Generating Indirect Answers to Yes-No Questions 
Nancy Green 
Department of Computer Science 
University of Delaware 
Newark, DE 19716, USA 
Internet: i green@udel.edu 
Sandra Carberry 
Department of Computer Science 
University of Delaware 
Visitor: Inst. for Research in Cognitive Science 
University of Pennsylvania 
Internet: carberry@udel.edu 
Abstract 
An indirect answer to! a Yes-No question conversation- 
ally implicates the speaker's evaluation of the truth of 
the questioned proposition. We present the approach 
to generation used in our implemented system for gen- 
erating and interpret~ing indirect answers to Yes-No 
questions in English. Generation of a discourse plan is 
performed in two phases: content planning and plan 
pruning. During content planning, stimulus conditions 
are used to trigger speaker goals to include appropriate 
extra information wit h the direct answer. Plan prun- 
ing determines what parts of this full response do not 
need to be stated explicitly - resulting in, in appropri- 
ate discourse contexts, the generation of an indirect 
answer. 
1. Introduction 
J 
Imagine a discourse context for (1) in which R's use i 
of just (ld) is intended to convey a No, i.e., that R 
is not going shopping tonight. (By convention, square 
brackets indicate that the enclosed text was not ex- 
plicitly stated.) The part of R's response consisting of 
(ld) - (le) is what we 'call an indirect answer to a Yes- 
No question, and if (lc) had been uttered, (lc) would 
have been called a direct answer. 
l.a. O: I need a ride to the mall. 
b. Are you going shopping tonight? 
c. R: \[no\] 
d. My car's not running. 
e. The timing belt is broken. 
According to one study of spoken English \[Ste84\], 
13 percent of responses to Yes-No questions were in- 
direct answers. Thus, the ability to interpret indirect 
answers is required for robust dialogue systems. Fur- 
thermore, there are good reasons for generating indi- 
rect answers in a dialogue system. First, a direct yes 
or no alone may be misleading if extra information is 
needed to qualify the answer. Second, an indirect an- 
swer may contribute to a more efficient dialogue. For 
example, in addition to providing the requested infor- 
mation, it may anticipate a follow-up question from 
Q, or it may allow R to respond immediately with- 
out asking for clarification of the question. That is, 
the increased efficiency is more the result of avoid- 
ing the follow-up question or clarification subdialogue 
than the result of omitting the direct answer. Third, 
an indirect answer may be preferable for politeness 
considerations, as in (1). 
We have developed a computational model for the 
interpretation and generation of indirect answers to 
Yes-No questions in English. More precisely, by a Yes- 
No question we mean one or more utterances used as 
a request by Q (the questioner) that R (the respon- 
der) convey R's evaluation of the truth of a proposi- 
tion p. An indirect answer implicitly conveys via one 
or more utterances R's evaluation of the truth of the 
questioned proposition p, i.e. that p is true, that p is 
false, that there is some truth to p, that p may be true, 
or that p may be false. Our model presupposes that 
Q's question has been understood by P~ as intended 
by Q, that Q's request was appropriate, and that Q 
and I:t are engaged in a cooperative goal-directed di- 
alogue. Both the interpretation and generation com- 
ponents have been implemented in Common Lisp on a 
Sun SPARCstation. 
A language user's pragmatic knowledge of how lan- 
guage is used to answer Yes-No questions in English 
can be used to constrain the problem of generating 
and interpreting indirect answers. This knowledge is 
encoded in our model as a set of domain-independent 
discourse plan operators and a set of coherence rules, 
described in section 2.1 Our system is reversible in that 
the same pragmatic knowledge is used by the genera- 
tion and interpretation modules. Generation is mod- 
eled as discourse plan construction and interpretation 
as discourse plan inference. The output of generation 
is a discourse plan which can be realized by a tacti- 
cal generation component \[McK85\], and the output of 
interpretation is the speaker's inferred discourse plan. 
We found that while the above pragmatic knowledge is 
sufficient for interpretation, it is not sufficient for the 
1 Our main sources of data were previous studies \[Hir85, 
Ste84\], transcripts of naturally occurring two-person-dia- 
logue lAme92\], and constructed examples. 
189 
7th International Generation Workshop * Kennebunkport, Maine • June 21-24, 1994 
problem of content-planning during generation. This 
paper describes our approach to generation, including 
our solution to this problem. 
Generation of a discourse plan, described in sec- 
tion 4, is performed in two phases: content planning 
and plan pruning. During content planning, stimulus 
conditions, described in section 3, are used to trigger 
new speaker goals to include appropriate extra infor- 
mation with the direct answer. Plan pruning deter- 
mines what parts of this full response do not need to 
be stated explicitly. In appropriate discourse contexts, 
a plan for an indirect answer is generated. In other 
contexts, i.e. those in which the direct answer must 
be given explicitly, a plan for a direct answer with ap- 
propriate extra information is generated. Note that, 
according to the study mentioned earlier \[Ste84\], 85 
percent of direct answers are accompanied by such in- 
formation. Thus, it is important to model this type 
of response as well. Section 5 discusses other factors 
in generating indirect answers, and section 6 surveys 
related work. 
2. Pragmatic Knowledge 
A language user's knowledge of how to answer a Yes- 
No question appropriately can be described by a set 
of discourse plan operators representing full answers. 
