Rhetorical Relations: Necessary But Not Sufficient 
Sandra Carberry Jennifer Chu Nancy Green 
carberry@cis.udel.edu jchu@cis.udel.edu green@cis.udel.edu 
Department of Computer Science 
University of Delaware 
Newark, Delaware 19716 
alld 
Lynn Lambert 
lambert@pcs.cnu.edu 
Department of Physics and Computer Science 
Christopher Newport University 
Newport News, Virginia 23606 
It is generMly agreed that coherent (lis('ourse consists of segments tha,t are related to one 
another. A number of researchers have a, rgm~d lbr the use of rhetorica,l\[(~ri75\] or coherence relations 
\[I-Iol)79\], and the rhetorical relations specified by R S'I? \[MT87\] have I)e(,~, used in structuring text 
\[tlov88, MP90\]. In this l)a.l)el" , we exa.mine rh(q.ori(:al relations in the cold.~,xt of dia.logue, ra, tlmr 
than single-speaker te×t. We argue that rea,s(millg about relational l)r()lmsit, ions is necessary but 
not sufficient ior structuring (lialogtm, 1)oiHt out s(,v(q'al prol)h,ms of a, pplyiug RST to dialogue, a, nd 
argue tbr the necessity of recognizillg the intentions underlying utteratlces a,lld I, he rich relationshil)s 
among these intentions. 
Our research on recognizing expressions of (hml)t and interpreting ilMirect replies provides 
evidence that what Moore a,nd Polla,ck call i lfformal,itmal level rela, ti(ms \[Ml'!)2\] play a,n illlporta, llt 
role in identil~ing intentions ill (lia.h)gue. (r~,llsi(l(~l' s()me ('ontiilua.l.i~),s of I.h(~ Iblh~wing (liah)gm, 
sequence: 
(l) S l: "Who i.~ leaching (,'.5':?60 (a sy.~tcms cowrm )'./'" 
(2) $2: "Dr. ,5'mith is t~'achi,,g (,',5';760. " 
(3) SI: "Isu't Dr. Smith on sabbatical:e" 
(4) $2: "No, hc 's not on .s,bbalical. " 
If $1 responds with "Wasn't Dr. Smith aw,trdcd a ~;albright?', then since it is plausilfle that. 
Dr. Smith having a Full)right would result ill his going on sa,l)l)a, tical, Sl's utterance should Im 
interpreted as exl)ressing doubt at the l)rOlmsition that Dr. Smith is ii(:)l, (m sabbatical. Similarly, 
if heing a systems person is necessa.ry for teaclfillg C,$360 and Sl reslmll(Is with "Lsn't Dr. Smith a 
theory person?", then since Dr. Smith being a, theory person in an alterna, l, ive t,o Dr. Smith being a. 
systems person, Sl's utterance shonhl be interpreted as exl)ressing (h)ul)t at the l>roposition that l)r. 
Smith is teaching CS360. But if SI had instead asked "l.~u't C'$240 a prcrcq'ui.~itc for CS360~', Sl's 
utterance should be interpreted as seekillg new itlforma.tion since there is ,,o pla.usibh~ intbrmational 
relation suggesting that the new utterance contril)utes to (leteu'mitfiug wht,tlmr Dr. Smith is ~)nt 
sabbatical or to identifying the instructor of (',S3(i0. Thus, identilying the illformational relationship 
between utterances is essential for responding apl~rt>l)ria,tely. \[LC,92\] I~rovi(tes a,n algorithm tha.t 
utilizes such relations in recognizing expressions of doul)t. 
