Investigating Discourse Relations 
Meg~n Moser Johauna, D. Moore 
University of Pittsl)urgh 
Learning Research and Development Center 
Pittsburgh, PA 15260 
moser~J~pogo, isp.pitt, cdu 
University ot" Pittsburgh 
Depa.rtment of (',Omlmter Science, and 
Lea,rning Resea.rch a, nd Development Center 
Pittsburgh, PA 15260 
j'moolr'@c.~, pitl. rd'u 
Ill this paper, we present the theoretica.1 foun(la.tions which underlie our current research. We 
emphasize a methodologicaJ point: connections among proposed theories of discourse structure must 
be carefiflly explicated. By making these connections, we ca.n (letermilm whether theories ma.ke 
equiva.lent claims, consistent 1)ut unrelated cla.ims or differ iu substa.ntive issues. A synthesis of 
current theories is required, a descriptive fl'a.mework which assumes the common ground and which 
highlights issues to be resolved. Further, there a.re two distinct :lh.ctors to guide us ill developing 
such a synthesis. First, the descriptive fl'a.mework nmst facilita.te the fi:)rmula.tion of controversi;d 
issues in terms of empirical predictions. Then, the synthesis ca.n I)e developed a.nd refined by 
linguistic studies. Second, the fr~uuework must be a.pplicaJ)le to computer processing of na.tural 
language, both understa.nding and gener~tion. Our particula.r interest is the computer generation 
of explanations in a. tutoring system. Idea.lly, the fi'a.mework will provide a. me~ns through which 
the results of discourse a.nalysis and computatiorml linguistics ca.n inform one another. 
In its broadest outline, the goaJ is to uu(lersta.ad the precise i~H:era.ction between fea.tures of 
tbrm, meaning a.nd ffim(:tion ill the crea.tioll of discourse coherem:e. Wha.t kiuld o1' tbrm, mea.ning 
~tnt(i flmction links occur \])etween uttera.Jlces a.nd how a.re these thre~, ki,l(Is of links recognized? In 
a tirst step towards the synthesis we wouhl like to see, we will discuss the l~erspective which the 
G&S (Grosz and Sidner 1986) and RST (=RhetoricaJ Structure Theory, Ma.nn and Thompson 1988) 
theories take on links of meaning a.nd function. We conclude with a brief description of a.n empiricaJ 
study suggested by this theory compa.rison. Note tha.t we consider only monologic discourse at this 
time, believing genera.lizz~tions between this a.nd multi-agent discourse to be prema.ture. 
In the study and discussion of rhetorica.l relations, the terminology ha.s become nonsta,ndard 
a.nd confusing. Here, we a.dol)t the term "discourse rela.tions" to mea.n all the conn(,ctions a.mong 
the exl)ressions ill (liscourse which, ta.keu tog(,th(~r, a.ccount 1or its ('()\]lerence. So, in order to be 
a. (liscourse relation in the sense used here, two criteri,~ a.re requir(,(I. One, the rebttion concerns 
elements of a sentence utterance a.nd other utteraJtces in the context. Two, the rel;Ltion must be 
recognized in order to understand the discourse, i.e., it contril)utes to coherence ra.ther than another 
concern such as style. As suggested above, SOlne fea.ture of form, mea.ning or fimction defines a 
discourse relation. These will be termed textual, informa.tiona.1 a.nd intentionaJ discourse relations 
respectively. Informational a.nd intentional discourse rela.tions a.re essentia.lly non-linguistic in the 
sense that they do not originate with la.nguage. Mffch recent work on discourse relations either 
explicitly discusses or implictly uses a. distinction between in|brma.tionaJ and intentiona.1 relatious 
(Schiffrin 1987, Redeker 1990, Hovy and Ma.ier 1992, Moore a.nd Polla.ck 1992, Moser 1992, Sanders, 
Spooren and Noor(hnan 1992, inter aJia.). The distinction, a. kind of sema.ntic-l)ragmatic distinction, 
concerns the source of discourse rela,tions, whether ,~ rela.tion orgina,tes with what is being talked 
about (informationa.1) or with why we a.re taJking about it (intentionaJ). Inibrma.tionzd discourse 
relations arise bec,~use the meanings of exl~ressions in utt(.~ra.n('(-,s, the things 1)eing ta.lked about, 
stand in some relation ill tlle domain of discourse. CAUSE, \['or exa.mple, is a.n iuforma.tiona.l rela.tion 
because it is a rela.tion between things tha.t a,r~, being ta.lked a.l)oul., tim fa('l, of on(~ situa,tion or 
94 
ew:nt causing another. Intentional relations, ill coiltrast, concern how one spa ll is intellded t() affect 
the bearer's mental attitude about another, i.e., why the speaker included some span of text. For 
example, the EVIDENCE relation holds between two text spans if one is int(~ll(led to increase the 
heater's belief in the other. 
