Conditional Responses in Information-Seeking Dialogues
Elena Karagjosova and Ivana Kruijff-Korbayov´a
Computational Linguistics, Saarland University, Saarbr¨ucken, Germany
felka,korbayg@coli.uni-sb.de
Abstract
The paper deals with conditional responses of the
form “Not if c/Yes if c” in reply to a question “?q”
in the context of information-seeking dialogues. A
conditional response is triggered if the obtainability
of q depends on whether c holds: The response in-
dicates a possible need to find alternative solutions,
opening a negotiation in the dialogue. The paper
discusses the conditions under which conditional
responses are appropriate, and proposes a uniform
approach to their generation and interpretation.
1 Introduction
The goal of this paper is to provide a basic account
of conditional yes/no responses (CRs): We describe
the conditions under which CRs are appropriate, and
how these conditions translate into a uniform ap-
proach to understanding and producing CRs.1
We focus on information-seeking dialogues be-
tween a human user and a dialogue system in the
travel domain. We allow for mixed initiative and ne-
gotiation to let a dialogue be more collaborative than
“quizzing”. In this context CRs arise naturally (1).
(1) U.1: Do I need a visa to enter the U.S.?
S.1: Not if you are an EU citizen.
(2) S.1’: Yes, if you are not an EU citizen.
(1:S.1) is an example of a negative CR, asserting If
you’re an EU citizen, then you do not need a visa to
enter the U.S. An alternative, positive CR is (2:S.1’),
asserting If you’re not an EU citizen, then you do
need a visa to enter the U.S..
In both cases, the system answers the question
(1:U.1), but it makes the answer conditional on the
value of a particular attribute (here, citizenship).
1This work was done in SIRIDUS (Specification, Interac-
tion and Reconfiguration in Dialogue Understanding Systems),
EC Project IST-1999-10516. We would like to thank Geert-Jan
Kruijff for detailed discussion and comments .
Moreover, the CR suggests that, for another value,
the answer may be different (2).
The CRs in (1:S.1) and (2:S.1’) are elliptical utter-
ances. Intuitively, they can be expanded to the com-
plete propositions in (3) and (30). The material for
resolving the ellipsis comes from the immediately
preceding context. In the approach we work with,
ellipsis is resolved with respect to the current ques-
tion under discussion (QUD, (Ginzburg, 1996)).
(3) No, you don’t need a visa to enter the U.S. if you are an
EU citizen.
(30) Yes, you do need a visa to enter the U.S. if you are not
an EU citizen.
The dialogue move of a CR depends on the con-
text. Consider (4) and (5). Similarly to (1), in (4)
the system does not know an attribute-value (A/V)
on which the positive or the negative answer to the
yes/no question is contingent (here, whether the user
wants a business or economy class flight).2
(4) U.1: A flight from K¨oln to Paris on Sunday.
S.1: I’m sorry, there are no flights from K¨oln to Paris
on Sunday.
U.2: Can I fly on Monday?
S.2: Not if you want business class.
S.20: Yes, if you want economy class.
(5) U.1: I want a business class flight from K¨oln to Paris
on Sunday.
S.1: I’m sorry, there are no business flights from K¨oln
to Paris on Sunday.
U.2: Can I fly on Monday?
S.2: Not if you want business class.
S.20: Yes, if you want economy class.
The system’s CR (4:S.2) is a request for further in-
formation: whether the user wants a business flight
(Monday is out), or does not (she is able to fly on
Monday). Likewise, (4:S.20) is a request for fur-
ther information whether the user wants an economy
flight (Monday is available), or not (Monday is out).
Dialogue (5) is different. Now the user indi-
cates that she is interested in a business class flight
2We realize that intonation might play a role. However,
given space restrictions we cannot address this issue here.
       Philadelphia, July 2002, pp. 84-87.  Association for Computational Linguistics.
                  Proceedings of the Third SIGdial Workshop on Discourse and Dialogue,
(5:U.1). The system by default assumes that this re-
mains unchanged for another day of travel.
