Preventing False Temporal Implicatures: 
Interactive Defaults for Text Generation* 
Jon Oberlander and Alex Lascarides 
Centre for Cognitive Science and Iluman Communication l{csearch Centre 
University of Fdinburgh 
2 Buccleuch Place, F, dinburgh Ell8 91,W Scotland 
Introduction 
Given the causal and temporal relations between 
events in a knowledge base, what are the ways they 
can be described in text? 
Elsewhere, we have argued that during interpreta- 
tion, the reader-hearer H must iufer certain tempe, 
ra\[ information from knowledge about the world, lan- 
guage use and prugmatics. It is generally agreed that 
processes of Gricean implicature help determine the 
interpretation of text in context. But without a no- 
tion of logical consequo_nce to underwrite them, the 
infercnccs~ftcn defea~sib\]e in nature will appear 
arbitrary, and unprincipled, llence, we have explored 
the requirements on a formal model of temporal im- 
plicature, and outlined one possible nonmonotouic 
framework for discourse interpretation (La.scarides & 
Asher \[1991\], Lascarides & Oberlander \[1992a\]). 
ttere, we argue that if the writer-sllcakcr S is to 
tailor text to H, then discourse generation can be 
informed by a similar formal model of implicaturc. 
We suggest two ways to do it: a version of \[\[obbs et 
al's \[1988, 1990\] Generation as Abduction; and the 
Interactive Defaults strategy introduced by aoshi et 
al \[1984a, 1984b, 1986\]. In investigating the latter 
strategy, the basic goal is to determine how notions 
of temporal reliability, precision and coherence call 
be used by a nonmonotonic logic to constrain the 
space of possible utterances. We explore a defea.sible 
reasoning framework in which the interactions be- 
tween the relative knowledge bases of S and H helps 
do this. Finally, we briefly discuss limitations of the 
strategy: in particular, its apparent marginalisation 
of discourse structure. 
The paper focuses very specitically on implicatures 
of a temporal nature. ~lb examine tile relevant exam- 
pies in sufficient detail, we have had to exclude dis- 
cussion of many closely related issues in the theory 
of discourse structure, rib motivate tbis restriction, 
let us therefore consider first why we might want 
to generate discourses with structures which lead to 
temporal complexities. 
*The ~uthors gratefully acknowledge the support of 
the Science and Engineering Research Council through 
project number on/G22077, tIORO is supported by the 
Economic and Social Research Council. We thank our 
anonymous reviewer for their helpful comments. Email 
contact: jonQcogaci.ed, ac.uk 
Getting Things Out of Order 
Consider tile following suggestion for generating tex: 
tuul descril)tions of causal-temporal structures, I)e- 
scribe things in exactly the order in which they hap- 
pened. If textual order is maAe to match eventual or- 
der, then perhaps little can go wrong; for the bearer 
can safely a.ssume that all the texts she bears are 
narrative. Under these circumstances, the problem 
of selecting adequate regions in the space of utter- 
ances pretty much (lis~solw~s. We do not believe that 
this suggestion will work, in general, and consider 
here two 0.rgunlelltS against it. 
Hovy's argument 
Basically, the generation strategy snggested above 
fails to emphasise the force of some eventualities over 
others (cf. the nncleus-satellite distinction in RST). 
A useful device for emphasis is the topic-couuuent 
structure: we mention the important event first, and 
then the others, which till out or give further detail 
about that important event. These 'comments' on 
the 'topic' may be elfects, but they could also bc the 
cause of the topic. If the latter, then textual order 
and temporal order mismatch; the text is a cmlsal 
explanation in such cases, and having only narrativc 
discourse structure available would preclude its gen-- 
era, ion. Compare (1) and (2), modified from Ilovy 
\[199o\]. 
(1) First, Jim bumped Mike once and hurt him. 
Then they fought. Eventually, Mike stabbed 
him. As a result, aim died. 
(2) aim died in a fight with Mike. After Jim Inunped 
Mike once, they fought, and eventually Mike 
stabbed him. 
The textual order in (1) matches temporal order, 
whereas ill (2) there is mismatctL And yet (2) is 
nmch better than (1). This is bccause the 'impor- 
tant' event is Jim's death. Everything mentioned 
in (1) leads up to this. But because, the events are 
mentioned in their temporal order, the text obscures 
the fact that all the events led to Jim's death, even 
though syntactic markers like and then and as a re- 
suR are used. 
The causal groupings are clearer in (2) because 
it's clear during incremental processing that the text 
following tile mention of Jim's death is a description 
of how it came about. This is so even thougtl ,to 
Ac'II~,S DE COLlNG-92, NANWI..'S, 23-28 AOt~-r 1992 7 2 1 Pnoe. or: COLING-92, NnrctT.s, Auo. 23-28, 1992 
syntactic markers indicate this causal structure. By 
contrast, in (1) the reader realises what's going on 
only at tile last sentence. The discourse structure is 
therefore unclear until the whole text is heard, for 
the narrative requires a common topic which is only 
stated at the end. 
So (2)'s a better discourse than (1); but we would 
never generate it, if textual order had to mirror even- 