Our research on interpreti~g iudirect rel)lies \[GC'92\] l)rovid~'s a.(hlitio,~al evidence that in- 
h)rmational relations play an iml)ortant role in discourse iutel'l)r(,ta.tio~. (',onsider the tbllowing 
dia.|ogue sequence: 
(5) SI: "Did Harry scc his girlJ)'icnd last u,~.t'kcm, dW' 
(6) S2: "He: h:ft for New York on Friday. " 
We have shown \[GC92\] that S1 and S2's shaxed a.ssumptiol~ tha.t S2's iutlirect answer in (5) is 
coherelDtly related to sonte possible dir(~ct a.usw~r is uecessary 1.o (,n;ibl(~ S 1 I.o identify S2's intended 
(but implicit) answer of "Yes" or "No". l;k~r example, SUl~pt)se Sl I)(dieves tha.t it is mutually be- 
lieved that Harry and his girlfriend I)oth live iu Deliver. Then, S1 w(mld interl)ret ((i) as describing 
a.n alternative to Ha,rry seeing his girlfriend a.,M thus would iul'(~r a. n(,ga.l.iv(' a.nsw(~r to (5). ()n 
the other hand, SUl)l)ose SI believes tha.t it is mutually believe(I that l\[arry lives in Denver but his 
girlfriend lives in New York. Then, $1 would iuterpret ((i) as (lescril)ing a.u action enabling Harry 
to see his girlfriend and thus would iufer a l)ositive answer. In ol.\]mr words, Sl's recognition of 
the relational l)roposition which $2 intended to convey is nec(,ssa.ry lk)r iuterl)reting S2's intended 
answer. If $1 failed to recognize such a rela.tion, then SI would I)e unable to answer a question 
such a.s "Why didu't Ilarry st:t: his girlfriend la.~t 'weekend?" Thus in our mod(-d \[GC92\], the al)pli- 
cability conditions of discourse plan operators tbr a.nswering a Yes/No question include conditions 
requiring that COml)onents of the response (whether iml)licit or explicit) be related by coherence 
relations similar to informational-level rhetorical relations. Since the sa.me set of operators is used 
in generating answers, inforlnational relations also play a role in ensuring that an indirect answer 
is approl)riate. In other woMs, the intbrmationa.1 relations coustrain wha.t ::xtlu iuibrmation (infor- 
mation not specifically requested by the questioner) may be included in au a.l)prol)riate resl)onse. 
Alk indirect answer is genera.ted by use of the ~'xlra iM'ormati(m ~l.h)lJ~. 
In our view, this extr;~ intbrmation m;~.y serve a rhetori('a.I t'u ncl.iol, such as to increase belief. 
For example, $2 might have decided to include the extra iuforma.tion in ((i) (tha,t I-larry left for 
New York on Frida.y) in his negative a.nswer because $2 anticipate(I tha.t SI would have doubted 
a simple "No". In our current research, we a.re defining stimulus (:(:mditioJls for use of the extra 
iJfformation in an answer. A stimulus condition describes a situa.ti(m i,~ which it may 1)e beneficial 
to inch,de iMbrmation that was neither requested m)r wha.t tim Sl)(,ak(~r Ii~.(l a. prior intention to 
convey. Thus, stimulus Coliditions are used to s(,lect ~.u a,I)l)r~)t)ria.t(' it,l~)rlual.ioua.I r(,l;J.tion. 
\]Iowever, we contend that rhetorical r(~la.ti()ns, eslwcia.lly a.s d(,liu(,(I I)y RST, axe insul\[icient 
tbr characterizing dialogue. Although (~ne might consider (level()piug a l;u'ger s~t of rhetoric;d 
relations \[Hob79\], the tollowing diah)gue can n(.)t b(, COml)letely ('ha.ra.~'t.(,riz(,(I by rlmtorical rebttions 
1)etween segments. 
(7) $1: "The AI class must b(' ¢'ancclh'd today." 
(8) "John is taking the: course: and h,~' .j'u.~l "wcnt home. " 
(9) $2: "John oJ'tcn cut.~ ¢.la.~'s,..~." 
Although one might argue that there is a Motiwl.te-Volitiona.l-Actiou rela.tion between John cutting 
class ill (9) and John going home in (8), the fimction of uttera,nce (9) is to attack the evidence 
relationship between utterances (8) and (7), which RST cannot account for since RST captures 
rela.tions between spans of text whereas (9) is rel'uting the ilnpli('it l)rol)(,sitiou between (7)-(8), not 
the text itself. 
