Anchoring a Lexicalized Tree-Adjoining Grammar for Discourse 
Bonnie Lynn Webber and Aravind K. Joshi 
Department of Computer & Information Science 
University of Pennsylvania 
Philadelphia PA USA 19104-6389 
\[bonnie, joshi\]~central.cis.upenn.edu 
Abstract 
We here explore a "fully" lexicalized Tree-Adjoining 
Grammar for discourse that takes the basic ele- 
ments of a (monologic) discourse to be not simply 
clauses, but larger structures that are anchored on 
variously realized discourse cues. This link with 
intra-sentential grammar suggests an account for dif- 
ferent patterns of discourse cues, while the differ- 
ent structures and operations suggest three sepa- 
rate sources for elements of discourse meaning: (1) a 
compositional semantics tied to the basic trees and 
operations; (2) a presuppositional semantics carried 
by cue phrases that freely adjoin to trees; and (3) 
general inference, that draws additional, defeasible 
conclusions that flesh out what is conveyed compo- 
sitionally. 
1 Introduction 
In the past few years, researchers interested in ac- 
counting for how elements combine in a discourse, 
have taken to using the adjoining operation found 
in Tree-Adjoining Grammar (TAG) (Gardent, 1994; 
Gardent, 1997; Polanyi and van den Berg, 1996; 
Schilder, 1997; van den Berg, 1996; Webber, 1991). 
More recently, Cristea and Webber (1997) have ar- 
gued that a Tree-Adjoining Grammar for discourse 
would also need the substitution operation found in 
a lexicalized TAG (Schabes, 1990). Here we move 
further and explore a fully lexicalized TAG for dis- 
course, allowing us to examine how the insights of 
lexicalized grammars - that the basic elements of a 
clause are not simply words, but structures that re- 
flect a word's role and syntactic/semantic scope - 
carry over to discourse. We show how this suggests 
explanations for such phenomena as the following: 
• that arguments of a coherence relation can be 
stretched "long distance" by intervening mate- 
rial; 
• that multiple discourse cues can appear in a sin- 
gle sentence or even a single clause; 
• that when discourse cues appear in the middle 
of clauses, they contribute to coherence in more 
specific ways; 
• that coherence relations can vary in how and 
when they are realized lexically. 
One way of understanding the current work is 
that it extrapolates from lexically-based views of 
how structure and meaning are associated within a 
sentence to how aspects of discourse structure and 
meaning might be associated in similar ways. While 
the idea that discourse-level mechanisms might re- 
semble intra-sentential mechanisms has long been 
an undercurrent within discourse research, we have 
come to believe that the framework of lexicalized 
grammar can be effecti'.'ely used to demonstrate the 
validity of this intuition. While we present the ideas 
in terms of one well-known formalism - Lexicalized 
TAG - other lexicalized formalisms such as CCG 
(Steedman, 1996b) might prove equally useful for 
expressing the same theoretical insights and imple- 
menting them for discourse generation and/or inter- 
prey ation. 
A superficial reading of the current proposal might 
suggest that it ismerely a simple embedding of RST 
(Mann and Thompson, 1988) in TAG. That would 
be incorrect. First, the primary feature of a fully 
lexicalized TAG is that each elementary tree in the 
grammar has an anchor that indexes the tree and 
defines its syntactic/semantic scope. Here, we posit 
a set of inilial (non-recursive) trees, whose anchor is 
a discourse cue. Structurally, some initial trees re- 
semble the nucleus-satellite structures of RST, and 
some, its joint schema. But the resemblence is only 
superficial, as initial trees have a purely composi- 
tional semantics that makes no assumptions about 
what the speaker is trying to do. I 
Secondly, there is a single auzil~ary tree whose 
semantics corresponds simply to continuing the de- 
scription conveyed by the structure to which it is 
adjoined. Any additionalinferences that a listener 
draws from the resulting adjacency are defeasible, 
1 The LTAG formalism itself allows an elementary tree to 
be associated with a meaning that is not compositional with 
respect to its sub-parts. This is used, for example, for asso- 
ciating meaning with syntactically-flexible idioms. However, 
we have not found the need to exploit this possibility for dis- 
course, though we leave open the possibility. 