tual order. If a generation system were permitted to 
generate (2), however, a price must be paid. The 
proper interpretation of (2) relies on the recruitment 
of certain causal information, left implicit by the ut- 
terance. The generator thus bas some responsibil- 
ity for ensuring that the interpreter accomplishes the 
required inferences. A formal model of impticature 
must be folded into the generation process, so that 
the appropriate reasoning can proceed. 
States |nteract with causal information 
Ill La.scarides and Oberlander \[1992\], we considered 
in detail the following pair of examples: 
(3) Max opened the door. The room was pitch dark. 
(4) Max switched off the light. The room was pitch 
dark. 
Now, no-one would want to say that (3) involved a 
room becoming pitch dark immediately after a door 
was opened. Rather, most accounts (such as those 
based in or around DRT, such as ttinrichs \[1986\]) will 
take the state of darkness to overlap tile event of 
door-opening. That's how one might say states are 
dealt with in a narrative: events move things along; 
states leave them where they are. But if we have a 
piece of causal information to hand, things axe rather 
different. In (4), it seems that the state doesn't over- 
lap the previously mentioned event. 
If one wishes to preserve the assumption about the 
role of states in narrative, it would have to be weak- 
ened to the constraint that states either leave things 
where they are, or move them along. This is not a 
very convincing move. An alternative is to formalise 
the role of the additional causal knowledge. Infor- 
mally, the basis for the distinct interpretations of (3) 
and (4) is that the interpretation of (4) is informed 
by a causal preference which is lacking in the case 
of (3): if there is a switching off of the light and a 
room's being dark that are connected by a causal, 
part/whole or overlap relation, then normally one 
infers that the former caused the latter. This knowl- 
edge is defeasible, of course. In generation, such 
knowledge will constrain the space of adequate utter- 
ances; if H lacks the defeasible causal knowledge that 
switching off lights cause darkness, then (4) won't be 
adequate for H, who will interpret (4) in the same 
way as (3), contrary to S's intentions. Given this, 
S must contain a defeasible reasoning component to 
compute over such knowledge. 
The important point for now is that even if we de- 
scribe things in the order in which they are assumed 
to happen, this doesn't necessarily make the candi- 
date utterance a good one. if the speaker and the 
hearer possess differing world knowledge, there may 
be problems in retrieving the correct causal-temporal 
structure. 
Two Methods of Generating with 
Defeasible Knowledge 
Generation by Defeasible Reasoning 
There is a very general way in which we might view 
interpretation and generation in terms of defensible 
reasoning. Consider the process of discourse inter- 
pretation as one of KB extension. The K8 contains 
an utterance-interpretation, and a set of knowledge 
resources; the latter may include general knowledge 
of the world, knowledge of linguistic facts, knowledge 
about tire discourse so far, and about the speaker's 
knowledge state. We then try to extend the KB so 
as to include the discourse interpretation. Consider 
now the process of generation; it too can be thought 
of as KB extension. Tills time, the KB contains a 
temporal-causal structure, and a set of knowledge 
resources, perhaps identical to that used in interpre- 
tation. We now try to extend the KB so as to in- 
clude the realization of a linguistic structure's seman- 
tic features (with predicates, arguments, connectives, 
orderings), where these features ensure that the final 
linguistic string describes the causal structure in the 
KB. This view might be described as generation by 
defeasible reasoning. 
Modulo more minor differences, these notions are 
close to the ideas of interpretation as abduction 
(Hobbs et al \[1988\]) and generation as abduction 
(ltobbs et al \[1990:26-28\]), where we take abduc- 
tion, in the former case for instance, to be a process 
returning a temporal-causal structure which can ex- 
plain the utterance in context. Correspondences be- 
tween a defensible deduction approach and an ab- 
ductive approach have been established by Konolige 
\[1991\]; he shows that the two are nearly equivalent, 
tire consistency-based approach being slightly more 
powerful \[1991:15-16\], once closure axioms are added 
to the background theory. Lascarides & Oberlander 
\[1992b\] discuss ill detail how such a generation pro- 
cess produces temporally adequate utterances. 
Interactive defaults 
Here, we turn to another, less powerful but simpler, 
method of applying defensible reasoning: the Interac- 
tive Defaults (ID) strategy introduced by Joshi, Web- 
ber and Weischedel \[1984a, 1984b, 1986\]. Rather 
than considering the defeasible process a.s applying 
directly to the KS's causal network, we instead con- 
sider its role as constraining or debugging candidate 
linearised utterances, generated by some otimr pro- 
cess; here we will remain neutral on the nature of 
that originating process. 
A speaker S and a hearer H interact through a 
dialogue; a writer S and a reader tl interact through 
