Proceedings of the Interactive Question Answering Workshop at HLT-NAACL 2006, pages 1–8,
New York City, NY, USA. June 2006. c©2006 Association for Computational Linguistics
Contextual phenomena and thematic relations in database QA dialogues:
results from a Wizard-of-Oz Experiment
N´uria Bertomeu, Hans Uszkoreit
Saarland University
Saarbr¨ucken, Germany
uszkoreit|bertomeu@coli.uni-sb.de
Anette Frank, Hans-Ulrich Krieger and Brigitte J¨org
German Research Center of Artificial Intelligence
Saarbr¨ucken, Germany
frank|krieger|joerg@dfki.de
Abstract
Considering data obtained from a corpus
of database QA dialogues, we address the
nature of the discourse structure needed
to resolve the several kinds of contextual
phenomena found in our corpus. We look
at the thematic relations holding between
questions and the preceding context and
discuss to which extent thematic related-
ness plays a role in discourse structure.
1 Introduction
As pointed out by several authors (Kato et al., 2004),
(Chai and Ron, 2004), the information needs of
users interacting with QA systems often go beyond
a single stand-alone question. Often users want to
research about a particular topic or event or solve
a specific task. In such interactions we can expect
that the individual user questions will be themati-
cally connected, giving the users the possibility of
reusing part of the context when formulating new
questions.
That users implicitly refer to and even omit ma-
terial which can be recovered from the context
has already been replicated in several Wizard-of-
Oz experiments simulating natural language inter-
faces to databases, (Carbonell, 1983), (Dahlb¨ack
and J¨onsson, 1989), the most frequent contextual
phenomena being ellipsis, anaphora and definite de-
scriptions.
A big challenge for interactive QA systems is,
thus, the resolution of contextual phenomena. In or-
der to be able to do so a system has to keep track of
the user’s focus of attention as the interaction pro-
ceeds. The attentional state at a given point in the
interaction is given by the discourse structure. An
open issue, however, is the nature of the discourse
structure model needed in a QA system. Ahrenberg
et al. (1990) argue that the discourse structure in NL
interfaces is, given the limited set of actions to be
performed by the system and the user, simpler than
the one underlying human-human dialogue. Upon
Ahrenberg et al. (1990) this is given by the discourse
goals, rather than the overall goals of the user, as is
the case in task-oriented dialogues, (Grosz and Sid-
ner, 1986). Following Ahrenberg et al. (1990), the
QA discourse is structured in segments composed
by a pair of initiative-response units, like question-
answer, or question-assertion, in the absence of an
answer. A segment can be embedded in another seg-
ment if it is composed by a clarification request and
its corresponding answer. The local context of a
segment is given by the immediately preceding seg-
ment. Upon Ahrenberg et al. (1990), the latter re-
liably limits up the search space for antecedents of
anaphoric devices and ellipsis. However, as we will
see, there are few cases where the antecedents of
contextual phenomena are to be found beyond the
immediately preceding segments. This suggests that
a more complex approach to discourse structure for
QA systems is needed.
In more recent studies of interactive QA special
attention has been paid to the thematic relatedness of
questions, (Chai and Ron, 2004), (Kato et al., 2004).
Chai and Ron (2004) propose a discourse model-
ing for QA interactions in which they keep track
of thematic transitions between questions. Although
1
the applications of tracking thematic transitions be-
tween questions have not been investigated in depth,
Sun and Chai (2006) report on an experiment which
shows that the use of a model of topic transitions
based on Centering Theory improves query expan-
sion for context questions. However, these previous
studies on the thematic relations between questions
are not based on collections of interactive data, but
on questions centered around a topic that were col-
lected in non-interactive environments. This means
that they do not consider the answers to the ques-
tions, to which following questions can be related.
This paper presents data on different kinds of con-
textual phenomena found in a corpus of written nat-
ural language QA exchanges between human users
and a human agent representing an interactive infor-
mation service. We address two issues: the kinds
and frequencies of thematic relations holding be-
tween the user questions and the preceding context,
on the one hand, and the location of antecedents for
the different contextual phenomena, on the other.
