Intentionality and Human Conversations. 
Jaime G. Carbonell Jr. 
Department of Computer Science 
Yale University - New Haven, Connecticut 
Abstract 
This paper analyzes principles of human 
conversation based on the conversational goals of 
the participants. Several conversational rules 
are proposed that seem crucial to the process of 
interpreting and generating conversations. These 
rules, as well as other aspects of the 
conversation process, are embodied in MICS, a 
computer program that generates one side of a 
conversation. The process model underlying MICS, 
and some illustrative output, are presented. 
I) Formulating rules about human conversations. 
This paper is an empirical approach to 
understanding the processes that underlie human 
conversations. Since the task of codifying all 
the knowledge required for modeling human 
discourse is monumental, we confine our approach 
to formulating rules about the conversational 
intent of utterances in the course of a dialog. 
This approach leads us to investigate the effects 
of shared assumptions and knowledge between the 
speakers, the social and interpersonal relations 
of the speakers, and the inferences that must be 
made by both speakers in a conversation. We take 
a different approach to analyzing conversations 
than other research efforts, such as those 
adopting the speech-acts paradigm (Mann et al 
\[1977\]) or investigating task-specific dialogs 
(Grosz \[1977\])o in the hope that our new 
perspective will shed some light on otherwise 
obscure or neglected aspects of human discourse. 
Consider the following conversation fragment 
between Bill and John, two college students 
sharing an apartment: 
I) JOHN: Hi, what's new, Bill? 
BILL: l'm going to visit my folks tonight. 
We can analyze Bill's utterance in Conversation 
Fragment (I) in terms of its immediate meaning, 
that is, a representation of Bill's utterance in 
Conceptual Dependency or some other meaning 
representation. This, however, is a very 
incomplete analysis of what Bill said. Why did 
Bill say that he was visiting his folks? Bill 
could just as easily have said, "I'm going to 
brush my teeth tonight." This utterance, however, 
doesn't answer John's question; brushing one"s 
teeth is not "something new". Therefore, we could 
propose a rather simple conversational rule: 
RULE I : If a question is asked in the course of 
a conversation, the other participant 
should answer this question. 
Rule I, however, is a little too naive. 
Suppose Bill's answer was: "There are a few more 
microns of dust on the windowsill than the last 
time you asked me that question." This is indeed 
"something new", but we would think of Bill as a 
wise guy for answering the question literally 
rather than addressing what John "must have 
meant". What did John really mean? John must 
have been looking for something out of the 
ordinary and of some intrinsic importance. Let us 
propose a new rule to incorporate this principle: 
RULE 2: In the formulation of an answer, the 
speaker should address the true signifi- 
cance of the question, not just its 
literal meaning. 
What is the true significance of a question? 
In Conversation Fragment (I), Bill might have 
answered: "The J-particle angular momentum of 
+3/2 was confirmed today." John, a literature 
major who does not understand Physics, may not be 
inclined to continue the conversation. Therefore, 
Bill's answer is not what was called for, unless 
Bill intentionally wanted to end the conversation. 
This example suggests that Bill missed something 
in establishing the true significance of John's 
question. John did, indeed, explicitly ask to 
hear something new; implicitly he meant something 
important and out of the ordinary. The J-particle 
answer conforms to these requirements, but it is 
still an inappropriate response. Therefore, the 
true significance of John's answer must include 
John's conversational goal. Why did John ask 
"What's new"? The answer is, obviously, to start 
a conversation with Bill. Bill, being aware of 
this conversational goal, needs to choose an 
answer that attempts to initiate conversation. 
That is, Bill should choose a topic of 
conversation that John can talk about and that 
John may be interested in. Conversational Rule 
(3) summarizes this discussion: 
RULE 3: In introducing a new topic of conversa- 
tion, the topic should be chosen so that 
both speakers have some knowledge and 
interest in its discussion. 
