Redundancy in Collaborative Dialogue 
Marilyn A. Walker 
University of Peimsylvania, Computer Science Dept.* 
Philadelphia, PA 19104 
lyn@linc.cis.upenn.edu 
1 Introduction 
It seems a perfectly valid rule of conversation not to 
tell people what they already know. Indeed, Grice's 
QUANTITY lllaxim has often been interpreted this way: 
Do not make your contribution more informative than 
is required\[f\]. Stalnaker, as well, suggests that to assert 
something that is already presupposed is to attempt to 
do something that is already done\[14\]. Thus, the notion 
of what is informative is judged against a background 
of what is presupposed, i.e. propositions that all conver- 
sants assume are mutually known or believed. These 
propositions are known as the COMMON GROUND\[10, 5\]. 
The various formulations of this 'no redundancy' rule 
permeate many computational analyses of natural lan- 
guage and notions of eooperativity. However consider 
the following excerpt from the middle of an advisory 
dialogue between IIarry (h), a talk show host, and Ray 
(r) his caller 1. 
Example 1 : 
(6) r. uh 2 tax qunstionu. 
onu: since April 81 we have had an 
85 year old lother living ~ith us. 
her only income has been social security 
plus approximately $3000 from a 
certificate O~ deposit and i wonder 
whatJs the situation as far as 
claiming her as a dependent or does 
thag income from the certificate of 
deposit rule her out as a dependent? 
(7) h. yes it does. 
(8) r. IT DOES. 
(9) h. ¥UP THAT KHflCKS HER OUT. 
In standard information theoretic terms, both (8) and 
(9) are REDUNDANT. Harry's assertion in (9) simply 
paraphrases what was said in (7) and (8) and so it 
*This researcll was partially funded by AltO grmat DAAL03- 
89-00031PRI and DARPA grant N00014-90~J-1863 at the Uni- 
vernity of Pem~ylvania, by Hewlett Packard, U.K., and by art 
NSF award for 1991 Summer \[rmtitute in Jalmn 
1 Thee examples come ft~m the talk show for financial advice, 
6'peaking o/ Your Money, on WCAU in Philadelphia. This col~ 
pus w~s collected and transcribed by Marth~t Pollack anti Julia 
Hirschberg\[12\]. 
cannot be adding beliefs to the cmnmon ground 2. Fur- 
thermore, the truth of (9) cannot be in question, for in- 
stead of 19), \[larry could not say Yup, but lhat doesn't 
knock her out. So why does Ray (r) in (8) RF~PEAT 
Harry's (h) assertion of it does, and why does lfarry 
PARAPHRASE himself and Kay in (9)? 
My claim is that mformationally redundant utterances 
(IRU's) have two main discourse functions: (1) to pro- 
vide EVIUENCI~ to support the assumptions underlying 
the inference of mutual beliefs, (2) to CENTER a propo- 
sition, is. make or keep a proposition salient\[6\]. This 
paper will focus on (1) leaving (2) for future work. 
First consider the notion of evidence. One reason why 
agents need EVIDENCE for beliefs is that they only have 
partial information about: (1) the state of world; (2) 
the effects of actions; (3) other agent's beliefs, prefer- 
ences and goals. This is especially true when it comes 
to modelling the effects of linguistic actions. Linguistic 
actions are different than physical actions. An agent's 
prior beliefs, preferences and goals cannot be ascer- 
tained by direct inspection. This means that it is 
difficult for the speaker to verify when an action has 
achieved its expected result, and so giving and receiv- 
ing evidence is critical and the process of establishing 
mutual beliefs is carefully monitored by the conver- 
sants. 
The characterization of IRU's ms informationally re- 
dundant follows from an axiomatization of action in 
dialogue that I will call the DETERMIniSTIC MODEL. 
This model consists of a number of simplifying assump- 
tions such as: (I) Propositions are are either believed 
or not believed, (2) Propositions representing beliefs 
and intentions get added to tim context by the unilat- 
eral action of one conversant, (3) Agents are logically 
ormfiscient. (4) The context of a disconrse is an undif- 
ferentiated set of propositions with no specific relations 
between them. I claim that these assumptions nmst be 
dropped in order to explain the function of IRU's in 
dialogue. 
Section 2 discusses assumption (1); section 3 shows how 
assmnption (2) can be dropped; section 4 discusses (3); 
section 4.1 shows that some IRU's facilitate the infer- 
ence of relations between adjacent propositions. 
2\[8) is not realized with a rising question intonation. This 
will be discussed in sectiott 6.1. 
