Sentential Structure and Discourse Parsing 
Livia Polanyi, Chris Culy, Martin van den Berg, Gian Lorenzo Thione, David Ahn
1
 
FX Palo Alto Laboratory 
3400 Hillview Ave, Bldg. 4 
Palo Alto, CA 94304 
{polanyi|culy|vdberg|thione}@fxpal.com, ahn@science.uva.nl 
 
 
Abstract
1
 
In this paper, we describe how the LIDAS 
System (Linguistic Discourse Analysis Sys-
tem), a discourse parser built as an implemen-
tation of the Unified Linguistic Discourse 
Model (U-LDM) uses information from sen-
tential syntax and semantics along with lexical 
semantic information to build the Open Right 
Discourse Parse Tree (DPT) that serves as a 
representation of the structure of the discourse 
(Polanyi et al., 2004; Thione 2004a,b). More 
specifically, we discuss how discourse seg-
mentation, sentence-level discourse parsing, 
and text-level discourse parsing depend on the 
relationship between sentential syntax and dis-
course. Specific discourse rules that use syn-
tactic information are used to identify possible 
attachment points and attachment relations for 
each Basic Discourse Unit to the DPT.  
1 Introduction 
In this paper, we describe discourse parsing under 
the Unified Linguistic Discourse Model (U-LDM) 
(Polanyi et al. 2004). In particular, we describe the 
relationship between the output of sentential pars-
ing and discourse processing.  
The goal of discourse parsing under the U-LDM 
is to assign a proper semantic interpretation to 
every utterance in a text. In order to do so, the 
model constructs a structural representation of rela-
tions among the discourse segments that constitute 
a text. This representation is realized as a Dis-
course Parse Tree (DPT). Incoming discourse ut-
terances are matched with the informational con-
text needed for interpretation, and attached to 
nodes on the right edge of the tree.  
                                                      
1
 Current address:  
Language and Inference Technology Group,  
Informatics Institute 
Kruislaan 403 
1098 SJ Amsterdam, The Netherlands  
1.1 Discourse Parse Tree 
The U-LDM builds upon the insights and mecha-
nisms of the Linguistic Discourse Model (LDM) 
(Polanyi 1988). The DPT specifies which segments 
are coordinated to one another (bear a similar rela-
tionship to a common higher order construct), 
which are subordinated to other constituents (give 
more information about entities or situations de-
scribed in that constituent, or, alternatively, inter-
rupt the flow of the discourse to interject unrelated 
information), and which are related as constituents 
of logical, language specific, rhetorical, genre spe-
cific or interactional structures (n-ary relations). 
Importantly, the representation identifies which 
constituents are available for continuation at any 
moment in the development of a text and which are 
no longer structurally accessible.  
1.2 Discourse Parsing 
 While full understanding of the meaning of con-
stituent utterances, world knowledge, inference 
and complex reasoning are needed to create the 
correct Discourse Parse Tree to represent the struc-
ture of discourse under the LDM, in developing the 
U-LDM, it became apparent that most of the in-
formation needed to assign structural relations to a 
text can be recovered from relating lexical, syntac-
tic and semantic information in constituent sen-
tences to information available at nodes on the 
DPT. Largely formal methods involving manipu-
lating information in lexical ontologies and the 
output of sentential syntactic and semantic analysis 
are sufficient to account for most cases of dis-
course continuity, and give rise to only limited 
ambiguity in most other cases.
2
 The assignment of 
correct temporal, personal and spatial interpreta-
tion to utterances, which relies in large part on the 
relative location of referential expression and their 
                                                      
2
 Complex default logic based reasoning as in Struc-
tured Discourse Representation Theory (Asher 1993; 
Asher and Lascarides 2003), speculations about the in-
tentions or beliefs of speakers (as in Grosz and Sidner 
(1986)), or the intricate node labeling exercises familiar 
from Rhetorical Structure Theory (Mann and Thompson 
1988; Marcu 1999, 2000) are not necessary. 
referents in the DPT representation of the structure 
of the discourse, can then often be recovered. From 
a computational point of view, this is good news. 
Parsing consists of the following steps for every 
sentence of the discourse. 
1. The sentence is parsed by a sentence level 
parser, in our case the LFG-based Xerox Lin-
guistic Environment (XLE).  
2. The sentence is broken up in discourse relevant 
segment based on the syntactic information 
from the sentence parser.  
3. The segments are recombined into one or more 
small discourse trees, called Basic Discourse 
Unit (BDU) trees, representing the discourse 
structure of the sentence.  
4. The BDU trees corresponding to the sentence 
are each attached to the DPT tree.
3
  
