Intentions, hffornla,tion, and Inference: 
Two Rhetorical Questions 
.Ion ()l)erlander* 
Centre for Cognitive Science and Human Cominunicatiotl Research Centre, 
University of Edinburgh 
1 Introduction 
Moore and Pollack \[1992:540\], following Grosz and Sidner \[1986:201\] point to the possibility that 
interpreters can often infer intentional relations fi'om informational relations, and vice versa. I 
originally intended to explore the nn.ture of these inferences in the framework of SDRT and DICE, 
and show how they go through in some cases, but not in others) The details of this modelling turn 
out to be interesting, 1)ut not wholely surprising. I will theretbre take the inference-based model for 
granted, and pose instea.d two rhetorical qttesl.ions that arose whelL working it out. The questions 
are: first, when is an intentionM relation not a.n intentional r(da.tion? And secondly, when is an 
inforinational relation not an ilfforma,tional rela, tion? 
2 A question of intention 
(1) Nobody seems to ca.re al)ont any modern poet nowadays except .lohn Betjema,n, who 
writes agreeably in praise of buttered toast and railway sta.tions, and bet:ame a best seller 
almost By Appointment after Princess Margaret said she liked his verse. \[LOB A19:122- 
125\] 
*The support of the Science ;rod Engineering Research Council through projt:ct lllllllbel: GR/G22077 is gratefully 
acknowledged. HCRC is supported by the Econontic and Social Research Council. My thanks to R.obert Dale, Alex 
L;mcarides and Johanna Moore for helpfifl discussions. Email: J.0berlander~ed.ac.uk 
IOn SDKT, see Asher \[1993\]; on COltllllOltSel|Se Entailment, see Asher and Morreau \[1991\]; on DICE itself, see 
I,;Lscaridcs and Asher \[1991\] or most recently, L;Lu¢:;tridcs and Oberlamler \[1993\]. 
98 
(2) a. Betjeman writes agreea, bly in pra,ise of buttered toast a.nd railway stations, 
b. and became a best seller a.hnost By Appointment 
c. after Princess Margaret said she liked his verse. 
Which relations hold between the elements of (2)? If we are interested ill tile domain of time and 
causality, the informationM relations are easy enough: the eventuality of (2a,) temporally overlaps 
with those of both (21)) and (2c); (2b)'s event temporally succeeds (2c)'s, a.nd is caused by it. But 
which intentionM relations hohl between thenV, t 
In Lascarides and Oberlander \[1993\], we analysed the l)resuppositiomd beha.vi()ur of temporal con- 
nectives like after, and proposed a. model ill which subordinate cla.us(,s like (2(:) a.re accomntodated 
via discourse attachment. As usual, various defeasil)le inferences go through, aad the conclusion is 
that the following discourse relations hold between the SDR.Ss: Backg,vund(2a., 2c), Narlution(2c, 
2b), and Result(2c, 2b). Notice: the inference process delivers not one hut l',,o discourse relations 
holding between (21)) and (2c). So it, seems tha.t we have gone a. step further than Moore and 
Pollack: we not only agree with their Mullih'vrl cla.im; we a,lso go on to make a, Multirchttion cla,im: 
more than one intentiona.l-level r(,la, ti(m ca.n simulta.neously hohl Imtween l.w(~ discourse segments. 
In response, one could invent a new discourse relation, Narratiou'n'Rt'.~ull, which wouhl hohl in 
just these cases. But its tack of independent motiva.tion, a.nd its resemblance to Knott ~rn(I Dale's 
\[1992:7\] inform-accidcnt-and-mcntion-fi'uit rela.tion is undermining. Alteruately, one could argue 
that the multiplicity of discourse relations arises because we alrea.dy have too many intentional 
relations, rather than too few. On this account, we should dipose of (sa.y) Result, and make do 
with the intentionM relation Narration, and various informational rela.tions, like cause, to cover the 
cases. There is something to be said for this view, but I won't say it here. 
Rather, I would defend the Multirela.tion cla.im by observing that it is the ua.tural concomita.nt of the 
multiple intentions served by single segments of discourse. We entertain as welt as describe; impress 
as well as convince. Grosz and Sidner \[198{i:178\] 1)oint out that the a,ssumption that each segment 
has but one purpose will "in the end prove too strong". So segments serve multil)le orthogonal 
goals; and this means that a. segment ma.y Ih.ll into multiple rela.tionships, (,v(,n with a single other 
segment. 
Thus, it shouhl be no surl)rise tha.t a. discourse sl.ru(:tllre tlmoi'y can (hdiver m.r~, tha.n one intentiona.l 
relation holding between two discourse segments. Indeed, on this stt~ry, it. is much more surprising 
that some theories do not. 