A~\]~s DE COLING-92, NAINrFES, 23-28 AO~" 1992 3 4 5 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
2 Mutual Beliefs in a Shared 
Environment 
The account proposed here of how the COMMON 
GROUND is augmented, is based is Lewis's SHARED EN- 
VIRONMENT model for common knowledge\[10, 2\]. In 
this model, mutual beliefs depend on evidence, openly 
available to the conversants, plus a number of under- 
lying assumptions. 
Shared Environment Mutual Belief In- 
duction Schema 
It is mutually believed in a population P that 
ql if and only if some situation ~q holds such 
that: 
1. Everyone in P has reason to believe that 
,q holds. 
2. 3 indicates to everyone in P that every- 
one in P has reason to believe that 8 
holds. 
3. S indicates to everyone in P that @. 
The situation ~q, used above in the mutual belief in- 
duction schema, is the context of what has been said. 
This schema supports a weak model of mutual be- 
liefs, that is more akin to mutual assumptions or mu- 
tual suppositions\[13\]. Mutual beliefs can be inferred 
based on some evidence, but these beliefs may depend 
on underlying assumptions that are easily defensible. 
This model can be implemented using Gallier's theory 
of autonomous belief revision and the corresponding 
system\[4\]. 
A key part of this model is that some types of 
evidence provide better support for beliefs than 
other types. The types of evidence considered are 
categorized and ordered based on the source of 
the evidence: hypothesis < default < inference 
< linguistic < physical(See \[2, 4\]). This ordering 
reflects the relative defeasibility of different assump- 
tions. Augmenting the strength of an assumption thus 
decreases its relative defensibility. 
A claim of this paper i8 that one role of IRU's is to en- 
sure that these assumptions are supported by evidence, 
thus decreasing the defensibility of the mutual beliefs 
that depend on them\[4\]. 
Thus mutual beliefs depend on a defensible inference 
process. All inferences depend on the evidence to sup- 
port them, and stronger evidence can defeat weaker ev- 
idence. So a mutual belief supported as an inference 
can get defeated by linguistic information. In addi- 
tion, I adopt an an assumption that a chain of reason- 
ing is only as strong as its weakest link: 
Weakest Link Assumption: The strength 
of a belief P depending on a set of under- 
lying assumptions al,...an is MIN(Strength (a,, ...4,)) 
This seems intuitively plausible and means that the 
strength of belief depends on the strength of underly- 
ing assumptions, and that for all inference rules that 
depend on multiple premises, the strength of an in- 
ferred belief is the weakest of the supporting beliefs. 
This representation of mutual belief differs from 
the common representation in terms of an iterated 
eonjunction\[ll\] in that: (1) it relocates information 
from mental states to the environment in which utter- 
anees occur; (2) it allows one to represent the different 
kinds of evidence for mutual belief; (3) it controls rea- 
soning when discrepancies in mutual beliefs are discov- 
ered since evidence and assumptions can be inspected; 
(4) it does not consist of an infinite list of statements. 
3 Inference of Understanding 
This section examines the assumption from the DETER- 
MINISTIC MODEL that: (2) Propositions representing 
beliefs and intentions get added to the context by the 
unilateral action of one conversant 3. This assumption 
will also be examined in section 5. 
The key claim of this section is that agents monitor the 
effects of their utterance actions and that the next ac- 
tion by the addressee is taken as evidence of the effect 
of the speaker's utterance 4. That the utterance will 
have the intended effect is only a hypothesis at the 
point where the utterance has just been made, irrespec- 
tive of the intentions of the speaker. This distinguishes 
this account from others that assume either that utter- 
ance actions always succeed or that they succeed unless 
the addressee previously believed othecwise\[ll, 8\]. 
I adopt the assumption that the participants in a dia- 
logue are trying to achieve some purpose\[7\]. Some as- 
pects of the structure of dialogue arises from the struc- 
ture of these purposes and their relation to one another. 
The minimal purpose of any dialogue is that an utter- 
ance be understood, and this goal is a prerequisite to 
achieving other goals in dialogue, such as commitment 
to future action. Thus achieving mutual belief of un- 
derstanding is an instance of the type of activity that 
agents must perform as they collaborate to achieve the 
purposes of the dialogue. I claim that a model of the 
achievement of mutual belief of understanding can he 
extended to the achievement of other goals in dialogue. 
Achieving understanding is not unproblematic, it is 
a process that must be managed, just as other goal 
achieving processes are\[3\]. Inference of mutual under- 
standing relies upon some evidence, e.g. the utterance 
that is made, and a number of underlying assumptions. 
