Ellipsis Resolution by Controlled Default Unification for Multi-modal and
Speech Dialog Systems
Michael Streit
German Research Center for
Artificial Intelligence - DFKI
Stuhlsatzenhausweg 3
Saarbr¨ucken
Germany
streit@dfki.de
HansUlrich Krieger
German Research Center for
Artificial Intelligence - DFKI
Stuhlsatzenhausweg 3
Saarbr¨ucken
Germany
krieger@dfki.de
Abstract
We present a default-unification-based approach to
ellipsis resolution that is based on experience in
long running multimodal dialog projects, where it
played an essential role in discourse processing. We
extend default unification to non-parallel structures,
which is important for speech and multimodal di-
alog systems. We introduce new control mecha-
nisms for ellipsis resolution by considering dialog
structure with respect to specification, variation and
results of tasks and combine this with the analysis
of relations between the information elements con-
tained in antecedent and elliptic structures.
1 Introduction
The application of default unification (Carpenter,
1993) or priority union (Kaplan, 1987; Calder,
1991) to discourse is attractive, because these re-
lated concepts meet the intuition that new informa-
tion extends, corrects, or modifies old information,
instead of deleting it, by keeping what is consis-
tent.1 The use of default unification as a means
for ellipsis resolution has been discussed in the first
half of the nineties (Pr¨ust et al., 1994; Grover et al.,
1994). Later, the discussion silted up, perhaps be-
cause the conditions on parallelism that have been
imposed occured to be too strong (cf. (Hobbs and
Kehler, 1997)).
1.1 Applications in Dialog Systems
Since this discussion, default unification-based el-
lipsis resolution has been applied in working sys-
tems of at least two projects, where it played an es-
sential role in discourse processing.
The first implementations have been provided in
the second half of the nineties at Siemens, where the
DIAMOD project developed a serial of prototypes
for multi-modal human machine dialog (cf. (Streit,
2001)).
1Priority union as introduced by Kaplan bears essentially
the same idea as Carpenter’s default unification.
The DIAMOD project realized applications for
appointment management and driver assistance, but
also for controlling machines and other technical
devices (Streit, 1999). The DIAMOD systems pro-
vided the user with a quite natural dialog, includ-
ing clarification and repair dialogs and even multi-
tasking (e,g,. the user could deal with different ap-
pointments in parallel).
Applied under appropriate conditions, default
unification turned out to be a robust and efficient
method for VP ellipsis and even NP ellipsis. It was
also successfully used to inherit information in situ-
ations without any syntactically recognizable ellip-
sis.
Later in the SmartKom project (Wahlster et al.,
2001), default unification and its application on
ellipsis resolution was reinvented under the label
overlay (Alexandersson and Becker, 2003). Over-
lay basically consists of priority unions applied to
frame-like semantic representations without consid-
ering reentrencies. This inhibits providing sloppy
readings of an ellipsis (without impairing the dialog
performed in the systems very much).
1.2 Problems with Default Unification
Default unification without additional control
shows an inherent tendency to over-accumulate in-
formation. Even worse, the method may accumulate
information that is semantically inconsistent (but
not recognized as such) or at least is practically ab-
surd. Such inconsistencies or absurdities typically
arise from dependencies between information ele-
ments that are not expressed (or not expressible) in
the type hierarchy, domain model, or ontology that
is underlying the default unification process. For
instance, it may occur that by introducing a name
of a new object, the address of an old object of the
same kind is inherited, which is pragmatically ab-
surd. Or a numeric date or time specification may
be wrongly combined with a deictic reference to an-
other date or time, which is semantically inconsis-
tent. On the other hand, the intrinsic parallelism
of default unification does inhibit the handling of
fragmentary and other non-parallel ellipsis, which
is very common in spoken dialog.
1.3 Problems with Ellipsis Resolution in
Information Seeking Dialog
In dialog systems that serve for browsing informa-
tion sources, it is significant that the user modifies
and varies her query, be it spontaneously or system-
atically. As is discussed mainly in section 5, re-
moving and inheriting old information are equally
important for ellipsis resolution in this type of in-
teraction. We want to notice, that the removal of
information is independent from the question if in-
formation has been grounded (Traum, 1994) or not.
