From Elementary Discourse Units to Complex Ones 
Holger Schauer 
Computational Linguistics Division 
Freiburg University 
D-79085 Freiburg, Germany 
s c\]~auer@coling, uni-freiburg, de 
Abstract 
Coherence relations have usually 
been taken to link clauses and larger 
units. After arguing that some 
phrases can be seen as discourse 
units, a computational account for 
such phrases is presented that inte- 
grates surface-based criteria with in- 
ferential ones. This approach can be 
generalized to treat intra-sentential 
cue-phrases. Since cue-phrases are 
not always present, referential rela- 
tions between nominal expressions 
are additionally used to derive a 
text's discourse structure. 
:1 Introduction 
It is widely acknowledged that texts are not 
just collections of sentences, but have a struc- 
ture of their own. There has been substan- 
tial work in order to account for the different 
phenomena of discourse structure by apply- 
ing theories of coherence relations, e.g. (Mann 
and Thompson, 1988; Asher, 1993) among 
others. Coherence relations represent rich se- 
mantic linkage (like Cause t or Evaluation) 
between text segments of varying size. 
However, what the minimmn size of text 
segments to relate should be, is still left open 
to debate. As common approaches argue 
that coherence relations relate events or situ- 
ations (e.g. (Hobbs et al., 1993; Asher, 1993)) 
and that such events are usually introduced 
1Coherence relations in this paper are basically 
tulc~n from taken Rhetorical Structure Theory (Mann 
and Thompson, 1988) will appear emphasized and 
Capitalized. 
through the means of verbs, it has become 
standard practice to consider clauses to be 
the appropriate size for elementary discourse 
units. It has, however, also been observed 
(Vander Linden and Martin, 1995; Grote et 
al., 1997) that sometimes phrases may serve 
as very condensed forms to express elaborate 
contents. Recently, (Schauer and Hahn, 2000) 
provided a more detailed analysis when prepo- 
sitional phrases (PPs) may serve as elemen- 
tary discourse units. 
Cursorily viewed, the claims of another re- 
cent study stand in contrast to the idea of 
intra-clansal discou~e units: (Schauer, 2000) 
examined the interplay of coreferential ex- 
pressions and discourse structure and con- 
cluded that referential relations are a good 
indicator of the discourse structural configu- 
rations in case the units examiued are entire 
sentences. This poses the question whether 
not entire sentences are the appropriate grain 
size for elementary discourse units. 
I will argue that these results - i.e. the dif- 
ferent levels of granularity for discourse units 
- are not incompatible with each other. The 
approach used in (Schauer and Hahn, 2000) 
to derive the coherence relation governing a 
prepositional phrase neatly carries over to the 
computation of coherence relations signaled 
by sentence-internal cue-phrases. This then 
allows an integration with the algorithm us- 
ing referential relations that was proposed in 
(Schauer, 2000). 
2 Adjuncts as Discourse Units 
The question at what size of textual expres- 
sions one should start looking for discourse 
units has not been suftlciently answered yet. 
46 
(Mann and Thompson, 1988), for example, 
state that "Unit size is arbitrary, but the di- 
vision of the text should be based on some 
theory-neutral classification. That is, for in- 
teresting results, the units should have inde- 
pendent functional integrity. In our analy- 
ses, units are basically clauses \[...\]" \[p.248\]. 
It has been noted previously that sometimes 
phrases already seem to have enough "func- 
tional integrity" to give some "interesting re- 
sults", however, until recently there has been 
no detailed proposal what kind of phrases 
should be considered. 
Consider the following (German) example: 2 
(1) a. Mit dem P6LXZ-A versucht Elite- 
group neue Kunden zu gewinnen. 
\[With the P6LXZ-A -- Elitegroup 
tries to attract new customers.\] 
b. Mit einem PCI-Slot bietet das Mother- 
board wenig Platz fiir Erweiterungen. 
\[With one PCI slot (only) -- the moth- 
erboard provides only little space for 
extensions.\] 
A straightforward coherence analysis with 
relations from RST (Mann and Thompson, 
1988) takes (l-b) as a single unit and links 
it to (l-a), probably via an Evaluation re- 
lation. Paraphrasing sentence (l-b) reveals, 
however, a plausible decomposition into two 
interdependent discourse units: 
(2) a. The motherboard has one PCI slot, 
b. so it provides only little space for ex- 
tensions. 