We also discuss the question whether thematic rela-
tions can contribute to determine discourse structure
and, thus, to the resolution of the contextual phe-
nomena.
In the next section we present our data collection
and the aspects of the annotation scheme which are
relevant to the current work. In section 3 we present
data regarding the overall thematic cohesion of the
QA sessions. In section 4 we report on data regard-
ing the co-occurrence of discourse phenomena and
thematic relations and the distance between the phe-
nomena and their antecedents. Finally, we discuss
our findings with regard to their relevance with re-
spect to the nature of discourse structure.
2 Corpus and methodology
2.1 Experimental set-up
In order to obtain a corpus of natural QA inter-
actions, we designed a Wizard-of-Oz experiment.
The experiment was set up in such a way that the
exchanges between users and information system
would be as representative as possible for the inter-
action between users and QA systems. We chose an
ontology database instead of a text based closed do-
main QA system, however, because in order to simu-
late a real system short time responses were needed.
30 subjects took part in the experiment, which
consisted in solving a task by querying LT-WORLD,
an ontology containing information about language
technology1, in English. The modality of interaction
was typing through a chat-like interface.
Three different tasks were designed: two of them
concentrated on information browsing, the other one
on information gathering. In the first task sub-
jects had to find three traineeships at three different
projects in three different institutions each on a dif-
ferent topic, and obtain some information about the
chosen projects, like a contact address, a descrip-
tion, etc. In the second task, subjects had to find
three conferences in the winter term and three con-
ferences in the summer term, each one on a differ-
ent topic and they had to obtain some information on
the chosen conferences such as deadline, place, date.
etc. Finally, the third task consisted of finding infor-
mation for writing a report on European language
technology in the last ten years. To this end, subjects
had to obtain quantitative information on patents, or-
ganizations, conferences, etc.
The Wizard was limited to very few types of re-
sponses. The main response was answering a ques-
tion. In addition, she would provide intermediate
information about the state of processing if the re-
trieval took too long. She could also make state-
ments about the contents of the database when it did
not contain the information asked for or when the
user appeared confused about the structure of the
domain. Finally, she could ask for clarification or
more specificity when the question could not be un-
derstood. Yet the Wizard was not allowed to take
the initiative by offering information that was not
explicitely asked for. Thus all actions of the Wiz-
ard were directly dependent on those of the user.
As a result we obtained a corpus of 33 logs (30
plus 3 pilot experiments) containing 125.534 words
in 2.534 turns, 1.174 of which are user turns.
2.2 Annotation scheme
The corpus received a multi-layer annotaton2 con-
sisting of five levels. The levels of turns and part-of-
speech were automatically annotated. The level of
turns records information about the speaker and time
1See http://www.lt-world.org.
2We employed the annotation tool MMAX2 developed at
EML Research, Heidelberg.
2
stamp. For the other levels - the questions level, the
utterances level, and the entities level - a specific an-
notation scheme was developed. For these, we only
explain the aspects relevant for the present study.
2.2.1 Questions
This level was conceived to keep track of the
questions asked by the user which correspond to
queries to the database. With the aim of annotating
thematic relatedness between questions we distin-
guished two main kinds of thematic relations: those
holding between a question and a previous ques-
tion, quest(ion)-to-quest(ion)-rel(ation), and those
holding between a question and a previous answer,
quest(ion)-to-answ(er)-rel(ation).
Quest-to-quest-rels can be of the following types:
• refinement if the current question asks for the
same type of entity as some previous question,
but the restricting conditions are different, ask-
ing, thus, for a subset, superset or disjoint set
of the same class.
(1) US: How many projects on language tech-
nologies are there right now?
US: How many have been done in the
past?
• theme-entity if the current question is about the
same entity as some previous question.
(2) US: Where will the conference take place?
US: What is the dead-line for applicants?
• theme-property if the current question asks for
the same property as the immediately preced-
ing question but for another entity.
(3) US: Dates of TALK project?
US: Dates of DEREKO?
• paraphrase if the question is the rephrasing of
some previous question.
• overlap if the content of a question is subsumed
by the content of some previous question.
We distinguish the following quest-to-answ-rels:
• refinement if the current question asks for a
subset of the entities given in the previous an-
swer.