3 A. question of information 
99 
(4) 
This may seem to carry the iml)lication that the knowledge ill question is a.cquired without 
observation. The fact, if it be a t'a.ct, that I take lol,ger steps left l'oot forward wouhl not 
\]lave any bearing ()n the ca.t'e with which I might illvestiga.te the ma.tter; l might make 
my measurements carelessly and gel; the wrong answer. But where I intend something it 
seems to be guaranteed tha.t I couh\[ not get a wrong answer, so it seems as though we 
nmst know our own intentions independently of obserw~.ti(m. \[LOB G63:96-103\] 
a. But where I intend something 
b. it seems to be guaranteed that I could not get a wrong answer, 
c. so it seems as though we must know our own intentions iztdependently of ohservation. 
Which intentionM rela.tions hold between (41)) a.lM (4c)? (This datum is a real-liik: version of Moore 
and Pollack's Bush example.) The stra.ight|brwa.rd answer al~l)ea.rs to b(,: (.'on.~cqu~:ucc(4b, 4c) and 
Backyround(4b , 4c) (cf. Oberlander \[1993\]). The (.'on,~cque:ncc relation is an evidential relation, in 
which the first-mentioned seglnent is taken to supply evidence tot the secolM-mentioned seglnent. 
It's a different relation froln Evidcn, ec, because the direction of evi(h~ntial SUl)port is reversed. 
Now: suppose that, like the original Bush example, our Incorrigil~le example lacked the explicit 
connective so. What difference would this make? The a.nswer is: we would lose the clue as to which 
of (41)) and (4c) was the directive part (of'. Elhadad \[1992:204\], who adapting Halli(lay, distinguishes 
between directive and subordinate segments). Thus, even with exactly the same causal structure 
beneath it, the argument could be inverted: we could be mentionillg (4c) to l)rovide evidence for 
(41)), rather than vice versa.. But 'when wollld th(~ argument be iHverted? In Grosz ~md Sidner 
\[1986:201\], there is an equivalence between the intbrmational relati(m supports and the intentional 
rel;ttion Dominates. Under DICE's inference regime, the flip I)etwe(m (,'on.~cquencc and Evidence 
tracks the flip in the direction of supports. 
But how can a doma.in rela.tion like supports "change direction"? We agree that it's "not iml)lica- 
tion"; but it's not even just defeasiblt: intplication. It's like implicatioll, in that it concerns transfer 
of degrees of belief; but unlike most implications, there need 1)e nothing "out there" to make one 
direction right, and the other wrong. A spea.ker and hearer may di\[:l'er as to whether p causes q 
or vice versa. Only one of them will he right on each occasion. But if they dift'er as to whether 
p supports q, they can both be right; it just depends on their individual networks of prior beliefs. 
And if this is so, supports is not an informatio~,a.l---(lomain---rela.tiol. (:m a pa.r with causes. If it is 
a (lomMn relation, a theory of discourse structure must esta.I)lish whi,'h doma.in it's a relation on. 
4 Rhetorical Answers 
Q1 When is an intentional relation not aa intelktional rela.tiotl? 
100 
- When it's two intentional relations. 
Q2 When is ,~n informa.tional relation not a,n informational rela.ti(m? 
- When it's supports. 

References 
Asher, N. \[1993\] Reference to A b.stract Objects in Engli.~h: A Philosophi,'al ,¢h'manties for Natural 
Language Metaphysics. Dordrecht: Kluwer Aca.demic Publishers. 
Asher, N. and Morreau, M. \[1991\] ConmLon Sense Entaihnent: A Modal Theory of Non- 
monotonic Reasoning. In P~w:eedings of the 12th lnternatioual Joint C'onJ~.:rence on Artificial 
Intelligence. 
Grosz, B. and Sidner, C. \[1986\] Attention, Intentions, and the Structure of Discourse. Com- 
putational Linguistics, 12, 175-2()4. 
Knott, A. and Dale, It. \[1992\] Using Linguistic Phenomena to Motiva.te a Set of Rhetorical 
Relations. HCRC/ltP-39, Huma.n C, ommunication Research Centre, December, 1992. 
Lascarides, A. and Asher, N. \[1991\] Discourse Relations and Defea.sible Knowledge. In P1o- 
ceedings of the 29th Annual Meeting of A.~.sociatio~ for Comp'utatio'ual Linguistics, 1)p55--63. 
Lasearides, A. and Oberlander, J. \[1993\] Temporal Connectiw,s in a Discourse Context. In 
Proceedings of the 6th Confe'rcncc of the E'm'opcau Chapter of the A.~.sociulion for C'omlmlational 
Linguistics, Utrecht, Netherlands, April 2 lst--23rd, 1993, pl)2(i0-2(i,~. 
Moore, J. and Pollack, M. \[1992\] A Problem tor RST: The Need \['or Multi-Level Disc(mrse 
Analysis. Comlmtational Linguistics , 18,537-544. 
Oberlander, J. \[1993\] Beliefs a.nd Intentions iu the Abductive Gellel'~l.tion of Discourse. To 
appear in Proceedings of the Thild h~te'rnational Colloquium o'n Cognitive .%icuce, DonostbL-Sa,n 
Sebastian, May 4th-8th, 1993. 