 
In the remainder of this paper, we will describe 
how the LIDAS System (Linguistic Discourse 
Analysis System), a discourse parser built as an 
implementation of the U-LDM, uses information 
from sentential syntax and semantics along with 
lexical semantic information to build up the struc-
tural representation of source texts
4
. Specifically, 
we will discuss discourse segmentation, sentence-
level discourse parsing, and text-level discourse 
parsing—phases of the discourse parsing process 
that depend on the relationship between sentential 
syntax and discourse.  
2 Discourse Segments and Basic Discourse 
Units 
U-LDM discourse segmentation is based on the 
syntactic reflexes of the semantic content of the 
linguistic phenomena making up discourse. Since 
elementary discourse units must be identified to 
build up discourse structure recursively, discourse 
segments under the U-LDM are identified as the 
syntactic units that encode a minimum unit of dis-
course function and/or meaning that must be inter-
preted relative to a set of contexts to be under-
stood. Minimal Functional units include greetings, 
connectives, discourse PUSH/POP markers and 
other “cue phrases” that connect or modify content 
segments. Minimal meaning units are units that 
express information about not more than one event, 
                                                      
3
 If a sentence is sufficiently complex, it may consist 
of two or more completely different independent dis-
course units. Such cases are treated as if the two parts of 
the sentence were really two different sentences. One 
consequence of this is that a syntactic coordination of 
two sentences may correspond to a subordination in the 
DPT, because they are treated independently. 
4
 The LiveTree environment for discourse parsing is 
described in detail in Thione 2004b. 
event-type or state of affairs in a possible world. 
Roughly speaking they are “elementary proposi-
tions” or “event-type predicates” corresponding in 
(neo-) Davidsonian semantics to an elementary 
statement that contains at most one event-
quantifier. Structurally, a minimum meaning unit 
does not contain a proper subpart that itself com-
municates content and has a syntactic correlate that 
can stand on its own.  
Under the U-LDM, segments can be discontinu-
ous (if there is overt material on both sides of an 
intervening segment) or fragmentary. A single 
word answer to a question is a complete segment, 
whereas the same word uttered in an incomplete 
and unrecoverable phrase is a fragment. 
The U-LDM defines segmentation in purely sen-
tence syntactic terms. However, the choices of 
definitions are motivated by semantic considera-
tions. For example, since auxiliaries and modals do 
not refer to events distinctly from the main verb, 
they do not form segments separate from the corre-
sponding main verbs. By the same reasoning, other 
modal constructions that involve infinitives (e.g. 
have to, ought to, etc.) also constitute a single 
segment with their complements as do Cleft con-
structions, despite the presence of two verbs.
7
 On 
the other hand, Equi verbs (e.g., try, persuade) and 
Raising verbs (e.g., seem, believe, etc) form sepa-
rate BDUs from their verbal complements, since 
both eventualities can be continued. In contrast, 
even though event nominals, including gerunds, 
refer to eventualities (possibly) distinct from the 
verbs of which they are arguments or adjuncts, 
                                                      
5
 For example, “Chris served the chapter | as Social 
Chairman” 
6
 For example: “Finally, | the audio is summarized.” 
7
 For example: “It is segmentation that we are dis-
cussing.” 
Content segments  
(BDUs) 
Simple clauses, Subordinate 
clauses and participial phrases, 
Secondary predications,
5
 Inter-
polations (e.g. parentheticals, 
appositives, interruptions, etc.), 
Fragments (e.g. section head-
ings, bullet items, etc.) 
Operator segments Conjunctions that conjoin seg-
ments, Discourse operators 
(e.g. “scene-setting” preposed 
modifiers)
6
Discourse connec-
tives 
Table 1: Examples BDU and Operator Seg-
ments. 
those eventualities can not (easily) be continued 
and therefore are not segments.
8
 