Clause (2-a) gives an Explanation for the 
subordinate clause (2-b). This might be at- 
tributed to the impact of the cue word "so" 
in (2-b). More generally, it has been claimed 
that whenever an implicit coherence relation 
can be made explicit by a paraphrase incor- 
porating a specific cue word, then this coher- 
ence relation can always be assumed to hold. 3 
So, from a methodological point of view it 
2In the following, I will summarize the arguments 
from (Schauer and Hahn, 2000) that focus on prepo- 
sitional phrases. 
S"As a test for the presence of an implicit connec- 
tion it can be required that the connection could have 
been explicit... "(Martin, 1992, p.184) 
cannot be justified to analyze Sentence (2) 
as being composed of two elementary units, 
while the prepositional phrase "with one PCI 
slot" should be an indistinguishable part of 
the whole Sentence (l-b). 
Obviously, not all kind of phrases may serve 
as elementary units. As a first criterion to de- 
termine whether a phrase is a discourse unit, 
we propose to consider only those phrases 
that are not syntactically mandatory com- 
plements of their governing syntactic head. 
Complements are assumed to be specified in 
terms of subcategorization frames, valency 
hsts, etc. Adjuncts are phrases which do 
not match such a complement specification of 
their governing syntactic head. 
In Example (l-b), while it is both possi- 
ble to omit the optional prepositional phrase 
"with one PCI slot" and to decompose the 
sentence such that the prepositional phrase 
turns into a proper clause, this is only hardly 
possible for any other phrase in this example, 
say the subject "the motherboard". 
The second major criterion operates at the 
semantic level. Semantic specifications of lex- 
emes, independent of the specific semantic 
theory one subscribes to, are confined to "typ- 
ical" properties, e.g., events are character- 
ized by agents, patients, instruments, loca- 
tions, time frames, etc. The interpretation 
of phrases referring to these attributes can 
be taken care of by straightforward semantic 
interpretation mechanisms. As long as the 
interpretation makes reference only to such 
"typical" properties it is not necessary to con- 
sider such phrases as discourse units. But 
when an analyst thinks that the interpreta- 
tion involves (additional) non-typical, unpre- 
dictable information, coherence relations may 
capture their value-adding meaning. 4 E.g., 
the prepositional phrase of (l-a) only refers 
to a such typical property, namely the instru- 
ment of an action. This typicality considera- 
tion does not carry over to an "explanation" 
of events, which is the interpretation of "with 
4The interpretation of complements is assumed to 
involve only these typical attributes. The interpreta- 
tion of adjuncts, in contrast, may refer to both typical 
and non-typical information. 
47 
one PCI slot" from Sentence (l-b). 
Ultimately, however, there is also a third 
criterion that needs to be considered: namely, 
that a coherence relation can be found by 
which the phrase may be related to the re- 
mainder of the clause it occurs in (Marcu et 
al., 1999). We propose that this search for 
an applicable coherence relation may be per- 
formed by trying to paraphrase the phrase 
and its matrix clause in such a way that it 
results in two clauses that are related by a 
cue-phrase. 
(Schauer and H~.hn, 2000) evaluated the 
applicability of the criteria by' hmnan anno- 
tators using two corpora. We found that 
11.0% of the prepositional phrases in a Ger- 
man text corpus can be seen as discourse units 
and 8.0% in an English control set. These 
PPs are responsible for 14.8% and 12.1%, re- 
spectively, of the relations fomad in the texts. 
The discourse units triggered by prepositional 
phrases always occur in satellite (i.e. subordi- 
nate) position, which is not surprising, given 
their status of optional information. 
3 From Prepositional Phrases to 
Coherence Relations 
For an automated coherence analysis, we will 
basically take the preposition as a trigger for 
computing a coherence relation. To make 
this discussion more concrete, it is embed- 
ded in the framework of SYNDIKATE, a text 
analysis system under development in our lab 
(Hahn and Romacker, 1999). 
