Towards stratification of R,ST 
Tanya Korelsky~a.lM Richard Kittredge:~ 
tCoGenTex, Inc., Ithaca, NY 14850 
e-mail: tanya@cogentex.com 
{llniversit6 de Montr6al, Montrda.1, QC, Canada H3C 337 
e-mail: kit tredg(i)iro.umont rea.l.ca 
1 Introduction 
Moore and Pollack have recently given an analysis of R.ST relations in tel'mS of intentional ver- 
sus informational levels of discourse \[5\]. Accol'ding to their (convincing) analysis, presentational 
RST relations correspond to the intentional level of discourse and subject matter relations to the 
inlbrmational level. For each text there shouhl exist two RST analyses: one containing only pre- 
sentational relations and the other containing only subject matter relaLtions. Mann and Thoinpson 
had discussed multiple analyses in RST (Ill, l)P. 2(i-30), stating that a presentational analysis 
is the only analysis when l)oth presentational and sul)ject-matter relation definitions are satisfied 
between a given pair of text spans. In their view, the l)resentational analysis is chosen as correct 
since it describes the changes in the bearer's system of belie£s, i.e. provides deeper insights into 
discourse goals. Moore and Pollack, however, argue that these two a.nalyses should not compete; 
instead, "a complete model of discourse structure lnust maintain both levels of relation". 
In this paper we take this argument one sma.ll step further and suggest a way these two levels 
might be organized into a stratified structure. Our discussion here ha.s a very narrow focus anti does 
not attempt to answer such iml)ortaut (luestions a.s whether the RST collection of presentational 
relations is exhaustive and adequate \[br describing a.ll possible lilt(rational structures. 
2 Analogy with "Meaning" vs. "Means" in MTT 
In Meaning-Text theory (MTT), when there are two (o1' more) alternative analyses of text which 
(according to intution) belong to distinct levels of representation, one should ask whether one of 
tlmm can be seen as a means of expressing the other. Moore and Pollack argue for co-existence of 
intentional (presentational) and infornlatiolLa\] (subject matter) allalys(,s, but they do not attempt to 
describe the relationship between them. They show that there is no one-to-one mapping between 
presentational and subject matter relations, and moreover, that the presentational and subject 
matter analyses may have different structures (e.g., the i)resentational and subject matter relations 
may have opposite directionality, \[5\], pp. 542- 543). Their conclusion is that there is no easy way 
to relate these two levels. 
Froln the viewpoint of MTT, however, the intentional vs. inibrma.tiona.l dichotomy observed by 
Moore and Pollack appears quite natural. An a.nalogy ca.n be seen with the relationshil~ between 
the semantic and deep syntactic levels in MTT: semantic relations are in a many-to-many corre- 
Sl)ondence with (leep syntactic relations. Moreover, there is no isomorl~hism or even preservation 
52 
of structure between a semantic graph and a deep syntactic tree which expresses it. For exaxnple, 
on the semantic level, an adverb such as oftcn is typically analyzed a.s a predicate whose argument 
(i.e., dependent) is the remaining sentence graph. On the (deep) syntactic level, the same adverb 
is a dependent of the main verb. Thus the direction of dependency can easily change as one passes 
from semantics to syntax. 
Following this analogy, we propose to consider the informational level of discourse as a means of 
expressing the intentional level. Intuitively, this seems quite l)lausil)le because inlbrming the hearer 
about informational relations 1)etween discourse elements can hardly be considered a self-sufficient 
goal. Moore and Pollack stress that the intentional level is the primary one since it describes the 
speaker's strategy to achieve her intentional goal. 
In keeping with MTT methodology, to give substance to a claim of separate strata, one must 
produce a system of rules that map the relations of the "meaning" level to the relations of the 
"means" level. Each "meaning" relation can map to many "means" relations (or comi)inations of 
such relations) and vice versa, one and the same "means" relation can al~pe~.r in more than one 
mapping rule. Ea.ch rule describes the contextual conditions tbr its a.pl)lic;l.tion in enough detail to 
justify the distinction between various "means", if such exist. 
At the moment, we can give only a few examl)les of such real)ping rules, olle of which we \[)resent 
below. Nevertheless, we present this very preliminary ana.lysis ti)," l),H'l)()ses of discussion, in the 
belief that work in discourse rel)resentation is crucial for text generation the¢)ry and apl)lications. 
A stratificational view has significant iml)licatioHs. ()lie of these is to eliminate lineal" order 
from the description of intentional RST relations. Note th~l.t in RST a givel, intentional relati~m 
may hold between two text segments a.l)pearing in either order provided th~l.t apl~tol)riate discourse 
connectives are used. In our view, linear order and, correspondingly, discourse connectives should 
appear only on the informational level. A1)stract intentional relations express only the dependellcies 
between propositions, not the lexical or "syntactic" artifacts associated with the means of their 
expression. 
3 Example of One-to-Many Mapping 
The following example, taken fi'om Moore and Pollack, shows a. one-to-many ma.pl~ing from an 
intentional relation to various inibrmational relations. 
(la) George Bush supports big 1)usiness. 