The U-LDM, in contrast to other discourse theo-
ries, has a fine-grained taxonomy of minimal units. 
The most prominent distinction of discourse seg-
ments is between Basic Discourse Units (BDUs) 
and Operator segments (see Table 1). BDUs are 
segments realized in a linguistic utterance that can 
independently provide the context for segments 
later in the discourse. Operator segments can not 
do so, they can only provide context for segments 
in their scope. 
In the following section, we explain more fully 
how sentence syntax is used to segment sentences 
into discourse segments.  
2.1 The role of the XLE in discourse segmen-
tation 
Dividing a text into discourse segments begins 
by applying a sentence breaking algorithm to de-
termine the tokens to be segmented. These tokens 
are normally complete sentences, but may also be 
fragments in some cases such as titles or elliptical 
phrases such as “Yes”. The tokens are processed 
by a discourse segmenter. After the segmenter has 
completed its work, segments are passed to a U-
LDM BDU-Tree parser, which constructs one or 
more BDU-Trees from the BDUs identified in each 
sentence.  
The LIDAS segmenter first sends the input 
chunk to be parsed by the Xerox Linguistic Envi-
ronment parser (XLE) (Maxwell and Kaplan 
1989). The XLE is a robust Lexical Functional 
Grammar (LFG) parser. The XLE tries to parse the 
input, and returns either a packed representation of 
all possible parses or the most probable parse as 
selected by its stochastic disambiguation compo-
nent (Riezler et al. 2002). If the XLE can not find a 
parse of the input as a complete sentence it tries to 
construct a sequence of partially parsed fragments.
9
 
The LIDAS segmenter segments the most probable 
parse and, as a backup, the first parse from the 
packed representation, if the first and the most 
probable parse differ. 
                                                      
8
 Note that the contrast between modal verbs on the 
one hand and Raising and Equi verbs on the other 
clearly shows that the surface form of a phrase (e.g., the 
infinitival complements) is not always sufficient to de-
termine segment status. Similarly, a finite verb is not a 
sufficient indicator of a segment, as seen in the cleft 
constructions. Crucially, and appropriately, we need to 
refer to the discourse property of potential continuation. 
9
 The XLE has a failure rate (i.e., no parse whatso-
ever) of approximately 0.5% on our corpus of technical 
reports. 
The parse information consists of a c(onstituent) 
structure (essentially a standard parse tree) and a 
f(unctional) structure containing predicate-
argument information, modification information, 
and other grammatical information such as tense 
and agreement features. F-structures make up the 
primary source of linguistic information used in 
discourse parsing.  
To identify the discourse segments within the 
sentence, seven syntactic configurations in c-
structure and f-structure are examined. The rela-
tively small set of configurations in Table 2 ac-
counts for the full range of discourse segments. 
According to these segmentation rules, it is pos-
sible for a segment to be embedded in another 
segment. Because on the tree-projection of U-
LDM structures terminal nodes represent a con-
tiguous textual span, we recombine the two parts 
of a non-contiguous segment in the BDU-tree (sec-
tion 4) and DPT (section 5) using the concatena-
tion relation (+). Concatenated nodes contain the 
complete f-structure information of the completed 
segment and are full-fledged BDUs, available for 
further “anaphoric anchoring”. 
All segments are returned to the discourse parser 
in an XML format, which includes the f-structure 
information as well as the textual spans for each 
sentential segment. 
3 Discourse Parsing 
The next step after segmentation is combining the 
segments into a BDU-tree according to the rules of 
discourse parsing, resulting in a discourse tree rep-
resenting the sentence. After that, the BDU-tree is 
                                                      
10
 F-structures with SUBJ (subject) attributes are con-
sidered possible discourse segments since they typically 
encode an eventuality. However, because they do not 
introduce independent anchor points for future attach-
ment, the f-structures corresponding to modal and auxil-
iary verbs are excluded. 
Content Segments 
Certain F-structures with subjects
10 
Fragments 
Parentheticals 
Headers 
Syntactic coordination (except conjunct itself)  
Operator segments 
Conjuncts in coordination 
Initial comma separated modifiers 
Subordinating conjunctions 
Table 2. Segment classification configura-
tions. 
combined with the Discourse Parse Tree, again 
according to the discourse parsing rules.  
In discourse parsing, units of discourse are at-
tached to an emergent discourse tree. Attachment 
always takes place on the right edge of the tree. 
The parser has to make two decisions: what unit to 
attach to and what relationship obtains between the 
incoming unit and the selected attachment site. The 
types of evidence are used to determine this in-
clude: 
 
• syntactic information 
o subordinate/complement relations, parallel 
syntactic structure, topic/focus and centering 
information, syntactic status of re-used lex-
emes, preposed adverbial constituents, etc.  
• lexical information 
o Information from lexical ontology: re-use of 
same lexeme, synonym/antonym, hypernym, 
participation in the same lexical frame as 
well as specific discourse connectives tem-
poral and adverbial connectives indicating 
set or sub-set membership of any type for 
example, specifically, alternatively
11
. 
o Modal information: realis status, modality, 
genericity, tense, aspect, point of view
12
),  
• Structural information of both the local at-
tachment point and of the BDU-tree 
• Presence of constituents of incomplete n-ary 
constructions on the right edge 
o questions, initial greetings, genre-internal 
units such as sections and sub-sections, etc. 
 