The process of understanding the 
conversational import of an utterance may be 
conceptually divided into two primary 
subprocesses: I) determine the conversational 
goal of the utterance, and 2) establish the real, 
often implicit, meaning of the utterance. Lehnert 
\[1977\] analyzes the process of establishing the 
real meaning of questions. Our analysis focuses 
on the conversational goals of the participants 
and the establishment of a shared knowledge base 
between the participants. It is this shared 
cultural, personal, and factual knowledge that the 
conversational participants leave implicit in each 
commtmication. To illustrate this fact, consider 
Conversational Fragment (2): 
2) JOHN: Do you want to go out and try the bar 
at Monument Square? 
BILL: I'm going to visit my folks tonight. 
Real significance of Bill's utterance: 
i) No, I do not want to go to the M~nument 
Square bar. 
ii) My reason for not wanting to go is that 
I made a previous commitment, and I can- 
not be in two places at once tonight. 
141 
ill) The previous commitment is a visit to 
my folks. 
iv) I am telling you about the reason why 
I cannot go drinking with you rather 
than just saying "no" because I do not 
want you to get angry at me. 
v) I may also wish to shift the topic of 
conversation to a discussion about my 
fam il y. 
Bill knows that John will interpret his 
answer so as to conclude its real significance; 
otherwise Bill would have chosen to explicitly 
state the real significance. How does Bill know 
that John will understand him correctly? Clearly 
Bill and John must share some common sense 
knowledge such as: 
a) A person cannot be in two places at once. 
b) Previous commitments should be honored. 
c) If X's invitation or suggestion is turned 
down by Y without apparent reason, then X 
is likely to get upset at Y. 
d) If a person introduces a new topic in a 
conversation, he may want to discuss the 
current topic further, 
Both Bill and John are aware that they share a 
common cultural knowledge base. This knowledge is 
very crucial in determining what is said in the 
conversation. Bill must have considered (1) 
through (iv) before deciding that is was 
sufficient to say only (ill). How did Bill decide 
to say only (ill)? He must have concluded that 
John would infer (i) , (ii) and (iv) without 
difficulty. Thus, Bill knew about John's general 
knowledge because of their common cultural 
background and their personal relation. Bill used 
this knowledge to decide what to say in the 
c onv er satlon. 
In the course of a conversation, people make 
assumptions about each other's knowledge. It is 
sometimes easier to see what these conversational 
assumptions are when they turn out to be 
incorrect, as in the following example: 
3) PETE: How are you going to vote on 
Proposition 13? 
MARY: On what? 
PETE: You know, the property tax limitation. 
MARY: Oh yeah. I'm not registered to vote. 
Which way were you trying to convince 
me to vote? 
PETE: I was hoping you would help me make 
up my mind. 
MARY: Actually, I don't give a damn about 
politics. 
At the beginning of the conversation Pete ass~ed 
that Mary knew what Proposition 13 was, that she 
was able to vote, that she would vote, and that 
she had already decided how to vote on Proposition 
13. All of these assumptions turned out to be 
incorrect, and the course of the conversation 
turned towards clarifying the incorrect 
ass~ptions. This example is an instance of a 
more general rule of conversation: 
RULE 4: If a participant in a conversation 
discovers that his assumptions about 
the shared knowlege between the two 
speakers is incorrect, then he will 
steer the conversation to 
I) establish a common knowledge base on 
a specific topic, or 
2) discover what their shared knowledge 
is in general, or 
3) shift the conversational topic to 
some matter where a common knowledge 
base is more likely to exist, or 
4) end the conversation. 
The assumptions discussed thus far have been 
of a factual nature, but assumptions are also made 
about the conversational intent of the 
participants and about their interest in the 
conversational topic. Mary inferred Pete" s 
conversational intent incorrectly: He was seeking 
advice, not trying to lobby for or against 
Proposition 13. Pete started the entire 
conversation on the wrong topic by assuming that 
Mary was interested in politics or taxes. A 
conversation about a topic that one of the 
participants finds uninteresting will usually 
digress to other topics or fizzle out as the 
uninterested party volunteers no new information, 
finds an excuse to do something else, or states 
outright that the conversation is boring (as was 
the case in our example). 