What both the negative and positive CR in (5)
do is to start a negotiation to either confirm or re-
vise the user’s decision for business class. The sys-
tem’s response (5:S.2) or (5:S.20) indirectly proposes
a change (to economy class) to achieve the higher-
level goal of finding a flight from K¨oln to Paris on
Monday. If the user insists on business class, this
goal cannot be achieved.
If we want a dialogue system to understand and
appropriately produce CRs, we need to describe
their semantics in terms of the contextual condi-
tions and communicative goals under which these
responses occur, and the effects they have on the di-
alogue context. We aim at providing the basis of
an account that can be implementated in the GoDiS
dialogue system. GoDis is an experimental sys-
tem in the travel domain, using the information-state
approach to dialogue developed the TRINDI and
SIRIDUS projects (Cooper et al., 1999; Lewin et al.,
2000). We focus on aspects that can improve its flex-
ibility and functionality.
Overview. In x2 we discuss the uses of positive
and negative CRs in terms of their appropriateness
conditions and their interpretation. Inx3 we discuss
dialogue moves. We end the paper with conclusions.
2 Uses of conditional responses
In this section we present two different types of CRs
and discuss in what contexts they are appropriate.
CRs can be used as answers to yes/no-questions.3
A CR does not provide a yes/no answer simpliciter,
though: It provides an answer that is contingent on
the value of some attribute. Consider (1). The sys-
tem’s reply (1:S.1) provides an answer that is con-
tingent on the value of the citizenship attribute. If
the value is (or implies) EU citizenship, the answer
is negative: If the user is an EU citizen, she does not
need a visa to enter the U.S.
The CR in (1) also seems to suggest the contra-
positive that if the value is “non-EU-citizen”, the
answer is positive. (2) illustrates the opposite case.
We consider this additional suggestion an implica-
ture. The assertions and implicatures that arise from
3Corpora show also occurrences of CRs in response to state-
ments, cf. (Karagjosova and Kruijff-Korbayov´a, 2002).
CRs are summarized in Figure 1.
Green & Carberry (1999) characterize CRs in
terms of the speaker’s motivation to provide infor-
mation “about conditions that could affect the ve-
racity of the response”. However, they only consider
cases like (4) in which the A/V on which the CR is
contingent has not yet been determined in the pre-
ceding context (or cannot be assumed). Cases like
(5) where the A/V has been determined are left un-
noticed.4 We discuss each of the cases below.
Not-determined A/V. The A/V on which a CR is
contingent can be one that has not yet been deter-
mined in the preceding context, as in (1) and (4).
We call this type of CR a non-determined A/V CR
(NDCR). Besides the assertion and the implicature
that answer the yes/no question as specified in Fig-
ure 1, the NDCR amounts to indirectly giving rise to
the question “whether c holds”.
Consider the user’s utterances in (6) as continu-
ations of (1). They show that the implicitly raised
question cannot be answered just by “yes” or “no”.
Rather, it requires some content that matches with c.
(6) U.2: Yes.jNo.
U.20: Yes, I am.jNo, I am not.
U.200: Yes, I have German citizenship.j
No, I have Czech citizenship.
The responses in (6:U.2) could be interpreted as
acknowledgments, but certainly not as answers to
whether the user is an EU citizen. This is corrobo-
rated by the following continuation of (6:U2’) where
the system does answer the pending question.
(7) S.2: Then you do (not).
S.20: Then you do (not) need a visa.
(7:S.2) is elliptical for (7:S.20). Correct resolution of
the ellipsis is possible only if the question whether
the user needs a visa is the topmost QUD.
The need to answer the implicitly raised ques-
tion depends on what goals the participants try to
achieve. “Do I need a visa?” in (1) is satisfactorily
answered with either a yes or a no, or when enough
information is provided so the asker can find out the
answer herself. On the other hand, consider (8).
(8) U.1 Can I fly to Paris tomorrow?
S.1 Not if you want to fly economy class.
4Both cases are attested in corpora (Karagjosova and
Kruijff-Korbayov´a, 2002).