The combined weight of the evidence determines 
the attachment point and the attachment relation. 
Interestingly, the weight given to each type of in-
formation is different for attachment site selection 
and relationship determination. Lexical ontological 
information, for example, is generally more impor-
tant in determining site, while semantic, syntactic 
and lexical “cue” information is more relevant in 
determining relationship. 
In Polanyi et al. 2004, a small set of robust rules 
is given for determining the attachment site and 
relationship of an incoming BDU-tree to the exist-
ing parse tree of the discourse. In the present pa-
per, which has as its focus understanding the rela-
tionship of sentential syntax to discourse structure, 
we concentrate on describing some fundamental 
                                                      
11
 These expressions are used as input to specific 
lexically driven rules that indicate language or genre 
specific binary relations between BDUs as suggested by 
Webber, Joshi and colleagues in their work on D-
LTAGs. (Webber and Joshi, 1998; Webber, Knott and 
Joshi, 1999; Webber, Knott, Stone and Joshi 1999.) 
12
 See discussion Wiebe, 1994. 
aspects of the relationship between the sentential 
syntactic structure of an incoming sentence and its 
corresponding sentential discourse structure. 
3.1 The Sentential Syntax - Discourse Struc-
ture Interface 
In discourse parsing with LIDAS, the output of 
the XLE parser supplies functional constraints and 
roles for syntactic structures. The syntactic struc-
ture of a sentence accounts for the discourse-
relevant relationships between segments within a 
sentence. LIDAS grammars exploit this informa-
tion by mapping syntactic relations onto discourse 
relations. The LIDAS grammar formalism permits 
the parser to leverage the XLE’s output by (1) 
checking positive or negative constraints (equality 
or inequality operators) (2) recursively searching f-
structures for specified attributes (* and ? wild-
cards), (3) enforcing dependent constraints,
13
 and 
(4) using Boolean connectives to combine con-
straints. (5) applying constraints universally or ex-
istentially to the set of matching f-structures. While 
LIDAS grammar rules can incorporate constraints 
that operate on all supported types of linguistic 
evidence, Table 3 gives three examples of rules 
that specify attachments based on the syntactic re-
lationship between constituents. 
 
1. 
BDU-1,BDU-2:  
BDU-1/phi = BDU-2/ADJUNCT/link;  
  � Right-Subordination. 
2. 
BDU-1,BDU-2:  
BDU-1/*/ADJUNCT/link = BDU-2/phi;  
  � Subordination. 
3. 
BDU-1,BDU-2:  
BDU-1/*/{XCOMP|COMP|COMP-EX}/link 
  = BDU-2/phi;  
  � Context. 
Table 3: Rules based on syntactic relationship. 
 
Rule 1 captures one general case for preposed 
modifiers. Prepositional and adverbial phrases, 
often temporal modifiers, that precede the main 
clause they modify can either elaborate on the 
main clause or modify the context in which the 
main clause is interpreted. Lexical information is 
used to distinguish among different types of modi-
fiers. Rule 2 expresses the general case of subordi-
nate adjunct clauses and shows that the syntax for 
                                                      
13
 For example, the following constraints show a de-
pendency between (1) and (3),  
(1) BDU-1/(*)/ADJUNCT/link = BDU-2/phi; 
(2) BDU-2/ADJUNCT-TYPE = “relative”; 
(3) $1/SPEC/DET/DET-TYPE = "indef"; 
One or more sub-f-structures that match the wildcard 
in the first constraint are tested in the third constraint. 
This constraint is part of a rule of sentential discourse 
syntax used to identify non-restrictive relative clauses. 
LIDAS’ discourse rules allows for recursive search 
over f-structures by seeking adjunct phrases match-
ing the incoming BDU anywhere within the f-
structure of the attachment point. Rule 3, which 
shows a compact syntax for disjunctive constraint, 
builds a sentence level discourse relation, the Con-
text Binary, that forms a complex context unit 
from its child constituents. Context Binaries are the 
general case for clausal complementizers. 
In the next section, we discuss one of the basic 
discourse parsing rules. 
3.2 Discourse Subordination  
We will assume the following extended hierarchy 
of grammatical functions:
14
 PRED > SUBJ > OBJ 
> COMP > ADJUNCT > SPEC.
15
 Given this hier-
archy we propose as a general principle of dis-
course construction that promotion in the hierar-
chy means demotion in the discourse. For exam-
ple, if the SUBJ of the incoming unit of discourse 
refers to the same entity
16
 as OBJ at the attachment 
node, then the relationship between them will be a 
subordination. In general, if an expression with 
grammatical function GF in a BDU B refers to the 
same (or a subcase of the same) entity as an ex-
pression with grammatical function GF’ in an ac-
cessible antecedent BDU A, where GF > GF’, then 
B will be attached as a subordinate structure to A 
on the DPT. This principle is expressed as a rule in 
the grammar that fires if it is not superseded by 
other rules. For example, the Narrative rule, which 
coordinates event clauses, takes precedence over 
the Discourse Demotion Rule. 
If the grammatical function hierarchy rule does 
not apply, but the BDU refers to a subclass
17
 of the 
antecedent BDU, there is evidence for a subordina-
tion relation. For example, if the subject of the 
BDU stands in a part-of relationship with the sub-
ject of the antecedent BDU, we can conclude that 
the relationship is a subordination. 
                                                      