A full answer consists of a direct answer (which we 
call the nucleus) and, possibly, extra relevant informa- 
tion (satellites). An indirect answer can be modeled 
as the result of R's providing one or more satellites 
without expressing the nucleus explicitly. Due to co- 
herence constraints holding between nucleus and satel- 
lite, Q may infer R's discourse plan by inferring which 
coherence relation between an indirect answer and a 
candidate direct answer is plausible. The coherence 
relations are similar to the subject-matter relations of 
Rhetorical Structure Theory (RST) \[MT87\]. ~ 
The discourse plan operators used to generate and 
interpret R's response in (1) are given in Figures 1 
and 2. Figure 1 depicts the top-level operator for con- 
structing a No answer. To explain our notation, s 
and h are constants denoting speaker (R) and hearer 
(Q), respectively. Symbols prefixed with "?" denote 
propositional variables. The variable in the header 
of Answer-no will be instantiated with the questioned 
proposition, e.g. in (1) that R is going shopping. Ap- 
plicability conditions are necessary conditions for ap- 
propriate use of the operator. For example, for a No 
answer to be appropriate, R must believe that the 
questioned proposition p is false. Also, an answer is 
not appropriate unless s and h share the expectation 
2The terms nucleus and satellite have been borrowed 
from RST. Note that according to RST, a property of the 
nucleus is that its removal results in incoherence. However, 
in our model, a direct answer may be removed without 
causing incoherence, provided that it is inferrable from the 
rest of the response. 
(Answer-no s h ?p): 
Applicability conditions: 
(discourse-expectation (informif s h ?p)) 
(believe s (not ?p)) 
Nucleus: 
(inform s h (not ?p)) 
Satellites: 
(Use-otherwise s h (not ?p)) 
(Use-obstacle s h (not ?p)) 
(Use-contrast s h (not ?p)) 
Primary goals: 
(BMB h s (not ?p)) 
Figure 1: Discourse plan operator for No answer 
(Use-obstacle s h ?p): 
;; s tells h of an obstacle explaining 
;; the failure ?p 
Existential variable: ?q 
Stimulus conditions: 
(explanation-indicated s h ?p ?q) 
(excuse-indicated s h ?p ?q) 
Applicability conditions: 
(believe s (cr-obstacle ?q ?p)) 
(Plausible (cr-obstacle ?q ?p)) 
Nucleus: 
(inform s h ?q) 
Satellites: 
(Use-elaboration s h ?q) 
(Use-obstacle s h ?q) 
(Use-cause s h ?q) 
Primary goals: 
(BMB h s (cr-obstacle ?q ?p)) 
Figure 2: Discourse plan operator for Obstacle 
that s will provide s's evaluation of the truth of p, 
which is denoted as (discourse-expectation (informif s 
h p)). Primary goals describe the intended effects of 
the operator. We use (BMB h s p) to denote that h be- 
lieves it to be mutually believed with s that p \[CM81 \]. 
In general, the nucleus and satellites of a discourse 
plan operator describe primitive or non-primitive com- 
municative acts. Our plan formalism allows zero, one, 
or more occurrences of a satellite in a full answer. The 
expected (but not required) order of nucleus and satel- 
lites is the order they are listed in an operator. (inform 
s h p) denotes the primitive act of s informing h that 
p. The satellites in Figure 1 refer to non-primitive 
acts, described by discourse plan operators which we 
have defined (one for each coherence relation used in 
a full answer). For example, Use-obstacle, a satellite 
of Answer-no, is defined in Figure 2. 
To explain the additional notation in Figure 2, 
(cr-obstacle q p) denotes that the coherence relation 
190 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
Operator: 
Answer-yes 
Answer-no 
Answer-hedged 
Answer-maybe 
Answer-maybe-not 
Satellites: 
Use-condition 
~Use-elaboration 
Use-cause 
Use-otherwise 
~se-obstacle 
Use-contrast 
Use-contrast 
Use-result 
Use-usually 
Use-possible-cause 
Use-result 
Use-usually 
Use-possible-obstacle 
Table 1: Satellites of top-level operators 
cr-obstacle holds between q and p.S Thus, the first ap- 
plicability condition can be glossed as requiring that s 
believe that the coherence relation holds. In the sec- 
ond applicability condition, (Plausible (cr-obstacle q 
p)) denotes that_, given what s believes to be mutually 
believed with h, the coherence relation (cr-obstacle q 
p) is plausible. In our model, this sort of condition is 
evaluated using a set of coherence rules based on the 
relation definitions of RST. 
In summary, we provide our system with a set of 
discourse plan operators representing shared domain- 
independent knowledge of the informational content 
of a full answer. (Operators whose names begin with 
Answer are top-level Operators.) Table 1 shows the 
satellites of each top-level discourse plan operator. In 
addition, we provide a set of coherence rules defining 
the plausibility of the Coherence relations generated by 
the satellite discourse plan operators. 4 
3. Stimulus Conditions 
Applicability conditions prevent inappropriate use of a 
discourse plan operator. However, they do not model 
a speaker's motivation for choosing to provide extra 
information. A full answer might provide too much 
information if every satellite whose applicability con- 
ditions held were included in the full answer. On the 
other hand, at the time when he is asked a question, R 
may not have the primary goals of a potential satellite 
operator. Thus, a goal-driven approach to selecting 
satellites would provide insufficient information. To 
overcome these limitations, we have incorporated stim- 
ulus conditions into the discourse plan operators in our 
3Although this is not one of the original relations of 
RST, it is similar to other causal relations defined in RST. 
4In addition to RST Subject-matter relations, we have 
identified the following coherence relations in our corpus: 
cr-obstacle, cr-possible-cause, cr-possible-obstacle, and cr- 
usually. 
model. Playing a key role in content determination, 
stimulus conditions describe conditions motivating a 
speaker to include a satellite during plan construction. 
They can be thought of as triggers which give rise to 
new speaker goals. In order for a satellite to be in- 
cluded, all of its applicability conditions and at least 
one of its stimulus conditions must be true. While 
stimulus conditions may be derivative of deeper prin- 
ciples of cooperativity \[Gri75a\] or politeness \[BL78\], 
they provide a level of precompiled knowledge which 
reduces the amount of reasoning required for content 
determination. 