a text. Joshi et al argue that it is inevitable that both 
S and H infer more from utterances than is explicitly 
contained within them. Taking Griee's \[1975\] Maxim 
of Quality seriously, they argue that since both .5' and 
H know this is going to happen, it is incumbent upon 
S to take into account the implicatures II is likely 
to make on the basis of a candidate utterance. If S 
detects that something S believes to be false will be 
among H's implicatures, S must block that inference 
somehow. The basic way to block it is for S to use 
ACRES DE COLING-92, NANrF~, 23-28 AotYr 1992 7 2 2 PROC. OF COLING-92, NANTES, AUG. 23-28. 1992 
a different utterance; one which S does not believe 
will mislead H. 
In terms of defeasible reasoning, the point is that 
S must use it to calculate the consequences of the 
candidate utterance; if the process allows the deriva- 
tion of something S believes to be false, the utterance 
should not be used in its current form. Joshi et al 
illustrate with tile following example; given the KB 
in (5), and the question in (6), they want the process 
to show why the answer in (7b) is preferred to that 
in (7a): 
(5) Sam is an associate professor; most associate 
professors are tenured; Sam is not tenured. 
(6) ls Sam an associate professor? 
(7) a. Yes. 
b. Yes, but he is not tenured. 
We wish to elaborate this interactive defaults strat- 
egy 0D), and consider in greater formal detail the de- 
feasible reasoning al)out causal-temporal strnctures 
that S and H are assumed by S to indulge itl; and to 
consider which candidate utterances arc eliminated 
on this basis. 
Discourse Structure and Temporal 
Constraints 
ID requires a theory of implicatnrc in terms of de- 
faults, and an underlying logical notion of nonrnono- 
tonic or defensible inference. We also require a for- 
mal eharacterisation of the properties an adequate 
candidate utterance must possess; we define these 
below in terms of temporal coherence, reliability and 
precision. Fnrthermore, we assume a model of dis- 
course structure is required. For certain discourse 
relations, such as Narration and Explanation, are 
implicated from candidate utterances (cf. texts (1) 
and (2)), and these impose certain temporal relations 
on tile events described. We turn to this latter issue 
first. 
Discourse Structure and Inference 
The basic model in which we embed ID assumes that 
candidate discourses possess hierarchical structure, 
with units linked by discourse relations modelled 
after those proposed by Hobbs \[1985\]. Lascarides 
& Asher \[1991\] use Narration, Explanation, llaek- 
ground, Result and Elaboration. They provide a log- 
ical theory for determining the discourse relations 
between sentences in a text, and the temporal rela- 
tions between the events they describe. The logic 
used is the nonmonotonic logic Common Sense En- 
tailment (CE) proposed by Asher & Morreau \[1991\]. 
Implieatures are calculated via default rules. For 
example, they motivate the following rules as man- 
ifestations of Gricean-style pragmatic maxims and 
world knowledge, where the clauses a and/3 appear 
in that order in the text. Informally: 
• Narration 
If clauses ~ and/3 are discourse-related, then nor- 
mally Narration(c~,/~) holds. 
• Axiom on Narration 
If Narration(c~,/3) holds, and c~ and /3 describe 
events e 1 and e7 respectively, then el occurs before 
e2. 
* Explanation 
If claus~ ~ and fl are discourse-related, and tile 
event described in fl caused that described in ¢v, 
then normally Ezplanalion(e~,\[t) holds. 
* Axiom on Explanatloxt 
If Ezplanation(c~,\[t) holds, then event el de- 
scribed by c~ does not occur bcfore event e2 de- 
scribed by/3. 
• Causal Law 
If clauses c~ and fl are discourse-related, and (~ de- 
scribes the event c I of x fidling and fl the event e2 
of y pushing x, then normally c2 causes el. 
• Causes Precede Effects 
If event e~ eanses el~ t\]lell c I doesn't occur bcfore 
e2. 
The rules for Narration and l"xplanation constitute 
defe~iblc tingnistic knowledge, and the Axioms on 
them, indefeasible linguistic knowledge. Thc Causal 
Law is a mixture defea-sible linguistic knowledge 
and worhl knowledge: given that tim clauses are 
diseourse-rclated somehow, the events they describe 
must he commetcd in a causal, part/wholc or over- 
lap relation; here, given the events in question, they 
must staud illa causal relation~ if things are nor- 
real. That Causes Precede the.it Etfcets is in(lethtmi- 
ble world knowledge. These rules arc used under the 
cE inference regime to infer the discourse structures 
ofcandidate texts. Two i)atterns of inference are par- 
ticularly relevant: Defensible Modus Ponens (birds 
normally fly, Twecty is a bird; sn Tweety flies); and 
the Penguin Principle (all penguins are birds, birds 
normally fly, penguius normally don't fly, q'weety is 
a penguin; so Tweety doesn't fly). 
For example, in thc absence of information to the 
contrary, the only one of the rules whose antecedent 
is satisfied in interpreting text (8) is Narration. 
(8) Max stood up. John greeted hinl. 
Other things being equal, wc infer via Defeasible 
Modus Ponens that the Narration relation holds be- 
tween (8)'s clauses, thus yielding, assuming logical 
omniscience, an interprctation where the descriptive 
order of events matches their temporal order. On 