14
 PRED denotes the tensed verb. It plays a role in the 
following discussion because verbs can be nominalized. 
15
 For the purposes of this hierarchy, grammatical 
function COMP includes the features COMP-EX and 
XCOMP. An element inside an expression with 
a grammatical function GF is itself in that position with 
respect to the elements that are not in that expression, 
although a separate ordering might exist between ele-
ments within the same expression. C.f. Obliqueness 
command in HPSG (Pollard and Sag, 1994). 
16
 In LIDAS, no reference resolution is done. Identity 
of reference is approximated using lexical semantics. 
17
 The notion of subcase as used here covers a num-
ber of different notions. An expression e is a subcase of 
f if (1) e is a set that is a subset of f, (2) e is a subtype of 
f, or (3) e is a part of f, among other relations. 
Table 4 gives the interaction of this rule with the 
hierarchy rule and the resulting relationships: G 
denotes whether a shift in the grammatical function 
hierarchy occurred, and L whether the shifted ele-
ment refers to the same entity, part of that entity or 
to an entity that is larger. If more than one such 
expression can be found, we consider the expres-
sion in the incoming unit with the highest gram-
matical function. 
 
 G
+
 G
0
 G
-
 
L
+
  S/C
18
 C S 
L
0
 S C N/A 
L
-
 S S N/A 
Table 4. Interactions of the hierarchy 
and subcase rules. 
The table is read as follows: take the expressions in 
the incoming unit that have a relationship with an 
expression in the attachment point. Let e be the 
expression among these that has the highest gram-
matical function, and f be the corresponding ex-
pression in the attachment point. If the grammati-
cal function of e is higher than that of f, we write 
G
+
, if it is the same G
0
, if it is lower G
-
. Similarly, 
if e is a supercase of f, we write L
+
, if e and f refer 
to the same entity L
0
, and if e is a subcase of f, L
-
. 
For example 
 
1. U1: The man was wearing a coat. 
U2: The Burberry looked brand new. 
 
In this case we notice that the words coat and 
Burberry are such that L<coat, Burberry> = L
-
 
and we also notice that Burberry gets promoted to 
the subject of the incoming unit, while coat was 
the direct object of U1. Therefore G<coat, Bur-
berry> = G
+
. From Table 4 we correctly identify 
that U2 does indeed subordinate on U1. 
Both L
-
 
and G
+
 
give evidence for discourse sub-
ordination. Grammatical function demotion G
-
 is a 
less clear case.
19
 Some mixed combinations sug-
gest discourse coordination (as for the preservation 
case <L
0 
, G
0
>) while others contribute too little 
                                                      
18
 Semantics would help to disambiguate this case. If 
the antecedent f is more specific than the anaphoric 
element e, two cases are possible. Either e is used as a 
definite description referring to the same entity as f, in 
which case the relation is a coordination, or e is used to 
denote a larger class of entities than f, in which therela-
tion is likely to be a subordination. 
19
 As is understanding precisely what is involved 
from a discourse relation point of view with complex, 
mixed promotion/demotion phenomena (from the sub-
ject of an XCOMP to the adjunct of an OBJ, for exam-
ple) 
significant information to independently determine 
the discourse attachment (N/A cases). In those 
cases, evidence from other rules in the grammar 
determines the result.  
4  BDU-Trees 
Identifying the relationship of a BDU to the dis-
course in which it occurs is a complex parsing 
process involving lexical, semantic, structural and 
syntactic information. The first step in understand-
ing the relationship of a given BDU to the extra-
sentential discourse context is to understand the 
role a BDU plays within the sentence in which it 
occurs. As a first step of constructing a discourse 
parse-tree of the sentence, the XLE parse and sen-
tential discourse rules are used to identify relation-
ships between the BDUs within the sentence re-
sulting in one or more BDU-trees. These small 
sentence-level discourse trees consist of one main 
clause and its subordinated clauses or preposed 
modifiers.
20
 