3.1 Semantic Interpretation 
Crucial for the discussion here is that the syn- 
tactic analysis provided by the system allows 
to distinguish betweenadjuncts and manda- 
tory complements: for Sentence (l-b), a de- 
pendency relation ppadj between the prepo- 
sitional phrase and its governing syntactic 
head is derived, that results from the valency 
specification for the main verb "bietet" (pro- 
vide). For computing a semantic interpre- 
tation s , the corresponding conceptual corm- 
lares of the content words are then checked 
SWe assume the framework of description logics 
(Woods and Sclmaolze, 1992). 
for role compatibility. 
In Example (l-b), the major interpretation 
constraints derive fzom the main verb "bi- 
etet" (provide) which is represented by the 
concept PROVIDE. It has three major concep- 
tual roles, PROVIDE-PATIENT, PROVIDE-CO- 
PATIENT, and INSTRUMENT. The PROVIDE- 
PATIENT and PROVIDE-CO-PATIENT roles can 
be filled by some instance of MOTHERBOARD 
and SPACE, respectively. 
Focusing on the analysis of the PP, each 
preposition specifies allowed semantic rela- 
tions (Romacker et al., 1999). In the case 
of "mit" (with) they allow an interpretation 
in terms of the conceptual INSTRUMENT role, 
so the corresponding role of PROVIDE is frilled 
with PCI-SLOT. This causes additional con- 
ceptual interpretation processes: a production 
rule checks whether the filler of the INSTRU- 
MENT role may be a PART-OF the concep- 
tual correlate of the syntactic subject, link- 
ing SPACE and MOTHERBOARD. Along the 
line, a HAS-PART-STATE is established that 
corresponds to the situation that "the moth- 
erboard has one PCI slot". 6 
3.2 Discourse Interpretation 
As discussed above, a preposition does not 
always signal a coherence relation. Also, a 
preposition may give rise to different coher- 
ence relations depending on context, so one 
can not simply paste a label of "the" coher- 
ence relation into the representation. Instead, 
the computation of coherence relations is per- 
formed by testing relation-specific constraints 
on the configuration of the text-knowledge 
base resulting from semantic interpretation. 
Following standard practice, coherence rela- 
tions are assumed to connect situations. So, 
we try to derive the relation between the 
conceptual correlate PROVIDE of the main 
verb and the inferred state HAS-PART-STATE, 
which stems from the conceptual interpreta- 
tion of the prepositional phrase. 
6In case that the sentence ~he mo~he.rboard has 
one PCI slot" would have been parsed, the PART-OF 
relation would have been derived via a production 
rule operating on the interpretation of the main verb 
"has". Thus, establishing the HAS-PART-STATE in F.,x- 
ample (l-b) yields a uniform analysis. 
48 
EXPLANATION. 12 
-~ HAS-PERT-STA~. 12 SATELLITE ' 
HAS-i~UKP OSE 
~ Fs-P~r<4> 
~_co _P~TiENT-~ PCI" SLOT" 3 i 
r''l 
Figure 1: Discourse Interpretation for Example (l-b) 
The preposition "mit" (with) may signal an 
Explanation. The basis for computing this re- 
lation consists in recognizing that the matrix 
clause gives some kind of evaluation: namely, 
that the "motherboard provides only little 
space.". The degree expression SPACE, with 
its associated negative POL--MARKER, (Staab 
and Hahn, 1997) is the trigger for recognizing 
the evaluative status of the matrix clause. We 
can now start looking for the signalled Expla- 
nation for this negative judgment. 
Explanation can be found quite easily: usu- 
ally, a MOTHERBOARD has up to four PCI- 
SLOTS (see the default value that is denoted at 
the PART-OF relation in Figure 1). But actu- 
ally the situation HAS-PART-STATE is a state 
in which only one is present, which is obvi- 
ously "little". The computations that derive 
the Explanation-relation are all performed by 
means of production rules. 
As a preposition might give rise to differ- 
ent coherence relations, the constraints of all 
those relations have to be checked. None of 
the constraints that are checked for other co- 
herence relations signaled by 'knit" (e.g. Con- 
dition) are satisfiable in this example. Sim- 
ilarly, for Example (l-a) no coherence rela- 
tion would be derived, because the constraints 
would not be appicable. So, the compu- 
tational approach is complementary to the 
semantic criterion of typicallity discussed in 
Section 2. 