The assumptions are given with the inference rule be- 
low. 
say(it, B, u, p) --A-> 
aThis is an utterance action version of the STRIPS 
assumption. 4Except for circumstances where it is clear that the flow of 
the conversation has been interrupted. 
AcrEs DE COLING-92, NANTES, 23-28 AOI~T 1992 3 4 6 Pgoc. OF COLING-920 NANTES, AUG. 23-28, 1992 
Next Assumption 
Utterance addressed 
PROMPT attention 
REPEAT hearing., attention 
PARAPHRASE realize, hearing, attention 
INFERENCE license, realize, hearing, attention 
IMPLICATURE license, realize, hearing, attention 
ANY Next copresence linguistic 
Utterance license, realize, hearing , attention default~ 
Evidence 
qype~ 
lingu~ 
linguistic I 
linguistic \] 
linguist)c~ 
linguistic \] 
Figure 1: How tbe Addressee's Following utterance upgrades the evidence underlying assumptions 
understand(B, u, p) \[evidence-type\] 
Assumptions = 
eopresent(h, B, u) \[evidence-type\] 
attend(B, U) \[evidence-type\] 
hear(B, u) \[evidence-type\] 
bel(B, realize(u, p)) \[evidence-type\] 
This schema means that when A says u to B intending 
to convey p, that this leads to the mutual belief that 
B understands u as p under certain assumptions. The 
assumptions are that A and B were cnpresent, that B 
was attending to the utterance event, that B heard the 
utterance, and that B believes that the utterance u 
realizes tim intended meaning p. 
The \[evidence-typeJ annotation indicates the 
strength of evidence supporting the assumption. All of 
the assumptions start out supported by no evidence; 
their evidence type is therefore hypothesis. It isn't 
until after the addressee's next action that an assump- 
tion can have its strength modified. 
The claim here is that one class of IRU's addresses 
these assumptions underlying the inference of mutual 
understanding. Each type of IRU, the assumption ad- 
dressed and the evidence type provided is given in Fig- 
ure 1. Examples are provided in sections 3.1 and 3.2. 
It is also possible that A inteuds that BY saying u, 
which realizes p, B should make a certain inference q. 
Then B's understanding of u should include B making 
this inference. This adds an additional assumption: 
bel(B, license (p, q)) \[evidence-typeJ 
Thus assuming that q was inferred relies on the as- 
sumption that B believes that p licenses q in the con- 
text. 
Figure 1 says that prompts, repetitions, paraphrases 
and making inferences explicit all provide linguistic ev- 
idence of attention. All that prompts such as sh huh 
do is provide evidence of attention. However repeti- 
tions, paraphrases and making inferences explicit also 
demonstrate complete hearing. In addition, a para~ 
phrase and making an inference explicit provides lin- 
guistic evidence of what proposition the paraphraser 
believes the previous utterance realizes. Explicit infer- 
ences additionally provide evidence of what inferences 
tile inferrer believes the realized proposition licenses in 
this context. 
Ill each case, the IRU addresses one or more 
sumptions that have to be made in order to infer 
that mutual understanding has actually been achieved. 
The assumption, rather than being a hypothesi~ 
or a default, get upgraded to a support type of 
linguistic as a result of the IRU. The fact that 
different II~U's address different assumptions leads to 
the perception that some 1KU's are better evidence 
for understanding than others, e.g. a PARAPHRASE i8 
stronger evidence of understanding than a REPEAT\[3\]. 
In addition, any next utterance by the addressee can 
upgrade the strength of the underlying assumptions to 
default (See Figure 1). Of course default evidence is 
weaker than linguistic evidence. The basis for these 
default inthrences will be discussed in section 5. 
3.1 Example of a Repetition 
Consider example 1 ill section 1. Ray, in (8), repeats 
IIarry's assertion from (7). This upgrades the evidence 
for tile assumptions of hearing and attention associated 
with utterance (7) from hypothesis to Xinguistic. 
The assumption about what proposition p7 is realized 
by u7 remains a default. This instantiates the infer- 
ACRES DE COL1NG-92, NANTI~S, 23-28 Aotrr 1992 3 4 7 Paoc. OF COLING-92, NAIVrr~s, Auc;, 23-28. 1992 
ence rule for understanding as follows: 
say(harry, ray, uT, pT) --I-> 
understand(Ray, u7, p7) \[default\] 
Assunptiona : 
{ eoprseant(harry, ray, u7) \[linguistic\] 
strand(ray, u7) \[linguistic\] 
hear(ray, uT) \[linguistic\] 
bel(ray, realize(uT, pT)) \[default\] 
} 
Because of the WEAKEST LINK assumption, the belief 
about understanding is still a default. 