The root node of a BDU-tree represents the non-
subordinated material (often referred to as active 
material) within that BDU-tree, including at least 
the information of the main clause. The projection 
of the root node on the active leaves is referred to 
as the M-BDU (Main BDU). Only syntactic infor-
mation from the M-BDU is used to attached the 
BDU-tree to the DPT. In general, a sentence yields 
as many BDU-trees as top-level coordinated 
clauses.  
Consider, Example 2, a sentence taken from our 
corpus of Technical Reports: 
 
2. As a consequence, any hypertext link followed 
opened a new browser window, which we think of 
as a "Rabbit Hole" because the new window indi-
cates to users that they are no longer navigating in-
side the slideshow, but are instead navigating the 
Web. 
 
The U-LDM segmentation of this sentence:  
 
As a consequence | any hypertext link | followed | 
opened a new browser window | which we think of as a 
"Rabbit Hole" | because | the new window indicates to 
users | that they | are no longer navigating inside the 
slideshow | but | are instead navigating the Web. 
                                                      
20
 BDU-trees are very similar to the D-LTAG dis-
course segments (D-LDSs). However, BDU-trees in-
clude subordinated material, so they are typically larger 
than D-LDSs. Furthermore, because U-LDM sentence 
and discourse grammars are different, two BDU-trees 
may stand in a coordinated relationship in sentence 
grammar and be subordinated on the discourse level (cf. 
example 1). 
 
For compound sentences, members of the top-
level coordinate structures are attached independ-
ently, reflecting the fact that top-level coordinated 
clauses can escape the boundaries of the sentence 
when attaching to the main discourse structure. For 
example, the discourse parse of the following pas-
sage, illustrated in Example 3, consists of two sen-
tences, five segments, four BDUs and three BDU-
Trees that eventually form one DPT. 
 
3. S1: B1: The man soaked himself in the water.  
S2: B2: It was warm and soothing B3 and he 
decided to linger a little longer than usual.  
 
Despite the apparent syntactic coordination be-
tween the two main clauses, the two BDU-Trees in 
S2 show the independence of BDU-Tree/DPT at-
tachments. B1 describes a punctual event on a 
main story line. B3 describes the next event. The 
semantic and aspectual information of the verb to 
soak in BDU-1 and of the copula in BDU-2 com-
municates a switch from an event-line to an em-
bedded elaboration. The syntactic promotion of 
water from object of the preposition in the adjunct 
phrase in the water to the subject (through ana-
phoric reference) in the next segment indicates dis-
course subordination. Given L
0
<water, it> and 
G
+
<ADJUNCT/OBJ/PRED, SUBJ/PRED>, we 
find from table X, <L
0
 ,G
+
>  � S. As a conse-
quence, B2 is subordinated to B1. In the DPT, B3 
is coordinated with B1 at a point above B2, despite 
being syntactically coordinated to it at the senten-
tial level.  
5 Global Discourse Construction 
BDU-trees are attached to the DPT of the entire 
discourse as single units by computing the rela-
tionship between their M-BDUs and the accessible 
nodes aligned along the right edge of the DPT. 
Rules of discourse attachment that include those 
discussed for BDU construction as well as other 
genre and structural level rules are used in global 
DPT construction. 
The parsing process at the Discourse Parse Tree 
(DPT) level works as follows. Once BDU-Trees 
have been constructed and are ready to be attached 
to the DPT, each node along the right edge is ex-
amined, and, through a set of discourse rules, an 
ordered set of active Discourse Constituent Units 
(DCUs) is produced, representing possible attach-
ment points.
21
 This set can then be pruned of its n 
                                                      
21
 Lexical information is the main source of evidence 
in attachment site determination while other sources 
contribute to a lesser degree. The opposite is true for 
determining attachment relations 
lowest scoring constituents, according to a preset 
threshold or other criteria. 
 