Up to now, studies hardly consider these problems.
1.4 Overview
In this paper we will present a revised and fairly ex-
tended version of the methods developed in the DI-
AMOD project.2 We discuss two main problems.
On the one hand we show how default unification
can be applied to non-parallel or fragmentary struc-
tures, on the other hand we discuss dependencies of
ellipsis resolution from the structure of information
and tasks, that are rarely addressed in the literature.
Especially we discuss the following problems.
† The extension of default unification to frag-
ments.
† Control of ellipsis resolution by considering
dialog structure w.r.t. specification, variation
and results of tasks.
† Control of ellipsis resolution by considering
relations between old and new information.
† Handling of set-valued features in specifying
vs. varying dialog phases.
We will couch our approach in terms of typed
feature structures. Thereby features correspond to
slots, and types correspond to frames.
We assume feature structures being well-typed,
but not totally well-typed (Carpenter, 1992). This
means that for every type it is defined which fea-
tures it can have, and which types come into ques-
tion as values of these features. But it is not re-
quired that every possible feature is present or has
a value. In order to use feature structures for the
purpose of encoding semantic representations, we
need set-valued features (which are semantically in-
terpreted as conjunctions of its elements). For in-
stance, a movie may be described by a conjunction
2The quite simple but effective means for controlling de-
fault unification that have been introduced in the DIAMOD
systems have not been published yet.
of genres (e.g., crime and science fiction), and an
appointment usually has more than one participant.3
In this paper we will mainly refer to examples
taken from DIAMOD, but also make use of material
taken from SmartKom user input. We note here that
the methods described in this paper are not imple-
mented in SmartKom. We also consider examples
as they are discussed in the literature.
2 Default unification
Default unification is a method to inherit defeasible
(in our case old) information which does not contra-
dict strict (in our case new) information. As already
mentioned, the consistency criterion is to weak, but
the basic approach is useful. There are two forms of
default unification: credulous and skeptical default
unification. Credulous default unification tries to
maintain as much old information as possible. Due
to structure sharing, there are often different alter-
natives for achieving a maximal amount of old in-
formation. Skeptical default unification takes only
the information that is common to all credulous so-
lutions. We are interested in getting every maximal
solution, which correspond to the strict, sloppy or
mixed readings of ellipsis. By mixed readings we
mean readings that contain a strict reading in one
part, and a sloppy reading in another.
We follow the definition of credulous de-
fault unification provided by Carpenter (Carpenter,
1992). But we take the most general type as the top
element of the type lattice, while Carpenter takes it
as the bottom element.
If O is the old, defeasible information and N is
new, strict information, then the credulous default
unification of O and N is the unification of O’ with
N, where O’ is a minimal structure that subsumes O
such that O’ and N unify:
O >uc N = fO0uNjO0 w O minimal s.t. O0uN 6= ?g
The following example, shows how default
unification can be used in ellipsis resolution.
1. John revises his paper.
2
66
4
revise
AGENT 1 john
OBJECT
• paper
AUTHOR 1
‚
3
77
5
3The first case concerns a closed class of values. In prin-
ciple, this case could technically be solved without set-valued
features by introducing a highly differentiated (rather artificial)
type system with a type for every value combination. The sec-
ond case concerns an open class of entities that cannot be com-
piled in a type system.
2. And Bill does too.
• event-agentive
AGENT bill
‚
The analysis of these utterances is slightly
simplified. John would get a more complicated pre-
sentation with john being the value of the NAME
feature of the type person. The verb do is consid-
ered as being the most general verb with an agent.
We use here event-agentive as a supertype of activi-
ties.
In this example the types of the top nodes are on
a comparable “level”. By being on a comparable
level we mean that the top node of the one item is
a supertype of the top node type of the other item.