4 Interclausal Coherence 
The proposed mechanism essentially treats 
the preposition as a kind of cue-phrase that 
determines a set of possible coherence rela- 
tions, combining it with a inference mecha- 
nism. The main advantage of using such a 
combination is that it is possible to disam- 
biguate cases in which cue-phrases may give 
rise to different relations without being con- 
fined to testing the complete range of co- 
herence relations. This problem is not re- 
stricted to prepositions: (Knott and Sanders, 
1998) actually build a multi-taxonomy of cue- 
phrases in which elements may give rise to 
several relations, depending on context. E.g, 
"as" has three connotations: "as" may signal 
temporal, causal or slml\]arity relationship. 
49 
The presented approach can'ies over neatly 
to such cue-phrases with multiple relations: 
indeed, the process described in Section 3 
can be seen as a formal reconstruction of the 
paraphrase. The conceptual representations 
that is derived for Sentence (2) is basically 
the same as for (l-b): in both cases a HAS- 
PART-STATE and a PART-OF relation are com- 
puted that connects the MOTHERBOARD and 
its PCI-SLOT. Finally, the computation of 
the coherence relation for Sentence (2) differs 
only with respect to the trigger: in Sentence 
(2), the cue-phrase "so" triggers the computa- 
tion instead of the preposition ~mit" in (l-b). 
The presented approach is thus able to han- 
dle coherence relations that operate on the 
intra-sentential level. Still, this is only a first 
step to account for the discourse structure 
of entire texts: cue-phrases are by no means 
always present. And even if there is a cue- 
phrase in a new unit to attach, it is still often 
not clear from the cue alone to which unit the 
new one should connect: 
(3) a. Apple presented a new member of the 
MessagePad organ/zer thmily. 
b. A new display and a new version of the 
operating system are introduced. 
c. However, the old MessagePad 120 is 
still available. 
The cue "however" alone does not give 
enough information to decide whether Sen- 
tence (3-c) should connect to (3-b) or to (3-a): 
further information is needed, like that there 
is a referential relation between the old Mes- 
sagePad 120 and the MessagePad family. As 
a step towards connecting larger units, a re- 
cent study (Schauer, 2000) examined the in- 
teraction of discourse structure and referen- 
tim relations among nominM expressions. 
4.1 Structural Dependencies 
What seems intuitively clear, is that a theory 
of discourse structure should reflect the struc- 
ture of the text under consideration. Rhetor- 
ical Structure Theory, however, concentrates 
on the effects intended by the author of a text. 
RST focuses on a '~pre-ree~|izational" struc- 
ture and is not primarily concerned with text 
phenomena. The results are analyses that do 
not account for the structural constraints in a 
text, here referential constraints: the depen- 
dency of an anaphoric expression on its an- 
tecedent - that it is resolvable and hence in- 
terpretable - can hardly be captured in RST. 
While this is of course not an issue for RST 
as a theory, it is a prerequisite for any sys- 
tem that wants to account automatically for 
a text's discourse structure. As an example 
consider the following fragment: 
(4) a. The Vaio F190 comes with convincing 
equipment. 
b. It features a DVD-ROM and a 6 GB 
harddisk. 
In classical RST, example (4-a) could be said 
to stand in an Evaluation relation to (4-b). 
The definition of Evaluation requires that the 
satellite evaluates the nucleus, see Figure 2. r 
However, this would not capture a structural 
dependency: the pronominal anaphora "it" 
ca.nnot be interpreted correctly without the 
antecedent, so (4-b) depends on (4-a). 
Vaio F 190 It 
Figure 2: Structure of Evaluation 
In RST, one could reflect this referential 
dependency by analyzing (4-b) as giving Evi- 
dence for (4-a) (see Figure 3), as an Evidence 
relation may be applied when the hypothesis 
is the nucleus. Such an analysis would ne- 
glect the "convincing" in (4-a) which is quite 
an explicit evaluation. 
In order to account for the structural 
dependency and the appropriate semantic 
linkage, we propose to use a new kind of 
Evaluation-N(ucleus) relation. More gener- 
ally, this loosens the relation between nucleus- 
satellite assignment and semantic content of 
coherence relations. 
7The depicted structures reflect standard RST 
schemata. The target of the arrow marks the nucleus. 