GaS take a spea.ker's plan to be the source of discourse structur(~. A general intention may 
dominate several subintentions which may theutselves be fiH'ther ,'~,lil,~,(I. At the bottom of the 
hierarchicM structure are intentions which are rea,lize(I by 1)roduci,,g iltter;,.llC(~S ill the discourse,. 
Intentions higher in the intention structure are realized by the subintentions they dominate and, 
possibly, additional utterances which express the higher level intention. Intentions are related I)y 
dominance, when one intention generates one or more sul)intentions, or by satisfa.ction-preceden(:e, 
when the realization of one intention is a. precondition ior the realization of a.llother, hi this theory, 
spans of text are related indirectly by the relatiolls I)etweell the intc~lll.iolls they rea.lize. 
In the original formulation of RST, the inibrmational-intentional (listillction wa.s noted (th(,re 
called "subject matter" and "l)resenta.tiona.1 '' relations), but wa.s not fully a,l)precia.ted. Text is 
hierarchically structured and exa.ctly one R.ST relation is postulated 1)etwe('ll contiguous spans. 
Moore and Pollack (1992) 1)ropose that, iH th.ct, two contiguous spans of text may be ill both 
an informational and an intentional relation simultaneously and that recognition of one kind of 
relation can facilitate recognition of the other. They note that inteHtiona.1 relations arise because 
"consecutive discourse elements are rela,te(I to one a.~lother 1)y mea.~s (~\[' tlm ways in which they 
participate ill \[a spea.ker's\] plan." (1). 2). A~ i~tentioHal rela.tiol~ i~(lic:~.tes I)()th ~, direction of elrect 
and a kind of effect. Both elements ()f' a.n intention~d rela,tioll iJl R ST c(~rt'esl~Oll(I 1,o elements in th(, 
relations among intentions in G&,S. 
First of all, one span is contributed with the intention of affecting the purpose of another 
span. In the original RST with only a single relation between Sl)ans, the direction of the rela.ti,~n 
(from satellite to nucleus) always represented the direct|oH of efiect, h, i l,troducing simultaneous 
intentional and informationa.1 relations, nucle;:l.rity is an a.spect of intenti(:)md rela.tiolls only. The 
direction of an intentional relation ill I(.ST corresl~onds to dominance ill G£;S. Tha.t is, the satellite 
span, S, affects ttle 1)urpose of tlle nuclells sinful, N, ollly if the intenti()ll that: S re~l.lizes in (lonlim~.ted 
by the intention that S and N (and possibly others) realize together. 
