RECOGNIZING ADVICE, WARNINGS,PROMISES AND THREATS 
Kevin Donaghy 
School of Computer Science and Information Technology 
Rochester Institute of Technology, Rochester, New York 14623 
hkd@cs.rit.edu 
It is argued here that utterances in the imperative 
mood typically are embedded either explicitly or 
implicitly in Aristotelean practical arguments, i.e., 
arguments whose conclusions specify an action to be 
performed by an agent and whose premises provide 
justification for that action. It is argued further that 
the illocutionary force of an imperative, e.g., advice, 
warning, request, etc., is determined by the structure 
of the practical argument in which it is embedded. 
Algorithms for reconstructing partial practical argu- 
ments are developed. Emerging from the discussion is 
a set of heuristics for identifying advice, warnings, 
conditional promises and conditional threats in natural 
language settings. Sample output from a test program 
employing these heuristics is presented. Finally, it is 
suggested that the techniques outlined in this paper 
point to the possibility of dialogue and story under- 
standing systems which are more general and require 
significantly less domain specific "knowledge than do 
current systems. 
I. Practical Arguments 
Consider the utterance "Don't go near the stove". Is 
this a warning, an order, a request or perhaps an 
instance of some other speech act category? Apart 
from context, it is impossible to tell. But once context 
is supplied the answer is typically quite evident. 
1. If you touch the stove, you will burn yourself. So 
don't go near the stove. (warning) 
2. The player who avoids touching the stove usually 
wins the game. So don't touch the stove. (advice) 
3. I can't take another one of your casseroles. If you 
want to live don't touch the stove. (threat) 
What these cases have in common is that they are all 
examples of what philosophers since Aristotle have 
called practical arguments, that is, arguments whose 
conclusions name an action to be performed by an 
agent and whose premises provide reasons for the 
agent to perform that action. Speech acts such as 
advice, warning, instruction, and moral exhortation are 
conceptually linked to practical arguments in the fol- 
lowing way. In classifying an utterance as advice, 
warning, etc., we specify the type of practical argu- 
ment in which that utterance is either implicitly or 
explicitly embedded. 
Consider advice, for example. To advise X to do A 
is, among other things, to imply that X's interests will 
be served by doing A. When advice comes packaged 
This project is funded by Universal Energy Systems, hac. and 
the United States Air Force, Contract No. F49620-8g-c- 
0053/SB5881-0378. 
in the form of an imperative, that implication func- 
tions as the premise of a general practical argument 
whose conclusion is "X, do A." To warn X not to do 
A, on the other hand, is to imply that X's interests 
will suffer if X does A. That implication in turn func- 
tions as the premise of a general practical argument 
whose conclusion is "X, do not do A." To morally 
exhort X to do A is to imply that some moral end will 
be served should X do A. Once again, that implica- 
tion is the premise of the general practical argument 
being advanced by the speaker. 
A fundamental assumption of this paper is that argu- 
ments of the form "If X then Y. So (don't) do Z." 
comprise a small but important subset of practical 
arguments, for the reason that many if not all practical 
arguments with imperative or quasi-imperative conclu- 
sions can be recast in this form without loss of mean- 
ing or structure. This assumption is based on the 
Aristotelian means-end model of practical arguments 
as deliberations which "assume the end (viz. a desire 
need, interest or goal of the agent) and consider how 
and by what means it is to be attained." (Aristotle, 
1915, 1112b15-31). Consider the following example. 
The stove is hot. So don't touch it. 
While readily understandable, this argument is incom- 
plete. Fleshed out, it becomes 
1. The stove is hot. 
2. Hot things cause burns when touched. 
3. If you touch the stove, you will burn yourself. (1,2) 
4. (You wish to avoid burning yourself.) 
5. So don't touch the stove. (3,4) 
In the short version, the hearer's interests as well as 
the implications of the stove's being hot are so obvi- 
ous that they are not mentioned. Note that in the long 
version, I is not even a premise of the main argument. 
Its role is to provide evidence for 3. If this example 
is typical, the form "If X then Y. So (don't) do Z" 
may well capture the deep structure of a large and 
significant class of practical arguments. 
How does one go about reducing practical arguments 
to the form "If X then Y. So (don't) do Z"? To con- 
tinue the example, suppose "Hot things cause burns 
when touched" has been stored in a knowledge base. 
The reduction of "The stove is hot. So don't touch it" 
can then be carried out as follows. 
1. Assume that the real premise (RP) of the the argu- 
ment is of the form "If X then Y" where X is the 
negation of the propositional content of the conclusion 
and Y is some as yet unspecified harm to H. 
2. Also assume that the role of the stated premise 
(SP) is to provide evidence for RP. 
1 
336 
3. The consequent of RP (viz. "you will burn your° 
:~elf') can now be deduced fiom SP and the known 