Example 
Our group is developing new techniques for helping 
manage information for enhanced collaboration. We 
explore solutions for seamlessly connecting people to 
their personal and shared resources. Our solutions in-
clude services for contextual and proactive information 
access, personalized and collaborative office appli-
ances, collaborative annotation, and symbolic, statisti-
cal, and hybrid processing of natural language. 
Our team includes researchers with diverse back-
grounds including: ubiquitous computing, computer-
supported collaboration, HCI, IR , and NLP. 
Segmented Text 
1. Our group is developing new techniques for helping 
2. manage information for enhanced collaboration.  
3. We explore solutions for seamlessly connecting 
people to their personal and shared resources.  
4. Our solutions include services for contextual and 
proactive information access, personalized and col-
laborative office appliances, collaborative annota-
tion, and symbolic, statistical, and hybrid process-
ing of natural language. 
5. Our team includes researchers with diverse back-
grounds  
6. including:  
7. ubiquitous computing, computer-supported col-
laboration, HCI, IR , and NLP. 
Rules 
1-2 Intrasentential XCOMPS --> CX 
CX(1,2)-3 Demotion 2  
 [ Pres. Progressive to Simple Pre-
sent] 
 [ Our group --> We ] 
 [ Same Verb Class ] --> S 
3-4 Promotion  
 [Solutions from OBJ to SUBJ] --> S 
5-6 Relative Adjunct with Null Deter-
miner --> S 
6-7 Colon + OBJ linked to NP segment 
 --> CX 
CX(1,2)-5 Synonym Subjects, Same 
Tense --> C
22
 
Table 5. Analyzed Webpage Example 
In the second stage, attachment rules are checked 
against possible attachment sites. Rules that fire 
successfully attach the BDU-Tree to the DPT at the 
chosen site with the relationship specified by the 
rule. Local semantic, lexical and syntactic informa-
tion is then percolated into the parent DCU. If mul-
tiple attachments at different sites are possible, 
                                                      
22
 Discourse parsing is not unambiguous and different 
analyses may apply. So, here the same subject, change 
in progressive feature, and different verb class, which 
also apply would create an S. Future research is needed 
to understand precisely which rules take precedence. 
ambiguous parses are generated; less preferred at-
tachments are discarded and the remaining attach-
ment choices generate valid parse trees. 
Once a BDU-tree has been attached, its leaves 
become terminal nodes of the DPT and nodes on 
its right edge are accessible for continuation and 
may serve as contextual anchors for subsequent 
sentences. Table 5 shows an example text taken 
from the FXPAL Webpage describing our research 
group, a segmentation of the text and the DPT con-
structed following the rules. Figure 1 gives a s-
creenshot of the resulting tree. 
 
 
Figure 1: Tree of Analyzed Webpage Example 
6 Comparison with D-LTAG 
LIDAS is a purely symbolic discourse parser most 
similar to the D-LTAG parser described in (Forbes 
et. al. 2003). The overall structures of the LIDAS 
and D-LTAG parsers are almost identical. There 
are a number of apparently significant differences 
between the underlying theories although some of 
these may turn out to be notational variants after 
more extensive analysis. 
An important difference between D-LTAG and 
U-LDM derives from the fundamental question of 
segmentation; D-LTAG segments are larger than 
our segments, corresponding more or less to BDU-
trees (but cf. footnote 20). In D-LTAG, because 
only one grammar formalism governs both dis-
course and sentence level parsing, continuation can 
also take place on parts of segments as defined by 
sentence-level syntactic relations. Under the U-
LDM, which employs independent sentential and 
discourse grammars, only segments are potential 
anchors for continuation. Not only because we use 
an external syntactic parser as an oracle that in-
forms segment attachment on the BDU-tree, but 
more importantly because sentential syntax can be 
overridden by discourse syntax in some cases. 
Another basic difference between the two ap-
proaches is that D-LTAG builds its initial and aux-
iliary trees around connectives. Every discourse 
relation is expressed by a, possibly empty, connec-
tive. In U-LDM, connectives give evidence about 
the possible discourse relations, as do other parts 
of the sentence, but they do not solely determine 
discourse relation. We can thus account correctly 
for cases in which the wrong connective is selected 
to express a semantic relationship among seg-
ments.  
Lastly, our parser is meant to be incremental at 
the discourse level, whereas the D-LTAG parser 
appears to operate on the discourse as a whole.
 23
 
7 Conclusion 
In this paper we described a novel approach to dis-
course segmentation and discourse structure analy-
sis that unifies sentential syntax with discourse 
structure and argued that most of the information 
needed to assign a structural description to a text is 
available from sentential syntactic parsing, senten-
tial semantic analysis and relationships among 
words captured in a lexical ontology such as 
WordNet. The U-LDM discourse rules and parsing 
strategies presented here are a first step. We have 
tested out these rules in analyzing a corpus of over 
300 Technical Reports that have been summarized 
under the PALSUMM System that operates on top 
of LIDAS. (Polanyi et al 2004; Thione et al 2004) 
Much work remains to be done. Understanding the 
complex inter-relationships between rules is a for-
midable task. Critically important, too, is to unify 
the semantically motivated structural analyses pre-
sented here with an explicit S-DRT type formal 
semantic account of discourse semantics. How-
ever, we believe that the results presented here rep-
resent an important advance in understanding the 
nature of natural language texts. 