86 
and may be cancelled or corrected by material in the 
subsequent discourse. Our proposal thus factors the 
combinability of elementary discourse clauses from 
inferences that may then be drawn, thus providing 
a tool for sorting out different semantic processes in 
discourse, instead of lumping them into a single cat- 
egory. Many of these inferences have been given the 
status of discourse relations in RST. However, we 
argue in Section 2.3 that one can gain from distin- 
guishing what is derived compositionally from what 
is derived inferentially. 
Thirdly, there are auxiliary trees for other dis- 
course cues, that can adjoin to either initial trees 
or auxiliary trees. These discourse cues contribute 
meaning (and coherence) through their presupposi- 
tions or assertions or both. They can thereby serve 
to constrain the range of inferences that a listener 
might draw when a description is extended, limit- 
ing them to ones compatible with the contribution 
of the discourse cue. Similarly, a discourse cue ad- 
joined to an initial tree can either further specify the 
compositional meaning of the related units or con- 
strain how that initial tree can be used in extending 
another description. This will explain how several 
discourse cues can appear in the same sentence or 
even the same clause, each contributing to either 
the compositional or presuppositional semantics of 
the discourse (Section 2.2). 
This is still a "work in progress", with many open 
questions. However, it may still pique the interest 
of two historically distinct groups: it may stimu- 
late people working on syntax to look beyond the 
clause for phenomena familiar to them within it, 
while it may help people working on discourse to 
ground their claims and insights in more traditional 
varieties of linguistic formalisms. 
2 Elements of a Lexicalized TAG for 
Discourse 
A lexicalized TAG begins with the notion of a lexi- 
cal anchor, which can have one or more associated 
tree structures. For example, the verb likes anchors 
one tree corresponding to John likes apples, another 
corresponding to the topicalized construction Ap- 
ples John likes, and a third corresponding to the 
passive construction Apples are liked by John. All 
in all, there is a tree for each minimal syntactic 
construction in which likes can appear, all sharing 
the same predicate-argument structure. This syn- 
tactic/semantic encapsulation is possible because of 
the extended domain of locality of LTAG. Trees in 
such a tree family may differ in being licensed by dif- 
ferent states of the discourse (i.e., information struc. 
ture (Steedman, 1996a)). 
A lexicalized TAG contains two kinds of elemen- 
tary trees: initial (non-recursive) trees that reflect 
basic functor-argument dependencies and auxiliary 
trees that introduce recursion and allow elementary 
trees to be modified and/or elaborated. In our Iex- 
icalized discourse TAG, we have so far found the 
need to posit only two types of initial tree families 
(Section 2.1) and two types of auxiliary trees (Sec- 
tion 2.2). While the resulting grammar is thus very 
simple - only one type, only binary predicates- it 
so far appears expressively adequate. 
2.1 Initial Trees 
Subordinate conjunctions are one major class of dis- 
course cues, and clause-level LTAG already provides 
an account of subordinate clauses with overt sub- 
ordinate-conjunctions. Its "verb-centric" account 
in (XTAG-Group, 1995) appropriately treats sub- 
ordinate clauses as adjuncts - i.e., auxiliary trees. 
However, from a discourse perspective, we treat two 
clauses connected by a subordinate conjunction as 
an initial tree whose compositional semantics re- 
flect the subordinate conjunction as predicate (or 
"functor") and the clauses as arguments. There is 
an initial tree for each minimal structural pattern 
of main clause and subordinate clause, including 
those shown in Figure 1. All such trees share the 
same predicate (functor) argument structure. As in 
clause-level tree families, each pattern may have dif- 
ferent preconditions on its use that reflect the cur- 
rent state of the discourse (i.e., information struc- 
ture). For example, it has been noted that a "when" 
clause in initial position presupposes that the sit- 
uation described therein is in the heater's discourse 
model (or can be so accommodated), while a "when" 
clause coming after the main clause is not so con- 
strained. 
In Section 3, we discuss reasons for taking the lex- 
ical anchors of these initial trees to be feature struc- 
tures that may correspond to one or more subordi- 
nate conjunctions such as "if" and "when". Here we 
just take them to be specific lexical items. 
