Rhetorical structure in dialog* 
Amanda Stent 
Computer Science Department 
University of Rochester 
Rochester;:N'Y 14~27 ...... 
stent~cs, rochester, edu 
Abstract 
In this paper we report on several issues arising 
out of a first attempt to annotate task-oriented spo- 
ken dialog for rhetorical structure using Rhetorical 
Structure Theory. We discuss an annotation scheme 
we are developing to resolve the difficulties we have 
encountered. 
1 Introduction 
In this paper we report on several issues arising out 
of a first attempt to annotate complex task-oriented 
spoken dialog for rhetorical structure using Rhetor- 
ical Structure Theory (RST): 
o Relations needed (section 3.1) 
o Identification of minimal units for annotation 
(section 3.2.2) 
• Dialog coverage (section 3.2.3) 
• Overlap due to the subject-matter/presenta- 
tional relation distinction (section 3.3) 
We discuss how we are dealing with these issues in an 
annotation scheme for argumentation acts in dialog 
that we are developing. 
2 Previous work 
We are engaged in tim construction and inlplemen- 
ration of a theory of content-planning for complex, 
mixed-initiative task-oriented dialogs based on cor- 
pus analysis, for use in dialog systems such as the 
TRIPS system (Allen et al., 2000) 1 . Our basic 
premise is that a conversational agent should be able 
to produce whatever a human can produce in simi- 
lar discourse situations, and that if we can explain 
why a human produced a particular contribution, 
" This work w~ supported by ONR research grant N00014- 
95-l-1088, U.S. Air Force/Rome Labs research contract no. 
F30602-95-1-0025, NSF research grant no. IRI-9623665 and 
Columbia University/NSF research grant no. OPG: 1307. We 
would like to thank the armuymous reviewers and l)r. Jason 
Eisner for their helpful comments on earlier drafts of this 
paper. 
IWe are using the Monroe corpus (Stent, 2000), with ref- 
erence t.o the TRAINS corpus (Heeman and Allen, 1995) and 
the HCRC Mapta~sk corpus (Anderson et al., 1991). 
247 
we can program a conversational agent to produce 
something similar. Therefore, in examining our di- 
alogs the question we must answer is "Why did this 
speaker produce this?". 
RST is a descriptive theory of hierarchical struc- 
ture in discourse that identifies functional relation- 
ships between discourse parts based on the inten- 
tions behind their production (Mann and Thomp- 
son, 1987). It has been used in content plan- 
ning systems for text (effectively text monolog) (e.g. 
(Cawsey, 1993), (How, 1993), (Moore and Paris, 
1993)). It has not yet been used much in content 
planning for spoken dialog. 
Because the dialogs we are examining are task- 
oriented, they are hierarchically structured and so 
provide a natural place to use RST. In fact, in or- 
der to uncover the full structure behind discourse 
contributions, it is necessary for us to use a model 
of rhetorical structure. Certain dialog contribu- 
tions are explained by the speaker's rhetorical goals, 
rather than by task goals. In example 1, utterance 3 
is justification for utterance 1 but does not directly 
contribute to completing the task. 
Example 1 
A 1 They can't fix that power line at five 
ninety and East 
B 2 \Veil it 
A 3 Because you got to fix the tree first 
The details of how to apply RST to spoken dialog 
are unclear. If we mark rhetorical structure only 
within individual turns (as has generally been the 
case in annotations of text dialog, e.g. (Moser et 
al., 1996),(Cawsey, 1993)), we miss the structure in 
contributions like example 1 or example 2. There 
is also tile question of how to handle dialog-specific 
behaviors: grounding utterances and back-channels 
(utterances that maintain the comnmnication), and 
al)andoned or interrupted utterances. 
Example 2 (simplified) 
A 1 Bus C at irondequoit broke down. 
B 2 Before it. even got started? 
A 3 ~'eah, but we convinced some people to 
loan US sonic vans. 
Initial annotation 
Dialog-specific Subtypes of Elaboration Other 
Comment Particularize, Generalize Comparison 
Correction Instantiate Counter-expectation 
Cue i Exemplify Agent, Role 
Argumentation acts 
• Question~response.:: 
Proposal-accept 
Greeting-ack. 