the other band, in interl)reting text (9), in the ab- 
sence of further information, two defanlt laws haw~ 
their antecedents satisfied: Narration and the Causal 
Law. 
(9) Max fell. John pusbed him. 
The consequents of these default laws cannot both 
hold in a consistent Ks. By the Penguin Principle, 
the law with the more specific antecedent wins: the 
Causal Law, because its antecedent logically entails 
that of Narration. \]\[lence, (9) is interpreted a.s a ca.se 
where the pushing caused the falling. In turn, this 
entails that the antecedent to Explanation is veri- 
fied; and whilst conflicting with Narration, it's more 
specific, and hence its consequent--Ezplanation-- 
follows by the Penguin Principle. Compare this with 
(8): similar logical forms, bnt different discourse 
structures, and different temporal structures: 
IThe formal details of how the logic oE models these 
AOrES DE COLING-92, NANTES, 23-28 Aour \] 992 7 2 3 I)ROC. OF COLING-92. NANTES, AtJ(J. 23-28, 1992 
Temporal Constraints 
So against this background, what are tile proper- 
ties we require of cmldidate utterances? We concen- 
trate on those constraints that are central to tem- 
poral import. Following Bach \[1986\], we take 'even- 
tualities' to cover both events and states. We de- 
fine lemporal coherence, temporal reliabilily and lera- 
petal precision--the notions that will characterise 
the adequacy of an utterauce--iu terms of a set 
C of relations between eventualities. This set in- 
tuitively describes when two eventualities are con- 
nected. The relations ill C are: causation, the 
part/whole relation, 2 temporal overlap, and the im- 
mediately precedes relation (where 'et immediately 
precedes e2 ' means that el and e 2 stand ill a causal or 
part/whole relation that is compatible with el tcm 
porally preceding e2). s The definitions are: 
• Temporal Coherence 
A text is temporally coherent if the reader can in- 
fer that at least one of tile relations in C holds be- 
tween the eventualities described m tile sentences. 
• Temporal Reliability 
A text is temporally reliable if one of the rcla- 
tions in C which the reader infers to hold does 
in fact hold between tile eventualities described in 
the sentences. 
• Temporal Precision 
A text is temporally precise if whenever the reader 
infers that one of a proper subset of the relations 
in C holds between the eventualities described in 
the sentences, then she is also able to infer which. 
A text is temporally incoherent if the natural inter- 
pretation of the text is such that there are no in- 
ferrable relations between the events. A text is tem- 
porally unreliable if tim natural interpretation of the 
text is such that the inferred relations between tile 
events differ from their actual relations in the world. 
In addition, a text is temporally imprecise, or as we 
shall say, ambiguous, if the natural interpretation of 
tile text is such that the reader knows that one of 
a proper subset of relations in C holds between the 
eventualities, but the reader can't infer which of this 
proper subset holds. 
It follows from the above definitions that a text call 
be coherent but unreliable. On the other hand, there 
may be no questiou about reliability simply because 
we cannot establish a temporal or causal relation be- 
tween the two eventualities. At any rate, a generated 
utterance is adequate only if it is temporally coher- 
ent, reliable and precise. We intend to apply tile ID 
strategy to eliminate candidate utterances that are 
inadequate in this sense. 
interpretations, and those of (3) versus (4), are given in 
Lascarides & Asher \[1991\]. Note that although double 
applications of the Penguin Principle, as in (9), are not 
valid in genera\], they show that for the particular case 
considered here, o~ validates the double application. 
2We think of 'el is part of e2 ~ in terms of Moens and 
Steedman's \[1988\] event terminology, as 'el is part of the 
preparatory phase or consequent phase of e2'. 
aWe a~SUllle that an eYeut el precedes an event e2 
if el's culmination occurs before e2's. So there are 
part/whole relations between el and e~ that are com- 
patible with el temporally preceding e2. 
Applying the ID strategy 
Before applying ID with temporal constraints, we 
must consider the possible relations between the 
knowledge of speaker S and that which speaker S 
has about hearer H's knowledge state. Notice, in- 
cidentally, that Joshi et al explicitly adopt the view 
that \]D is for debugging candidate utterances. In 
principle, their framework, however, is more general. 
Although the idea of debugging is intuitive, we shall 
sometimes talk in terms of constraining the space of 
possible utterances, rather than of debugging specific 
utterances. The definitions of temporal constraints 
are relevant either way. 
Relative KBs 
Let B(S) be S's beliefs about the KS, linguistic 
knowledge (LK) and world knowledge (WK). Let 
B+(H) be S's beliefs about what H believes about 
ttle KB, LK and WK. And let B-(H) be S's beliefs 
about what H doesn't know about the KIL LK and 
WK (so B+(H) and B-(H) are mutually exclusive). 
Problems concerning reliability and precision arise 
when B(S) and B+(H) are different, and when S's 
knowledge of what H believes is partial (i.e. for some 