50 
Vaio F190 It 
Figure 3: Structure of Evidence 
4.2 Structural Configurations 
Analyzing texts using the outlined notion 
of structural dependencies as a basic guide- 
line, we determined the structural configu- 
rations involving coreference relations. Ba- 
sically, when two units contain coreferentiai 
expressions, they are usually connected by 
a coherence relation which subordinates the 
anaphoric unit. In the simplest case, the re- 
lation is one of Elaboration. However, further 
linguistic cues or inferences might give rise to 
semantically "richer" relations. 
When more than one coreference relation 
is involved, issues become more complex. For 
instance, consider the following example: 
(5) a. The Panasonic LC90S is a 19"-display. 
b. Although its screen size of 482 ram 
corresponds to a conventional 21"- 
monitor, 
c. considerably less space is required. 
d. The device can be attached to a video- 
card via a USB-comaector. 
Obviously, the nominal anaphor '2he device" 
in (5-d) requires an antecedent. One possi- 
bility would be the pronominal "it" in (5-b), 
leading to a resolution to "the LC90S". How- 
ever, this is not reflected in the discourse 
structure that seems most appropriate, of. 
Figure 4. The topic of (5-b) and (5-c) (the 
size) is not further elaborated on in (5-d), so 
one might say there is a mini-segment bound- 
ary between these two sentences. Hence, it 
would also be not correct to analyze (5-d) as 
an Elaboration of (5a-c), because (5-d) elab- 
orates only (5-a). So, unless further connec- 
tions (stemming from cue-phrases, inferences 
or whatever) license a topic continuation, a 
new unit containing anaphora will connect to 
the upmost unit bearing their antecedents. 
Elaboration 
~----~aboration ~, 
5a ~¢f_~ssion 5d 
5b 5c 
Figure 4: RST analyses for Example (5) 
5 From References to Coherence 
The configurations just described naturally 
lead to a combined account of deriving a 
text's discourse structure and resolving its re- 
ferring expressions, repeated from (Schauer, 
2000) in Figure 5. Basically, the algorithm 
uses the successful resolution of anaphoric ex- 
pressions as a guideline for determining the 
target unit to which a new unit should be 
connected which in turn restricts the set of 
units which should be searched for resolving 
further referring expressions. 
The algorithm assumes the availability of 
several capabilities for anaphora resolution. 
First, for a given unit candidate s a set of noun 
phrases needs to be identified that may be 
anaphoric expressions. Second, some resolu- 
tion process is necessary that checks whether 
an anaphoric expression can be resolved in a 
given list of possible antecedents. This pro- 
cess makes heavy use of underlying (domain) 
knowledge (Hahn et al., 1996): in Exam- 
ple (1), "the P6LXZ-A" is an instance of a 
MOTHERBOARD. Since ~the motherboard" in 
(l-a) is a definite noun phrase and syntactic 
as well as conceptual information match with 
the plausible antecedent "the P6LXZ-A', a 
referential link can be established, see the IS- 
COREFERENTIAL relation in Figure 6. 
Abstracting away from the actual imple- 
mentation, the algorithm operates over a tree 
structure that manages lists of accessible an- 
SFor the sake of argument and because the algo- 
rithm is taken from (Schauer, 2000), clauses will be 
taken as units for the moment. The issue will be dis- 
cussed below in more detail. 
51 
tree := tree(centers_forward(first (clauses)),NIL) 
clauses := rest (clauses) 
forall clause := clauses do 
ana_nodes := array of lists of nodes. 
forall ana_cand ::= anaphoric_expressions(clause) do 
node := lowest_right_node(tree) 
while node do 
if match(ana_cand,ante_list (node)) then 
ana_nodes \[ ana_cand \] := append ( ana_nodes \[ ana_cand \], 
node := predecessor (node) 
done 
done 
target_node := find_highest_node_matching_all (ana_nodes) 
/, found at least one antecedent node ,/ 
if target_node t:hen 
/, connect new unit to old node ,/ 
connect (target_node , tree (centers_forward (clause), NIL)) 
done 
Figure 5: Algorithm integrating Coreferences and Discourse Structure 
node) 
tecedent candidates and reflects the discourse 
structure in terms of coherence relations that 
has been derived so far. 9 
The algorithm now loops through all 
clauses of a text, building up both the tree 
and the antecedent lists incrementally. When- 
ever a new clause has to be considered, its left 
context consisting of the right frontier of the 
tree is checked for plausible antecedents. 