New manual 
Subtypes of Elaboration Schemas 
~Set~member ......... Joke, List 
Process-step Make-plan 
Object-attribute Describe-situation 
Figure 1: Examples of other relations 
In our first attempt to annotate, we removed 
abandoned utterances, back-channels, and simple 
acknowledgments such as "Okay". We used utter- 
ances as minimal units; utterances were segmented 
using prosodic and syntactic cues and speaker 
changes (see 3.2.2). We did occasionally split an ut- 
terance into two units if it consisted of two phrases or 
clauses separated by a cue word such as "because". 
Two annotators, working separately, marked one 
complete dialog using Michael O'Donnell's RST an- 
notation tool (1997). They used the set of relations 
in (Mann and Thompson, 1987), and some addi- 
tional relations specific to dialog or to our domain. 
Examples of the additional relations are given in fig- 
ure 1. When we compared the results, the tree struc- 
tures obtained were similar, but the relation labels 
were very different, and in neither case was the entire 
dialog covered. Also, the annotators found structure 
not covered by the relations given. As a result, we 
stopped the annotation project and started develop- 
ing an annotation scheme that would retain rhetor- 
ical relations while dealing with the difficulties we 
had encountered. The rest of this paper describes 
this new annotation scheme. An example of the type 
of analysis we are looking for appears in figure 3. 
3 Issues and proposals 
The issues we encountered fall into three areas, 
which we will examine in turn: issues related to in- 
dividual relations, dialog-specific issues, and issues 
related to the well-known presentational/subject- 
matter distinction in RST. 
3.1 Relations 
The key in any annotation project is to have a set 
of tags that are mutually exclusive, descriptive, and 
give a useful distinction between different behaviors. 
The set of relations we used failed this test with 
respect to our corpus. 
As in earlier work (Moore and Paris. 1992). our 
annotators found some of the relations ambiguous. 
In particular, the differences between the motivate 
and justify relations and between the elaboration and 
motivation relations were unclear (partly because 
248 
we did not distinguish between presentational and 
subject-matter relations). 
Some of the relations we used overlapped. The 
elaboration relation is too broad; in some sections 
of our dialogs almost every utterance is an elabora- 
tion of the first one, but the utterances cover a wide 
variety of different types of elaborations. Anticipat- 
ing this, we had given the annotators several more 
specific relations (see figure 1), but we also allowed 
them to use the elaboration tag in case a type of elab- 
oration arose for which there was no subtype. As a 
result of the overlap, use of the elaboration tag was 
inconsistent. The joint relation is also too broad. 
Other relations were never used, although one an- 
notator went on to look at several more dialogs. In 
short, the set of relation-tags we used did not effec- 
tively partition the set of relations we saw. 
In our annotation scheme, we are taking several 
steps to define relations more clearly, reduce over- 
lap, and eliminate too-broad relations. Instead of 
giving annotators an semi-ordered set of relations 
with their definitions, we are giving them decision 
trees, with questions they can use to clarify the dis- 
tinctions between relations at each point (figure 2). 
The annotators did not find the relation definitions 
in (Mann and Thompson, 1987) particularly help- 
ful, but we are including simplified definitions, and 
annotators are instructed to test against the defini- 
tions before labeling any relation. We are including 
several examples with each definition, so that anno- 
tators can obtain an intuitive understanding of how 
the relations appear. Finally, we are providing any 
useful discourse cues that signal the existence of a 
relation. 
We are eliminating relations that overlap with 
others. Where a relation appears to cover a variety 
of different phenomena, as in the case of elaboration, 
we are using more specific relations instead. We are 
eliminating the joint relation, as it gives no help- 
ful information from a content-planning perspective 
and annotators are tempted to over-use it. 
One of the criticisms of RST is that there is an 
infinite set of relations (Grosz and Sidner, 1986). 
The goal is to arrive at a mutually-exclusive, clearly- 
defined set of relations with" discriminatory power in 
each domain, so we expect that for each new do- 
main, it may be necessary to start with an initial 
set of high-level relations selected from different cat- 
egories, examine a small set of texts or dialogs in that 
domain, and then revise the set of relations by mak- 
• ing relevant high-leve! .relations more.specific.._We.. 
used this process to develop our annotation scheme. 
In the manual we include instructions for moving to 
new domains. Our examples come from a variety of 
domains and types of discourse, to add generality. 
3.2 Dialog-specific issues 
3.2.1 Dialog-specific relations, schemas and 
conversational games 
Task-oriented dialog is a complex behavior, involv- 
ing two participants, each with their own beliefs 
and intentions, in a collaborative effort to inter- 
act to solve some problem. There is a whole set 
of behaviors related to maintaining the collabora- 
tion and synchronizing beliefs that does not arise 
in monolog \[(Clark, 1996), (Traum and Hinkelman, 
1992)\]. These include answering questions, agree- 
ing to proposals, and simply acknowledging that the 
othe r participant has spoken. 