Our methodology for identifying stimulus condi- 
tions was to survey linguistic studies (described be- 
low), as well as to analyze the possible motivation 
of the speaker in the examples in our corpus. Ac- 
cording to StenstrSm \[Ste84\], the typical motivation 
for providing extra information is to answer an im- 
plicit wh-question. Other reasons cited are to provide 
an explanation justifying a negative answer, to qual- 
ify the answer, or "social reasons". Levinson \[Lev83\] 
claims that the presence of an explanation is a struc- 
tural characteristic of dispreferred responses, such as 
refusals and unexpected answers. Brown and Levin- 
son \[BL78\] show how various uses of language are mo- 
tivated by politeness. For example, use of conversa- 
tional implicature is one strategy for a speaker to avoid 
doing a face-threatening act, s and as we discuss in sec- 
tion 6, an indirect answer conversationally implicates 
the direct answer. Hirschberg \[Hir85\] claims that an 
indirect scalar response, rather than a Yes or No an- 
swer, is necessary in certain cases to avoid misleading 
the hearer. 
In the rest of this section, we describe each of the 
stimulus conditions used in our system. For each stim- 
ulus condition, we give one or more axioms defining 
sufficient conditions for it. Axioms are encoded as 
Horn clauses, where symbols prefixed with "?" are 
variables, all variables are implicitly universally quan- 
tified, and the antecedent conditions are implicitly 
conjoined. The head, i.e. the consequent, of an axiom 
is a stimulus condition appearing in one or more of our 
discourse plan operators. Table 2 summarizes which 
stimulus conditions appear in which operators. For ex- 
ample, the stimulus condition (explanation-indicated s 
h ?p ?q) occurs in the discourse plan operator shown 
in Figure 2. Keep in mind that additional constraints 
on .¢p and ?q arise from the applicability conditions of 
the operator containing the stimulus condition. 
3.1 Definitions 
Explanation-indicated 
This stimulus condition appears in all of the oper- 
SThey claim that another strategy is to give the direct 
answer last, i.e. in our terminology, to give the satellite(s) 
before the nucleus. 
191 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
Operator: 
Use-cause 
Use-condition 
Use-contrast 
Use-elaboration 
Use-obstacle 
Use-otherwise 
Use-possible-cause 
Use-possible-obstacle 
Use-result 
Use-usually 
Stimulus conditions: 
Explanation-indicated 
Clarify-condition- 
indicated 
Appeasement-indicated 
Answer-ref-indicated 
Clarify-extent- 
indicated 
Substitute-indicated 
Answer-ref-indicated 
Clarify-concept- 
indicated 
Explanation-indicated 
Excuse-indicated 
Explanation-indicated 
Excuse-indicated 
Explanation-indicated 
ExPlanation-indicated 
Excuse-indicated 
Explanation-indicated 
Explanation-indicated 
Table 2: Stimulus conditions of discourse plan opera- 
tors 
ators for providing causal explanations. For example 
in (2), 6 which indirectly conveys a No, R gives an ex- 
planation of why R won't get a car. 
2.a. q: actually you'll probably get a car 
won't you as soon as you get there 
b. R: can't drive 
Note that this stimulus condition may contribute to 
greater dialogue efficiency by anticipating a follow-up 
request for an explanation. Formally, the condition is 
defined by the following axiom. 
((explanation-indicated s h ?p ?q) 
<- 
(wbel s (wbel h (not ?p))) 
(unless (wbel s (accepts-authority h s)))) 
This may be glossed as, s is motivated to give h an 
explanation q for p, if s suspects that h suspects that 
p is not true, unless it is the case that s has reason 
to believe that h will accept p on s's authority. (wbel 
agent p), which we usually gloss as agent suspects that 
p, denotes that agent has some confidence in the belief 
that p. Note tha, t R may acquire the suspicion that Q 
doubts that p is true by means of syntactic clues from 
the Yes-No question, e.g. the form of the question in (2@ 
Excuse-indicated 
Although this stimulus condition appears in some 
of the same causal operators as Explanation-indicated, 
~Stenstr6m's (110) 
it represents a different kind of motivation. A Yes-No 
question may be interpreted as a prerequest \[Lev83\] 
for a request, i.e. as an utterance used as a preface to 
another request. 7 Prerequests are often used to check 
whether a related request is likely to succeed, or to 
avoid having to make the other request directly. Thus, 
a negative answer to a Yes-No question used as a pre- 
request may be interpreted as a refusal. To soften the 
refusal, the speaker may give an explanation of the 
negative answer, as illustrated in (1). Formally, the 
condition is defined by the following axiom. 
((excuse-indicated s h (not ?p) ?q) 
<- 
(~bel s (prerequest h s (informif s h ?p)))) 
This may be glossed as, s is motivated to give h an 
excuse q for (not p), if s suspects that h's request, 
(informif s h p), is a prerequest. Techniques for inter- 
preting indirect speech acts \[PA80, Hin89\] can be used 
to determine whether the antecedent holds. 
Answer-ref-indicated 
This condition appears in Use-Elaboration, illus- 
trated by (3), s and in Use-Contrast, illustrated by 
(4) .9 
3.a. Q: 
b. 
c. E: 
d. 
Did you have a hotel in mind? 
\[What hotel did you have in mind?\] 
\[yes\] 
There's a Holiday Inn right near 
where I'm working. 
4.a. Q: You're on that? 
b. \[Who's on that?\] 
c. R: no no no 
d. Dave is. 
In (3), R has interpreted the question in (a) as a pre- 
request for the wh-question shown in (b). Thus, (d) 
not only answers the question in (a) but also the an- 
ticipated wh-question in (b). Similarly in (4), R may 
interpret the question in (a) as a prerequest for the 
wh-question in (b), and so gives (d) to provide an an- 
swer to both (a) and (b). Formally, the condition is 
defined by the following axiom. 