fact that hot things cause burns when touched. 
This and sinfilar algorithms have been implemented in 
?ASA (Practical Arguments & Speech Acts), a test 
program which accepts practical arguments as input 
and identifies their principal speeeh acts. 
:~. PASA 
~everal hem'istics in PASA assist in the identification 
of speech acts and reduce substantially the need for 
domahl specific or nonlinguistic knowledge that would 
otherwise be necessary. As a start, consider the fol- 
ilowing examples. 
~. If you finishyour homework, l will give you more 
c~Lstor oil to drink. So finish your homework. 
2. If you don't finish your homework, I will give you 
more castor oil to drink. So finish your homework. 
In neither c~e is there any difficulty in deducing S's 
views on C~LStor oil. In the first example, it is prom- 
i~ed as reward, and in the second is threatened as pun- 
ishment. What makes these deductions possible is the 
relationship between the propositional contents of the 
imirerative and the antecedent of the conditional. In 
the fu'st hlsumce they are identical; in the second, the 
one is the negation of the other. PASA utilizes both 
heuristics to identify speech acts and to deduce and 
record the evaluative stance of the speaker towards a 
given state of affairs. The former is instrumental in 
the identification of promises and advice, and the 
latter of threats and warnings. "\[he next pair of exam- 
pies illustrate another helpful heuristic which in fact is 
a genelalization of the above. 
3. If you finish your homework, I will give you more 
castor oil to drink. So get started. 
4. If you don't finish your homework, I will give you 
more castor oil to drink. So get started. 
Since getting started on a project increases (dramati- 
c;ally) the likelihood of finishing iL it is obvious that 
example 3 is a promise and example 4 a threat. 
Whenever possible, PASA makes similar inferences. 
3, Sample I/O 
PASA was designed as a testbed for the ideas 
p~e.sented in sections 1 o 2. As such, it is not a full- 
llcdged natural language system, nor even a mature 
prototype for such a system. Inputs to the program 
me one premise practical arguments with imperative 
conclusions. In the initial set of examples, the princi- 
pal speech act is determined by examining the struc- 
ture of the argument and the syntactic form of 
premise's consequent. Program output is in boldface. 
C-Prolog version 1.4 
I% \[pasa\]. 
pasa consulted 33004 bytes 7.449997 sec. 
I?- talk. 
I: If you finish your homework then I will let you 
watch television. So finish that math. 
S promised to let you watch television if you finish 
your homework. 
I: I will ground you for a week if you fail the test 
tomorrow. So do not fail. 
S threatened to ground you for a week if you fail 
the test tomorrow. 
I: If you study for tile exam then you will pass. So 
study haxd. 
S advised you to study hard since if you study for 
the exam you will paw. 
In the next exmnple, PASA first has to be taught the 
relationship between starting arid finishing a project. 
I: Start on your math right away. I will let you watch 
television tonight if you finish all your homework 
before six. 
Let Y be the state of affairs described in the conse° 
quent of the premise° Which of the following most 
accurately describes the viewpoint of the speaker? 
A. Y is in the interests of the hearer. 
B. Y is not in tile interests of the hearer. 
C. Y does not affect interests of the hearer one way 
or tile oilier. 
l: A. 
Let X be the state of affairs d~cribed in the 
antecedent of the premise and Z the action 
specified in the conclusion. Which of. the following 
most accurately describes the viewpoint of the 
speaker? 
A. The hearer should do Z in anticipation of Y. 
B, By doing Z hearer would increase likelihood of X. 
C. Neither of the above. 
k B. 
S promised to let you watch television tonight if 
you finish all your homework before six. Starting 
on your ninth right away will make it more likely 
that you will finish all your homework before sLx. 
PASA now knows what it needs to know in order to 
paraphrase similzLr cases. The next example is an 
enthymeme. PASA must generate the hidden premise 
before it can haz~ard a paraphrase. 
I: The stove is hot. So do not touch the stove. 
S warned you not to touch the stove since if you 
touch the stove you will be burned. 
4. Towards Language Driven Understanding 
PASA is a modest example of a language driven 
understanding system in which the need for domain 
specific knowledge is minimized. A methodological 
decision was made early on to appeal to nonlinguistic 
information only as a last resort. The motivation for 
this was twofold, In the first place, domain driven syso 
terns are inherently limited by the vast amounts of 
domain specific information required to process even 
the simplest texts. There appears little hope of gen- 
eralizing these systems so that they are capable of 
exploiting structural commonalities between stories 
and dialogues from different donmins. Secondly, reli- 
ance on domain knowledge for quick fixes to text pro- 
cessing problems tends to deaden sensitivity to 
linguistic information present in those texts. 
337 
I am convinced that linguistic cues play a far richer 
and more powerful role in natural language under- 
standing than has been commonly supposed and that 
speech act analysis will prove a useful tool for sys- 