p, p ¢ 13+(H) and p q B-(II)). Suppcme that S's 
goal is to convey the content of a proposition corn 
tained in his KB, say q. Suppose also that a WFF p 
is relevant to generating a particular utterance de- 
scribing q. Then there are several possible relations 
between B(S), B+(H) and B-(H) that concern p: 
• Case 1 
S knows p and also knows that H does not: 
p C B(S) and p ~_ B-(H) 
• Case 2 
S knows p and isn't sure whether H does or not: 
p e 13(S) and p q B+(H) and p q B-(H) 
• Case 3 
H potentially knows more about p than S does: 
p f\[ B(S) and p ¢. B+(H) and p f\[ 11-(II) 
• Case 4 
S thinks H is mistaken in believing p: 
p ff 11(S) and p e B+(H) 
Of course, the cases where both S and H both believe 
p (p E B(S) and p e B+(II)) and where neither 
do (p q B(S) and p C B-(H)) are unproblematic, 
and so glossed over here. We look at each of the 
above cases in turn, considering tile extent to which 
tile definitions of reliability, coherence and precision 
hell) us eonstraiu the utterance space (or alternately, 
debug candidate utterances). 
Case 1: ~'¢ knows more about p than H 
We now examine the problems concerning reliability 
that arise when p E B(S) and p E B-(H). There 
are two possibilities: either p represents defeasible 
knowledge of tile lmlguage or the world, or p is some 
fact in the Km We investigate these in turn. 
p is defeaslble knowledge Let p be a dcfeasible 
law that represents knowledge that S has and which 
S knows H lacks. ~lb illustrate, take the case where 
p is the causal preference introduced earlier: 
ACrEs DE COL1NG-92, NANTES, 23-28 ho~r 1992 7 2 4 PROC. OF COL1NG-92, NANTES, AUG. 23-28, 1992 
• Causal Law 
If the clauses ,~ and fl are discourse related, anti 
c~ and /7 describe respectively the events el of x 
falling aud e2 of y pushing x, then normally e~ 
caused e I • 
Consider the ca.qc where S intends to convey the 
proposition that John's pushing Max caused the lat- 
ter to fall. Suppose S has a KB which will allow her to 
generate the description in (9), among others. 
(9) Max fell. John pushed llim. 
We have argued that this text is coherent, precise 
and reliable for S because tile causal law (about the 
usual causal relation between pushings and failings) 
is more specific than the linguistic ride (Narration). 
But since H lacks the causal law, (9) will trigger a 
different inference pattern in H; one in which Nar- 
ration wins after all. S must block this pattern by 
changing the utterance; she has eascntially two op- 
tions. If clause order is kept fixed, then ,5' could shift, 
tense into the pluperfect im in (10); or else S can 
insert a clue word, such as because, into tile surface 
form, to generate (11): 
(10) Max fell. John \]lad pushed him. 
(11) Max fell because John pushed him. 
The success of tile latter tactic requires ,'5' and H to 
nmtually know a new linguistic rule, more specific 
ttlan Narralion, such as the following: 4 
• Non-evldential ~Becaus& 
If c~ and/3 are discourse-related, and the text seg- 
ment is (r because fl, then normally tile ewmt lie 
scribed in ~ caused that described in/7. 
On tile otimr hand, if clause order is not taken to be 
fixed, then 5' can simply reorder (9): 
(12) John pushed Max. Max fell. 
So, when 5" bclicves H lacks the relevant causal law, 
5" can simply reorder, and let Na~'atiou do the rest. 
However, recalling the above discussion, in some 
cases a discourse structure that invokes Explanation 
is better than one that invokes Narration. So simply 
reordering events and letting the rule for Narration 
achieve tile correct inferences won't work successfidly 
in all cases. Furthermore, recaning tile iliscussion 
about states and causation above, it becomes appar. 
ent that this tactic of always letting Narration do 
tile work will lead to problenls with texts like (3) 
and (4). 
(3) Max opened the dnor. The roonl was pitcll (lark. 
(4) Max switched off the light. The room was pitch 
dark. 
q'lle reason is that, in the absence of the causal law 
which relates light switching to darkness, (4) will be 
analysed exactly as (3), giving the wrong result. A 
solution would be to replace the state expression with 
all event expression: 
(4/) Max switched off the light. The room wen~ pitch 
dark. 
4This is a pragmatic, rather than semantic rule; it's 
not obvious tll~t this is tile best choice of representation. 
An obviolts alternative is to introduce further clue 
words, and appropriate linguistic rules for reasoning 
about them. This means exploiting linguistic knowl- 
edge to overemne tile gaps in H's world knowledge. 
This tlelps explain tile observation that texts which 
(lescribc events ill reverse to temporal order, with- 
out marking the reverse, may bc quite rare. It's easy 
enough 1o interpret such texts, when we have the all: 
l>ropriate WK. lhlt if a considerate speaker or writer 
ha~ reason to believe that some or all of her audi- 
ence lacks that WK, then she will either avoid such 
descriptive reversals, or mark them with thc type of 
clues we have discussed. 
p is a fact in the Krl We now turn to the case 
where p is a fact about tim Kn which S knows and 
which S knows H lacks. Suppose that p asserts a 
causal relation between two events (lilt does not rep- 
resent an excelltion to any (lefi;p~qible causal prefer- 
ences, and that S wishes to convey tile information 