3.2 Example of a Paraphrase 
This assumption is challenged by a number of cases 
in naturally occurring dialogues where inferences that 
follow from what has been said are made explicit. I 
restrict the inferences that I discuss to those that are 
(a) based on information explicitly provided in the di- 
alogue or, (b) licensed by applications of Gricean Max- 
ims such as scalar implicature inferences\[9\]. 
For example the logical omniscience assumption would 
mean that if l(a) and (b) below are in the context, then 
(c) will be as well since it is entailed from (a) and (b). 
(1) a. You can buy an I It A if and only if you do 
NOT have an existing pension plan, 
b. You have an existing pension plan. 
c. You cannot buy an I It A, 
Consider the following excerpt: 
Exawple 2: 
(18) h. i see. are there any uther children 
beside your gila? 
(19) d. no 
(20) h. YOUR WIFE IS AN OILY CHILD 
(21) d. right, and uh wants to give 
her some security .......... 
Harry's utterance of (20) is said with a falling intona- 
tional contour and hence is unlikely to be a question. 
This utterance results in an instantiation of the infer- 
ence rule as follows: 
say(harry, ray, u20, p20) --l-> 
understand(Ray, u20, p20) \[linguistic\] 
Assumptions = 
{ copretent(haxry, ray, uT) \[linguistic\] 
ate end(ray, u7) \[linguistic\] 
hear(ray, u7) \[linguistic\] 
bel(ray, realize(.7, pT)) \[linguistic\] 
} 
In this ease, the belief about understanding is sup- 
ported by linguistic evidence since all of the sup- 
porting assumptions are supported by linguistic evi- 
dence. Thus a paraphrase provides excellent evidence 
that an agent actually understood what another agent 
meant. 
In addition, these IItU's leave a proposition salient, 
where otherwise the discourse might have moved on to 
other topics. This is part of the CENTERING function 
of IKU's and is left to future work. 
4 Making Inferences Explicit 
This section discusses assumption (3) of the determistic 
model, namely that: Agents are logically omniscient. 
The following excerpt demonstrates this structure. Ut- 
terance (15) realizes la, utterance (16) realizes lb, and 
utterance (17) makes the inference explicit that is given 
in lc for the particular tax year of 1981. 
Example 3: 
(18) h. oh no. 
I E A'e were available 
as long as you are not a participant 
in an existing pension 
(as) j. oh i see. 
well i did uork i do uork for a 
company that has a pension 
(17) h. ahh. THEN YOU'RE NOT ELIGIBLE 
FOR EIGHTY ONE 
(18) j. i see, but i am for 82 
After (16), since the propositional content of (17) is 
inferrable, the assumption that Harry has made this 
inference is supported by the inf arance evidence type: 
bel(H, lieense(p16, p17)) \[inference\] 
According to the model of achieving mutual under- 
standing that was outlined in section 3, utterance (17) 
provides linguistic evidence that Ilarry (h) believes 
that the proposition realized by utterance (16) licenses 
the inference of (17) in this context. 
hal(H, license(pl6, p17)) \[linguistic\] 
Furthermore, the context here consists of a discussion 
of two tax years 1981 and 1982. Utterance (17) selects 
*ighLy one, with a narrow focus pitch accent. This 
implicates that there is some other tax year for which 
Joe is eligible, namely 198219\]. Joe's next utterance, 
but I am for 82, reinforces the implicature that Harry 
makes in (17), and upgrades the evidence underlying 
the assumption that (17) licenses (18) to linguistic. 
ACIES DE COLING-92, NANTES, 23-28 AO~r 1992 3 4 8 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
4.1 Supporting Inferences 
A subcase of ensuring that certain inferences get made 
involves the juxtaposition of two propositions. These 
cases challenge the assumption that: (4) The context 
of a discourse is an undifferentiated set of propositions 
with no specific relations between them. While this 
assumption is certainly not made in most discourse 
models, it is often made in semantic models of the 
context\[14\]. In the following segment, Jane (j) de- 
scribes her financial situation to Iiarry (h) and a choice 
between a setthment and an annuity. 