(4) LT: 3810.
US: How many of them do research on
language technology?
• theme if the current question asks about an en-
tity first introduced in some previous answer.
(5) LT: Semaduct, ...
US: What language technology topics
does the Semaduct project investigate?
Although Chai and Jin (2004) only consider tran-
sitions among questions in dialogues about events,
most of our relations have a correspondence with
theirs. Refinement corresponds to their constraint
refinement, theme-property to their participant-shift,
and theme-entity to their topic exploration.
2.2.2 Utterances
Utterances are classified according to their
speech-act: question, answer, assertion, or request.
Our annotation of discourse structure is identical in
spirit to the one proposed by Ahrenberg et al. (1990).
A segment is opened with a user question to the
database and is closed with its corresponding an-
swer or an assertion by the system. Clarification
requests and their corresponding answers form seg-
ments which are embedded in other segments. Re-
quests to wait and assertions about the processing of
a question are also embedded in the segment opened
by the question.
Fragmentary utterances are annotated at this level.
We distinguish between fragments with a full lin-
guistic source, fragments with a partial source,
and fragments showing a certain analogy with the
source. The first group corresponds to fragments
which are structurally identical to the source and
can, thus, be resolved by substitution or extension.
(6) US: Are there any projects on spell checking in
Europe in the year 2006?
US: And in the year 2005?
Fragments with a partial source implicitly refer to
some entity previously introduced, but some infer-
ence must be done in order to resolve them.
(7) US: How is the contact for that project?
US: Homepage?
3
The last group is formed by fragments which show
some kind of parallelism with the source but which
cannot be resolved by substitution.
(8) US: Which conferences are offered in this win-
ter term in the subject of English language?
US: Any conferences concerning linguistics in
general?
2.2.3 Reference
We distinguish the following types of reference
to entities: identity or co-reference, subset/superset
and bridging.
Co-reference occurs when two or more expres-
sions denote the same entity. Within this group we
found the following types of implicit co-referring
expressions which involve different degrees of ex-
plicitness: elided NPs, anaphoric and deictic pro-
nouns, deictic NPs, and co-referent definite NPs.
Elided NPs are optional arguments, that is, they
don’t need to be in the surface-form of the sentence,
but are present in the semantic interpretation. In (9)
there is an anaphoric pronoun and an elided NP both
referring to the conference Speech TEK West 2006.
(9) US: The Speech TEK West 2006, when does it
take place?
LT: 2006-03-30 - 2006-04-01.
US: Until when can I hand in a paper [ ]?
Bridging is a definite description which refers to
an entity related to some entity in the focus of at-
tention. The resolution of bridging requires some
inference to be done in order to establish the con-
nection between the two entities. In example (2) in
subsection 2.2.1 there is an occurrence of bridging,
where the dead-line is meant to be the dead-line of
the conference currently under discussion.
Finally, subset/superset reference takes place
when a linguistic expression denotes a subset or su-
perset of the set of entities denoted by some previ-
ous linguistic expression. Subset/superset reference
is sometimes expressed through two interesting con-
textual phenomena: nominal ellipsis3, also called se-
mantic ellipsis, and one-NPs4. Nominal ellipsis oc-
curs within an NP and it is namely the noun what
3Note, however, that nominal ellipsis does not necessarily
always denote a subset, but sometimes it can denote a disjoint
set, or just lexical material which is omitted.
4One-NPs are a very rare in our corpus, so we are not con-
sidering them in the present study.
is missing and must be recovered from the context.
Here follows an example:
(10) US: Show me the three most important.
3 Thematic follow-up
When looking at the thematic relatedness of the
questions it’s striking how well structured the in-
teractions are regarding thematic relatedness. From
1047 queries to the database, 948 (90.54%) follow-
up on some previous question or answer, or both.
Only 99 questions (9.46%) open a new topic. 725
questions (69.25% of the total, 76.48% of the con-
nected questions) are related to other questions, 332
(31.71% of the total, 35.02% of the connected ques-
tions) are related to answers, and 109 (10.41% of the
total, 11.49% of the connected questions) are con-
nected to both questions and answers. These num-
bers don’t say much about how well structured the
discourse is, since the questions could be far away
from the questions or answers they are related to.