Now one reason for taking something to be an ini- 
tial tree is that it has local dependencies that can 
be stretched long-distance. For example, the depen- 
dency between apples and likes in both John likes 
apples and apples John likes is localized in all the 
trees for likes. It can be stretched, however, long- 
distance as in Apples..Bill thinks John may like. In 
(Cristea and Webber i997), we have shown that 
the same long-distanc,~ stretching of dependencies 
occurs with both subordinate clauses (Ex. 1) and 
parallel constructions (Ex. 2) - e.g. 
1. a. Although John is very generous, 
b. giving money to whoever asks, 
c. when you actually need it, 
d. you'll see that he's a bugger to find. 
2. a. On the one hand, John is very generous. 
b. For example, suppose you needed some 
money. 
87 
S @~ NP 
\[\] ~ ~//X VP 
subconj (a) subconj (b) \[ (c) 
subconj 
(d) 
Figure 1: Initial Trees (a-c) belong to the tree family for a subordinate conjunction. The symbol I indicates 
a substitution site, while \[ \] stands for a particular subordinate conjunction and its feature structure. (d) is 
the initial tree for a parallel construction. 
c. You would just have to ask for it. 
d. On the other hand, he's a bugger to find. 
Thus here we also posit an initial tree for paral- 
lel structures (Figure ld). Since there are different 
ways in which entities are taken to be parallel, we 
currently assume a different initial tree for contras~ 
("on the one hand" ... "on the other hand" ... ), dis- 
junction ("either" ... "or" ... ), addition ("not only" 
... "but also" ...), and concession ("admittedly" 
..."but" ...). Such trees have a pair of anchors 
with two main properties. 
The first is that their lexical realization seems 
optional. In contrastive cases, a medial anchor 
such as "on the other hand" often appears lexical- 
ized without an initial phrase such as "on the one 
hand". In fact, there are more cases of this in the 
Brown Corpus than of the two appearing together. 
Also optional is the realization of the initial anchor 
in disjunction (omitting "either"), addition (omit- 
ting "not only"), and concession (omitting "admit- 
tedly"). But we have recently noted cases where 
only the initial anchor is realized lexically but not 
the medial anchor, although this is less common: 
Not only have they \[Rauschenberg's blueprints\] 
survived. The process of their creation was 
recorded by Life magazine in April 1951. (New 
York Review of Books, 6 November 1997, p.8) 
The second property is that the medial anchor ap- 
pears realizable in multiple ways. Cristea and Web- 
bet (1997) report that, of the eleven instances of "on 
the one hand" found in the Brown Corpus, four have 
their contrasting item cued by something other than 
"on the other (hand)" - including "but" and "at the 
same time" : 
3. On the one hand, the Public Health Service de- 
clared as recently as October 26 that present 
radiation levels resulting from the Soviet shots 
"do not warrant undue public concern" or 
any action to limit the intake of radioactive 
substances by individuals or large population 
groups anywhere in the Aj. But the PHS con- 
ceded .... (cb21) 
4. Brooklyn College students have an ambivalent 
attitude toward their school. On the one hand, 
there is a sense of not having moved beyond 
the ambiance of their high school. This is par- 
ticularly acute for those who attended Mid- 
wood High School directly across the street 
from Brooklyn College. They have a sense of 
marginality at being denied that special badge 
of status, the out-of-town school. At the same 
time. there is a good deal of self-congratulation 
at attending a good college ... (cf25) 
Other examples occur with "on the other extreme" 
and "at the other extreme" - cf. 
5. On the one hand we have the "All you have to 
do is buy it" brigade who seem to think the only 
problem is that we haven't gone and "done it". 
On the other extreme there are groups who 
think if it has been explored theoretically then 
it's been done. 
In Section 3, we will argue that both these properties 
can be accommodated by treating the lexical anchors 
of these initial trees as feature structures. 
2.2 Auxiliary Trees 
Discourse cues other than subordinate conjunctions 
are either adverbs (adverbial phrases) or conjunc- 
tions. In XTAG (1995), adverbials are handled as 
simple auxiliary tree s (Figure 2a-b). We do the same 
here, associating each cue with a feature structure 
that indicates its semantic properties for reasons to 
be discussed in Section 3. Semantically, such auxil- 
iary trees can be used to elaborate or clarify the dis- 
course relation holding between two discourse units. 