Erroneous assumptions about conversational 
intent lead to misunderstandings because each 
speaker will address the perceived intent of the 
other speaker's utterance. It is, therefore, 
imperative to correctly infer the other speaker's 
conversational intentions in order for the 
conversation to proceed naturally. The type 
misunderstanding that often results from 
incorrectly perceived conversational intentions 
is, on occasion, exploited in creating certain 
types of jokes, as in example 4: 
4) SON: Dad, I robbed the liquor store yesterday. 
DAD: How could you ever do such a thing, son? 
SON: Well, I got me this gun, and I pointed it 
at the cashier... 
To illustrate the importance of the implicit 
conversational goals and shared knowledge between 
the participants in a conversation, we present a 
few more dialog framents between Bill and John, 
the two college students sharing an apartment. In 
each example, as in conversations (I) and (2), 
Bill utters the same response, but its meaning is 
significantly different, depending on the context 
of the conversation. 
5) JOHN: Are you broke again? You are going to 
have to come up with your share of the 
rent this month. 
BILL: I'm going to visit my folks tonight. 
142 
Meaning of Bill's utterance: 
(i) Yes, l'm broke again. 
(il) Yes, I'll try to contribute my share of 
the rent. 
(iii) My parents might give me some money if 
I ask them. 
(iv) If I visit them and ask them in person I 
have a better chance of getting some money. 
(v) I'll visit them tonight and then I'll ask 
them for money. 
When we read Conversation Fragment (5), we 
infer that Bill may be going to ask his parents 
for money. How do we do this? We do not share 
knowledge with Bill to the effect that his parents 
have money or that Bill is willing to ask them for 
money. The answer is based on a conversational 
rule: 
RULE 5: The utterances in a conversation should 
be connected by continuity of topic, 
common conversational goals, and each 
participant addressing the intent of the 
utterances of the other participant. 
Since the reader assumes that Rule (5) is true for 
Conversation Fragment (5), he concludes that there 
must be a connection between Bill needing money 
and the visit to his parents. The reader then 
infers the most likely connection: Bill will ask 
his parents for money. John must also make this 
inference based on Rule (5), unless he knows that 
Bill regularly visits his parents to ask for 
money. The significant point illustrated in 
example 5 is that the conversation focused the 
inference mechanism to find a connection between 
the respective utterances. Therefore, 
conversational principles can play an important 
role in focusing human reasoning processes. The 
principle of focusing inference processes on 
significant or interesting aspects of 
conversational utterances and events is developed 
into a theory of human subjective understanding in 
Carbonell \[1978\]. 
Let us continue with the conversational 
fragments between Bill and John: 
6) JOHN: How come you never see your family? 
BILL: I'm going to visit my folks tonight. 
Meaning of Bill's utterance: 
(i) I do visit my family. 
(ii) Supporting evidence: I'm going to visit 
them tonight. 
(iii) Therefore, what you just said is not true. 
7) JOHN: Can I borrow your car? I got this heavy 
date tonight. 
BILL: I'm going to visit my folks tonight. 
Meaning of Bill's utterance: 
Alternative I. 
(i) No, you cannot borrow my car tonight. 
(il) I am going to visit my folks tonight. 
(iii) I need to drive there. 
(iv} The car cannot be fn two places at once. 
Alternative II, 
(i) Yes, you can borrow my car tonight. 
(ii) I am going to be at my folk's place, 
where I don't need to use it. 
8) JOHN: Can I have the apartment to myself? 
I got this heavy date tonight. 
BILL: I'm going to visit my folks tonight. 
Meaning of Bill's utterance: 
(i) Yes, you can have the apartment. 
(ii) What you want is for me to be elsewhere. 
(iii) I was planning on that any~my, since I 
am visiting my folks tonight. 
Conversation fragments (6), (7) and (8) 
illustrate the degree to which the understanding 
o f cony er sat ional utterances i s 
expectation-driven. The expectations are 
generated from previous utterances according to 
rule 5; the toplc, intent, and conversational 
goals introduced earlier in the conversation will 
be addressed by later utterances. In each case 
the same utterance on Bill's part is understood 
differently, depending on the context established 
by John' s previous utterance. Utterances in a 
conversation do not usually have a meaning 
independent of the rest of the conversation; 
their meaning is part of the context of the entire 
conversation. Thus, it is easy to see why quoting 
only a short passage from a conversation (or a 
political speech) can give that passage has an 
entirely different meaning from what was 
originally intended. 