(lb) He's sure to veto House Bill 1711. 
IlL Moore and Pollack's analysis, (la) and (tl,) are connected I)y the illtetltional relatit,ll Evi- 
dence. In particular, satellite (la), is given as Evidence, to increase th(, heater's belief ill nucleus 
(lb). The hearer is assumed to believe (la) or tind it credil)le. 
Moore and Pollack show that the same example a.(hnits a.n analysis on the infi)rnlational level in 
terms of the Volitional Cause relation. In our terms, this wouhl mean that the intentional relation 
Evidence can 1)e realized 1)y the informatiollal relation Volitional Cause under certain conditions. 
If there are other informationa.l relations that are cal)al)le of realizing the same intentional relati()n, 
these conditions should clearly i(lentity the apl)lication context tbr each it~fc)rmational relation. 
Another informationa.l relation capal)le of realizing Evidence is Non-Volitional Cause, as 
shown in text (2a-21)): 
(2a) Winters in Montreal are so cold. 
(2b) (Therefore,) I need a fltr coat (faux, of course). 
53 
To increase the hearer's belief in nucleus (2b), the speaker uses satellite (2a) as Evidence. On 
the informational level these two propositions are connected l)y a Non-Volitional Cause relation. 
Another example of an informational rela.tion realizing Evidence comes fi'om Mann and Thomp- 
soil, who admit that some texts can l)e given analyses in terms of both Evidence and Elaboration 
\[1\]. Consider a Inodification of (la-lb) above: 
(3a) George Bush definitely supports big business. 
(31)) (Look,) He just vetoed House Bill 1711. 
Here, (3a) is nucleus and (3l)) is satellite, given as Evidence to increase the hea.rer's belief in 
(3a). On the informational level these two propositions can I)e seen as connected \])y Elaboration, 
where (3l)) is a concrete instance of a more general 1)roposition (3a). In contrast, in I)oth examples 
1 and 2 the satellite (la,2a) was a general proposition. 
We offer the following sketch of an Intentional-to-informational ma.I)ping rule: 
If the intentional relation Evidence holds between two propositions 
PI and P2, where P1 is a nucleus and P2 is a satellite, 
then, 
if P1 is a general proposition (i.e. the equivalent of a corrlrnoi, set,se 
"law" is given as Evidence), 
then 
if there is a conscious agent such that I)oth P1 and P2 ret~r to her acticJns, 
then the Volitional Cause informational r~la.tiozJ can 1)~ chos~ll; 
else (if there is no agent in PI and P2 as (lescril)e(l above), 
then the Noll-Volitiolml Cause iJktbrmational relation 
can be chosen; 
if P2 is a general proposition, 
then the Elaboration information relation can 1)e chosen. 
4 Example of Many-to-One Mapping 
Moore and Pollack also give an exa.nll)l(:' of 1.(~xt Sl~l.lls whi(:h at(' il, the Condition relation on 
the informational level and, del)ending oil the reading, eithel" ilk the Enablement or Motivation 
relation oil the intentional level. 
(5a) Come home 1)y 5:00. 
(5b) Then we can go to the hardware store before it closes. 
First reading (Condition/Enablement): the speaker is ilLter(,sted ill increasing tile hearer's 
ability to perform the action described in (5h). Second reading: the speaker is just interested in 
motivating the hearer to do (Sa) (because, say, a surl)rise party is l)lanned). 
Viewing this examl)le fi'om our 1)erspective, both intentional relatioHs, Enablement a,n(l Moti- 
vation, can 1)e realized on the informational level I)y Condition. This is quite analogous to the way 
in which an aml)iguons sentence may be the common syntactic realization ot" two distinct semantic 
structures. Also, there is no linear order (or t.he'n colDnective) on the intelltiona\] level. But there 
is a direction of del)en(lency of satellite Oll m~cleus. Tim directiollality of the (lel)en(lency relation 
is preserved in the Enablement/Condition pair, and is revers~d ill the Motivation/Condition 
pair. As noted above, (lel~en(lency reversal is not surl)rising wh(,ll ()lke maps between distinct levels 
of description. 
54 

References 
\[1\] Mann, W. and S. Thompson (1987) Rhetorical Struct'ar¢: Theory: A Theory of ?~xt 07yani- 
zation, Technical Report No. ISI/RS-87-190, University of Southern (~a.lilbrnia, hfformation 
Sciences Institute. 
\[2\] Mel'~:uk, I. (1981) "Meaning-Text Models", Annual Review off Anthropology, vol.10, 1)p.27-62. 
\[3\] Mel'~:uk, I. (1988) Dependency Syntax: Theory and P~uetice, State University of New York 
Press. 
\[4\] Moore, J. and C. Paris (1989) "Planning Text for Advisory Dialogue.~", Proc. of the 27th 
Annual Meeting of the Association for CTmqmtational Linguistics, Va.l|COllver, 1)1).203-211. 
\[5\] Moore, J. and M. Pollack (1992) "A Problenl for RST: Tile Need for Multi-Level Discourse 
Analysis", Computational Linguistics, vol.18, no.4, 1)p.537-544. 