In example 3, utterance 3 provides motivation for 
utterance 1. However, A would not have produced 
utterance 3 without B's question. If we simply mark 
a motivation relation between utterances 1 and 3 we 
will be losing dialog coverage, the spans involved 
in the relation will not be adjacent, and we will be 
ignoring the important relationship between utter- 
ances 2 and 3. A better analysis would be to mark 
a question-answer relation between utterances 2 and 
3, and a motivation relation between utterance 1 and 
the unit consisting of utterances 2 and 3. 
Example 3 
A 1 Then they're going to have to 
basically wait 
B 2 Why? 
A 3 Because the roads have to be fixed before 
electrical lines can be fixed 
The question-answer relation is not in Mann and 
Thompson's original list of relations 2. It is an "ad- 
jacency pair ''a, and is a type of conversational game 
(ClarM 1996). Adjacency pairs, like other relations, 
are functional relationships between parts of dis- 
course, but. they are specific to multi-party discourse. 
In our annotation scheme, we include relations for 
different kinds of adjacency pairs (figure 1). We have 
2They do. however, include requests for information in the 
solutionhood relation 
aAn adjacency pair is a pair of utterances, the first of which 
imposes a cognitive preference for the second, e.g. question- 
answer, proposabaeeept. 
249 
1. In this set of spans, is the speaker attempting to 
affect the hearer's: 
o belief- go to question 2 
• attitude - go to question 3 
o ability to perform an action - enablemen~ 
...... .2.. Is:t:he_speaker..tryi.ug..to.inccrease.the.hearer'.s belief 
in some fact, or enable the hearer to better under- 
stand some fact? 
• Belief- evidence 
® Understanding- background 
3 .... 
Figure 2: Partial decision tree for presentational re- 
lations, expressed as a list of questions 
tentatively categorized adjacency pairs with subject- 
matter relations, although they may eventually be- 
come a third category of relation. 
Some of these relations are bi-nuclear. For in- 
stance, although usually the answer is the only part 
required for discourse coherence, at times both ques- 
tion and answer may be needed, as in example 4. 
Example 4 
A 1 And the last one was at the where 
on the loop? 
B 2 Four ninety. 
It would seem that these relations can only apply 
at the lowest levels of an RST analysis, with a dif- 
ferent speaker for each span. However, example 5, 
in which turns 2-7 are the answer to the question in 
utterance 1, shows that this is not the case. 
Example 5 (slightly simplified) 
A 1 What's "close"? 
B 2 "Close". Um I don't know. I I'm pretty 
sure that 
A 3 So Mount Hope and Highland would be. 
B 4 Yeah. 
A 5 Well what about like 252 and 383'? 
B 6 It says "next". 
A 7 Oka~v. So I guess it has to be adjacent. 
It might seem that .the simplest approach would 
be to annotate adjacency pairs between turns, and 
mark other rhetorical relations only within turns. 
However, we have found many instances of rhetori- 
cal relations, or even units (section 3.2.2), spanning 
turns. The two examples below illustrate a cross- 
speaker elaboration and a cross-speaker sequence re- 
lation. 
Example 6 
A i So that.takes care of the ill guy 
and the handicapped guy. 
B 2 " Okay 
B 3 And that takes two hours. 
A 1 
A 2 
B 3 
A 4 
B 5 
B 6 
Summary 
Make-fla~ \ (6) 
...... Object-attribute, Enablement 
,/ \ 
$olutionhood, Quesffon-answer (nun~er), 
Motivation , / ,~ 
, / \ (3) Assert-ack. 
(~) (2) , / \ 
(4} (5} 
We have to send buses to the Lake. 
There are people there to evacuate. 
How many are we sending? 
Two. 
Okay. 
So 1 ambulance to Pittsford and 2 
buses to the Lake. 
Figure 3: Sample analysis of part of a constructed 
dialog. Nuclei are marked with *; non-RST relations 
are in italics. 
Example 7 
A 1 So they can ta- to- take out the power. 
B 2 And then we have to wait ... 
With a model of adjacency pairs,_we can-now han- 
dle grounding acts such as acknowledgments. If an 
utterance is clearly a back-channel or abandoned, 
annotators are instructed to so mark it and leave it 
out of further annotation. 