((Answer-re:f-indicated s h ?p ?q) 
<- 
(wbel s (want h (knowref h ?t ?q))) 
This may be glossed as, s is motivated to provide h 
with q, if s suspects that h wants to know the referent 
Tin analyses of discourse based on speech act theory, 
e.g. \[PA80\], a question such as (lb) might be described as 
an indirect request. However, since it may be responded 
to appropriately by a Yes or No, it is included in the class 
of Yes-No questions addressed in our model. 
SAmerican Express tape 1 
9StenstrSm's (102). Note that a No answer may be 
conversationally implicated by use of (4d) alone. 
192 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
of a term t in q. As in Excuse-indicated, techniques 
for interpreting indirect speech acts can be used to 
determine if the antecedent holds. 
Substitute-indicated 
This condition appears in Use-contrast, illustrated 
by (5). 
S.a.Q: Do you have Verdi's Otello or Aida? 
b.R: \[no\] 
c. We have Rigoletto. 
Although Q may not have intended to use (5a) as a 
prerequest for the qu#stion What Verdi operas do you 
have.C, R suspects that the answer to this wh-question 
might be helpful to Q. Formally, the condition is de- 
fined by the following axiom. 
((Substitute-indicated s h ?p ?q) 
<- 
(wbel s (need h (knowref h ?t ?q)))) 
This may be glossed as, s is motivated to provide h 
with q, if s suspects that it would be helpful for h to 
know the referent of a term t in q. The antecedent 
would hold wl~enever obstacle detection techniques 
\[AP80\] determine that h's not knowing the referent 
of t is an obstacle to an inferred plan of h's. How- 
ever, note that not all helpful responses, in the sense 
described in lAPS0\], can be used as indirect answers. 
For example, even if the clerk (R) at the music store 
believes that Q's not knowing the closing time could 
be an obstacle to Q's buying a recording, tl/e response 
(6), given instead of (5c), would not convey (55) since 
it cannot be coherently related to (55). 
6. R: We close at 5:30 tonight. 
Clarify-concept-indicated 
This stimulus condition appears 
Use-elaboration, as illustrated by (7)3 0 
in 
7. Q: Do you have a pet? 
R: We have a turtle. 
In (7), R was motivated to elaborate on the type of 
pet they have, since turtles are not prototypical pets. 
Formally, the condition is defined by the following ax- 
iom. 
((Clarify-concept-indicated s h ?p ?q) <- 
(concept ?p ?c) 
(has-atypical-insgance ?q ?c)) 
This may be glossed as, s is motivated to clarify p 
to h with q, if p has a concept c, and q provides an 
atypical instance of c. Stereotypical knowledge is used 
to evaluate the second antecedent. 
1° Hirschberg's (177). 
Clarify-condition-indicated 
This stimulus condition appears in the operator 
Use-condition, as illustrated by (8). 11 
8.a. Q: Um let me can I make the reservation 
and change it by tomorrow 
b. R: \[yes\] 
c. if it's still available. 
In (8), a truthful Yes answer depends on the truth of 
(c). Formally, the stimulus condition is defined by the 
following axioms. 
((Clarify-condition-indicated s h ?p ?q) 
<- 
(ignorant s ?q)) 
((Clarify-condition-indicated s h ?p ?q) 
<- 
(wbel s (not ?q))) 
These may be glossed as, s is motivated to clarify a 
condition q for p to h 1) if s doesn't know if q holds, 
or 2) if s suspects that q does not hold. 
Clarify-extent-indicated 
This stimulus condition appears in Use-contrast, 
as illustrated by (9). 12 
9.a. Q: Have you gotten the letters yet? 
b. R: I've gotten the letter from X. 
On the strict interpretation of (9a), Q is asking 
whether R has gotten allof the letters, but on a looser 
interpretation, Q is asking if R has gotten any of the 
letters. Then, if R has gotten some but not all of 
the letters, a Yes would be untruthful. However, if 
Q is speaking loosely, then a No might lead Q to er- 
roneously conclude that R has not gotten any of the 
letters. R's answer circumvents this problem, by con- 
veying the extent to which the questioned proposition 
(on the strict interpretation) is true. Formally, the 
condition is defined by the following axioms. 
((clarify-extent-indicated s h 
(some-truth ?p) ?q) 
<- 
(wbel s (ignorant h ?q)) 
(believe s (highest-true-exp ?q ?p))) 
((clarify-extent-indicated s h (not ?p) ?q) 
<- 
(wbel s (ignorant h ?q)) 
(believe s (highest-true-exp ?q ?p))) 
These may be glossed as, s is motivated to clarify to h 
the extent q to which p is true, or the alternative q to 
p which is true, if s suspects that h does not know if q 
n American Express tape 10ab 
a2 Hirschberg's (59) 
193 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
holds, and s believes that q is the highest expression al- 
ternative to p that does hold. According to I-Iirschberg 
\[Hir85\] (following Gazdar), sentences Pi and pj (rep- 
resenting the propositional content of two utterances) 
are expression alternatives if they are the same except 
for having comparable components ei and e j, respec- 
tively. Further, Hirschberg claims that in a particular 
discourse context, there may be a partial ordering of 
values which the discourse participants mutually be- 
lieve to be salient, and that the ranking of ei and ej 
in this ordering can be used to describe the ranking of 
pi and pj. In the above example, (9b) is a realization 
of the highest true expression alternative to the ques- 
tioned proposition, p, i.e. the proposition that R has 
gotten all the letters, is 
Appeasement-indicated 
This stimulus condition appears in Use-contrast, 
as illustrated by (10). 14 
10.a. Q: Did you manage to read that section 
I gave you? 
b. R: I've read the first couple of pages. 