This may result in the phenomenon of there being 
more than one discourse cue in a sentence, as in 
6. Stephen Brook, in Class: Knowing your place 
in modern Britain, begins promisingly with 
88 
z /s, vP} \[\] 
- . \[\] 
cue (a) (b) cue 
(A 
cue 
Figure 2: Auxiliary Trees. (a) and (b) are auxiliary trees in the tree family for adverbial discourse cues, 
which serve to modify or constrain the relation holding between discourse units. Trees in the family may be 
rooted on S or VP. The symbol * indicates the foot node of the auxiliary tree. (c) is the auxiliary tree for 
basic elaboration. 
the proposition that "class distinction and 
class consciousness - they are both with 
us".... Brook then, however, runs into trou- 
ble because he feels obliged to provide a theory 
7. Although the episodic construction of the 
book often makes it difficult to follow, it nev- 
ertheless makes devastating reading. 
We will discuss the semantics of such examples 
shortly, and also the conjunctions "so" and "but". 
As noted in Section 1, an auxiliary tree (here 
shown in Figure 2c) is used to adjoin to a structure 
and continue the description of the entity (object, 
event, situation, state, etc.) that it conveys. Such 
a tree would be used in the derivation of a simple 
discourse such as: 
8. John went to the zoo. He took his cell phone 
with him. 
Here, the foot node of the auxiliary tree in Figure 2c 
would be adjoined to the root of the tree for the first 
clause, and its substitution site filled by the tree for 
the second. The tree's anchor may have no lexi- 
cal realization (as here, between main clauses), or it 
may be realized by "and" (as in embedded clauses 
- e.g. "Fred believes that John went to the zoo and 
that he took his cell phone with him"). The compo- 
sitional meaning associated with adjoining this tree 
is simply that the meaning of the second clause con- 
tinues the description of the same entity as the first. 
Other aspects of meaning - such as there being a 
causal connection or temporal relation between its 
sub-parts, or an evidential relation between them - 
would be derived inferentially, and hence possibly be 
found inconsistent, given the subsequent discourse. 
When an adverbial discourse cue is adjoined to a 
clause, it can constrain how the clause can be inter- 
preted as continuing the already-started description 
- for example, 
9. John went to the zoo. However, he took his 
cell phone with him. 
Following Knott and others (1996), we take the se- 
mantics of such discourse cues to be presupposi- 
tional. For example, according to Knott, "however" 
presupposes the existence of a (shared) defeasible 
rule, some or all of whose antecedents are licensed 
by the previous discourse, but which fails to hold ei- 
ther because the clause so marked contradicts either 
the conclusion or an antecedent. In Example (9), 
the defensible rule might be something like 
When people go to the zoo, they leave their 
work behind. 
So the clause marked by "however" in (9) both con- 
tinues the description of the event of John's going to 
the zoo (compositional semantics) and conveys that 
the above rule fails to hold because its conclusion is 
contradicted (presuppositional semantics). 
Of course, since these relation-modifying auxiliary 
trees are adverbials, they can, at least in English, be 
adjoined elsewhere in the structure, not just at the 
anchor - e.g. 
10. Cracked and broken plastic tableware will at- 
tract germs, so it should be thrown away, never 
mended. Plastic furniture and toys. however, 
can be repaired successfully with the appropri- 
ate adhesive. 
We speculate that such medially-occuring discourse 
cues (of which we are acquiring a growing corpus of 
naturally-occuring examples) occur at the boundary 
between a sentence's topic or theme (i.e., the ques- 
tion under discussion) and its comment or rheme 
(i.e., the contribution made towards that question) 
(Steedman, 1996a). There are then three possibili- 
ties: 
• The cue merely makes the boundary explicit, 
while its presuppositional semantics remains 
tied to the proposition as a whole; 
• The presuppositional semantics of the cue 
is grounded in whichever informational unit 
(theme or theme) occurs to its left; 
• The presuppositional semantics of the cue is 
grounded in the theme (wherever it occurs), 
specifying how the theme links to the discourse 
(i.e., how it is the question under discussion). 