The shared knowledge between two speakers 
depends on many different factors. Two speakers 
share a large amount of basic knowledge by merely 
being members of the human race (e.g. the basic 
drives that motivate humans such as hunger, 
self-preservation, etc .). More knowledge is 
shared if the two speakers are members of the same 
culture. (Much of the cultural and more basic 
human knowledge necessary to understand natural 
language is discussed in Schank and Ableson 
\[1977\].) If the two participants hold the same 
type of job, are professional colleages, or have 
the same special interests, then they will share 
some rather specific knowledge. Two people with 
the same special interests (such as football or 
radio-astronomy) will usually steer the 
conversation to a discussion of their common 
interest s. 
RULE 6: The topic of a conversation may drift to 
a subject where the conversational partici- 
pants share a great amount of knowledge. 
Another factor that determines the knowledge 
shared by the participants in a conversation is 
their interpersonal relation, i.e., how well they 
know each other. In conversational fragment (7), 
Bill's response can be interpreted in two 
different weys by the reader, but_ John will 
i~terpret his response unambiguously. John must 
know whethe~ Bill's response means tha~ Bill needs 
143 
the car or whether John is free to use it ; 
otherwise, Bill would have been more specific in 
his answer. 
Social relations and the perceived goals of 
conversational participants play an important role 
in interpreting the meaning of conversational 
utterances. Let us first consider the influence 
of the social relations between the two 
participants : 
9) ARMY GENERAL: I want a juicy hamburger. 
STAFF AIDE: Right away, sirl 
i0) 7-YEAR-OLD: I want a juicy hamburger. 
MOTHER: Maybe next wsek. We are having 
chicken today. 
II) PRISON INMATE I: I want a juicy hamburger. 
PRISON INMATE 2: Me tool Everything here 
tastes like cardboard. 
The utterance "I want a juicy hamburger" is 
interpreted differently in each dialog fragment. 
The difference in the interpretations is based on 
the different social relations existing between 
the two conversational participants. In Dialog 
(9) the utterance was interpreted to mean a direct 
order to the staff aide: "Get me a hamburger and 
make sure it is julcyl" In Dialog (I0), the 
7-year-old was expressing a request to his mother, 
hoping that his mother might comply. In Dialog 
(II), the same statement was interpreted as 
nothing more than wishful thinking. The fir st 
inmate made no order or request to the second 
inmate. Hence, the first utterance of each dialog 
fragment implies a different conversational goal 
depending upon the differences in the social 
relations of the conversational participants. The 
social context and the relationship between the 
two speakers generate expectations that guide the 
course of the conversation. A staff aide e~pects 
to be ordered about by a general. A mother 
expects her son to ask her for favors. Prison 
inmates cannot expect each other to do thinks that 
are made impossible by their incarceration. These 
expectations lead to a formulation of different 
conversational goals for the utterance, "I want a 
juicy hamburger ," in each conversational fragment. 
The conversational principle exemplified in our 
discussion is summarized as Conversational Rules 
(7) and (8): 
RULE 7: The social relationship between the 
participants in a conversation generates 
expectations about the intentional meaning 
of utterances in the conversation. These 
expectations are used to determine the 
conversational goals of each participant. 
RULE 8: Each speaker's perception of the conver- 
sational goals of the other speaker 
determines his interpretation of the 
other speaker" s utterances. 
Di f ferences in understanding o f 
conversational goals lead to different responses 
in a dialog, as illustrated in Conversation 
Fragments (9), (I0) and (Ii). We saw how a social 
relationship between two people can influence 
their interpretation o f each other" s 
conversational goals. Two strangers can also make 
assumptions about each other's conversational 
goals based on appearances, social circumstances 
and each other" s occupation. Consider, for 
instance, the various responses to John's question 
in the example below: 
Scenario: John walked up to a parson in the corner 
and asked : "DO you know how to get to 
Elm Street?" 