Second of all, in RST, one span is intended to a.fi'(:(:t another in ol~e of sevel'al ways. For exa.mp\]e, 
a satellite span, S, may be intended to ;~.fli~ct a he~.rel"s I)elief in the ml(:le~s spa.n, N, (the EVIDENCE 
relation); or S may be intended to a.trect a. bearer's desire to perform th(, ~.(:tiou i~t(lical.ed by N 
(the MOTIVATION relation). (.',orr(~sl~on<lihg\]y ill (;X:S, intelltio~s ol" w~.~'i~)~ts kinds m~l.y domim~.l,e 
other intentions. Roughly sl)ea.kihg , intentio:ts re,dAzed 1)y sl~ea.kin~ ~.r(~ I o (,ithe~" a.lrect a he;~.rer's 
beliefs or her actions. All RST EVII)ENC. E relatio~ c;~,n occur onlywhe~ th(, d(~mina,ting intentio~ 
is to affect another's belief. Similarly, a.n R ST MOTIVATION rela.tion ('a.n occur only when the 
dominating intention is to affect a.nother's a.ctio~. Thus the different kinds of intentional efl~ects in 
RST correspond to different kinds of dominating intentions ill G~S. However, RST makes more 
distinctions among kinds of effects thal~ G(~S, e.g., 1)oth EVIDENCE and .IUSTIFICATION are ways of 
affecting beliefs and both MOTIVATION and ENABLEMENT are ways of a.ffecting actions. 
Now we turn to the status of intbrmational relations in the two the(~ri(:s, h~ R.ST, a singh' 
informational relation is assigned 1)etween sl)ans of text. As a result, RST iHformational rela.ti(ms 
concern the sorts of entities denoted \])y entire spa~s, such as situ;~.l,it~ns and events. If we ad(~pt 
the view of Moore and Pollack, recognizing that inlbrmationa.1 ~l.tl(I intentional relati(~ns occur 
simultaneously, the informational relations are silH1)lified. As noted a.bove, m~clearity, which was 
an element of tile unique RST relation between spans N and S, is now a.n element of the intentiol~a/ 
relation and is independent of the intbrm~.tional rela.tion. Tha.t is, informatiom~.l relations no long~r 
conflate the sem~tntic link between situa.tio~s eXl)r(~sse(I I~y two sl)a.~s ;~n(I the domin~Lnce of Sl)~a.k~r 
95 
intentions realized by those spans. In fact, without this conflation, it is unclear whether the semantic 
relations hetween situations and events is in principle diffe.reut fl'om semantic links hetween other 
kinds of entities. Is the informational relation between two spans of text necessarily a relation 
hetween the entire spans? Or, is it l~ossibh, that the infornmtiolml relation is a series of links 
between wl.rious expressions contairmd in th(, Slm.lls? Examl~les of 1.1,,se links I~etween smaller 
constituents are would are coreference anl(I ;i. ca.usal link I)etw(,(~J~ (,×l~r¢,ssi¢~Jls such as "l'Ul~ a red 
light" and "the ticket." 
G&,S stresses that intentions of the speaker ar¢~ the primary sour((, of (lis~'ours(~ structure. Do- 
main knowledge plays a role in recognizing the intentional structure, but it is (Ioubtful whether any 
special distinction wouhl be made betweezl khowle(Ige about events a.tkd sit.u~l.tions arid knowledge 
about other kinds of domain entities. 
In comparing the status of intentional a.lld intbrmation~d discourse rela.tiozLs in I)oth RST and 
G,~,S, at least two issues were specified whose resolution in currently pl'esume(l by I)oth theories 
in isolation. First, what types of clominalw.e are distillguished by the theory- just dominance, 
dominance hy beliet:affecting intention versus dcuHina.uce by actiou-alli~ctittg intention, or the full 
range of RST intentional relations? Second, do semantic links between whole spans of text play 
a role in the theory, or do semantic relations between all sorts of eud.ities have the same status? 
Both these questions can be answered sel)ar,a.tely del)ending on whether they are meant as a ques- 
tion ahont linguistic theory or ahout its al~l)lication to language processing. Space prechtdes a 
discussion of textual discourse relations as well as questions about how the three kinds of relations 
interact. Further, proposals from linguists (Schiffriu 1987; R.e(leker I992) must be integrated into 
the synthesis. 