89 
Deciding among these alternatives requires more 
time for thought and analysis of both constructed 
and such "naturally-occuring" examples as 
1I. A soldering iron is a much more specialized tool, 
which you will rarely need. If the occasion 
does arise when you need to solder two pieces of 
metal together, however, choose a large elec- 
tric soldering iron with a tapered head. 
and Examples (6) and (7) above. In (11), the sub- 
ordinate clause itself is the theme. Such examples 
as (7) and (11) call into question RST's assump- 
tion that satellites, which these subordinate clauses 
would be taken to be, can be omitted without a great 
change in meaning to a discourse. These certainly 
cannot. 
Another open question (but more of a technical 
detail) is the appropriate handling of conjunctions 
such as "so" and "but". On the one hand, their se- 
mantics can best be seen as presuppositional - pre- 
supposing a defensible rule grounded in the previous 
discourse that succeeds in the case of "so" and fails 
in the case of "but" (Knott, 1996). On the other 
hand, they can only occur in the same position as 
"'and", which we treat as a possible lexical realiza- 
tion of the anchor of the description-extending aux- 
iliary tree, but which is not presuppositional. It is 
not. yet clear to us which is the more appropriate 
way to treat them. 
2.3 Compositional vs. Inferential 
Semantics 
One consequence of this approach is that clauses 
linked by an explicit subordinate conjunction have 
a different structural analysis than do clauses that 
are simply adjacent. This might appear problematic 
because the perceived meaning of such discourses is 
usually the same. For example, 
12. The City Council refused the women a permit 
because they feared violence. 
13. The City Council refused the women a permit. 
They feared violence. 
In our approach, (12) derives from the initial tree 
given in Figure 3a, while (13) derives from adjoining 
an auxiliary tree (Figure 3b) to the tree for the first 
clause and substituting the tree for the second clause 
at I. Herein lies the difference between the two: In 
(12), the causal connection is derived composition- 
ally, while in (13), one infers from the second utter- 
ance continuing the description started in the first, 
that the speaker intends the situation described in 
the second utterance to serve as an explanation for 
that described in the first. Thus, the causal connec- 
tion is defensible in (13) but not in (12). This can 
be seen by trying to continue each with "But that. 
wasn't the reason for their refusal." The extended 
version of (12) seems ill-formed, while the extended 
version of (13) seems perfectly coherent. 
Another reason for distinguishing a limited com- 
positional semantics from an open inferential seman- 
tics is illustrated by the following example: 
14. My car won't start. It may be out of gas. 
An RST analysis would simply decide what rela- 
tion held between the two clauses - perhaps non- 
volitional cause. However, non-volitional cause does 
not capture the different modal status of the two 
clauses, which in turn affects the modal status of 
the perceived relation: it is the car's possibly being 
out of gas that is possibly the cause of its not start- 
ing. We believe it is more systematic to just decide 
what description is being continued (here, the one 
begun in the first clause) and then derive further 
inferences that reflect the different modal status of 
the two clauses. That the above inference is defen- 
sible can be seen by continuing the discourse in (14) 
with "But that's not a possible reason for its not 
starting". 
2.4 Brief Example 
Here we illustrate our approach by considering Ex- 
ample 9 (repeated below) in more detail. 
9. John went to the zoo. However, he took his cell 
phone with him. 
Three types of elements participate in the analysis: 
(1) the syntactic analyses (trees) of the two clauses 
("John went to the zoo, "he took his cell phone with 
him") labelled a and fl in in Figure 4., along with 
their respective meanings (call them Pl(j) and P2(J)); 
(2) the auxiliary tree for the discourse cue "how- 
ever", labelled 7, along with its feature structure; 
and (3) the description-extending auxiliary tree la- 
belled 6. 
As the derivation in Figure 4 (below the arrow) 
shows, 3' adjoins at the root of a, fl substitutes into 
6, and ~5 adjoins at the root of ~. The semantics is 
as described earlier in Section 2.2. 
90 
3 Cue Phrases as Feature Structures 
(Knott, 1996) 
Earlier we noted that there was benefit to be gained 
from taking the anchors of elementary trees to be 
feature structures into which discourse cues (whose 
semantics was also in terms of feature structures) 
could substitute. Here we briefly argue why we be- 
lieve this is so. 