However, this is very seldom the case. In 60% of
the cases where the questions are thematically con-
nected, they immediately follow the question they
are related to, that is, the two questions are consecu-
tive5. In 16.56% of the cases the questions immedi-
ately follow the answer they are related to. 74.58%
of the questions, thus, immediately follow up the
question or/and answer they are thematically related
to6.
Table 1 shows the distribution of occurrences
and distances in segments for each of the rela-
tions described in subsection 2.2.1. We found that
the most frequent question-to-question relation is
theme-entity, followed by the question-to-answer re-
lation theme. As you can see, for all the relations ex-
cept theme, most occurrences are between very close
standing questions or questions and answers, most
of them holding between consecutive questions or
questions and answers. The occurrences of the re-
lation theme, however, are distributed along a wide
range of distances, 29.70% holding between ques-
tions and answers that are 2 and 14 turns away from
5By consecutive we mean that there is no intervening query
to the database between the two questions. This doesn’t imply
that there aren’t several intervening utterances and turns.
69 questions are consecutive to the question and answer they
are related to, respectively, that’s why the total percentage of
related consecutive questions is not 76.56%.
4
REF. Q. THEME E. Q. THEME P. Q. PARA. Q. OVERL. Q. REF. A. THEME A.
TOTAL 74 338 107 174 29 29 303
(7.80%) (35.65%) (11.29%) (18.35%) (3.06%) (3.06%) (31.96%)
1 SEGM. 88.73% 81.65% 100% 60.92% 78.57% 83.34% 46.39%
2 SEGM. 5.63% 1.86% 0% 8.09% 21.43% 13.33% 10.20%
Table 1: Occurrences of the different thematic relations
REL. / PHEN. THEME E. Q. THEME P. Q. THEME A. REF. Q. REF. A. CONNECTED TOTAL
FRAGMENT 53 (54.08%) 17 (16.32%) 3 (3.06%) 21 (21.42%) 0 97 (85.08%) 114
BRIDGING 40 (74.07%) 0 3 (5.55%) 1 (1.85%) 0 54 (58.69%) 92
DEFINITE NP 26 (78.78%) 0 4 (12.21%) 2 (6.10%) 0 33 (66%) 50
DEICTIC NP 19 (51.35%) 0 13 (35.13%) 2 (5.40%) 1 (2.70%) 37 (78.72%) 47
ANAPHORIC PRON. 13 (39.39%) 2 (6.06%) 10 (30.30%) 0 5 (15.15%) 33 (39.75%) 83
DEICTIC PRON. 2 (75%) 0 1 (25%) 0 0 3 (25%) 12
ELIDED NP 9 (69.23%) 0 2 (15.38%) 0 0 13 (61.90%) 21
NOMINAL ELLIPSIS 0 1 (7.69%) 6 (46.15%) 1 (7.69%) 5 (38.46%) 13 (81.25%) 16
Table 2: Contextual phenomena and the thematic relations holding between the questions containing them
and the questions or answers containing the antecedents.
each other. This is because often several entities
are retrieved with a single query and addressed later
on separately, obtaining all the information needed
about each of them before turning to the next one.
We found also quite long distances for paraphrases,
which means that the user probably forgot that he
had asked that question, since he could have also
scrolled back.
These particular distributions of thematic rela-
tions seem to be dependent on the nature of the
tasks. We found some differences across tasks: the
information gathering task elicited more refinement,
while the information browsing tasks gave rise to
more theme relations. It is possible that in an in-
teraction around an event or topic we may find ad-
ditional kinds of thematic relations and different
distributions. We also observed different strategies
among the subjects. The most common was to ask
everything about an entity before turning to the next
one, but some subjects preferred to ask about the
value of a property for all the entities under discus-
sion before turning to the next property.
4 Contextual phenomena: distances and
thematic relatedness
There are 1113 user utterances in our corpus, 409 of
which exhibit some kind of discourse phenomenon,
i.e., they are context-dependent in some way. This
amounts to 36.16% of the user utterances, a pro-
portion which is in the middle of those found in the
several corpora analyzed by Dahlb¨ack and J¨onsson
(1989)7. The amount of context-dependent user ut-
terances, as Dahlb¨ack and J¨onsson (1989) already
pointed out, as well as the distribution of the dif-
ferent relations among questions and answers ex-
plained above, may be dependent on the nature of
the task attempted in the dialogue.