RST in its original formulation does not cover en- 
veloping or parallel structures or conventional forms. 
However, even in task-oriented dialogs speakers oc- 
casionally tell jokes. Furthermore, there are fixed, 
structural patterns in dialog, such as form-filling 
behaviors. These are frequently domain-specific, 
and resemble schemas \[(McKeown, 1985), (Cawsey, 
1993)\]. While it may be possible to give an RST 
analysis for some of these, it is more accurate to 
identify, what is actually going on. Our annotation 
scheme includes four of these, make-plan, describe- 
situation, list and joke. It also includes an adjacency 
pair for greetings, a conventional form. 
An annotated dialog extract illustrating most of 
these issues is shownin figure 3. 
3.2.2 Identifying and ordering units 
In spoken dialog, both participants often speak at 
once, or one speaker may complete what another 
speaker says, as in examples 8 and 9. 
Example 8 (+'s mark overlapping speech) 
:\ 1 And + he's done + with that at one thirty 
B 2 + Okay + 
Example 9 
A 1 So it'll take them 
B 2 Two nmre hours 
250 
Our original use of utterances as minimal units 
splits a cross-turn completion from the utterance it 
completes (example 9), and says nothing about how 
to order units when one overlaps with another. We 
have altered our segmentation rules to take care of 
these difficulties. Our definition is that a minimal 
~unit.must.be one~.~f tthe following,~.with:eadier pos- 
sibilities taking precedence over later ones: 
1. A syntactic phrase separated from the immedi- 
ately prior phrase by a cue word such as "be- 
cause" or "since" 
2. A syntactically complete clause 
3. A stretch of continuous speech ended by a 
pause, a prosodic boundary or a change of 
speaker 
One unit will be considered to succeed another if 
it starts after the other. 
This means that the standard segmentation of a 
dialog into utterances may have to be modified for 
the purposes of an RST analysis, although a segmen- 
tation into utterances and one into minimal units 
will be very similar. Annotators will start with a 
dialog segmented into turns and utterances, and are 
encouraged to re-segment as needed. 
3.2.3 Dialog coverage 
When one gets higher in the tree resulting from an 
RST annotation, the spans typically begin to fol- 
low the task structure or the experimental structure. 
In the Monroe corpus, usually one partner tells the 
other about the task, then the two collaborate to 
solve it, and finally one partner summarizes the so- 
lution (following the experimental structure). In the 
TRAINS corpus usually one subtask in the plan is 
discussed at a time (following the task structure). 
Given the length and complexity of a typical dia- 
log, it may not be possible to achieve complete cov- 
erage, even with our expanded relation set and the 
use of schemas. If we can identify useful sub-dialogs 
or can associate parts of a dialog with parts of the 
task, finding annotations for each part may suffice. 
For our domain, we have established heuristics about 
when an annotator can stop trying to achieve cover- 
age. An annotator can stop when: 
o The top level of the annotation tree has one 
relation label covering the whole dialog. 
o The structure between the spans at the top level 
is identical to the task structure. 
* Tim structure between the spans at the top 
level is identical to a domain-dependent or 
expe.riment-dependent schema. 
o There is consensus between annotators that no 
more relations can be marked. 
3.3 The subject-matter/presentational 
relation distinction 
The relations in RST fall into two classes. Subject- 
matter relations such as summary are intended to 
be recognized by the hearer. Presentational rela- 
tions such as motivation are supposed to "increase 
some inclination" in the hearex~ ~LtCh .as. the. inclina- 
tion to act (Mann and Thompson, 1987). As Moore 
and associates have explained in (1992) and (1993), 
while the intentions of the speaker are adequately 
represented in the case of presentational relations 
by the relations themselves, in the case of subject- 
matter relations the intentions of the speaker may 
vary. Furthermore, these two types of relations ac- 
tually come from different levels of relationship be- 
tween discourse elements: the informational level 
(subject-matter relations), and the intentional level 
(presentational relations). RST conflates these two 
levels. 
Mann and Thompson said that, in the case where 
a presentational relation and a subject-matter re- 
lation were both applicable, the subject-matter re- 
lation should take precedence. However, we would 
like to have information about both levels when pos- 
sible. In our annotation scheme the presentational 
relations are split from the subject-matter relations 
and annotators are instructed to consider for each 
set of spans whether there is a subject-matter rela- 
tion, and also whether there is a presentational rela- 
tion. If there are two relations, both are marked. If 
one covers a slightly different span than the other, 
at the next level of annotation the span that seems 
more appropriate is used. 