that p. Then S can simply state p by exploiting 
H's available LK. Clue words may not be needed. 
For example, if p is ttle fact that Max stood up and 
ttlen John greeted him, S can tell H ttlis by uttering 
(8); Na*~'alion will make (8) reliable and precise for 
L/. 
(8) Max stood Ul). John greeted him. 
Similarly, if p is tile fact that Max opened tile door, 
and wtlile this was going on tile room was pitch dark, 
then (3) is reliable and precise for 11 : 
(3) Max opened tim door. The room was pitch dark. 
But what if p asserts a causal relation between 
two events that violates a dcfe~mible causal prefer- 
ence that H has? Snppose p asserts that Max's fall 
iumlediately preceded aolln's puslling hinl. And sup- 
pose that S knows that H has tile defeasible causal 
law mel~tioned above, but lacks p. Then neitller (9) 
nor (12) are reliable for //, indicating that S cannot 
generate all atomic text, to assert p. 
(9) Max fell. John pushed him. 
(12) Jobn pushed Max. lie fell. 
tf wouhl interpret (9) a~ an explanation; and (12) 
as a narrative, for nothing will eontliet with Narra- 
(inn in that case: tile causal preference for pushings 
causing failings would simply reinforce the temporal 
structure imposed by Narration. The obvious option 
is to nlove from (9) to 113); anotber option is to re- 
cruit tile pluperfi:et, ms ill 114); note that 115) is not 
a sohttion, since so can be read evidentially. 
(13) Max fi~ll. And then John pushed Ilint. 
(14) J(?a~ pushed Max. lie had fallen. 
(15) Max fell. So John pushed him. 
The seed it) utter (13) rather tllan (9) explains why 
it c~Ltl be necessary to use and then, even tllough 
th,'~ thll-stop is always available and, by Narration, 
has the default effect of temporal progression. So, 
ill general, one might wish to paraphrase Joshi et 
all if a relation CtUl be defeasibly referred to hohl 
between two eventualities, and S' wants solnething 
dilferent, it is essential to mark the desired relation 
with sometlling strunger. 
ACRES DF. COL1NG-92, NANTES, 23-28 AOt'rl 1992 7 2 5 PROC. OF COLING 92, NANq ES. AUG. 23-28, 1992 
Case 2: 6: knows p but isn't sure if H 
does 
In general, S will have only partial knowledge about 
H's beliefs. This has its drawbacks. 
p is a defeasible causal preference Suppose 
that S isn't sure whether or not H believes the defea- 
sible causal law relating falling and pushing. Then 
there are at least two ways in whicb S's model of 
H's knowledge can be expanded to a complete state- 
ment of H's knowledge. The first, B1, contains the 
causal law. The second, B2, does not. Now uppose 
that S wishes to convey the proposition that John's 
pushing Max caused Max to fall. Then, if S assumes 
H's knowledge corresponds to Bl, then H will find 
a reliable interpretation for (9). 
(9) Max fell. John pusbed him. 
On the other hand, if S assumes that H's knowl- 
edge corresponds to BT, then H will interpret (9) in 
an undesirable way, witt, the falling preceding the 
pushing; as we said before, Narration would win. 
Under this model, S isn't sure how H will interpret 
(9), because S doesn't know if H's knowledge correo 
sponds to B! or B2. Hence the ambiguity of(9) man- 
ifests itself to the generator S, if not to the hearer H, 
because S doesn't haw. ~ sufficient information about 
H to predict wbich of the two alternative temporal 
structures H will infer for (9). This is slightly differ- 
ent to the previous case where S actually knows H 
lacks the causal law, making (9) unreliable. 
"lb avoid uttering unreliable text, S will have to 
utter something other thmJ (9). Indeed, it may be 
possible for S not to worry about tim ambiguity of 
(9) at all, if some 'safe' strategy can bc tbund that 
would guide S's expansion of H's knowledge in a way 
that wmdd ensure the generation of reliable text for 
H. A plausible strategy for S's reasoning about H 
would he the following: if S isn't sure whether or not 
H knows p, then assume H doesn't know p. On the 
face of it this seen~s plausible. But just how safe is 
it? 
We state it in terms of B+(H) and B-(H): 
• lfp q B+(H) and p q B-(H), assume p E B-(H) 
and generate-and-test under this assumption. 
But this won't work in general. If S wants to con- 
vey a violation of the causal law p, but H actually 
believes p, tben the strategy will suggest the use of 
(9), which will actually be unreliable for H. 
In fact, there is no safe strategy, save tim ouc 
where S considers several alternative expansions of 
H's knowledge. As a result, ambiguity of text will 
manifest itself to S in certain cases, because of her 
partial knowledge of H. This is perhaps somewhat 
surprising. Nonmonotonic reasoning is designed as a 
medium for reasoning witb partial kuowledge. And 
yet here we have shown S cannot maintain textual 
reliability on the basis of a partial statement of H's 
KB, even if nonmonotonic inference is exploited. 
p is a fact about the KB: Ambiguity Suppose 
that 5' wants to convey the information that Max's 
fall immediately preceded John pushing \]tim, and 
suppose S knows that H knows the causal law, but S 
doesn't know for sure if H knows already that Max 
fell before John pushed him. Then, for similar rea- 
sons as those mentioned earlier, S isn't sure if (9) is 
reliable or not. 
(9) Max fell. John pushed him. 
'17o be sure that text is reliable in this case, .q will 