References 
Nicholas Asher. 1993. Reference to Abstract Objects in 
English: A Philosophical Semantics for Natural Lan-
guage Metaphysics. Kluwer Academic Publishers. 
Nicholas Asher and Alex Lascarides. 2003. Logics of 
Conversation. Cambridge University Press. 
Katherine Forbes, Eleni Miltsakaki, Rashmi Prasad, 
Anoop Sarkar, Aravind Joshi and Bonnie Webber. 
2003. D-LTAG System - Discourse Parsing with a 
Lexicalized Tree-Adjoining Grammar, Journal of 
Language, Logic and Information, 12(3). 
Barbara Grosz and Candace Sidner. 1986. Attention, 
Intention and the Structure of Discourse. Computa-
tional Linguistics 12:175-204. 
William C. Mann and Sandra A. Thompson. 1988. Rhe-
torical Structure Theory: Towards a Functional The-
ory of Text Organization. Text 8(3)243-281. 
Daniel Marcu. 1999. Discourse trees are good indicators 
of importance in text. In Advances in Automatic Text 
Summarization. I. Mani and Mark Maybury (eds), 
123-136, The MIT Press. 
Daniel Marcu. 2000. The Theory and Practice of Dis-
course Parsing and Summarization. The MIT Press. 
Cambridge, MA. 
John Maxwell and Ronald M. Kaplan. 1989. An over-
view of disjunctive constraint satisfaction. In Pro-
ceedings of the International Workshop on Parsing 
Technologies, Pittsburgh, PA. 
Livia Polanyi. 1988. A Formal Model of Discourse 
Structure. Journal of Pragmatics 12: 601-639. 
Livia Polanyi. 2004. A Rule-based Approach to Dis-
course Parsing. In Proceedings of SIGDIAL ’04. 
Boston MA. 
Stefan Riezler, Tracy H. King, Ronald M. Kaplan, 
Richard Crouch, John T. Maxwell, and Mark John-
son. 2002. Parsing the Wall Street Journal using a 
Lexical-Functional Grammar and discriminative es-
timation techniques. In Proceedings of the 40th An-
nual Meeting of the Association for Computational 
Linguistics (ACL’02), Philadelphia, PA. 
Stefan Riezler, Tracy H. King, Richard Crouch, and 
Annie Zaenen. 2003. Statistical Sentence Condensa-
tion using Ambiguity Packing and Stochastic Disam-
biguation Methods for Lexical-Functional Grammar. 
In Proceedings of HLT-NAACL'03, Edmonton, Can-
ada.  
Radu Soricut and Daniel Marcu. 2003. Sentence Level 
Discourse Parsing using Syntactic and Lexical In-
formation. In Proceedings of HLT/NAACL’03, May 
27-June 1, Edmonton, Canada 
Thione, Gian Lorenzo, Martin van den Berg, Chris Culy 
and Livia Polanyi. 2004a. Hybrid Text Summariza-
tion: Combining external relevance measures with 
Structural Analysis. Proceedings ACL Workshop 
Text Summarization Branches Out. Barcelona. 
Thione, Gian Lorenzo, Martin van den Berg, Chris Culy 
and Livia Polanyi. 2004b. LiveTree: An Integrated 
Workbench for Discourse Processing. Proceedings 
ACL Workshop on Discourse Annotation. Barcelona. 
Bonnie Webber and Aravind Joshi, 1998. Anchoring a 
lexicalized tree-adjoining grammar for discourse. 
COLING/ACL Workshop in Discourse Realtions and 
Discourse Markers. Montreal, Canada. 86-92. 
Bonnie Webber, Alistair Knott and Aravind Joshi. 
1999a. Multiple discourse connectives in a lexical-
ized grammar for discourse. In 3rd Int’l Workshop on 
Computational Semantics. Tilburg. 309-325. 
Bonnie Webber, Alistair Knott, Matthew Stone and 
Aravind Joshi. 1999b. Discourse Relations: A Struc-
tural and Presuppositional Account Using Lexical-
ized TAGS. 37th ACL. College Park, MD. 41-48. 
Wiebe, Janyce M. 1994. Tracking point of view in 
narrative. Computational Linguistics 20 (2): 233-287. 