When all accessible antecedent nodes are 
determined, the highest node that provides 
antecedents for all resolvable anaphoric ex- 
pressions in the new unit is taken as the tar- 
get node, in accordance with the discussion in 
Section 4.2. 
If a new unit contains no referential expres- 
sion then the algorithm makes no prediction. 
If the target node has been found, the new 
unit is connected to it, i.e. the new unit is es- 
tablished as a satellite to the target unit. This 
means that the new unit opens a new right- 
most branch and hence becomes the lowest- 
right node of the tree. So, the new right fron- 
tier consists of the newly attached unit, the 
9The computed representation of the discourse 
structure is shown in Figure 6, while the storage of 
accessible referential objects is managed elsewhere in 
the system as it interacts with the syntactic parser. 
modified node and its predecessors. 
The evaluation in (Schaner, 2000) showed 
that the predictions made by the algorithm 
depend crucially on the size of the elementary 
units considered. When clauses are consid- 
ered to be the elementary unit size, the pre- 
dictions are correct in up to 81% of the cases 
in which the algorithm makes a prediction -- 
under the pre-condition that intra-sentential 
units axe handled first. Linking units by ref- 
erentiai makes surely no sense when the units 
considered axe phrases: of course, syntactic 
constraints always mandate that intra-clausal 
units are related with each other first. 
This algorithm is only a useful approxima- 
tion towards a complete account of a text's 
discourse structure. Relating to the highest 
unit providing antecedents works only when 
there is a mini segment boundary every time 
an anaphoric expression is used. Although 
the algorithm ignores other sources of rela- 
tions between units -- which are exactly the 
cases where the algorithm fails -- the results 
are surprisingly good. As such, enhancing 
the cue-phrase-only approach by additionally 
considering coreference relations is in the mid- 
die of cheap surface-only approaches and very 
52 
~s~z zs-cozczzmrrz~ ~ I - 
• ~ ,__~/~-POL-~n -~ 
I 
S'~Z..~-CO-P~IZI¢2 
Figure 6: Representation of the Complex Discourse Unit for Example (1) 
expensive inference-only methods. A com- 
plete approach will also try to account for 
inter-sentential cue-phrases and also further 
intercounections, based on the computed rep- 
resentation, see Figure 6) 0 
6 Related Work 
Rhetorical Structure Theory (Mann and 
Thompson, 1988) was basically used as the 
theory on which the presented arguments 
were grounded. Unfortunately, its definitions 
are intended to be applied by human anno- 
tators and have long resisted approaches to 
automation because of their vagueness. Pre- 
vious approaches to (automatically) deriv- 
ing a text's discourse structure either tried 
to rely on purely surface-based criteria (cue- 
phrases, (Marcu, 1998)) or concentrated on 
inferences about representations of discourse 
units (Asher, 1993; Hobbs et al., 1993). How- 
ever, neither of these approaches tries to ac- 
count for phrases as discourse units. Two re- 
1°The Evaluation relation in Figure 6 is due to the 
evaluative nature of Sentence (l-b), see the discussion 
in Section 3. 
cent studies mention the role of PPs as carri- 
ers of coherence relations (Vander Linden and 
Martin, 1995; Grote et al., 1997), but only for 
the purpose of text generation. 
As prepositions (and cue-phrases in gen- 
eral) can signal different coherence relations, 
the presented computational approach cou- 
ples a cue-phrase approach like (Marcu, 1998) 
with inferences using the computed seman- 
tic representation. Only recently, there have 
been some approaches (Webber et al., 1999) 
that acknowledge the need to incorporate 
both a treatment of cue-phrase approaches 
and inferences. However, how these inferences 
take place is not addressed, neither is the level 
of intraclansal coherence, that, as we have ar- 
gued, substantially relies on an integration of 
beth veins. Purely inferential approaches like 
(Asher, 1993; Hobbs et al., 1993) have also 
been criticized for being not really applica- 
ble because of computational costs. The pre- 
sented approach tries to minimize these costs 
by relying on surface-based criteria that re- 
duce the set of coherence relations that have 
to be considered. 
53 
Cue-phrases are not necessarily alone re- 
sponsible for the discourse structure of texts. 