Example %: 
(l) j. hello harry, my name is jane 
(2:) h. welcome jane 
(3) j. i just retired december first, 
and in addition to my pension and 
social security, I have a 
supplemental annuity 
(4) h. yes 
(5) j. which i contributed to 
while i was employed 
(8) h. right 
(7) j. from the state of IJ mutual fund. 
and ISm entitled to a lump sum 
settlement which would be betueen 
i6,800 and 17,800, or a lesser life 
annuity, and the choices of the annuity 
um would be $128.45 per month. 
That would be the maximum 
with no beneficiaries 
(8) h. You can stop right there: 
take your money. 
(9) j. take the money. 
(10) h. absolutely. 
YOU'RE ONLY GETTING 1500 A YEAR. 
at 17,000, no trouble at all ~o 
get 10 percent on 17,000 bucks. 
Iiarry interrupts her at (8) since he believes he has 
enough information to suggest a course of action, and 
tells her lake ~lonr money. To provide SUPPORT for this 
course of action he produces an inference that follows 
from what she has told him in (7), namely You're only 
gelling 1500 (dollars) a year. SUPPORT is a general 
relation that holds between beliefs and intentions in 
this model. 
Presumably Jane would have no trouble calculating 
that $125.45 a month for 12 months amounts to a little 
over $1500 a year, and thus can easily accept this state- 
ment that is intended to provide the necessary SUP- 
PORT relation, ie. the juxtapt~ition of this fact against 
the advice to lake the money conveys that the fact that 
she is only getting 1500 dollars a year is a reason for 
her to adopt the goal of taking the money, although 
this is not explicitly stated. 
5 Evidence of Acceptance 
In section 3, I examine the assumption that: (2) Propo- 
sitions representing beliefs and intentions get added to 
the context by the unilateral action of one conversant. 
I suggested that this assumption can be replaced by 
adopting a model in which agents' behavior provides 
evidence for whether or not mutual understanding has 
been achieved. I also discussed some of the effects of 
resource bounds, is. eases of ensuring that or providing 
evidence that certain inferences dependent on what is 
said are made. 
Achieving understanding and compensating for re- 
source bounds are issues for a model of dialogue 
whether or not agents are autonomous. But agents' au- 
tonomy means there are a number of other reasons why 
A's utterance to B conveying a proposition p might not 
achieve its intended effect: (1) p may not cohere with 
B's beliefs, (2) B may not think that p is relevant, (3) B 
may believe that p does not contribute to the common 
goal, (4) B may prefer doing or believing some q where 
p is mutually exclusive with q, (5) If p is about an ac- 
tion, B may want to partially modify p with additional 
constraints about how, or when p, 
Therefore it is important to distinguish an agent actu- 
ally ACCEPTING the belief that p or intending to per- 
form an action described by p from merely understand- 
ing that p was conveyed. Other accounts legislate that 
helpful agents should adopt other's beliefs and inten- 
tions or that acceptance depends on whether or not the 
agent previously believed ~ Pill, 8\]. But agents can 
decide whether as well as how to revise their beliefs\[4\]. 
Evidence of acceptance may be given explicitly, but 
acceptance can be inferred in sonm dialogue situations 
via the operation of a simple principle of cooperative 
dialogueS: 
COLLABORATIVE PRINCIPLE: Conversants 
must provide evidence of a detected discrep- 
ancy in belief as soon as possible. 
This principle claims that evidence of conflict should 
be made apparent in order to keep default infer- 
ences about acceptance or understanding from go- 
ing through. 1RU's such as PROMPTSp REPETITIONS~ 
PARAP|IRASES, and making an INFERENCE explicit can- 
not function as evidence for conflicts in beliefs or 
intentions via their propositional content since they 
are informationally redundant. If they are realized 
with question intonation, the inference of acceptance 
is blocked. 
In the dialogue below between tiarry (b) and Ruth (r), 
Ruth in (39), first ensures that she understood Harry 
correctly, and then provides explicit evidence of non- 
acceptance in (41), based on her autonomous prefer- 
ences about how her money is invested.. 
STiffs is a simplification of the COLLABOnATIVE PLANNING 
PRINC~PLE~ described in \[15\]. 
Ac'rEs DE COLING-92, NANTES, 23-28 Ao~-r 1992 3 4 9 PROC. OF COLING-92, NANTES, AUG. 23-28. 1992 
Exa~le 5 : 
(38) h. and I'd like 1K thouwand in a 
2 and a half year certificate 
(39) r. the full 18 in a 2 and a half? 
(40) h. that's correct 
(41) r. GEE. NOT AT MY AGE 
In the following example, Joe in (14) makes a statement 
that provides propositional content that conflicts with 
Harry's statement in (13) and thus provides evidence 
of non-acceptance. 