First, in viewing discourse cues in terms of feature 
structures, we are following recent work by Knott 
(1996; 1996). Knott's study of the substitutabil- 
ity patterns of discourse cues reveals that their four 
common patterns - synonymy, exclusivity, hyper- 
nymy/hyponymy and contingent substitutability - 
because 0 
± 
sem = pl(j) 
± 
sem = p20) 
Figure 3: Trees used in the derivation of Ex. 12 and Ex. 13 
however 7 '.' \[ \] \[\] 
.2 
'),' however sem = p20) 5" 
Figure 4: Derivation of Example 9 
can, by assuming inheritance (that, except for con- 
tingent substitutability, a substitution pattern that 
holds for a discourse cue also holds for all its hy- 
ponyms), follow from interpreting cues in feature- 
theoretic terms: 
• If cue a is synonymous with cue 3, they signal 
the same values of the same features. 
• If a is exclusive with /3, they signal different 
values of at least one feature. 
• If a is a hypernym of ~3,/3 signals all the features 
that o~ signals, as well as some features for which 
a is undefined. 
• If a and ,3 are contingently substitutable, c~ and 
fl signal some of the same features, but a is also 
defined for a feature for which /3 is undefined 
and /3 is defined for a feature for which a is 
undefined. 
Drawing on the extensive literature devoted to in- 
dividual cue phrases, Knott provided semantics for 
some of these features in terms of preconditions on 
their use and/or their communicative effects. 
Following Knott in treating discourse cues in 
terms of feature structures, it also appears benefi- 
cial to treat tree anchors as feature structures as 
well, distinct from those of discourse cues. 
The reason for treating the anchor of subordinate 
clause initial trees as feature structures is one of 
representational efficiency: we can posit fewer such 
trees if we take their anchors to be features struc- 
tures that allow the (possibly contingent ) substitu- 
tion of any subordinate conjunction with a compat- 
ible feature structure. For example, we can have 
one tree whose anchor has the feature restricted- 
situation, that can be realized as either "if" or 
"when" in some texts, but only "when" in others 
-- e.g.: 
15. Emergency parking regulations are in force 
\[when, if\] more than six inches of snow has 
fallen. 
16. I found 30 new messages had arrived \[when, *if\] 
I logged on this morning. 
(Knott and Mellish, 1996) distinguish "if" and 
"when" by their different values for the feature 
modal status: "when" has the value actual, while "if" 
has the value hypothetical. One can therefore say 
that other semantic features in Ex. I6 conflict with 
the value hypothetical, only allowing "when". (N.B. 
One could also take "'when" as being unmarked for 
modal status, its hypothetical reason begin synony- 
mous with "whichever". The conflict with "if" in 
Ex. 16 would still follow.) 
The argument for treating the pair of anchors of 
parallel structures as feature structures follows from 
the variability in the realization of the medial anchor 
noted in Section 2.1. One way to account for this is 
that the anchor has features separate from those of 
the discourse cues. Any cue can then be used to 
realize the anchor, as long as it is either 
• less specific than the anchor, as in Ex. 3 - "but" 
has few features in Knott's taxonomy; 
• more specific than the anchor, as in Ex. 5 - 
"on the other extreme", although it does not 
appear in Knott's taxonomy, intuitively appears 
to mean more than just "side". 
• partially overlapping with the anchor, as in 
Ex. 4 - "at the same time" has temporal 
91 
features, but does not seem intrinsically con- 
trastive. This corresponds to Knott's concept 
of contingent substitutability. 
It also appears as if the clause/discourse within the 
scope of an anchor can either reinforce its features 
(as in Ex. 17 below) or convey features of the anchor 
when it is not itself realized lexically, as in Ex. 18: 
17. On the one hand, according to Fred, John is 
very generous, On the other hand, according 
to everyone else, he will only give if he sees an 
angle. 
18. According to Fred, John is very generous. Ac- 
cording to everyone else, he will only give if he 
sees an angle. 
But this part of our work is more speculative and 
the subject of needed future work. 