In (10), R conveys that there is some (though not 
much) truth to the questioned proposition in an effort 
to soften his answer. Note that more than one stim- 
ulus condition may motivate R to include the same 
satellite. E.g., in (10), R may have been motivated 
also by clarify-eztent-indicated, which was described 
above. However, it is possible to provide a context for 
(9) where appeasement-indicated holds but not clarify- 
extent-indicated, or a context where the converse is 
true. Formally, the condition is defined by the follow- 
ing axiom. 
((appeasement-indicated s h (not ?p) ?q) <- 
(wbel s (undesirable h (not ?p))) 
(wbel s (desirable h ?q))) 
((appeasement-indicated s h 
(some-truth ?p) ?q) <- 
(wbel s (undesirable h (not ?p))) 
(wbel s (desirable h ?q))) 
This may be glossed as, s is motivated to appease h 
with q for p not holding or only being partly true, if s 
suspects that (not p) is undesirable to h but that q is 
desirable to h. The antecedents to this rule would be 
evaluated using heuristic rules and stereotypical and 
specific knowledge about h's desires. For example, two 
13Recall that additional constraints on p and q arise from 
the applicability conditions of operators containing this 
stimulus condition, namely Use-contrastin this case. Thus, 
another constraint is that it is plausible that cr-contrast 
holds. Note that cr-contrast also is defined in terms of 
such an ordering. 
a4 Hirchberg's (56) 
heuristics of rational agency that might lead to beliefs 
about h's desires are 1) if an agent wants you to per- 
form an action A, then your failure to perform A may 
be undesirable to the agent, and 2) if an agent wants 
you to do A, then it is desirable to the agent that you 
perform a part of A. 
4. Generation Algorithm 
The inputs to generation in our model consist of 1) a 
set of discourse plan operators augmented with stim- 
ulus conditions, 2) a set of coherence rules, 3) R's 
beliefs, and 4) the discourse expectation that R will 
provide R's evaluation of the truth of the questioned 
proposition p. The output of the generation algorithm 
is a discourse plan which can be realized by a tactical 
generation component \[McK85\]. We assume that when 
answer generation begins, the speaker's only goal is to 
satisfy the above discourse expectation35 
Our answer generation algorithm has two phases. 
In the first phase, content planning, the generator cre- 
ates a discourse plan for a full answer, i.e., a direct 
answer and extra appropriate information, e.g. (lc) 
given explicitly, followed by (ld) : (le). In the second 
phase, plan pruning, the generator determines which 
propositions of the planned full answer do not need to 
be explicitly stated. For example, given an appropri- 
ate model of R's beliefs, our system generates a plan 
for asserting only the proposition conveyed in (le) as 
an answer to (la) - (lb)36 Note that an advantage of 
our approach is that, even when it is not possible to 
omit the direct answer, a full answer is generated. 
4.1 Content planning 
Content planning is performed by top-down expansion 
of an answer discourse plan operator. The process be- 
gins by instantiating each of the operators with the 
questioned proposition until one is found such that its 
applicability conditions hold. Next, this (top-level) 
answer operator is expanded. A discourse plan op- 
erator is expanded by deciding which of its satellites 
to include in the full answer and expanding each of 
them (recursively). A satellite (e.g. Use-obstacle in 
Figure 2) is selected if, for some instantiation of its 
existential variable, all of the applicability conditions 
and at least one of the stimulus conditions of the in- 
stantiated operator hold. 
Note that testing of applicability conditions and 
stimulus conditions (including searching for an appro- 
aSThe plan that is output by our algorithm specifies an 
ordering of discourse acts based upon the ordering of co- 
herence relations specified in the discourse plan operators. 
However, to model a speaker who has more than the above 
initial goal, reordering may be required. 
16The tactical component must choose an appropriate 
referring expression to refer to R's car's timing belt, de- 
pending on whether (ld) is omitted. 
194 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
priate instantiation of the existential variable) is han- 
dled by a theorem prover. Our model of R's beliefs, 
represented as a set of Horn clauses, includes 1) general 
world knowledge presumably shared with Q, 2) knowl- 
edge about the preceding discourse, and 3) R's beliefs 
(including "weaker be~liefs" encoded using the wbel op- 
erator) about Q's beliefs. Much of the shared world 
knowledge needed to evaluate our coherence rules con- 
sists of knowledge from domain plan operators. 
4.2 Plan pruning 
The output of the ~ontent planning phase, an ex- 
panded discourse plan representing a full answer, is the 
input to the plan pruning phase of generation. The ex- 
panded plan is represented as a tree of discourse acts. 
The goal of this phase is to make the response more 
concise, 17 i.e., to determine which of the planned acts 
can be omitted while lstill allowing Q to infer the full 
discourse plan. To do this, the generator considers 
each of the acts in the frontier of the tree from right 
to left. (This ensures that a satellite is considered be- 
fore its related nucleus.) The generator creates a trial 
plan consisting of the original plan minus the nodes 
pruned so far and minus the current node. Then, us- 
ing the interpretation imodule, the generator simulates 
Q's interpretation of a response containing the infor- 
mation that would be given explicitly according to the 
trial plan. If Q couldlinfer the full plan (as the most 
preferred interpretation), then the current node can be 
pruned. Otherwise, itlis left in the plan and the next 
node is considered. 
For example, consider Figure 3 as we illustrate the 
possible effect of pruni;ng on a full discourse plan. The 
leaf nodes, representing discourse acts, are numbered 1 
- 8. Arcs labelled N and S lead to a nucleus or satellite, 
respectively. Node 8 corresponds to the direct answer. 