In the following example, utterance 3 is justifica- 
tion (presentational) for utterance 1, but it is also 
in a non-volitional cause (subject-matter) relation- 
ship with utterance 1. The annotator would be in- 
structed to label both relations. 
Example 10 (slightly simplified) 
A 1 I can't find the Rochester airport 
B 2 + I- it's + 
A 3 + I think I have + a disability with maps 
We would also like more information, at times, 
about the subject matter in the spans of a relation. 
The relation between a "When" question and an- 
swer is question-answer, as is that between a "Why" 
question and answer; but the first question-answer 
forms part of an elaboration and the second forms 
part of a justification or motivation. In our ammta- 
tion scheme, we supply a list of content types, such 
as time. location and number. The annotator adds 
the content type in I)arentheses after the relation tag 
when required. This means that the annotator may 
have to mark three items for a given set of spans: 'the 
presentational relation (if any), the subject-matter 
relation, and the content type (if required). We find 
25t 
this approach preferable to expanding the set of re- 
lations to include, for instance, temporal-question- 
answer and spatial-question, answer. Cawsey used a 
similar method in (1993). 
4 Current and future work 
• -"-We-:havean :amaotation ~manuat"that weare"refming " 
using TRAINS-93 dialogs 4. Shortly, we will begin 
annotating the Monroe corpus with the new manual 
and different annotators. We will also annotate a 
few dialogs from a different corpus (e.g. Maptask) 
to ensure generality. We plan to use the results of 
our annotation in the construction (ongoing) of new 
generation components for the TRIPS system at the 
University of Rochester (Allen et al., 2000). 
5 Related Work 
In recent years there has been much research on 
annotation schemes for dialog. Traum and Hinkel- 
man outline four levels of "conversational acts" in 
(1992). "Argumentation acts", including rhetorical 
relations, form the top level, but this level is not de- 
scribed in detail. DAMSL (Core and Allen, 1997) in- 
cludes speech acts and some grounding acts, but not 
rhetorical relations. The HCRC Maptask project an- 
notation scheme includes adjacency pairs, but not 
rhetorical relations (Carletta et al., 1996). 
The COCONUT project annotation manual al- 
lows the annotator to mark individual utter- 
ances as elaboration, and segments as summary, 
act:condition, act:consequence or otherinfo (DiEu- 
genio et al., 1998). This annotation scheme does 
not treat rhetorical structure separately from other 
types of dialog behavior. We have observed enough 
structure in the corpora we have looked at to jus- 
tify treating rhetorical structure as a separate, im- 
portant phenomenon. For instance, in a DAMSL- 
tagged set of 8 dialogs in our corpus, 40% of the 
utterances were statements, and many of these ap- 
peared in sequences of statements. The relationships 
between many of these statements are unclear with- 
out a model of rhetorical structure. 
In (1999), Nakatani and Traum describe a hierar- 
chical annotation of dialog for I-units, based on the 
.. domination and satisfaction-precedence relations of 
(Grosz and Sidner, 1986). Other researchers have 
shown that Grosz and Sidner's model of discourse 
structure (GST) and RST are similar in many re- 
spects \[(Moser and Moore, 1996), (Marcu, 1999)\]. 
However, RST provides more specific relations than 
GST, and this is useful for content planning. As 
well as helping to specify generation goals, content 
and ordering constraints, the rhetorical information 
is needed in case the system has to explain what it. 
has said. 
4A rough draft is available from the author. 
RDA is an annotation scheme for identifying 
rhetorical structure in explanatory texts in the 
SHERLOCK domain (Moser et al., 1996). We follow 
RDA in requiring annotators to consider both in- 
tentional and informational relations. However, be- 
cause of the dialog issues previously described, RDA 
is not sufficient for dialog. 
Marcu uses discourse-cuesto"automa~ically un- 
cover rhetorical relations in text (1997). Much of 
this work is applicable to the problem of uncovering 
rhetorical relations in dialog; however, many cues 
in dialog are prosodic and it is not yet possible to 
obtain accurate information about prosodic cues au- 
tomatically. 
6 Conclusions 
We have examined several issues arising from a first 
attempt to annotate spoken dialog for rhetorical 
structure. We have proposed ways of dealing with 
each of these issues in an annotation scheme we are 
developing. Much future work is certainly needed 
in this area; we hope that the results of our annota- 
tion may form a quantitative baseline for comparison 
with future work. 