Notice that due to the well-typing condition, types
and features may not be mixed arbitrarily . For in-
stance, the most general type of the type hierarchy
(and many others too), cannot be combined with the
feature agent. Otherwise our level condition would
be meaningless.
We find two minimal upper bounds of (1.) that
unify with (2.).
(1’)
2
64
revise
AGENT john
OBJECT
• paper
AUTHOR john
‚
3
75
(1”)
2
66
4
revise
AGENT 1
OBJECT
• paper
AUTHOR 1
‚
3
77
5
We get by unifying (1’) with (2) the strict reading
(2’), while we get the sloppy reading (2”) by using
(1”).
(2’)
2
64
revise
AGENT bill
OBJECT
• paper
AUTHOR john
‚
3
75
(2”)
2
66
4
revise
AGENT 1 bill
OBJECT
• paper
AUTHOR 1 bill
‚
3
77
5
3 Default Unification on Substructures
While classical studies focus on parallism, the im-
portance of non-parallel and fragmentary ellipsis is
shown by empirical analysis of spoken dialog (cf.
(Fernandz and Ginzburg, 2002)). The focus of an
elliptic utterance often has no direct counterpart in
the antecedent, which makes Rooth’s matching con-
dition not directly applicable (cf. (Rooth, 1992),
(Hardt and Romero, 2001)). Grammatically re-
quired verbs (e.g., the semantically weak verb do)
may be omitted in dialog ellipsis. In German spo-
ken language, this is also possible in single and se-
quential utterances of one speaker.
We take an example from TALKY, which is the
appointment management multimodal dialog sys-
tem that was developed in the framework of DI-
AMOD. The reaction of the system to the first ut-
terance of the user is not necessarily important, be-
cause users often proceed with (3) without waiting
for the system’s answer (i.e., by barge in) or with-
out paying much attention to the system’s reaction
(in case of an experienced user).
1. USER: Ich m¨ochte einen Termin eintragen. (I
want to enter an appointment)
2. SYSTEM: presents a new appointment entry
3. USER: mit Schmid (with Schmid)
We achieve the following two representations of
the user’s utterances.
(1)
2
66
66
4
want
AGENT user
TOPIC
2
4
enter
AGENT system
OBJECT appointment
3
5
3
77
77
5
(3)
• thing-with-participant
PARTICIPANT schmid
‚
“Matching” cannot be achieved by assuming that
there is a hidden attitude connected to very utterance
which could be inserted.
Instead, we search for “matching” nodes with
comparable types before normal default unification
is applied: thing-with-participant unifies with ap-
pointment, which leads to:
2
66
66
64
want
AGENT user
TOPIC
2
64
enter
AGENT system
OBJECT
• appointment
PARTICIPANT schmid
‚
3
75
3
77
77
75
In principle, it is quite possible that thing-with-
participant describes a certain (collective) type of
agents. In this case, the processing would produce
an ambiguity. In the DIAMOD system as in many
other dialog systems, the agent role is usually re-
stricted to the user and to incarnations of the system.
It is not alway posible to find a matching type.
In this case we try to find paths that connect termi-
nal nodes of the antecedent structure with the top
node of the elliptic structure. It is important, that
such connection paths do not introduce new struc-
tures corresponding to verbal complements or sub-
ordinated sentences.
If no match is achieved we get simply the new
structure back, which is the normal result of apply-
ing default unification to non-matching structures.
4 Task Completion as a Barrier for
Elliptic reference
In the following example (taken from Talky), the
user performs her specification in a stepwise man-
ner by extensively using ellipsis.
1. USER: Ich m¨ochte am Montag ein Treffen ein-
tragen. (I want to enter a meeting at monday)
2. SYSTEM: Presents an empty appointment en-
try
3. USER: Im Bananensaal (In the “banana
room”)
4. SYSTEM: Presents appointment entry with
banana room
5. USER: Ich meine im Raum Leibniz (I mean in
room Leibniz)
6. SYSTEM: Presents appointment entry with
room Leibniz
7. USER: um sechs Uhr (at six o’clock)
8. SYSTEM: Presents appointment with room
and begin time 6 a.m
9. USER: abends (at the evening)
10. SYSTEM: Presents appointment with room
and begin time 6 p.m
Some information has been corrected or clarified,
but there was no information removed implicitly.