12.1) The stranger replied: "You go two blocks 
toward that tall building and turn right." 
12.2) The cab driver in the corner replied: "Sure; 
Hop in. Where on Elm do you want to go?" 
12.3) The person, who was holding up a map and a 
piece of paper with an Elm Street address on 
it, replied : "No, could you tell me how to 
get there?" 
12.4) The child answered: "Yes, I know how to get 
there! " 
The question was interpreted to mean four 
different things, depending on whom John spoke to. 
If a stranger asks, "Do you know how to get to X," 
the listener usually interprets this to mean "I 
want to go to X, but I do not know how to get 
there. Please give me directions." Since the 
occupation of a cab driver is to take people to 
their destination it is perfectly legitimate for 
him to interpret the question as: "If you know 
how to get to X please take me there." The person 
who is visibly lost and trying to find his way may 
interpret John's question as: "You seem to be 
lost. Can I help you find your way?" Response 
(12.3) illustrates that the responder did not 
infer that John's goal was to go to Elm street, in 
contrast with the two previous responses. A child 
often interprets questions of the form: "Do you 
know Y" literally, possibly inferring that the 
person asking the question is quizzing him. As in 
our previous examples, the differences in 
interpretation can be ~plained in terms of 
differences in the perceived goals of the 
participants in the conversation. 
II) MICS: A process model of human conversation. 
The phenomenon of human conversation is too 
comple~ for any single study to do justice to more 
than a narrow aspect of the problem. In order to 
fully understand human conversations we may have 
to understand all human cognitive reasoning 
processes. Our research approach can be outlined 
as follows: I) Study many sample conversations; 
2) try to establish some relatively general rules 
of conversation; 3) encode these rules into a 
process model; 4) see if this model accounts for 
certain aspects of human conversation; 5) realize 
that we solved hardly more than a minute part of 
the problem, and 6) reiterate the research process 
in a (hopefully positive) feed-back loop. 
The conversational rules discussed in the 
first section address problems that need to be 
considered if one is to understand human 
144 
conversations. There is little doubt, as 
demonstrated by countless examples, that 
conversational goals, shared knowledge between 
speakers, social relationships between speakers ', 
and the conversational import of each utterance in 
a dialog are aspects of human discourse that need 
to be analyzed if one is to understand how human 
conversations work. Analyzing these aspects, 
however, solves only a small subset of the larger 
problem of how conversations function. For 
instance, the problem of topic selection in a 
conversation needs to be addressed. How does a 
person change the topic in a conversation? How 
are new topics chosen? These questions are 
analyzed in Schank \[1977\]. Here we propose some 
additional ideas on the impact of shared knowledge 
and interests on topic-selection. 
MICS (Mixed-Initiative Conversational System) 
is a fully implemented computer program that 
generates one side of a natural-language 
conversation. MICS embodies the conversational 
rules discussed in this paper, a topic transition 
mechanism based on Schank \[1977\], and the idea of 
a conversational syntax. Conversational syntax is 
a set of rules that help to characterize 
well-formed conversations. For instance, the 
following four rules are among the twenty 
conversational syntax rules in MICS: 
RULE 9: Do not repeat information in the course 
of a conversation unless explicitly 
r eque st ed. 
RULE I0: Do not say things that the other speaker 
already knows about. 
RULE II: If the other speaker says something that 
violates one's beliefs, then respond in 
kind by re-asserting the violated belief. 
RULE 12: If insulted, return the insult or end the 
conversation. 
MICS illustrates that the conversational rules, 
topic transition rules and conversational syntax 
rules provide a sound, if possibly incomplete 
approach to the modeling of human conversation. 
MICS is able to generate reasonable conversations 
about domains on which it has some knowledge, but 
there are still many short-comings, as illustrated 
in the examples that follow. 
