Words: the 
Whimperative Question, in Cole and Morgan (eds.), 1975. 
Grice, H.P. Meaning. Phil. Rev. 66, pp. 377-388, 1957. 
Grice, H.P., Utterer's Meaning, Speaker Meaning, and Word Mean- 
ing, Found. of Language, 4, pp. 225-242, 1968. 
Hintikka, J. Knowledge and Belief, Ithaca, New York: Cornell Uni- 
versity Press, 1962. 
McDonald, D.D., Natural Language Production as a Process of Deci- 
sion Making under Constraint, Ph.D. Dissertation, Dept. of E.E. 
and Comp. So., MIT, 1980. 
Moore, R.C. Reasoning about Knowledge and Action, Ph.D. Thesis, 
Massachusetts Institute of Technology, 1979. 
Morgan, J.L. Conversational Postulates Revisited, Language 53, 2, 
1977. 
Morgan, J.L. Towards a Rational Model of Discourse Comprehen- 
sion, in Prec. of Second Conf. Theoretical Issues in Natural Lan- 
guage Processing, Champaign-Urbana, 1978. 
Morgan, J.L. Two Types of Convention in Indirect Speech Acts, in 
P. Cole (ed.) Syntax and Semantics (vol. 9): Pragmatics, New 
York: Academic Press, 1978. 
Perrault, C.R., Allen, J.F. and Cohen, P.R. Speech Acts as a Basis 
for Understanding Dialogue Coherence, in Prec. of Second Conf. 
Theoretical Issues in Natural Language Processing, Champaign- 
Urbana, 1978. 
Perrault, C.R. and Cohen, P.R. It's for Your Own Good: a Note on 
Inaccurate Reference, to appear in Elements of Discourse 
Understanding, A.K. Joshi, I.A. Sag, and B.L. Webber (eds.), 
New York: Cambridge University Press. 
Quine, W.V.O. Quantifiers and Propositional Attitudes. Journal of 
Philosophy 53, pp177-187, 1956. Reprinted in L. Linsky (ed.) 
Reference and Modality, Oxford: Clarendon Press, 1971. 
Russell, B. Introduction to Mathematical Philosophy. London: 
George Allen and Unwin, 1919. 
Sacerdoti, E.D. Planning in a Hierarchy of Abstraction Spaces. 
Prec. of Third Int. Joint Conf. on Artificial Intelligence, Stanford 
CA, 1973. 
Sacerdoti, E.D. The Nonlinear Nature of Plans. Prec. of Fourth Int. 
Joint Conf. on Artificial Intelligence, Tbilisi USSR, 1975. 
Sadock, J.M. Whimperatives, in J.M. Sadock and A. Vanek Studies 
Presented to R.B. Lees by his Students, pp. 223-38, Edmonton, 
1970. 
Sadock, J. M. Towards a Linguistic Theory of Speech Acts, New York: 
Academic Press, 1974. 
Schiffer, S.R. Meaning. London: Oxford University Press, 1972. 
Schmidt, C.F., Shridharan N.S. and Goodson J.L., The Plan Recog- 
nition Problem: An Intersection of Artificial Intelligence and 
Psychology, Artificial Intelligence 11, pp. 45-83, 1978. 
Searle, J.R. Speech Acts, New York: Cambridge University Press, 
1969. 
Searle, J.R. Indirect Speech Acts. in Cole and Morgan (eds), 1975. 
Stampe, D.W. Meaning and Truth in the Theory of Speech Acts. in 
Cole and Morgan (eds), 1975. 
Strawson, P.F. Intention and Convention in Speech Acts. Phil. 
Rev., 73,4, pp. 439-460, 1964. 