Locally, most steps could be considerd as a case of
fragmentary elaboration of the preceding utterances
(cf. (Schlangen and Lascarides, 2003)). But this
classification depends on more general properties of
the dialog. When the task is finished, the availabil-
ity of old information has changed:
1 USER: Bitte das Treffen jetzt eintragen.
(Please enter now the meeting
2 SYSTEM: Indicates that the meeting is stored
3 USER
a Bitte jetzt ein Treffen am Freitag eintra-
gen please enter a meeting at Friday now
b Und am Freitag. And at Friday!
c Am Freitag. At Friday!
In case of 3a, the the old information is removed.
With 3b we recognize that the activity (entering a
meeting in a schedule) is still available for being in-
herited elliptically, while further information, accu-
mulated before, is no longer relevant. If the user
wants to keep the more elements of the old informa-
tion, she has to use anaphoric references, e.g.,
4 Und dasselbe am Freitag (And the same at Fri-
day).
The elliptic reading in (3b) is very clear, (3c)
is rather an incomplete utterance that has to be
clarified. This is also quite different from the
specification phase of the meeting.
Task completion is a barrier for fragmentary elab-
oration. 4 After task completion, an elliptic re-
lation has to be be marked (e.g., by clue words as
und (and). Even then, ellipsis does not refer to the
whole information accumulated before, but rather to
the utterance that introduced the specification phase
of the preceding task.
5 Information Browsing
Typically, information request are answered after
every user input without a lengthly specification
phase. As in the case of elliptic specifications,
clarification dialog does not affect the elliptic rela-
tions between subsequent user queries. If the system
actively proposes an action, this will be different.
Browsing means to vary requests either because it is
not clear in advance which information is relevant,
how exactly it can be obtained, or because the user
wants to gather broad information in some area.
In browsing dialog, ellipsis is controlled by re-
lations between the informational content of the an-
tecedent and the elliptic utterance. According to our
remarks at the beginn of the section, we omit the re-
actions of the system in the subsequent examples.
By a group, we understand a collection of infor-
mation that is orthogonal to other information. By
4The reader may recognize a certain similarity of the con-
siderations in this section with the approach of (Grosz and Sid-
ner, 1986). An example: We restrict ourself to some remarks:
Grosz & Sidner focus on the segmentation of discourse along
the hierarchical structure of a task, while we focus on problems
concerning repetition (this section) and variation of tasks (next
section). Grosz & Sidner are mainly concerned with anaphoric
reference while we are concerned with ellipsis and related im-
plicit inheritance of information. In our approach, structural
relations between information is as much important as aspects
concerning the processing of tasks. Furthermore, we discuss
problems in relation to a special resolution mechanism, i.e., de-
fault unification.
Figure 1: Ontology for Searching Information about
Performances (simplified)
orthogonal we mean independent and not “compet-
ing”. For instance, we consider TIME, LOCATION
and CONTENT as basic groups of the information
that belongs to a performance. Independence is not
a sufficient criterion. Actor and genre are indepen-
dent, but as our examples may show there are con-
sidered as competing. We have no formal means to
recognize a group. The knowledge about groups has
to be provided.
We use the term information element (IE) of a
feature structure as follows: An IE consists of two
parts: a role path and a semantic content. Differing
from the usual definition of paths (Carpenter, 1992),
a role path is a sequence of alternating types and
features (T1F1...TnFn with Types Ti and Features
fi). The semantic content is expressed by the sub-
structure which is identified by applying the subse-
quence of the features of the role path (accordingly
to standard definition). Role paths can be translated
directly in an obvious way into feature structures.
We speak of an terminal information element
(TIE), if the substructure is a type without further
specification. A TIE is atomic, if its semantic con-
tent is atomic. We represent TIEs as extended role
paths by taking the type which expresses their se-
mantic content as last element of the path.