Table 2 shows the distribution of the most fre-
quent thematic relations holding between the ques-
tions containing the contextual phenomena consid-
ered in our study and the questions or answers con-
taining their antecedents. The rightmost column
shows the number of occurrences of each of the con-
textual phenomena described in subsection 2.2.3.
The second column on the right shows the number
of occurrences in which the antecedent is located
in some previous segment and the question contain-
ing the contextual phenomenon is related through a
thematic relation to the question or answer contain-
ing the antecedent. The percentages shown for each
phenomenon are out of the total number of its oc-
currences. The remaining columns show frequen-
7They found a high variance according to the kind of task
carried out in the different dialogues. Dialogues from tasks
where there was the possibility to order something contained
a higher number of context-dependent user initiatives, up to
54.62%, while information browsing dialogues contained a
smaller number of context-dependent user initiatives, 16.95%
being the lowest amount found.
5
cies of co-occurrence for each of the phenomena and
thematic relations. The percentages shown for each
phenomenon are out of the total number of its con-
nected occurrences.
For the majority of investigated phenomena we
observe that most questions exhibiting them stand
in a thematic relation to the question or answer con-
taining the antecedent. Although there may be sev-
eral intermediate turns, the related questions are al-
most always consecutive, that is, the segment con-
taining the contextual phenomenon immediately fol-
lows the segment containing the antecedent. In the
remainder of the cases, the contextual phenomenon
and its antecedent are usually in the same segment.
However, this is not the case for deictic and
anaphoric pronouns. In most cases their antecedents
are in the same segment and even in the same utter-
ance or just one utterance away. This suggests that
pronouns are produced in a more local context than
other phenomena and their antecedents are first to be
looked for in the current segment.
For almost all the phenomena the most frequent
relation holding between the question containing
them and the question or answer containing the an-
tecedent is the question-to-question relation theme-
entity, followed by the question-to-answer relation
theme. This is not surprising, since we refer back to
entities because we keep speaking about them.
However, fragments and nominal ellipsis show a
different distribution. Fragments are related to their
sources through the question-to-question relations
theme-property and refinement, as well. Regarding
the distribution of relations across the three differ-
ent types of fragments we distinguish in our study,
we find that the relations refinement and theme-
property only hold between fragments with a full
source and fragments of type analogy, and their re-
spective sources. On the other hand, practically all
fragments with a partial-source stand in a theme-
entity relation to their source. Questions containing
nominal ellipsis are mostly related to the preceding
answer both through the relations theme and refine-
ment.
4.1 Antecedents beyond the boundaries of the
immediately preceding segment
As we have seen, the antecedents of more implicit
co-referring expressions, like pronouns, are very of-
ten in the same segment as the expressions. The
antecedents of less explicit co-referring expressions,
like deictic and definite NPs, are mostly in the im-
mediately preceding segment, but also often in the
same segment. About 50% are 2 utterances away,
20% between 3 and 5, although we find distances up
to 41 utterances for definite NPs.
However, there is a small number (11) of cases in
which the antecedents are found across the bound-
aries of the immediately preceding segment. This
poses a challenge to systems since the context
needed for recovering these antecedent is not as lo-
cal. The following example is a case of split an-
tecedents. The antecedent of the elided NP is to be
found across the two immediately preceding ques-
tions. Moreover, as you can see, the Wizard is not
sure about how to interpret the missing argument,
which can be because of the split antecedents, but
also because of the amount of time passed, and/or
because one of the answers is still missing, that is,
more than one segment is open at the same time.
(11) US: Which are the webpages for European
Joint Conferences on Theory and Practice
of Software and International Conference on
Linguistic Evidence?
LT: Please wait... (waiting time)
US: Which are the webpages for International
Joint Conference on Neural Networks and
Translating and the Computer 27?
LT: http://www.complang.ac, ... (1st answer)
US: Up to which date is it possible to send a
paper, an abstract [ ]?