Hence, referential relations bel~ween nominal 
expressions were suggested as means for find- 
ing attachment points of new traits that need 
to be incorporated. (Corston, Oliver, 1998) 
enlarges the cue-phrase approach of (Marcu, 
1998) in a vein sJmi\]ar to the one presented 
in Section 5. However, how several corefer- 
ence relations interact with the resulting dis- 
course structure, is not spelled out. Also, in 
his approach, it remains uncle~ how the cor- 
rect target node to attach to is identified. 
Segmented Discourse Representation The- 
ory (SDRT), as described in (Asher, 1993), 
provides a framework of discourse structure 
which interacts with referential accessibility 
constraints. Asher does not rely on coref- 
erences for establishing target units; instead 
the derivation of a coherence relation (and 
thereby of the target unit to connect a new 
unit to) relies on rather abstract connections 
between "events". While recognizing corefer- 
ence relations certainly also requires domain 
knowledge and inference capabilities, recog- 
nizing connections between events seems an 
even more impossible task. The presented ap- 
proach is hence more light-weight in nature. 
(Webber et al., 1999) apply operations 
on trees for incorporating cue-phrases in a 
grammar-driven approach to discourse struc- 
ture. The presented algorithm could be in- 
tegrated with their approach, to account for 
the cases of units that are not connected by 
referential relations but via cue-phrases. 
7 Conclusion 
Starting from the question what are the el- 
ementary units to consider for a text's dis- 
course structure, I presented an account for 
prepositional phrases with adjunct-status as 
discourse units. Prepositions can be seen as 
a kind of cue-phrase; however, a preposition 
does not necessarily signal a coherence re- 
lation and even if it does is often ambigu- 
ous with regard to the coherence relation sig- 
naled. Therefore accounting for prepositional 
phrases as disco~se units requires additional 
inferences operating on the semantic repre- 
sentation of both PP and its matrix clause. 
The approach neatly carries over to the 
phenomena of ambiguous cue-phrases. How- 
ever, this is still not sufficient to account for 
deriving the discourse structure of texts in 
general: cue-phrases are by no means always 
present and even if there is a cue-phrase, de- 
tecting the correct attachment point of a new 
unit is usually not straight-forward. 
As (one step towards) a solution, ref- 
erential relations between nominal expres- 
sions were suggested. The presented algo- 
rithm integrates the resolution of anaphora 
-- which also depends heavily on inferences 
and domain-knowledge -- with choosing the 
target unit to which a new unit should con- 
nect to: namely, the highest node provid- 
ing antecedents to all anaphoric expressions 
in the new unit. In order for this algorithm 
to operate successful, it is however necessary 
that this process is started only after all phe- 
nomena of intra-sentential coherence relations 
have been accounted for, which might be done 
using the combined approach outlined above. 
Returning to the question posed at the be- 
ginning of the paper - what is the appropri- 
ate size of elementary discourse units - the 
answer is twofold: first of all, coherence re- 
lations can be found to hold between phrases 
and the clause containing them, so one should 
indeed start looking for discourse units at the 
phrase level. However, Syntax requires that 
the components of sentences group together, 
and returning to what (Mann and Thompson, 
1988) said, sentences have a kind of functional 
integrity - one that operates on a level that 
is different from those of phrases. Once this 
level is reached, larger chunks can be formed, 
e.g. by referential means. 
The presented approach to cue-phrases as 
well as the use of referential relations will 
be implemented in the text understanding 
system SYNDIKATE (Hahn and Romacker, 
1999), in order to account for semantically 
rich relations between larger text chlmks and 
the discourse structure of texts in general. 
This, however, will require further under- 
standing of the conditions of the coherence 
relations involved. 
54 
Acknowledgments 
The author is a member of the Graduate 
Program on Human and Machine Intelligence 
at Freiburg University, Germany, funded by 
DFG. I would like to thank the staff at 
the Computational Linguistics Lab, Freiburg, 
and also the reviewers for their comments. 

References 
Nicholas Asher. 1993. Reference to Abstract Ob- 
jects in Discourse. Dordrecht: Kluwer. 
Simon H. Corston-OliveL 1998. Identifying the 
linguistic correlates of rhetorical relations. In 
Proceedings of the COLING-A CL '98 Workshop 
'Discourse Relations and Discourse Markers', 
pages 8-14. Montreal, Canada, August. 
Dan Cristea, Nancy Ide, and Laurent Romary. 