We conchtde by sketching a discourse analysis study which begins to address the first question 
cited above, one which we plan to do. Using a broad range of text types, we select as tokens the 
pairs of spans that are related by lexical markers of discourse relatioHs such as "so," "hecause" and 
"therefore." For each token, we code the intbrmatioiml a.nd inteHti(:mal re\];i.tions that co-occur with 
it. As emphasized by Moore and Paris (1992), there is not a, one-to-olLe nlapl)illg between intentional 
and informational relations. Though not COml)letely indel)en(lent, the i~ossibilities to1" intbrmational 
relations given the occurrence of a particul~.r intentional relatioxD ~l.t'e manhy, and vi('e versa. By 
investigating the range of combinations of informa.tional arid inteld:ioH;fi relations which occur 
with a lexical marker, we cast identity the miHim;d descriptioll of the marker in terms of discourse 
relations. That is, we can see whether a marker correlates highly with a l)articular discourse relation 
or combination of informational and intentional relations. Such a study will l)roduce results that 
are uselqd for coinl)utational models of hoth llatul'al la.ngu~lg(, uzk(lersl.aL~dil~g az~d geHeration. If 
a certain marker correlates highly with c(~rtai~l discourse relations, the u~dersta, n~der can form 
hypotheses about the discourse relations that are present whez~ it ellCOUld;ers a marker. Similarly, 
if a generator must express two utterances connected hy a certain cotubin~a.tion of relations, it can 
use the information about the correla,tion 1)etween markers and cond)inations of discourse relations 
to dmose the most appropriate marker. 
In addition, this study will allow us to determine whether the distinction among types of 
dominance plays a, role in accounting tbr the distribution of markers. If the study confirms the 
relevance of this distinction, then (lomina~,:e types have a role t() play i~ a linguistic theory of 
discourse coherence. If the study does not ctmtirm the distinction, ;I. ct'ucial theoretical question is 
whether there is any direct formal linguistic pattern other than these, markers that makes essential 
use of dominance types in its description. Should fiH'ther research fail to identify such a pattern, 
then the status of donfinance types in linguistic theory in called in~t(~ question. I\[owever, even if it 
turns out that dominance types are Hot forn~ally m~.rked by la.nguag(~, it remains a.l~ open question 
whether their application to computer generation an~d tm(lerstan(li~g will b(, useful. 
96 

References 
\[1\] Grosz, B., 1978. Focusing in Dialog, Tinlal)-2: Theoretical Issues in Na, tura.l Language I'ro- 
cessing 7: 96-103. New York: Association for Coml)uting Machinery at,I Associa.tion of Coml)U- 
tational Linguistics. 
\[2\] Grosz, B. J. and Si(iner, C. L., 1986. Attention, intention, al|(I the structure of discourse. 
Computational Linguistics 12(3):175-204. 
\[3\] Hovy, E. and E. Maier, 1992. Parsimonious or Profliga.te: How Many and Which Discourse 
Structure Relations. ms. Submitted for publication. 
\[4\] Mann, W. C. and Thompson, S. A., 1988. Rhetorical Structure Theory: Towards it flmctional 
theory of text organization. TEXT 8(3):243-281. 
\[5\] Moore, J. D. and Pollack, M. E., 1992. A problem tbr R ST: The need for multi-level discourse 
analysis. Computational Linguistics 18(4). 
\[6\] Moore, J. D. anti Paris, C. L., 1992. Planning Text For Advisory Dialogues: Capturing Inten- 
tional and Rhetorical Information. University of Pittsburgh, Departmellt of Computer ScieJlt:<~ 
Technical Report Number 92-22. Submitted l'<)r Iml)li(:ation. 
\[7\] Moser, M., 1992. Analysis of conversational arguments. Presentation to Wa.shiugton Linguistics 
Society, Washington, DC, November 12. Workiag pal)er. 
\[8\] Redeker, G., 1992. Coherence and Structure ia Text and Discourse. ms. Under review. 
\[9\] Redeker, G., 1990. Ideational a.nd Pragmatic Markers of Dis('t, urse Structure. .lourn,l of 
P1ugmatics 14: 367-81. 
\[10\] Sanders, T. J. M., W. P. M. S. Spooren alld L. G. M. Noordumn, 19!)2. 'lbwa.rds a Ta.xononly 
of Coherence Relations. Discours(: Plvccss~..s 15: 1-35. 
\[11\] Schiffrin, D., 1987. Disco'arsc' Markers. C,a.nd)ridge: Cambridge University Press. 