Plan pruning would process the nodes in order from 1 
to 8. The maximal set :of nodes which could be pruned 
in Figure 3 is the set containing 2, 3, 4, 7, and 8. That 
is, nodes 2 - 4 might be inferrable from 1, node 7 from 
5 or 6, and node 8 from 4 or 7. In the event that it is 
determined that no node can be pruned, the full plan 
would be output. Not:e that our interpretation algo- 
rithm (described in \[GC94\]) performs a subprocess, hy- 
pothesis generation, to recognize missing propositions 
other than the direct answer, i.e. the propositions at 
nodes 2, 3, 4, and 7. 
4.3 Example 
This section gives an ~example of how the discourse 
plan shown in Figure 4 would be generated, corre- 
sponding to the answer shown in (11), where only 
(lld) is explicitly stated. 
lrConciseness is not the only possible motive for omitting 
the direct answer. As mentioned in section 3, an indirect 
answer may be used to avoid performing a face-threatening 
act. 
7 6 5 4 -- 
3 
2 1 
Figure 3: Example of full discourse plan before prun- 
ing 
11.a. Q: You went to the party, didn't you7 
b. R: \[no\] 
c. \[The baby sitter could not sit.\] 
d. The baby sitter was sick. 
In phase one, the generator creates a discourse plan for 
a full answer. First, it must decide which top-level dis- 
course plan operator to expand. Each of the top-level 
operators would be instantiated with the questioned 
proposition, (occur (go-party s) past). The top-level 
operators would be instantiated and tested until one 
was found whose applicability conditions held, namely, Answer-no. 
Next, the generator must decide which of 
the satellites of Answer-no to include. So, it would 
search for all propositions q such that, when the exis- 
tential variable ?q of a satellite is instantiated with q, 
then all of the satellite's applicability conditions and 
at least one of its stimulus conditions hold. In this 
example, the proposition (not (in-state (can-sit sitter) 
past)) would be found to satisfy the applicability con- 
ditions and the explanation-indicated stimulus condi- 
tion of Use.obstacle. (Note that this proposition might 
be chosen because the model of R's beliefs includes a 
domain plan operator for going to a party, with a pre- 
condition that the agent's sitter can baby sit.) 
Next, this instantiation of Use-obstacle must be 
expanded, i.e., the generator must decide which of the 
satellites of Use-obstacle to include. By a procedure 
similar to that described above, it would decide to 
include (another) Use-obstacle, where the existential 
variable is instantiated with the proposition (in-state 
(sick sitter) past). (Note that this proposition might 
be chosen because the model of R,'s beliefs includes the 
shared belief that being sick typically renders a sitter 
unfit to perform his or her duties.) The stimulus con- 
dition explanation-indicated might be the motivation 
for including this satellite. (For example, suppose it 
is shared knowledge that R's baby sitter is rarely un- 
available.) If further attempts to expand the plan were 
unsuccessful, then the full plan would contain the dis- 
course acts shown in Figure 4. (The acts in the plan 
are labelled (b) - (d) corresponding to the part of (11) 
they represent.) Note that if this plan were output 
195 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
(Answer-no s h (occur (go-party s) past)): 
Nucleus: 
b. (inform s h (not (occur (go-party s) 
past))) 
Satellites: 
(Use-obstacle s h 
(not (occur (go-party s) past))) 
Nucleus: 
(inform s h 
(not (in-state (can-sit sitter) 
past))) 
Satellites: (Use-obstacle 
s h 
(not (in-state (can-sit sitter) 
past))) 
Nucleus: 
(inform s h (in-state (sick sitter) 
past)) 
C. 
d. 
Figure 4: Discourse plan for (11) 
without undergoing plan-pruning, it would be realized 
as a direct answer (llb), and extra information (llc) 
-(lld). 
In phase two, plan-pruning, the generator's over- 
all goal is to make the answer as concise as possible. 
Therefore, it would consider the primitive acts speci- 
fied in the full plan in the order (d), (c), (b). Since 
(d) is not inferrable from any other act, it would be 
automatically retained in the plan. Next, a trial plan 
with (c) pruned from it would be considered. Interpre- 
tation of the response consisting of (b) and (d) is sim- 
ulated. Since the full plan could be inferred, (c) would 
be pruned. Next, another trial plan would be consid- 
ered, where, in addition to (c), (b) has been pruned. 
Since the full plan could be inferred from the result- 
ing trial response consisting of just (d), (b) would be 
pruned too. Thus, the output of phase two would be 
the plan shown in Figure 4 with (b) and (c) marked 
as pruned. 
5. Other factors in generation 
The primary focus of our research has been on 1) use 
of pragmatic knowledge which is common to both the 
interpretation and generation of indirect answers, i.e. 
the so-called reversible knowledge, and 2) identifying 
stimulus conditions. In this section, we give a brief 
description of some other factors in generating indirect 
answers. 
First, note that we made the simplifying assump- 
tion that R's only initial goal is to answer the ques- 
tion (accurately, efficiently, and politely). However, it 
is possible for R to construct an indirect answer which 
satisfies multiple initial goals of R. For example, R 
might have decided to give the answer in (12c) not 
only as an explanation, but also as background for R's 
request in (12d). 
12.a. q: Are you going to the lecture? 
b. R: \[no\] 
c. I have to get my brakes fixed. 
d. Can you recommend a garage near 
c ampus ? 
Second, stylistic goals may affect the generation 
of indirect answers. In addition to affecting syntactic 
and lexical aspects of discourse \[Hovg0, MG91, DH93\], 
stylistic goals may affect which and how much infor- 
mation is given. For example, elaboration can be used 
to make an answer more lively, as shown in (13). 
13,a. Q: Do you have a car? 
b. R: \[yes\] 
c. I bought a British-racing-green 
Austin-Healey 3000 last week. 