Y/N-Question ?q ?q
Response Not if c Yes if c
Assertion If c, not-q If c, then q
Implicature Possibly, if not-c, then q Possibly, if not-c, then not-q
Figure 1: Patterns of conditional responses
In (8) the response is contingent on whether the
user wants to fly economy class. Before flight se-
lection can proceed further, the question whether
c holds must be answered. In order to satisfy its
goal of finding a flight which satisfies the user re-
quirements, the system does need to know whether
c holds to find out whether q holds. This is a differ-
ence between (8) and (1). In (1), the system’s goal
is merely to answer the user’s question.
To summarize, the interpretation of a CR in re-
sponse to a question whether q in a context where c
has not been established is that (i) it is still not de-
termined whether q, because (ii) the answer (speci-
fied in Figure 1) is contingent on c, and thus (iii) the
question whether c holds is implicitly raised.
As for production, it is appropriate for the system
to produce a NDCR when (i) answering a yes/no-
question whether q, where (ii) the answer is either q
or not-q, depending on some additional A/V c which
has not yet been established in the context. We con-
jecture that whether a positive or a negative CR is
more cooperative in a particular context depends on
what the preferred answer to the question “whether
q” is assumed to be.
Contextually-determined A/V. Another context in
which a CR is appropriate is when an answer to a
yes/no-question is contingent on an A/V that has
already been established in the preceding context,
as in (5). We call this type of CR a contextually-
determined A/V CR (CDCR).
What does a CDCR communicate besides the as-
sertion and implicature that answer the question as
specified in Figure 1? We suggested in x1 that it
initiates a negotiation about the already established
A/V. However, this cannot happen by simply rais-
ing the question whether c holds, because c has al-
ready been established. We suggest that a CDCR
implicitly proposes to consider changing the A/V: It
re-raises the question whether c holds. Re-raising
c differs from raising a “new” question at least in
two aspects: c must be negotiable, and re-raising
c means it cannot be answered simply by provid-
ing a sufficiently discriminative positive or negative
response. To see the difference, consider (5) with
(5:S.2) continued by the following utterances.
(9) U.3: Yes.jNo.
U.30: Yes, I do.jNo, I don’t.
U.300: Yes, I want business class.j
No, I don’t want business class.
(10) U.3: OK, I can fly economy.
U.30: But I do want business class.
(11) U.3: How about Tuesday?
Like the responses in (6), the response in (9:U.3)
cannot be interpreted as answers to whether the user
wants to change her mind from business to economy
class. It seems hard to interpret even as acknowledg-
ment. But then we observe a number of differences
from the NDCR in (6):
The responses in (9:U.30) and (9:U.300) are not ap-
propriate as answers to the implicitly re-raised c, be-
cause a revision of an A/V is involved. Hence, some
kind of acknowledgment of the revision is needed in
addition to answering whether or not the A/V is to
be revised (and how). Such acknowledgments are
present in (10). In (10:U.3), ‘OK’ can be seen as ac-
knowledging the revision from business to economy
class. In (10:U.30), ‘but’ acknowledges the contrast
between the proposed revision and the actual preser-
vation of the A/V (here, business class). The con-
tinuation in (11), on the other hand, refuses the pro-
posed revision only implicitly by proposing instead
to check the flight possibilities on another day.
Another observation concerning a CDCR is that
it cannot immediately follow after an utterance in
which the value is established, as the inappropriate-
ness of (12:S.1) and (12:S.10) shows.
(12) U.1: Can I fly business class from K¨oln to Paris on
Sunday?
S.1: Not if you want business class.
S.10: Yes if you want economy class.
Intuitively, the reason for this is that there needs
to be some degree of uncertainty (in the sense of
being assumed but not known to be shared) about
the A/V. For example, in (5), the business class re-
quirement is assumed to be maintained when the day
is revised. The inappropriateness of (12:S.1) and
(12:S.10) can also be explained on purely semantic
grounds. When both the assertion and the implica-
ture as specified in Figure 1 are taken into account,
a contradiction arises: Given that the elliptical an-
swer is resolved to the previous utterance, (12:S.1)
asserts If user wants business class, then a business
flight from SB to Paris on Sunday is not available,
and implicates If user does not want business class,
then a business flight from SB to Paris on Sunday is
available. Similarly for (12:S.10).