In addition, rhetorical relatious as (lelivwd iu IT ST rela.te s:ll.ellites l.(~ a mtcl(~us, which to- 
gether form an uninterrul)ted text spa.\]~, l\[()w(,v(,r, lit dia.logu(~, ;~.,~ ill.l.Ol';i.iic(, III;I.y rel~.t(~ to an earli~r 
2 
utterance U yet be separated fl'om U by other utterances that do m)t relat(, to U. In such ('a.ses, 
expectations about speaker iutentions pl;i,y ~. ma.jor role in us,des'sta.t,di,,g, a.s illustrated by tl,, 
following example. 
(10) Sl: 
(11) $2: 
(12) $1: 
(13) $2: 
(14) Si: 
(15) s2: 
"AI is the best course to take ncxt semester. " 
"You should take Comp'ulcr Graphics instead." 
"Dr. Smith is teaching A 1 aml he ',~ wo'ndc'rfM. " 
"Dr. Brown is teaching Graphics and hc',s a great tcuch.c'r." 
"Th.c A I projects sou'nd like' a lot of J'u'n. " 
"Computer G'raphic.~ ha.s you din'win 9 all kind.~ of objccl.~. " 
SI and $2 each declare a claim in utteran(:es (10) and (11). Although (:)n(~ can argue that an 
alternative relation holds between utterances (12) a.nd (13) and between utterances (14) and (15), 
the structure of the dialogue is not c.oml)letely cal)tured by rela.l.ilkg l:hese consecutive pa.irs of 
utterances. Utterances (12) and (14) SUl)l,ort Sl's claim in (\]0), and tltl;es'aslces (13) alld (\]5) 
support S2's clMm in (11). Although one might suggest (lrol)ping the requirement tha.t rhetorica.l 
relations relate a span of utterances, it is unclear how rhetorica,l rela.tions Mone couhl handle 
complex dialogues. We believe that in i(lentit:ving the structure o\[" dialogues such as the ab(we, 
expectations about the discourse goals of the particil)ants must be taken into a.ccount (in this case, 
a speaker supporting his/her own claim and indi'mctly attacking tha.t of the other) and a more 
elaborate intentional structure allowed. 
Furthermore, uttera.nces ca.n silmfltaneously serve more than ot,e function, which would 
require two distinct RST a.na.lyses. (',(:msider the following (liMogue se(iHelwo: 
(16) $1: "Can you com, c to a party at my ho'u.~c ,5'alurday ~tigl~t?" 
(17) $2: "No, I can'l. " 
(18) "I have to 'work. " 
(19) ",.5'0 I am very .~orry, b'ul I must dcelim: your in'vitali,m. " 
Utterance (16) conveys both a. litera.l question alLd an invita.tion. Two RST a.nalyses ;i.re required 
for S2's response, one relating (17) and (18) to the literal questiols aml the other rela.ting (19) to the 
invitation. However, RST requires a single a.mdysis a.nd a hierarchical structure. In addition, RST 
cannot account for the fact that if (19) is omitted from S2's response, it m:l.y still be imp\]ica:l.ed. 
P~ecognizing intentions is essential tot dialogue ullderstan(ling, since these intentions pro- 
vide expectations used in interpreting subsequent utterances and identif~qng the structure of the 
dialogue. For example, we show in \[GC92\] that expectations about discourse goals pla.y a role in 
the interpretation of indirect replies to Yes/No questions. After Sl's request for informa.tion in 
the exchange (5) - (6), $1 and $2 sha,re the exlmctation that S2's response will convey the re- 
quested information. This expectation is used to focus on a. certain set of discourse plan ol)erators 
representing mutuMly accessil)le knowledge of standard forms tbr giving a. positive or negative an- 
swer. Furtherlnore, our dialogue model \[LC91\] captures not only communicative intentions l)ut a.lso 
dolnMn and prol)lem-solving intentions, a.ud these intentions result it, a. set. of expectations tlmt 
facilitate understanding subse(lueld; uttera.ttcos and genera.tillg a.1)l~rol~ria.t(, responses. In \[C, hug:~\], 
we explore response generation in (olla.l)ora.tiv(, (li;dogue. 