Two TIEs (or IEs) are of the same sort, if their
role path has a common prefix. Two TIEs are of the
same terminal sort, if their role paths are identical.
TIE1 is more general as TIE2 if TIE1 subsumes
TIE2. TIE1 subsumes TIE2 if the subsumption re-
lation holds between their translations to feature
structures. It will turn out that this definition is to
narrow and does not cover the intuitive meaning of
being more general.
The TIEs in elliptic expressions are usually less
specific or have a shorter role path than the TIEs
in the antecedent. Subsequently we assume that
the matching process (as described in former sec-
tions) has already been applied and that the TIEs of
the elliptic expression are extendend by appending
the role path from the root of the antecedent to the
matching node. Otherwise we could not correctly
determine if an IE is subsumed by another or if they
belong to the same group etc.
We only consider readings of elliptic expressions
that amount to a new request, ignoring other read-
ings of elliptic expressions, e.g., as positive or neg-
ative feedback.
1 USER: Welch filme laufen heute Abend in
Saarbr¨ucken? (Which movies are on today
evening at Saarbr¨ucken? )
2 USER: Welche Krimis kommen? / Krimis! /
(?)Und Krimis! (Which crime (movies) are
on? / Crime (movies) / And crime (movies))
3 USER:
a Welche Sience fiction filme laufen? / Und
science fiction? / Science fiction! (Which
science fiction movies are on? / And sci-
ence fiction? / science fiction)
b Sind Science fiction filme dabei? (Are
science fiction movies among them?)
In (2) the general information movies
(i.e., the TIE informationSearch:TOPIC:-
performance¡in¡cinema:CONTENT:movie) is
replaced by the coresponding concrete information
crime movies (i.e., the TIE informationSearch:-
TOPIC:performance¡in¡cinema:CONTENT:-
movie:GENRE:crime). All other information
belongs to different groups and is retained. In (3)
the information crime movies is replaced by infor-
mation of the same terminal sort. The specification
crime movies is deleted. GENRE is a set-valued
feature. Note that set-valued features act quite
differently depending on the context (information
browsing vs. task specification). If the information
crime should be retained, this has to be indicated,
e.g. by an anaphorical relation to the result of query
(2) as is done in (3b). The reading of (2) and (3)
is not affected by the form of the ellipsis, but the
strong indication of parallelism that is expressed
with ”Und Krimis” (”And crime (movies)”) seems
not acceptable due to the proper subsumption
relation between movies and movies with genre
crime.
1 USER: Welche Science fiction laufen heute
abend in Saarbr¨ucken?. (Which science fiction
(movies) are on today evening at Saarbr¨ucken
)
4 USER:
a Mit Bruce Willis? (With Bruce Willis?)
b Und mit Bruce Willis / Welche filme
mit Bruce Willis laufen (And with Bruce
Willis? / Which movies with Bruce Willis
are on)
In (4), the new information element Bruce Willis
does not belong to the same terminal sort as any el-
ement in the antecedent, but by contributing to the
specification of movies it belongs to the same group
as science fiction. It is a competing element of ’sci-
ence fiction’, and its effect on the information ele-
ment ’science fiction’ is a mixture of the effect of
elements of the same sort and elements of a differ-
ent group (as may be expected). 4b is an an accept-
able utterance in this context and it has the effect of
deleting the genre information, while 4a without ex-
plicit ellipsis indication could also count as adding
a specification.
1 USER: Welche Krimis kommen heute abend
in Saarbr¨ucken (Which crime (movies) are on
today evening at saarbr¨ucken?)
5 USER:
a Und in Saarlouis (And at Saarlouis?)
b Welche filme laufen in Saarlouis? (Which
movies are on at Saarlouis?)
In (5) the information Saarbr¨ucken is replaced by
an information of the same terminal sort. 5a has
the reading crime movies in Saarlouis. In 5b crime
movies is replaced by a more general informa-
tion. This is an indication that the specification
crime should be removed. But Welche filme (which
movies) has two other (less preferred) readings: an
anaphoric reading which (of those) movies are (also)
running at Saarlouis, or even an elliptic (or E-type)
reading which crime movies are on at Saarlouis.