MICS "knows" about itself as a person (e.g. 
it believes itself to be a computer programmer) 
and uses a model of the other conversational 
participant that contains certain information 
about the speaker, such as his social status, his 
occupation, his relationship with other people and 
his areas of special interest and more detailed 
knowledge. The model of the other speaker is a 
necessary part of the conversational model, since 
it enables MICS to apply the conversational rules. 
When MICS talks to a stranger, it builds a model 
of the speaker; in fact~ one of its 
conversational goals is to learn about the person 
with whom it is conversing. In the dialog that 
follows (an actual computer run), MICS starts out 
knowing nothing about Dave. 
MICS CONVERSATION LOG ******** 10/11/77 ******* 
HELLO, WHAT IS YOUR NAME? 
I am Dave. 
WHAT'S NEW TODAY, DAVE? 
Nothing, What's your name? 
MY NAME IS MICS. 
What is your job? 
I AM A COMPUTER PROGRAMMER 
That's nice. 
WHAT DO YOU DO FOR A LIVING? 
I MICS takes the initiative by asserting 
I its conversational goal of learning 
I about Dave. 
I am still a student. 
DO YOU LIKE BEING A STUDENT? 
For the time being yes. 
WHAT WILL YOU DO AFTERWARDS? 
I want to be a professor some day. 
WHAT DO YOU DO FOR FUN? 
i MICS formed a person-frame for Dave. At this 
I point, the conversational goal of trying to 
I establish an area of mutual interest and 
I shared knowledge is activated. 
I smoke dope. 
YOU MIGHT 6ET STONED IF YOU SMOKE TOO MUCH. 
I know, that's the idea. 
YOU MI~T 6ET IN TROUBLE IF THE COPS CATCH YOU. 
All right, stop moralizing, what do you do for fun? 
I LIKE TO PLAY CHESS 
That sounds boring. 
NO, IT IS A LOT OF FUNI 
My brother Eric plays chess. 
HE MUST BE SMART. 
I have to go. Bye Mics. 
NICE TALKING TO YOU, DAVE 
As a result of the conversation, MICS knows 
something about Dave, in particular that one of 
his llfe goals is to becomle professor. MICS's 
conversational goals cause it to establish its own 
conversational topics. When a new topic is 
introduced by the human speaker, MICS stays on the 
chosen topic or chooses a closely related topic. 
Another rule of conversation is invoked in the 
above dialog : 
RULE 13: Each participant in the conversation has 
the conversational goal of saying things 
that are important to the other 
participant. 
When the human participant says that he likes to 
smoke marijuana, MICS finds in its memory that 
smoking marijuana is illegal and can lead to the 
smoker getting stoned. Accord ing to MIC S ' s 
145 
beliefs getting stoned is undesirable, and doing 
something illegal leads to undesirable 
consequences. MICS determines that it is 
important for Dave to know about the bad 
consequences of his acts. Therefore, applying 
Rule (13), MICS warns Dave about the evils of 
marijuana. It is evident that MICS should have 
known that Dave knows that smoking marijuana leads 
to getting stoned, and that, in fact, getting 
stoned is Dave's primary purpose for smoking 
marijuana. If MICS knew this fact, then 
Conversational Syntax Rule (I0) would prevent MICS 
from warning Dave about getting stoned • 
Therefore, MICS needs a better mechanism for 
deciding what knowledge it shares with the other 
speaker. It is very difficult to determine what 
the other person already knows and what he is able 
to infer, but this is a problem that cannot be 
circumvented if we are to model human discourse. 
The conversational syntax rule of countering 
a belief violation (Rule II) caused MICS to object 
when Dave said that playing chess (MICS's favorite 
pastime) was not exciting • Since the two 
participants could not find a common area of 
interest, the conversation ended without any 
s ubstantitive dialog. This illustrates 
Conversational Rule (4), discussed in the previous 
section. Let us analyze a conversation between 
MICS and Pete, someone who MICS believes to be its 
friend. 
HELLO, WHAT IS YOUR NAME? 
Pete. 
WHAT'S NEW TODAY, PETE? 
Professor Stevenson is giving us an exam friday. 
STEVENSON GIVES DIFFICULT EXAMS. 
Yeah~ do you know that Dave is probably going 
to flunk? 