1998. Veins theory: A model of global dis- 
course cohesion and coherence. In Proceedings 
of the 35th Annual Meeting of the Association 
for Computational Linguistics, pages 281-285, 
Montreal, Canada, August. 
Dan Cristea, Daniel Marcu, Nancy Ide, and 
Valentin Tablan. 1999. Discourse structure 
and co-reference: An empirical study. In The 
ACL99 Workshop on Standards and Tools for 
Discourse Tagging, pages 48-57, June. 
Brigitte Grote, Nils Lenke, and Manfred Stede. 
1997. Ma(r)king concessions in English and 
German. Discourse Processes, 24(1):87-118. 
Udo Hahn and Martin Romacker. 1999. SYN- 
DIKATE - generating text knowledge bases 
from natural  texts. In Proceedings 
of the 1999 IEEE International Conference 
on Systems, Man, and Cybernetics, volume 5, 
pages 918-923. Tokyo, Japan, October. 
Udo I-I~hn~ Katja Markert, and Michael Strube. 
1996. A conceptual reasoning approach to tex- 
tual ellipsis. In Proceedings of the 12th Euro- 
pean Conference on Artifical Intelligence, pages 
572-576. Budapest, Hungary, August. 
Jerry R. Hobbs, Mark E. Stickel, Douglas E. Ap- 
pelt, and Paul Martin. 1993. Interpretation as 
abduction. Artificial Intelligence, 63:69-142. 
Alistair Knott and Ted Sanders. 1998. The classi- 
fication of coherence relations and their linguis- 
tic markers: an exploration of two s. 
Journal of Pragrnatics, 30(2):135--175. 
William C. Mann and Sandra A. Thompson. 
1988. Rhetorical Structure Theory: toward a 
functional theory of text organization. Text, 
8(3):243-281. 
Daniel Marcu, Estibaliz Amorrortu, and Mag- 
dalena Romera. 1999. Experiments in con- 
structing a corpus of discourse trees. In Pro- 
ceedings of the ACL 'g9 Workshop 'Standards 
and Tools for Discourse Tagging', pages 48---57, 
University of Maryland, USA, June. 
Daniel Marcu. 1998. A surface-based approach 
to identifying discourse markers and elementary 
textual units in unrestricted texts. In Proceed- 
ings of the COLING-ACL'98 Workshop 'Dis- 
course Relations and Discourse Markers; pages 
1-7. Montreal, Canada, August. 
J.R. Martin. 1992. English Text. System 
and Structure. John Benjamins, Philadel- 
phia/Amsterdam. 
Martin Romacker, Katja Markert, and Udo Hahn. 
1999. Lean semantic interpretation. In Pro- 
ceedings of the 16th International Joint Confer- 
ence on Artificial Intelligence, volume 2, pages 
868-875. Stockholm, Sweden, August. 
Holger Schauer and Udo Hahn. 2000. Phrases as 
carriers of coherence relations. In Proceedings 
of the 22rid Annual Meeting of the Cognitive 
Science Society, pages 429--434, Philadelphia, 
USA, October. 
Holger Schaner. 2000. Using coreferences for co- 
herence relations. In Proceedings of the 38th 
Annual Meeting of the Association for Com- 
putational Linguistics, Student Research Work- 
shop, Hong Kong, China, October. 
Steffen Staab and Udo Hahn. 1997. Comparatives 
in context. In Proceedings of AAAI97- the 
lgth National Conference on Artificial Intelli- 
gence, pages 616-621. Providence, USA, July. 
Keith Vander Linden and James tI. Martin. 1995. 
Expressing rhetorical relations in instructional 
texts: A case study of the purpose relation. 
Computational Linguistics, 21(1). 
Wietske Vonk, Lettic, u G.M.M. Hustinx, and 
Wire H.G. Simous. 1992. The use of referential 
expressions in structuring discourse. Language 
and Cognitive Processes, 3/4(7):301-333. 
Bonnie Webber, Alistalr Knott, Matthew Stone, 
and Aravind Joshi. 1999. Discourse relations: 
A structural and presuppositional account us- 
ing lexicalised TAG. In Proceedings of the 37th 
Meeting of the Association for Computational 
Linguistics, University of Maryland, June. 
William A. Woods and James G. Schnaolze. 1992. 
The Kt,-ONE family. Computers ~4 Mathemat- 
ics with Applications, 23(2/5):133-177. 