Although current resea.rch has been com'er,md with recognizillg tit(, i|lloJd.ions tlsa.t a. spea.k(,r 
is trying to convey, we 1)elieve th;~.t an effective aim intelligent sysl,em must do l|lore. If a sysl,(,m 
is to handle naturally occurring (lia.l(~gue, which can ra.nge from COmlfletely COOl~er;dfive to non- 
COOl)erative in a single interaction, then the system must be ~l.l~le I.o recognize hidden intentions, 
such as the intention to lie or deceive. Although recognition of such hid(le\]! intentions may not be 
essential for identifying the structure of the (lisco!!rse, it is necessary for the system's responses to 
be intelligent, natural, and effective. 
So what makes a dialogue coherent? We believe that dialogue coherence depends upon both 
informational and intentional level properties of the dialogue. We contend that a natural language 
dialogue system must be able to recognize a, speaker's intentions, tha.t this recognition of intention 
is often aided by identification of intbrmational level relational 1)rol)ositions, and thaJ; in many cases 
;L speaker intends for these informationaJ level rela.tions to be recognized (as in the exchange giw~n 
in (5) - (6)). Our research has led us to conclude that rhetorical rela.tions as specified by RST 
are necessary but not sufficient for ha.ndling dialogues -- the rich relationshiI)s among discourse 
intentions must also he captured. 

References 
Jennii~er Chu. Responding to user queries in a. collaJ)(~!'a.tive (qkvironment. In Proceedings 
of the 31st Annual Meeting of tlu' A.~.~ocia.tion Jbr (,'~rm.lralatioual Linguistics, Student 
Session, 1993. To apl)ear. 
Nancy Green and Sandra (:arl)erry. Co!!versa.tional lml)licat!lres in h!direct Rel)lies. In 
Proceedings of the 30th Awwmd Meeting oJ" the Assm'itt/imt for (,'omlrutatio'ual Li1~gui,~tic,~, 
pages 64-71, 1992. 
Joseph E. Grimes. The Thread of Diseour.~(:. Mouto!!, Th(, l\[a.gue, Paris, 1975, 
Jerry R. Hol)bs. Coherence and C, oreference. Cognitivt ,S'cicncc, 3:(i7-90, 1979. 
Eduard H. Hovy. Planning Coherent Multisentential Text. Procecdiugs of the 26th Annual 
Meeting of the Association ftrr Co~lqratational Linguistic,~, pages 163-169, 1988. 
Lynn Lambert and Sandra Carberry. A Tripartite Pla.\]l-ba.sed Model of Dialogue. In 
P~veeedings of the 29th Annual Met!lug of the AC'L, pa.g(~s 47-54, Berkeley, CA, June 
1991. 
\[LC92\] • q ! Lynn Lambert and Sandra Carl~erry. Using Linguistic, World, a.\]!d (,ontext lal Knowledge 
in a Plan Recognition Model of Dia.logue. In Proceedings of tlu' 14th International Con- 
? ference on Computational Lingui.s'tics (COLING-92), pages 310-31(i, Nantes, \[ rance, July 
1992. 
Johann~L Moore and C'ecile Paris. Planaiug Text for A(lvis()ry Dialogues. In Proceeding~ of 
the 27th A'm~ual Meeting of the A.~,~m'ialion for (.'Omlrulaliolml l, ing'u, istic,~, l)a.ges 203-211, 
Vancouver, Canada, 1990. 
Johanna Moore and Martha. Polliwk. A Prol)lem fi)r RST: The Need for Multi-Level 
Discourse Analysis. Computatio~ud Lingui.~lies, 18(d):537 544, 1!)92. 
William C. Ma.nn and Sandra. A. Thoml)SOu. Rhetorica.I Structure Theory: A Theory of 
Text Organization. Tectmica.1 Report ISI/RS-87-190, ISI/USC,, June 1987. 