4 Summary 
One way of seeing a grammar for discourse is as 
a story grammar - i.e., a semantic grammar with 
components marked for the role they play in the 
story or some sub-part. Alternatively, a discourse 
grammar can, like a sentence-level grammar, merely 
specify how structural units fit together and how 
the semantics of the whole would be derived. This 
is one such grammar. While previous authors have 
adopted only certain aspects of TAG or LTAG, here 
we have explored the possibility of a "fully" lexi- 
calized TAG for discourse, which allows to examine 
how the basic insights of a lexicalized grammar carry 
over to discourse. 
Our proposal allows us to construct a smooth 
bridge between syntactic clauses and discourse 
clauses, each anchored on a lexical item (at times 
empty but always carrying the appropriate features). 
It also allows us to factor out three separate sources 
for elements of discourse meaning, thus providing a 
tool for sorting out different processes in discourse 
and modeling them individually. As such. we believe 
the approach provides some new insights and tools 
for investigating discourse structure and discourse 
relations. 
Acknowledgements 
Our thanks to Mary Dalrymple, Christy Doran, 
Claire Gardent, Laura Kallmeyer, Alistair Knott, 
Matthew Stone and Mark Steedman for their invalu- 
able comments and suggestions. An earlier draft of 
this paper was presented at the Workshop on Un- 
derspecification, Bad Teinach, Germany May 1998. 

References 
Dan Cristea and Bonnie Webber. 1997. Ex- 
pectations in incremental discourse processing. 
In Proceedings of the 35 th Annual Meeting of 
the Association for Computational Linguistics 
(ACL97/EACL97), pages 88-95, Madrid, Spain. 
Morgan Kaufmann, Palo Alto CA. 
Claire Gardent. 1994. Discourse multiple depen- 
dencies. Technical Report ILLC Report LP-94-18, 
University of Amsterdam. 
Claire Gardent. 1997. Discourse tree adjoining 
grammars. Claus report nr.89, University of the 
Saarlandes, Saarbriicken. 
Alistair Knott and Chris Mellish. 1996. A feature- 
based account of the relations signalled by sen- 
tence and clause connectives. Language and 
Speech, 39(2-3):143-183. 
Alistair Knott. 1996. A Data-driven Methodol- 
ogy for Motivating a Set of Coherence Relations. 
Ph.D. thesis, Department of Artificial Intelligence, 
University of Edinburgh. 
William Mann and Sandra Thompson. 1988. 
Rhetorical structure theory: Toward a functional 
theory of text organization. Text, 8(3):243-281. 
Livia Polanyi and Martin H. van den Berg. 1996. 
Discourse structure and discourse interpretation. 
In P. Dekker and M. Stokhof, editors, Proceedings 
of the Tenth Amsterdam Colloquium, pages 113- 
131, University of Amsterdam. 
Yves Schabes. 1990. Mathematical and Computa- 
tional Aspects of Lexicalized Grammars. Ph.D. 
thesis, Department of Computer and Information 
Science, University of Pennsylvania. 
Frank Schilder. 1997. Tree discourse grammar, or 
how to get attached to a discourse. In Proceedings 
of the Tilburg Conference on Formal Semantics, 
Tilburg, Netherlands, January. 
Mark Steedman. 1996a. Representating discourse 
information for spoken dialogue generation. In 
Proceedings of the Second International Sympo- 
sium on Spoken Dialogue, pages 89-92, Philadel- 
phia PA, October. 
Mark Steedman. 1996b. Surface Structure and In- 
terpretation. MIT Press, Cambridge MA. 
Martin H. van den Berg. 1996. Discourse grammar 
and dynamic logic. In P. Dekker and M. Stokhof, 
editors, Proceedings of the Tenth Amsterdam Col- 
loquium, pages 93-111, ILLC/Department of Phi- 
losophy, University of Amsterdam. 
Bonnie Webber. 1991. Structure and ostension in 
the interpretation of discourse deixis. Natural 
Language and Cognitive Processes, 6(2):107-135.. 
The XTAG-Group. 1995. A Lexicalized Tree Ad- 
joining Grammar for English. Technical Report 
IRCS 95-03, University of Pennsylvania. See 
http://www.cis.upenn.edu/xtag/manuals.html. 