again have to exploit linguistic knowledge; for exam- 
pie, by uttering (13) instead of (9). 
(13) Max fell and then John trashed him. 
Case 3: H as advisor, S as pupil 
Suppose that for a certain proposition p, p ¢ B(S), 
p q B4"(H) and p f/ B-(H). This corresponds to 
H potentially knowing more about p than S, but 
S not knowing what more. That's pretty much the 
position of the tutee in a tutorial dialogue, and the 
advice-taker in an advisory dialogue. 
Case 4: S thinks that H is mistaken 
Suppose that p f\[ B(S) and p E B+(H). Then S 
doesn't believe p even though site's aware H does. 
This implies that 5' thinks H is mistaken in believing 
p. 
The fact that p q B(S) and p E B+(tt) could 
entail that a text that's reliable for S isn't for H. 
For example, suppose that H believes, by some weird 
perception of social convention, that there is a defeat 
sible cansal preference that greetings cause standing 
ups. Suppose tbat S wants to describe the situation 
where Max stood up and then John greeted him (i.e. 
an exception to H's causal preference). Then this is 
like the exception case above concerning falling and 
pushing: (16) is reliable for S but not for H. 
(16) Max stood up. John greeted him. 
Again, S could compensate for this by explicitly 
marking the temporal relation. Alternatively, the 
fact that p ¢ B(S) and p E B+(H) could entail that 
a text that's unreliable for S is reliable for H. Again, 
let p be the causal law that says that greetings cause 
standing ups. But this time suppose that S wants 
to describe the situation where John's greeting Max 
caused him to stand up. So this time, S wants to 
describe an instance of the causal law. Then both 
(16) and (17) are reliable for H, but only the latter 
is reliable for S. 
(17) John greeted Max. tie stood up. 
(16) is unreliable for S. Arguably, it wouldn't be 
in the set of possible linguistic realisatious, but only 
if this set is assumed to be characterised by what 
S finds reliable. But we bare no argument for this 
assumption, and so we don't make it. 
Conclusions 
Ilere, we summarise the current state of the model, 
and briefly discuss two of its limitations. 
We admitted that that job of defeasible reasoning 
in generation could be very general; but ttlat we were 
going to look at it in the context of the Interactive 
Defaults strategy. ID applies to the candidate ut- 
terances (or tile space of utterances), and criticises 
the utterances (or the space), producing better ut- 
terances, or a smaller space. The notion of logical 
Ac-rEs DE COLING-92, NArm.;s, 23-28 Ao~r 1992 7 2 6 PROC. OF COLING-92, NANTES, AUG. 23-28. 1992 
consequence supported by CE was used to make pre- 
cise how utterances are constrainted by m. Crucially, 
we used Defensible Modus Ponens and the Penguin 
Principle. The grounds for criticism were the tempo. 
ral ramifications of the utterance; if it was incoherent 
for //, unreliable for H or dangerously ambiguous 
(for 5'), it was bad. 
One limitation of the model is that, although it 
permits reasoning about the knowledge or beliefs of 
interlocutors, it neglects their goals and intentions 
to do actions. ID does not deal with the phenom- 
ena which motivate the work following Cohen and 
Perrault \[1979\] and Allen and Perrault \[1980\], (cf. 
Cohen, Morgan and Pollack \[1990\]). In particular, 
ID does not let S take into account those inferences 
H will make in attempting to ascribe a plan to S. 
Hobbs et al \[1990:44-45\] argue that inferences lead- 
ing to plan recognition are less significant in inter- 
preting long written texts or monologues, llence, 
it might be argued that the generation of such dis- 
courses need not give H's plan recognition particular 
weight. Nonetheless, ID is incomplete, to the extent 
that such inferences inflncncc discourse generation. 
Secondly, discourse structure and temporal struc- 
ture have become somewhat detached. Sometimes, 
it's only the causal-temporal structure derivable front 
the candidate that is being criticized. It may there 
fore be thought that the discourse structure is aa 
idle wheel as things stand, and should be either elim- 
inated (el. Sibun \[1992\]), or bc trusted with a greater 
share of the work, enriching the discourse with useful 
clue words (cf. Scott and Souza \[1990\]). Our tenta- 
tive view is timt tire latter view is plausible, and any- 
way is closer to the idea of generation by defensible 
reasoning, canvassed early on. 
The |D strategy examined here seems to involve a 
lot of hard work generating simple eamlidates which 
almost always require debugging. It would be prefer- 
able if we could do this work in advance, by defanlt. 
The alternative is explored in Lascaridcs and Ober- 
lander \[1992b\], in which we abduce discourse struc- 
tures from event structures, mid then interleave de- 
duction and abduction to derive linguistic realisa- 
tions. But in turning to the more global approach, 
we should not lose sight of the fact that simple 
texts are sometimes best. (2) illustrates this point: 
the rhetorical relations inferred aren't syntactically 
marked, arid yet the text is more natural than (1), 
where the relations are marked. As might be ex- 
pected, there seems to be a trade-off between the 
naturalness of the output and its computational cost. 