Thus, the interpretation of a CDCR is that (i) it is
now determined whether q or not-q holds, because
(ii) the answer (specified in Figure 1) is contingent
on c and c is established. Also, (iii) the CDCR in-
dicates the reason for the answer, and (iv) proposes
to reconsider the earlier made decision by implicitly
re-raising the question whether c holds, and (v) mak-
ing a suggestion for it to be revised. A negotiation is
started in which the conflicting A/V is either revised
or confirmed. In the latter case a different solution
to the overall goal must be sought.
As for production, the system may produce
a CDCR when (i) answering a yes/no-question
whether q, where (ii) the answer is either q or not-
q, depending on some A/V c which has been estab-
lished in the context prior to the question whether q.
Again, what polarity of CR is more cooperative in
a particular context depends on what the preferred
answer to the question whether q is assumed to be.
3 Conditional response dialogue moves
According to the dialogue annotation scheme of
(Allen and Core, 1997), utterances in which “the
participant does not address the proposal but per-
forms an act that leaves the decision open pending
further discussion” are called hold moves. The di-
alogue moves of a NDCR seem similar to hold in
that the answer to q remains pending due to its con-
tingency on an unknown A/V c. Once c is deter-
mined, q is answered. Hence, we propose to char-
acterize a NDCR as a dialogue move combining
the backward-looking function of a partial yes/no-
answer and hold, and the forward-looking function
of a yes/no question whether the condition holds. A
CDCR is different in that it proposes to reconsider a
contexually-determined c. Allen & Core provide no
suitable characterization of this. We propose to char-
acterize a CDCR as a dialogue move that combines
the backward-looking function of a yes/no-answer
with the forward-looking function of an alternative
question whether c is preserved or revised.
4 Conclusions
We proposed an approach to dealing with condi-
tional responses (CRs), which arise naturally in dia-
logues allowing for mixed initiative and negotiation.
We proposed two types of CRs. One type describes
the case where the answer is contingent on an at-
tribute/value that has not yet been determined in the
context (NDCRs). The other type deals with an at-
tribute/value that has already been set in the context,
and which now needs to reconsidered (CDCRs). The
distinction properly clarifies the different effects on
dialogue context CRs may have. We are currently
developing an implementation of CRs in the GoDiS
system (Kruijff-Korbayov´a et al., 2002).

References
James Allen and Mark Core. 1997. Draft of
damsl: Dialogue Act Markup in Several Layers.
http://www.cs.rochester.edu/research/cisd/resources/damsl.
Robin Cooper, Staffan Larsson, Colin Matheson, Massimo Poe-
sio, and David Traum. 1999. Coding Instructional Dialogue
for Information States. http://www.ling.gu.se/projekt/trindi/,
February.
Jonathan Ginzburg. 1996. Interrogatives: Questions, Facts and
Dialogue. In Shalom Lappin, editor, The Handbook of Con-
temporary Semantic Theory, pages 385–422. Blackwell, Ox-
ford, UK/Cambridge, USA.
Nancy Green and Sandra Carberry. 1999. A Computational
Mechanism for Initiative in Answer Generation. User Mod-
eling and User-Adapted Interaction, 9(1/2):93–132.
Elena Karagjosova and Ivana Kruijff-Korbayov´a. 2002. An
Analysis of Conditional Responses in Dialogue. In Proceed-
ings of the 5th International Conference on TEXT, SPEECH
and DIALOGUE, Brno, Czech Republic. forthcoming.
Ivana Kruijff-Korbayov´a, Elena Karagjosova, and Staffan Lars-
son. 2002. Enhancing collaboration with conditional re-
sponses in information seeking dialogues. Under review.
Ian Lewin, C.J.Rupp, Jim Hieronymus, David Milward,
Staffan Larsson, and Alexander Berman. 2000. Siridus
System Architecture and Interface Report (Baseline).
http://www.ling.gu.se/projekt/siridus/.