LT: http://uwb.edu/ijcnn05/, ... (2nd answer)
LT: For which conference?
US: For all of the conferences I got the web-
pages.
In the following example the antecedent of the
definite NP is also to be found beyond the bound-
aries of the immediately preceding segment.
(12) US: What is the homepage of the project?
LT: http://dip.semanticweb.org
USER: What is the email address of Christoph
Bussler?
LT: The database does not contain this informa-
tion.
US: Where does the project take place?
6
Here the user asks about the email address of a per-
son who was previously introduced in the discourse
as the coordinator of the project under discussion
and then keeps on referring to the project with a def-
inite NP. The intervening question is somehow re-
lated to the project, but not directly. There is a topic
shift, as defined by Chai and Jin (2004), where the
main topic becomes an entity related to the entity the
preceding question was about. However, this topic
shift is only at a very local level, since the dialogue
participants keep on speaking about the project, that
is, the topic at a more general level keeps on being
the same. We can speak here of thematic nesting,
since the second question is about an entity intro-
duced in relation to the entity in focus of attention
in the first question, and the third question is again
about the same entity as the first. The project has not
completely left the focus, but has remained in sec-
ondary focus during the second segment, to become
again the main focus in the third segment. It seems
that as long as the entity to which the focus of atten-
tion has shifted is related to the entity previously in
focus of attention, the latter still also remains within
the focus of attention.
5 Conclusions
The possibility of using contextual phenomena is
given by certain types of thematic relatedness - espe-
cially theme-entity and theme, for co-reference and
bridging, and refinement, theme-entity and theme-
property, for fragments -, and contiguity of ques-
tions. As we have seen, the immediately preced-
ing segment is in most cases the upper limit of the
search space for the last reference to the entity, or
the elided material in fragments. The directions of
the search for antecedents, however, can vary de-
pending on the phenomena, since for more implicit
referring expressions antecedents are usually to be
found in the same segment, while for less implicit
referring expressions they are to be found in the pre-
ceding one.
These data are in accordance with what Ahren-
berg et al. (1990) predict in their model. Just to
consider the immediately preceding segment as the
upper limit of the search space for antecedents is
enough and, thus, no tracking of thematic relations
is needed to resolve discourse phenomena. How-
ever, there are occurrences of more explicit types
of co-reference expressions, where the antecedent
is beyond the immediately preceding segment. As
we have observed, in these cases the intervening
segment/s shift the focus of attention to an entity
(maybe provided in some previous answer) closely
related to the one in focus of attention in the pre-
ceding segment. It seems that as long as this rela-
tion exists, even if there are many segments in be-
tween8, the first entity remains in focus of attention
and can be referred to by an implicit deictic or defi-
nite NP without any additional retrieval cue. We can
speak of thematic nesting of segments, which seems
to be analogous to the intentional structure in task-
oriented dialogues as in (Grosz and Sidner, 1986),
also allowing for reference with implicit devices to
entities in the superordinate segments after the sub-
ordinated ones have been closed. It seems, thus, that
thematic structure, like the discourse goals, also im-
poses structure on the discourse.
These cases, although not numerous, suggest that
a more complex discourse structure is needed for
QA interactions than one simply based on the dis-
course goals. The local context is given by the dis-
course segments, which are determined by the dis-
course goals, but a less local context may encompass
several segments. As we have seen, reference with
implicit devices to entities in the less local context
is still possible. What seems to determine this less
local context is a unique theme, about which all the
segments encompassed by the context directly or in-
directly are. So, although it does not seem necessary
to track all the thematic transitions between the seg-
ments, it seems necessary to categorize the segments
as being about a particular more global theme.
In a system like the one we simulated, having spe-
cific tasks in mind and querying structured data, a
possible approach to model this extended context,
or focus of attention, would be in terms of frames.
Every time a new entity is addressed a new frame
is activated. The frame encompasses the entity it-
self and the properties holding of it and other enti-
ties, as well as those entities. This would already
allow us to successfully resolve bridging and frag-
ments with a partial source. If the focus of atten-
8We found up to five intervening segments, one of them be-
ing a subsegment.