References 

Allen, J. F. & Perrault, C. R. \[1980\] Analyzing inten- 
tion in Dialogues. Artificial Intelligence, 115, 143 178. 

Asher, N. & Morreau, M. \[1991\] Comnton Sense Entail- 
ment: A Modal Theory of Nomnouotonic Reasoning. 
In Proceedings o\] the 12th International Joint Con\]cr- 
ence on Artificial Intelligence, Sydney, Australia, Au- 
gust 1991. 

Bach, E. \[1986\] The algebra of events. Linguistics anti 
Philosophy, 9, 5-16. 

Cohen, P. It. & Perrault, C. R. \[1979\] Elements of a 
Plan-Based Theory of Speech Acts. Cognitive Science, 
3, 177-212. 

Cohen, P. R., Morgan, J. & Pollack, M. E. \[1990\] 
Intentions in Communication. Cambridge, MA: hilT 
press. 

Gtice, H. P. \[1975\] Logic and Conversation. In Cole, P. 
and Morgan, J. l,. reds.) Synlaz and Semantics, Vol- 
rune 3: Speech Acts, pp41-58. New York: Academic 
Press. 

ilinrichs, E. \[1086\] Temporal Anaphora in Discourses of 
English. Linguistics and Philosophy, 9, 63--82. 

llobbs, J. it. \[1985\] On the Coherence and Structure of 
1)iscourse. Report OSL1-85 ,37, Center for the Study 
of Language and htformation, Stanford, Ca., October, 
1985. 

Hobbs, 3., Stickel, M., Martin, P. & Edwards, D. \[1988\] 
Interpretation as Abduction. In Proceedings of the 
261h Annual Meetin 9 of the Association \]or Compu° 
rational Linguistics, suNY, lluffMo, N.Y., June, 1988, 
pp95 103. 

flobbs, J., Sticket, M., Appelt, D. & Martin, P. \[1990\] 
Interpretation as Abduction. Technical Note No. 499, 
sl~l lnternationM, Menlo Park, Ca., l)ecember 1990. 

llovy, E. \[1990\] Pragmatics and Natural Language Gen- 
eration. Artificial Intelligence, 43, 153-197. 

Joshi, A., Webber, B. & Weisehedel, R. M. \[19849\] 
Preventing I"Mse Inferences. In Proceedings of the 
lOth h~ternatlonal Con.\[vrence on Computational Lin- 
guistics and the $2nd Annual Meeting o\] the Associa- 
tion .\[or Computational Linguistics, Stanford Univer- 
sity, Stanford, Ca., 2-6 July, 1984, pp134~138. 

Joshi, A., Webber, B. & Weischedel, It. \[1984b\] 
Default reasoning in interaction. In Proceedings o.\[ the 
Non-Monotonic tleasoning Workshop, AAAI, N.Y., Oc 
tober, 1984, pp144 150. 

Joshi, A., Weblrer, B. & Weischedel, R. \[1986\] 
Some Aspects of Default Re~oning in Interactive Dis- 
course, lteport MS-C1S-86-27, University of Pennsyl- 
vasia. 

Konolige, K. \[1991\] Abduction vs Closure in Causal 
Theories. Forthcoming Research Note in A rtificial In. 
telligeacc. Page references to ms. 
I,ascarides, A. & Asher, N. \[1991\] 

Discourse Relations and Couuoon Sense Entailment. 
Submitted to Journal o\] Logic, Language and In.for- 
mation. DYANA deliverable 2.5b, Centre for Cognitive 
Science, University of Edinburgh. 

Lasearides, A. & Oberlander, J. \[19929\] Temporal Co- 
herence anti Defeasible Knowledge. Theoretical Lin- 
guistics, 18. 

Lascarides, A. & Oberlander, J. \[1992b\] Abducing 
Temporal Discourse. In Dale, R. tlovy, E. RSsner, 
D. and f<t,ck, O. reds.) Aspeets o\] Automated Natural 
Language Generation. Berlin: Springer-Verlag. 

Scott, l). R. & Souza, C. S. \[19911\] Getting the Message 
Across in nsT-based Text Generation. In It. I)Me, C. 
Mellish and M. Zock reds.) Current Research in Nat- 
oral Langua~le Generation. London: Academic Press. 

Sibun, P. \[1992\] Generating Text without Trees. To ap- 
pear in Computational Intelligence: Speciol Issue on 
Natural Language Generation, 8. 