THAT'S GOING TO MAKE IT HARD FOR DAVE TO BECOME 
A PROFESSOR. 
Too bad, Dave is not too smart anyways. 
YES, DAVE LIKES TO SMOKE DOPE. 
l've got to go and study. 
YOU WILL DO WELL IN THE EXAM. 
The above conversation illustrates the 
usefulness of shared knowledge between the two 
participants. Pete and MICS both know about 
taking ~ams, and they have a common acquaintance 
- Dave. MICS is able to use what it learned about 
Dave in its previous conversation. MICS' s 
response to Dave possibly flunking the exam 
illustrates Conversational Rules (3), (5) and (7). 
The continuity of the conversation is preserved. 
The chosen topic transition is probably of 
interest to Pete, as inferred by MICS in the 
following manner: Pete introduced Dave into the 
conversation; therefore, he probably wants to 
talk about Dave. Not fulfilling a life goal (e.g. 
becoming a professor) is an inherently important 
consequence to Dave; therefore, Pete is probably 
interested in knowing this consequence. Rule 7 
applies because in the social relationship between 
Pete and MICS (they are friends) it is OK to talk 
about mutual acquaintances • (It is NOT 
necessarily OK for a judge and a defendant, for 
instance, to discuss mutual acquaintances.) 
MICS's utterance about it being difficult for 
Dave to become a professor illustrates the 
inferenceing that proceeds in the course of a 
conversation. The inferences are directed at the 
most serious consequences of the preceeding 
utterance: flunking the exam. Focusing the 
inference process by importance criteria was found 
to be necessary in order to keep conversations 
from digressing into uninteresting topics. 
When Pete states that Dave is not too smart, 
MICS reacts by verifying this evaluation. Since 
in the previous conversation Dave chose to ignore 
MICS's warnings about smoking marijuana, MICS 
concluded that Dave was not too bright. This 
conclusion is used as supporting evidence to back 
Pete's evaluation. The process of responding to 
opinions is captured in conversational rule 14: 
RULE 14: If a speaker in a conversation states an 
opinion, the other speaker should consider 
whether or not he agrees with the opinion. 
The first speaker usually has the conver- 
sational goal of ~mnting feedback (usually 
agreement) on his oDinion. 
The last exchange in the above dialog 
illustrates a deficiency in MICS determination of 
the conversational goals. When Pete says, "I've 
got to go and study," he means, "I want to end the 
conversation," but inferring this conversational 
goal requires understanding what it means to 
study. One cannot infer from the phrase, "I've 
got to go ...", that the conversation should 
terminate. For instance, if Pete says, "I've got 
to go to Europe ," he probably wants to introduce 
his upcoming trip as a new topic of conversation. 
The knowledge that studying and talking cannot 
proceed simultaneuously is necessary in order to 
conclude that Pete wanted to end the conversation. 
The primary usefulness of the MICS system is 
that it sheds light on problems about the rules of 
conversation that might otherwise remain 
unnoticed • The feedback from testing the 
implementation of our conversational model forces 
us to be more precise and complete in our 
formulation of our ideas about conversation. The 
following is a conversation fragment from a later 
version of MICS which has a more sophisticated 
inference capability for determining the knowledge 
that it shares with the other speaker. For 
instance, MICS infers that if a person voluntarily 
chooses to pursue a certain activity, then that 
person is aware of the intended consequences of 
this activity. 
Ashtrays. 
WHAT ABOUT ASHTRAYS? 
Ashtrays are useful. 
DO YOU SMOKE? 
Yes. 
YOU MIGHT GET CANCER IF YOU SMOKE TOO MUCH. 
Who cares, anyway I smoke dope, not tobacco. 
I SEE. YOU LIKE TO GET HIGH. 
It is better than getting drunk. 
NOT NECESSARILY. YOU MIGHT GET CAUGHT BY THE COPS. 