That movie is more general than crime movie can
directly inferred from examining the ontology, i.e.
by subsumption.
1 USER: Welche Krimis kommen heute abend in
Saarbr¨ucken (Which crime (movies) are on?)
6 USER:
a Und im Scala (And at the Scala (movie
theater))
b Welche filme laufen im Scala (Which
movies are on at the Scala movie theater)
In (6), Saarbr¨ucken is replaced by a more concrete
information of the sort location. The Scala movie
theater is expected to be in Saarbr¨ucken except for
Scala is a aforementioned cinema in another town.
The readings are quite similar to (5). But there is
one difference: assume (1) gets an empty result.
Than (5a) is still appropriate while (6a) is quite odd.
(5b remains (slightly) ambiguous, while (6b) has
only one reading. The problem with these findings
is, that we cannot recognize by subsumption that
Scala is more specific than Saarbr¨ucken.
In information browsing, the relations between
the information elements contained in the an-
tecedent and the information elements provided by
the ellipsis expression are relevant for resolution.
Concrete Information Rule If the elliptic expres-
sion contains a more concrete TIE than the an-
tecedent, old specifications that belong to an-
other group are retained.
General Information Rule If the elliptic expres-
sion contains more general information than
the antecedent, then the general information
tends to be understood as deleting the corre-
sponding concrete Information. The more gen-
eral TIE introduces a choice points for default
unification. Default unification has to produce
a reading (usually the more likely one) that ac-
cepts general information elements as potential
barriers for default unification and removes old
information which is beyond the barriers.
Same Sort Rule If the elliptic expression contains
information of the same terminal sort, the old
information is deleted, even if the information
elements belong to a set-valued feature, except
it is made explicit that the feature should be
added.
Competing Information Rule If the elliptic ex-
pression indicates parallelism and contains
“competing” information of the same group,
but not the same terminal sort, the old infor-
mation is deleted. Otherwise, competing infor-
mation can be understood as adding a further
specification.
Negative Result Condition Ambiguous readings
are sensible for the result of the antecedent
query. Negative (empty) results excludes
readings that make the specification more
concrete.
We only consider relations between an antecedent
query and a subsequent elliptic query. We do not
discuss here relations that come into play if a longer
history is considered. The examples show, that de-
fault unification has to be controlled by relations be-
tween information elements.
6 Conclusion and Problems
We presented an approach for the resolution of non-
parallel ellipsis by default uni£cation, which is in-
herently a parallel method. We discussed the de-
pendence of ellipsis interpretation on the state of
the dialog in respect to task processing, but also
on relation between the informational content of
antecedent structures and elliptic structures, which
leads to a removal of inforamation, which is up to
now not considered in studies on ellipsis. We also
addressed the interplay of these dependencies with
indications of parallism that are customarily viewed
as the main factors of ellipsis interpretation. and We
demonstrated how these insights are realized by us-
ing default unification as efficient base processing.
A topic of further research is the relation of gen-
eral and concrete information. For instance, the on-
tology shown in figure 1 resembles the ontology
used in SmartKom. The location of a cinema is
specified by using a common format for addresses,
in which country and town are on the same level
and the name of the object not directly related to the
address. Formally (if groups are already defined)
these informations would considered as competing.
This would prevent the Scala movie theater to be
transferred to Saarlouis (in most cases the compe-
tition criterion would exclude this possibility, for a
certain type of elliptic expression it would recog-
nize an ambiguity). But the criterion would also
delete the specification that the Scala cinema is
in Saarbr¨ucken if the information element Scala is
introduced elliptical after asking a question about
Saarbr¨ucken. This kind of problems is not exclu-
sively a problem of locations.
† Wo l¨auft Matrix? (Where is Matrix on?)
† Western? / Wo laufen Western? (Where are
western (movies) on
The phrase “Western” shows no indications for par-
allelism, hence the competition criterion would in
this case accept the reading that the user is looking
for a western named Matrix.