7
tion then shifts to one of the related entities, the user
demanding particular information about it, then its
frame is activated, but the previous frame also re-
mains somehow active, although to a lesser degree.
As long as there is a connection between the enti-
ties being talked about and a frame is not explicitly
closed, by switching to speak about a different en-
tity of the same class, for example, frames remain
somehow active and implicit references will be ac-
commodated within the activation scope.
In principle, the closer the relation to the entity
currently in focus, the higher the degree of activation
of the related entities. Yet, there may be cases of
ambiguity, where only inferences about the goals of
the user may help to resolve the reference, as in (13):
(13) US: How is the contact for that project?
LT: daelem@uia.ua.ac.be
US: What is the institute?
LT: Centrum voor Nederlandse Taal en Spraak.
US: Homepage?
Here the property ”Homepage” could be asked about
the institution or the project, the institution being
more active. However, the Wizard interpreted it as
referring to the project without hesitation because
she knew that subjects were interested in projects,
not in organizations. In order to resolve the ambigu-
ity, we would need a system customized for tasks or
make inferences about the goals of the users based
on the kind of information they’ve been asking for.
Determining at which level of nesting some expres-
sion has to be interpreted may involve plan recogni-
tion.
However, for open domain systems not having a
knowledge-base with structured data it may be much
more difficult to keep track of the focus of attention
beyond the strictly local context. For other kinds
of interactions which don’t have such a structured
nature as our tasks, this may also be the case. For
example, in the information browsing tasks in (Kato
et al., 2004), there is not a global topic encompass-
ing the whole interaction, but the information needs
of the user are given by the information he is en-
countering as the interaction proceeds, that is, he is
browsing the information in a free way, without hav-
ing particular goals or particular pieces of informa-
tion he wants to obtain in mind. In such cases it
may be difficult to determine how long frames are
active if the nesting goes very far, as well as making
any inferences about the user’s plans. However, it
might also be the case, that in that kind of interac-
tions no implicit referring expressions are used be-
yond the segmental level, because there is no such
an extended context. In order to find out, a study
with interactive data should be carried out.
Acknowledgements
The research reported here has been conducted in
the projects QUETAL and COLLATE II funded by
the German Ministry for Education and Research,
grants no. 01IWC02 and 01INC02, respectively. We
are also grateful to Bonnie Webber for her helpful
comments on the contents of this paper.

References
Ahrenberg Lars, Dahlb¨ack Nils and Arne J¨onsson 1990.
Discourse representation and discourse management
for natural language interfaces. Proceeding of the
Second Nordic Conference on Text Comprehension in
Man and Machine, T¨aby, Sweeden, 1990.
Jaime G. Carbonell. 1983. Discourse pragmatics and
ellipsis resolution in task-oriented natural language
interfaces. Proceedings of the 21st annual meeting
on Association for Computational Linguistics, Cam-
bridge, Massachusetts, 1983
Chai Joyce Y. and Ron Jin. 2004. Discourse Status
for Context Questions. HLT-NAACL 2004 Workshop
on Pragmatics in Question Answering (HLT-NAACL
2004) Boston, MA, USA, May 3-7, 2004
Dahlb¨ack Nils and Arne J¨onsson. 1989. Empirical
Studies of Discourse Representations for Natural Lan-
guage Interfaces. Proceedings of the Fourth Confer-
ence of the European Chapter of the ACL (EACL’89),
Manchester.
Grosz Barbara and Candance Sidner. 1986. Attention,
Intention and the Structure of Discourse. Computa-
tional Linguistics 12(3): 175-204.
Kato Tsuneaki, Fukumoto Junichi, Masui Fumito and
Noriko Kando. 2004. Handling Information Access
Dialogue through QA Technologies - A novel chal-
lenge for open-domain question answering. HLT-
NAACL 2004 Workshop on Pragmatics in Question
Answering (HLT-NAACL 2004) Boston, MA, USA,
May 3-7, 2004.
Sun Mingyu and Joycie J. Chai. 2006. Towards Intel-
ligent QA Interfaces: Discourse Processing for Con-
text Questions. International Conference on Intelligent
User Interfaces, Sydney, Australia, January 2006.