Also, extra information that entails the direct answer 
(e.g. repetitions and generalizations) may be given for 
emphasis. Further, a stylistic goal of terseness may 
override the stimulus conditions we have provided for 
indirect answers. For example, depending upon the 
degree of Q's interest in R's affairs, the indirect answer 
in (14c) - (14f) may provide Q with more than Q cares 
to know. 
14.a. Q: Are you going to the movies tonight? 
b. R: \[no\] 
c. I can't afford to. 
d. I spent $900 on my car last week. 
e. It needed a new transmission. 
f. The car is 15 years old. 
Third, certain syntactic forms for expressing Yes- 
No questions require more than just yes~no to make 
clear whether R is confirming or disconfirming that a 
proposition p holds \[Ste84\]. For example, is (15) shows 
the use of amplification (i.e. supplying an auxiliary 
verb) in answer to a negative-polarity, positive-bias 
question. 
I5.a. Q: Didn't Ann get the letter? 
b. R: Yes she did. 
c. R: No she didn't. 
(A question is said to have positive or negative polarity 
depending on the presence or absence of negation, re- 
spectively. Bias describes the questioner's presumed 
belief in the truth of the questioned proposition.) 
In answer to (16a), a negative-polarity, negative-bias 
question, 19 R provides an indirect answer, giving the 
extent to which the proposition that R has been up- 
18(15) and (16) are Stenstrbm's (105) and (111), 
respectively. 
19Although it is expressed as a declarative sentence, 
Stenstr6m classifies (16a) as a request for confirmation, 
which we treat as a type of Yes-No question. 
196 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
stairs is true, in order to avoid the same type of con- 
fusion. 
16.a. Q: I don't think you've been upstairs 
yet. 
b. R: Um only just to the ioo. 
6. Related research 
Our work differs from most previous work in cooper- 
ative response generation in that, in our model, the 
information given in an indirect answer conversation- 
ally implicates \[Gri75a\] the direct answer. I-Iirschberg 
\[Hir85\], who provided general rules for identifying 
when a scalar conversational implicature is licensed, 
claimed that speakers may give indirect responses to 
Yes-No questions in order to block potential incor- 
rect scalar implicatures of a simple Yes or No. She 
implemented a system which determines whether the 
Yes/No 2° licenses any unwanted scalar implicatures, 
and if so, proposes alternative scalar responses which 
do not. This type of response is similar, in our model, 
to a speaker's use of Use-contrast motivated by clarify- 
extent-indicated, as illustrated in (9). (As noted ear- 
lier, our coherence rules for the relation cr-contrast 
as well as our axioms for clarify-extent-indicated make 
use of the notion of a salient partial ordering which 
was elucidated by Hirsehberg.) However, Hirschberg's 
model does not account for quite a variety of types 
of indirect answers which can be generated using the 
other operators in our model, nor for other motives for 
using Use-contrast. 
Rhetorical or coherence relations \[Gri75b, Hal76, 
MT87\] have been used in several text-generation sys- 
tems to aid in ordering parts of a text (e.g.\[Hov88\]) 
as well as in content-planning (e.g. \[McK85, MP93\]). 
The discourse plan operators based on coherence rela- 
tions in our model (i.e. the operators used as satellites 
of top-level operators)play a similar role in content- 
planning of an indirect answer. When coherence re- 
lations are used for content-planning, it is necessary 
to constrain the information which thereby may be 
selected. McKeown \[McK85\] uses discourse focus con- 
straints. In \[Moo89\], plan selection heuristics are used 
that maximize the heater's presumed familiarity with 
the concepts in the text, prefer general-purpose to less 
general operators, and minimize verbosity. Maybury's 
\[May92\] system uses "desirable" preconditions, pre- 
conditions that are not inecessary preconditions, to pri- 
oritize alternative operators. In contrast to the above, 
by the use of stimulus conditions, our model is able to 
trigger new, opportunistic speaker goals. 
Moore and Pollack \[MP92\] show the need to dis- 
tinguish the intentional and informational structure 
of discourse, where the latter is characterized by the 
2°Hirschberg did not addresss other possible types of di- 
rect answers, represented by the Answer-Hedged, Answer- 
Maybe, and Answer-Maybe-Not operators in our model. 
sort of relations classified as subject-matter relations 
in RST. Note that in our interpretation component 
\[GC94\], informational relations (i.e. plausible co- 
herence relations holding between indirect answers 
and candidate direct answers) are used to infer the 
speaker's goal to convey a particular answer. Moore 
and Paris \[MP93\] argue that it is necessary for gen- 
eration systems to represent not only the speaker's 
top-level intentional goal, but also the intentional sub- 
goals that a speaker hoped to achieve by use of a 
rhetorical relation so that, if a subgoal is not achieved, 
then an alternative rhetorical means can be tried. In 
our model, the operators used as satellites of top-level 
answer discourse plan operators are based on RST's 
subject-matter relations. The primary goals of these 
operators are similar to the effect fields of the cor- 
responding RST relation definitions. As Moore and 
Paris argue, goals based on these relation definitions 
are rhetorical goals, not intentional goals. However, 
in our model, it is possible to determine the inten- 
tional subgoals by checking which stimulus conditions 
motivated the speaker to include a particular satellite 
operator. For example, if explanation-indicated moti- 
vates R to include a Use-cause satellite, then one can 
view R as having the implicit intentional goal of giving 
an explanation. Thus, our model provides the infor- 
mation necessary to recover from the possible failure 
of (all or part of) an indirect answer. 
Finally, our use of simulated interpretation during 
plan pruning has some similarity to previous work. In 
Horacek's approach to generating concise explanations 
\[I-Ior91\], a set of propositions representing the full ex- 
planation i s pruned by eliminating propositions which 
can be derived from the remaining ones by contextual 
rules. Jameson and Wahlster \[JW82\] use an antici- 
pation feedback loop algorithm to generate elliptical 
utterances. 