tematieally investigating those cues. I conclude with 
an iUustration of how domain and language driven 
approaches to story understanding might differ. Con- 
sider the following story (Wilensky, 1978, pp. 2-3). 
"One day John went through a red light and was 
pulled over by a cop. John had just gotten a sum- 
mons for speeding the previous week, and was told 
that if he got another violation, his license would be 
taken away. Then John remembered that he had two 
tickets for the Giant's game on him. He told the cop 
that he would g!ve them to him if he forgot the whole 
incident. The cop happened to be a terrific football 
fan. He took Iohn's tickets and drove away. 
Q1 Why did John offer the cop a couple of tickets? 
A1 Because he was afraid he was going to lose his 
license if he got another summons." 
Wilensky has this to say about the story. 
"Consider what is involved in making the inference 
that John offered the cop his football tickets to 
prevent the loss of his license. First, the reader would 
have to infer that the cop was going to give John a 
traffic ticket. This inference requires the knowledge 
that a policeman is supposed to ticket people who 
break traffic rules... Now the reader must interpret 
John's statement to the cop as an attempt to prevent 
him from giving him a ticket. To interpret this sen- 
tence as an offer, a reader must know that one way to 
prevent someone from doing something is to persuade 
him not to do it....by offering him something desirable 
in exchange for his cooperation. The understander 
can (then) infer that football tickets are desirable to a 
football fan, since football tickets are necessary for 
getting into a football game." 
Wilensky is setting the stage for a domain driven 
theory of understanding in which large stores of non- 
linguistic knowledge are required for story 
comprehension. Ironically, this very type of 
knowledge often impedes rather than assists 
comprehension, a fact known to every writer who has 
employed the O'Henry formula to surprise his audi- 
ence. One can imagine such a writer adding a short 
paragraph to Wilensky's story in which it is revealed 
that John, desperately hoping to lose his license, had 
threatened the cop with football tickets, a threat which 
proved pathetically ineffective. A major drawback in 
the domain driven approach is that it limits the under- 
stander to one interpretation of a story when several 
may be possible. Consider the same story from a 
language driven perspective. The story is presented in 
schematic form to insure that the understander has lit- 
tle or no domain knowledge available to him. 
1. One day A did B and was approached by C. 
2. A had just been given a D for doing E the previous 
week, and was told that if he got another D, then F 
would happen. 
3. Then A remembered that he had a G with him. 
4. A told C that he would give him a G if C did not 
give him a D. 
5. C accepted the G and did not give A a D. 
A = John B = ran a traffic light 
C = cop D = ticket 
E = speeding F = John loses license 
G = football tickets 
An understander would make little headway with the 
schematic version of the story until reaching statement 
4, where it becomes evident that A has either prom- 
ised or threatened to give C a G if C does not give 
him a D. At this point the schema lends itself to two 
quite different interpretations. 
Suppose that 4 is a conditional promise. Given the 
purpose of such promises, it follows that A does not 
want C to give him a D. Thus, from A's point of 
view, getting another D is undesirable. Now some 
sense can be made of statement 2. There is a strong 
probability that the the reason why getting another D 
is undesirable is because it would lead to F. So F too 
is most likely undesirable from A's perspective. 
Given this interpretation of 4, the understander now 
knows all he needs to know about the schematic story 
to answer the sample question. 
Q Why did A say that he would give G to C if C did 
not give him a D? 
AI Because A was afraid that F would happen if C 
gave D to A. 
Now suppose that 4 is a conditional threat. In this 
case, it follows that A wants C to give him a D. 
Given statement 2, the likelihood is that the reason 
why getting another D is desirable from A's perspec- 
tive is that it would lead to F. So F too is likely 
desirable. The appropriate answer to Q in this case is 
A2 Because A hoped that F would happen if C gave 
DtoA. 
It is of some interest that on this method of analysis, 
the inherent ambiguity of the story is preserved. This 
example and others like it provide suggestive evidence 
that understanding the general structure of such stories 
requires far less domain specific knowledge than 
Wilensky would lead us to believe. Clearly, many of 
the necessary inferences can be drawn from linguistic 
cues gleaned from the text. Domain knowlexlge facili- 
tates and enriches comprehension. However, it may 
not be as fundamental to the task of understanding as 
some researchers have suggested. 

REFERENCES 

Aristotle (1915). Ethica Nichomachea. In W.D. Ross 
(Trans.), The Works of Aristotle Translated into 
English (Vol. ix). London: Oxford University Press. 

Cohen, P.R., & Levesque, H.J. (1985). Speech acts 
and rationality. Proc. ACL, 49-59. 

Searle, J. (1969). Speech Acts. Cambridge: Cam- 
bridge University Press. 

Wilensky, R. (1978). Understanding goal-based 
stories. Ph.D Thesis. YI40. 