References 
J. Allen, D. Byron, M. Dzikovska, G. Ferguson, 
L. Galescu, and A. Stent. 2000. An architecture 
for a generic dialogue shell, upcoming in the Nat- 
ural Language Engineering Journal special issue 
on Best Practices in Spoken Language Dialogue 
Systems Engineering. 
A. Anderson, M. Bader, E. Bard, E. Boyle, G. Do- 
herty, S. Garrod, S. Isard, J. Kowtko, J. MeAl- 
lister, J. Miller, C. Sotillo, H. Thompson, and 
R. Weinert. 1991. The HCRC Maptask corpus. 
Language and Speech, 34:351-366. 
J. Carletta, A. Isard, S. Isard, J. Kowtko, 
and G. Doherty-Sneddon. 1996. HCRC dia- 
log structure coding manual. Technical Report 
HCRC/TR-82, HCRC, Edinburgh University. 
A. Cawsey. 1993. Planning interactive explanations. 
International Journal of Man-Machine Studies, 
38:169-199. 
H. Clark. 1996. Using Language. Cambridge Uni- 
versity Press. 
M. Core and J. Allen. 1997. Coding dialogs with the 
DAMSL annotation scheme. In AAAI Fall Sym- 
posium on Communicative Action in Humans and 
Machines, pages 28-35, November. 
B. DiEugenio, P. Jordan. and L. Pylkkiinen. 1998. 
The COCONUT project: Dialogue annotation 
manual. Technical Report 98-I, ISP, University 
of Pittsburgh. 
B. Gross- and C. Sidner. 1986. Attention, inten- 
tions, and the structure of discourse. Computa- 
tional Linguistics, 12(3). 
252 
P. Heeman and J. Allen. 1995. The TRAINS-93 
dialogs. Technical Report Trains TN 94-2, Com- 
puter Science Dept., U. Rochester, March. 
E. Hovy. 1993. Automated discourse generation us- 
ing discourse structure relations. Artificial Intel- 
ligence, 63(1-2):341-385. 
W.. Mann_and S. Thompsom 19877. Rhetorical struc- 
ture theow: a theory of text organisation. In 
L. Polanyi, editor, The Structure of Discourse. 
Ablex, Norwood, NJ. 
D. Marcu. 1997. The rhetorical parsing, sum- 
marization, and generation of natural  
texts. Technical Report CSRG-371, Department 
of Computer Science, University of Toronto. 
D. Marcu. 1999. A formal and computational 
synthesis of Grosz and Sidner's and Mann and 
Thompson's theories. In The Workshop on Levels 
of Representation in Discourse, Edinburgh, Scot- 
land. 
K. McKeown. 1985. Text Generation: Using Dis- 
course Strategies and Focus Constraints to Gener- 
ate Natural Language Text. Cambridge University 
Press, Cambridge. 
J. Moore and C. Paris. 1992. Exploiting user feed- 
back to compensate for the unreliability of user 
models. UMUAI, 2(4):331-365. 
J. D. Moore and C. L. Paris. 1993. Planning text 
for advisory dialogues: Capturing intentional and 
rhetorical information. Computational Linguis- 
tics, 19(4):651-695. 
J. Moore and M. Pollack. 1992. A problem for RST: 
The need for multi-level discourse analysis. Com- 
putational Linguistics, 18(4):537-544. 
M. G. Moser and J. D. Moore. 1996. Toward a 
synthesis of two accounts of discourse structure. 
Computational Linguistics, 22(3):409-420. 
M. Moser, J. Moore, and E. Glendening. 1996. 
Instructions for coding explanations: Identifying 
segments, relations and minimal units. Technical 
Report 96-17, University of Pittsburgh. Depart- 
ment of Computer Science. 
C. Nakatani and D. Traum. 1999. Coding discourse 
structure in dialogue (version 1.0). Technical Re- 
port UMIACS-TR-99-03, University, of Maryland. 
Michael O'Donnell. 1997. RST-Tool: An RST 
analysis tool. In Proceedings of the 6th Eu- 
ropean Workshop on Natural Language Gener- 
ation, Gerhard-Mercator University, Duisburg, 
Germany. 
A. Stent. 2000. The Monroe corpus. Technical Re- 
port TR728/TN99-2, University of Rochester. 
D. Traum and E. Hinkehnan. 1992. Conversation 
acts in task-oriented spoken dialogue. Computa- 
tional Intelli.qenee, 8(3):575--599. 