7. Conclusion 
An indirect answer to a Yes-No question conversation- 
ally implicates the speaker's evaluation of the truth 
of a questioned proposition. The generation of indi- 
rect answers is important if a dialogue system is to re- 
spond efficiently, accurately, and politely. This paper 
has presented the approach to generation used in our 
implemented system for generating and interpreting 
indirect answers to Yes-No questions in English. Ours 
is the first system to generate a wide range of types of 
indirect answers (as well as full answers). The system 
is reversible in that the same pragmatic knowledge is 
used in generation and interpretation. Generation is 
performed in two phases: content planning and plan 
pruning. During content planning, stimulus conditions 
are used to trigger speaker goals to include appropriate 
extra information with the direct answer. Plan prun- 
ing determines what parts of this full response do not 
need to be stated explicitly - resulting in, in appropri- 
ate discourse contexts, the generation of an indirect 
197 
7th International Generation Workshop • Kennebunkport, Maine • June 21-24, 1994 
answer. 

References 
\[AP80\] James F. Allen and C. Raymond Perrault. An- 
alyzing intention in utterances. Artificial In- 
telligence, 15:143-178, 1980. 
\[Ame92\] American Express tapes. Transcripts of au- 
diotape conversations made at SRI Interna- 
tional, Menlo Park, California. Prepared by 
Jacqueline Kowto under the direction of Patti 
Price. 
\[BL78\] Penelope Brown and Stephen Levinson. Uni- 
versals in language usage: Politeness phe- 
nomena. In Esther N. Goody, editor, Ques- 
tions and politeness: Strategies in social in- 
teraction, pages 56-289. Cambridge University 
Press, Cambridge, 1978. 
\[CM81\] H. Clark and C. Marshall. Definite refer- 
ence and mutual knowledge. In A. K. Joshi, 
B. Webber, and I. Sag, editors, Elements of 
discourse understanding. Cambridge Univer- 
sity Press, Cambridge, 1981. 
Chrysanne DiMarco and Graeme Hirst. A 
computational theory of goal-directed style 
in syntax. Computational Linguistics, 19(3), 
1993. 
Nancy Green and Sandra Carberry. A hybrid 
reasoning model for indirect answers. To ap- 
pear in Proceedings of the 32nd Annual Meet- 
ing of the Association for Computational Lin- 
guistics, 1994. 
\[Gri75a\] H. Paul Grice. Logic and conversation. In 
P. Cole and J. L. Morgan, editors, Syntax and 
Semantics lII: Speech Acts, pages 41-58, New 
York, 1975. Academic Press. 
\[Gri75b\] J. E. Grimes. The Thread of Discourse. Mou- 
ton, The Hague, 1975. 
\[Hal76\] M. Halliday. Cohesion in English. Longman, 
London, 1976. 
\[Hin89\] Elizabeth Ann Hinkelman. Linguistic and 
Pragmatic Constraints on Utterance Interpre- 
tation. PhD thesis, University of Rochester, 
1989. 
\[Hir85\] Julia Bell Hirschberg. A Theory of Scalar Im- 
plicature. PhD thesis, University of Pennsyl- 
vania, 1985. 
\[Hor91\] Helmut Horacek. Exploiting conversational 
implicature for generating concise explana- 
tions. In Proceedings. European Association 
for Computational Linguistics, 1991. 
\[Hov88\] Eduard H. Hovy. Planning coherent multisen- 
tential text. In Proceedings of the 26th Annual 
Meeting of the Association for Computational 
Linguistics, pages 163-169, 1988. 
\[Hov90\] Eduard H. Hovy. Pragmatics and natural 
language generation. Artificial Intelligence, 
43:153-197, 1990. 
\[JW82\] Anthony Jameson and Wolfgang Wahlster. 
User modelling in anaphora generation: ellip- 
sis and definite description. In Proceedings of 
the European Conference on Artificial Intelli- 
gence, 1982. 
\[Lev83\] S. Levinson. Pragmatics. Cambridge Univer- 
sity Press, Cambridge, 1983. 
\[May92\] Mark T Maybury. Communicative acts for 
explanation generation. International Journal 
of Man-Machine Studies, 37:135-172, 1992. 
\[MeK85\] Kathleen R. McKeown. Text Generation. 
Cambridge University Press, 1985. 
r\[MG91\] Marzena Makuta-Giluk. A computational 
rhetoric for syntactic aspects of text. Technical 
Report CS-91-56, Master's thesis, Department 
of Computer Science, University of Waterloo, 
Waterloo, Ontario, Canada, December 1991. 
\[Moo89\] Johanna D. Moore. A Reactive Approach to 
Explanation in Expert and Advice-Giving Sys- 
tems. PhD thesis, University of California at 
Los Angeles, 1989. 
\[MP92\] Johanna D. Moore and Martha E. Pollack. 
A problem for rst: The need for multi-level 
discourse analysis. Computational Linguistics, 
18(4):537-544, December 1992. 
\[MP93\] Johanna D. Moore and Cecile Paris. Plan- 
ning text for advisory dialogues: Capturing in- 
tentional and rhetorical information. Compu- 
tational Linguistics, 19(4):651-694, December 
1993. 
\[MT87\] W. C. Mann and S. A. Thompson. Rhetorical 
structure theory: Toward a functional theory 
of text organization. Text, 8(3):167-182, 1987. 
\[PA80\] R. Perrault and J. Allen. A plan-based analy- 
sis of indirect speech acts. American Journal 
of Computational Linguistics, 6(3-4):167-182, 
1980. 
\[Ste84\] Anna-Brita StenstrSm. Questions and re- 
sponses in english conversation. In Claes 
Schaar and Jan Svartvik, editors, Lund Studies 
in English 68. CWK Gleerup, MaimS, Sweden, 
1984. 