As a practical solution, we introduced a rule that
comprises that names and other “identifiers” of indi-
viduals are considered as being more concrete than
any other information elements, but further explo-
ration of the problem is necessary.
Also, the study of larger pieces of dialog as con-
sidered here is an important topic of further re-
search.
At several occasions we noticed that anaphoric
relations interact with elliptic relation. The inter-
action of anaphors and ellipsis is another important
topic of research.

References
J. Alexandersson and T. Becker. 2003. The formal
foundations underlaying overlay. In Proceedings
of the Fifth International Workshop on Computa-
tional Semantics, Tilburg.
J. Calder. 1991. Some notes on priority unions. In
ACQILEX Workshop on Default Inheritance in
the Lexicon, Cambridge, England.
B. Carpenter, editor. 1992. The Logic of Typed Fea-
ture Structures. Cambridge University Press.
B. Carpenter. 1993. Skeptical and credulous de-
fault uni£cation with applicationns to templates
and inheritance. In T. Briscoe, A. Copestake, and
V. de Pavia, editors, Inheritance, Defaults and
the Lexicon, pages 13–37. Cambridge University
Press.
R. Fernandz and J. Ginzburg. 2002. A Corpus
Study of Non-sentential Utterances in Dialogue.
Tratement Automatique de Languages, 43(2).
J. Grosz and C.L. Sidner. 1986. Attention, inten-
tions, and the structure of discourse. Computa-
tional Linguistics, 12:175–204.
C. Grover, Ch. Brew, S. Manandhar, and M. Moens.
1994. Priority union and generalization in dis-
course grammars. In 32nd. Annual Meeting of the
Association for Computational Linguistics, pages
17 – 24, Las Cruces, NM. Association for Com-
putational Linguistics.
D. Hardt and Maribel Romero. 2001. Ellipsis and
the structure of discourse. In Sinn und Bedeutung
VI, Osnabr¨uck, Germany.
J.R. Hobbs and A. Kehler. 1997. A Theory of Par-
allelism and the Case of VP Ellipsis. In P. R. Co-
hen and W. Wahlster, editors, Proceedings of the
35th Annual Meeting of the ACL, pages 394–401.
Association for Computational Linguistics.
R. Kaplan. 1987. Three seductions of com-
putational psycholinguistics. In P. Whitelock,
H. Somers, P. Bennett, R. Johnson, and
M. McGee Wood, editors, Linguistic Theory and
Computer Applications, pages 149–188. Aca-
demic Press.
H. Pr¨ust, R. Scha, and M. van den Berg. 1994. Dis-
course grammar and verb phrase anaphora. Lin-
guistics and Philosophy, 17:261–327.
M. Rooth. 1992. A theory of focus interpreattion.
Natural Language Semantics, 1:75–116.
D. Schlangen and A. Lascarides. 2003. The
Interpretation of Non-Sentential Utterances in
Dialogue. In Proceedings of the 4th SIGdial
Workshop on Discourse and Dialogue, Sapporo,
Japan.
M. Streit. 1999. The interaction of speech, deixis,
and graphics in the multimodal of£ce agent talky.
In P. Dalsgaar, CH. Lee, P. Heisterkamp, and
R. Cole, editors, Proceedings of the ESCA Tuto-
rial and Research Workshop on Interactive Dia-
log in Multi-Modal Systems.
M. Streit. 2001. Why are multimodal systems so
dif£cult to build. In Harry Bunt and Robbert-Jan
Beun, editors, Cooperative Multimodal Commu-
nication, volume 2155 of Lecture Notes in Com-
puter Science. Springer.
D. Traum. 1994. A Computational Theory
of Grounding in Natural Language Conversa-
tion. Ph.D. thesis, Computer Science Dept., U.
Rochester.
W. Wahlster, N. Reithinger, and A. Blocher. 2001.
Multimodal Communication with a Life-Like
Character. In Proc. of the 7th Proc. European
Conf. on Speech Communication and Technol-
ogy.