146 
Figure (I) is a diagram of the control flow 
of MICS. As a process model of conversation, MICS 
is a useful tool, but it is far from a complete or 
even an entirely correct model of human 
conversation. MICS is a 3 level production 
system. The first set of productions is the 
conversational-syntax rules. These rules 
formulate contraints on the nature of the 
utterance that MICS will generate next in order to 
insure that the conversation follows a proper, 
"natural" course. Some of these constrains take 
the form of conversational goals used by the other 
two sets of productions rules. The second set of 
rules is the topic selection productions. These 
productions generate topic transitions guided by 
conversational goals and the amount of shared 
knowledge between the speakers • Several 
alternative things to say may be generated in this 
phase of the processing. These, as well as the 
conversational goals are stored in a short term 
memory (STM) and are used by the third, and 
theoretically most significant phase of the 
program: the topic evaluation rules. 
The third set o f productions encodes the 
conversational rules discussed throughout this 
paper. These rules use the STM information, the 
memory models of the self and the other speaker, 
CONV. SYNTAX 
RULES 
LTM 
MODEL OF / ' 'I 
I LISTENER i 
I 
I 
GOALS OF I 
SELF t 
I 
EVENT / 
MF~ORY 
CAC ' S / / r 
and an inferencer when necessary. The purpose of 
these rules is to choose the most interesting 
topics (to both the self and the other speaker) 
from the alternatives generated by the second set 
of production rules. The inferencer is used to 
determine what the other speaker is likely to know 
and which aspect of the topic he would be most 
interested in discussing. Thus, the menaing of an 
utterance is produced by the third set o f 
production rules. 
The utterances are generated in English by a 
rather crude phrasal English generator. The 
utterances from the other speaker are analyzed for 
their meaning and conversational form by a 
primitive, key-concept oriented analyzer. 
Disclaimer: MICS is a first-pass process 
model of a theory of conversation, not a theory of 
learning about other people. As such, its ability 
to learn about the other conversational 
participants is not as general as the dialogs 
presented in this paper may suggest. MICS learns 
about the other speaker by instantiating a 
prototypical-person frame - a data structure that 
encodes the more generally applicable facts about 
people and their social relations. 
INPUT SENTENCE 
ANALYSIS I 
TOPIC/C-TYPE 
I I 
/ / 
/// 
-/,,,/ 
EVALUATION 
RULES 
ENGLISH 
GENERATOR 
ENGLISH RESPONSE 
TOPIC SELECTION 
RULES 
TRANSITIO~ 
I 
W~OTHER TOPIC~ 
SELECTIONS 
F INTEREST 
MICS 
CONTROL-FLOW 
FIGURE 1 
STM 
,,I 
/ 
147 
Conclusion. 
believe that the best way to analyze a 
problem as difficult as modeling human discourse 
is to forge ahead by creating rules that capture 
important aspects of the conversation problem. 
The usefulness of these rules should be tested in 
a reactive environment such as an interactive 
computer program. Since conversation is not a 
problem that can be isolated from other aspects of 
human cognitive behavior, we are researching it in 
conjunction with other aspects of Artificial 
Intelligence. A process-based theory o f haman 
conversation should give some insight into other 
Natural Language Processing issues in particular, 
and AI modeling of human reasoning processes in 
general. 
Re fer ences. 
Carbonell, J. G. 1978. Computer Models of Social 
and Political reasoninB, PhD. Thesis, 
Yale University, New Haven, Conn. 
Grosz, B. J. 1977. The Representation and Use of 
Focus in a System for Understanding 
Dialogs, Proc. of the fifth IJCAI., MIT, 
Cambridge, Mass. 
Lehner t, W. 1977. The Process of Question 
Answering, PhD Thesis. Tech. Report 88, 
Yale University, New Haven, Conn. 
Mann W., M~ore J., Levln J. 1977. A Comprehension 
Model for Human Dialogue, Proc. of the 
fifth IJCAI. MIT, Cambridge, Mass. 
Schank, R. C. and Abelson R. P. 1977. Scripts, 
Goals, Plans and Understanding, lawrence 
Lawernce Erlba~n. Hillside, NJ. 
Schank, R. C. 1977. Rules and topics in 
conversation, Cognitive Science, 
Vol. I, No. 4. 
148 