Exmaple 6 
(13) h. and -- there's no reason why you 
shouldn'~ have an I It h for last year 
(14) j. WELL I THOUGHT TSEY JUST ST£RTED 
THIS YEaR 
Joe's statement is based on his prior beliefs. In both 
of these cases this evidence for conflict is given im- 
mediately. However when there is no evidence to the 
contrary s, and goals of the discourse require achieve- 
ment of acceptance, inferences about acceptance are 
licensed as default. They can he defeated later by 
stronger evidence. 
Without this principle, a conversant might not bring 
up an objection until much later in the conversation, 
at which point the relevant belief and some inferences 
following from that belief will have been added to the 
common ground as dtafanlts. The result of this is that 
the retraction of that belief results in many beliefs be- 
ing revised. The operation of this principle helps con- 
versants avoid replanning resulting from inconsistency 
in beliefs, and thus provides a way to manage the aug- 
mentation of the common ground efficiently. 
6 Other hypotheses 
The first point to note is that the examples here are 
only a subset of the types of IRU's that occur in dia- 
logues. I use the term antecedent to refer to the most 
recent utterance which should have added the proposi- 
tion to the context. This paper has mainly focused on 
cases where the IRU: (1) is adjacent to its antecedent, 
rather than remote; (2) realizes a proposition whose an- 
tecedent was said by another conversant, (3) has only 
one antecedent. It is with respect to this subset of the 
data that the alternate hypotheses are examined. 
A distributional analysis of a subBet of the corpus (171 
IKU's from 24 dialogues consisting of 976 turns), on the 
relation of an IRU to its antecedent and the context, 
shows that 35% of the tokens occur remotely from their 
antecedents, that 32% have more than one antecedent, 
that 480£ consist of the speaker repeating something 
that he said before and 52% consist of the speaker re- 
peating something that the other conversant said. So 
sThls displaying of evidence to the contrary was called sat 
interruption in \[15\]. 
the data that this paper focuses on accounts for about 
30% of the data. 
6.1 Indirect Question Hypothesis 
In example (1) of section 1, an alternative account of 
Ray's repetition in (8) is that it is a question of some 
kind. This raises a number of issues: (i) Why doesn't it 
have the form of a question?, (2) What is it a question 
about?, and (3) Why is it never denied?. 
Of 171 IRU's, only 28 are realized with rising ques- 
tion intonation. Of these 28, 6 are actually redundant 
questions with question syntax, and 14 are followed by 
affirmations. 
If these are generally questions, then one possible an- 
swer to what the question is about is that Ray is ques- 
tioning whether he actually heard properly. But then 
why doesn't he use an intonational contour that con- 
veys this fact as Ruth does in example 5? On an ef- 
ficiency argument, it is hard to imagine that it would 
have cost Ray any more effort to have done so. 
Finally, if it were a question it would seem that it 
should have more than one answer. While 50 of these 
IRU's are followed by an affirmation such as that's cot. 
reef, right, yup, none of them are ever followed by a 
denial of their content. It seems an odd question that 
only has one answer. 
6.2 Dead Air Hypothesis 
Another hypothesis is that IRU's result from the radio 
talk show environment in which silence is not tolerated. 
So agents produce IRg's because they cannot think of 
anything else to say but feel as though they must say 
something. 
The first point to note is that IRU's actually occur 
in dialogues timt aren't on the radio\[l\]. The second 
question is why an agent would produce an IRU, rather 
than some other trivial statement such as I didn't know 
thai. Third, why don't these utterance correlate with 
typical stalling behavior such as false starts, pauses, 
and filled pauses such as uhhh. 
The dead air hypothesis would seem to rely on an as- 
sumption that at unpredictable intervals, agents just 
can't think very well. My claim is that IRU's are re- 
lated to goals, that they support inferencing and ad- 
dress assumptions underlying mutual beliefs, is. they 
are not random. In order to prove this it must be pos- 
sible to test the hypothesis that it is only important 
propositions that get repeated, paraphrased or made 
explicit. This can be based on analyzing when the 
information that is repeated has been specifically re- 
quested, such as in the caller's opening question or by 
a request for information from Harry. It should also be 
possible to test whether the IRU realizes a proposition 
that plays a role in the final plan that Harry and the 
caller negotiate. However this type of strong evidence 
AcIxs DE COLING-92, NANTES, 23-28 Aour 1992 3 5 0 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
against the dead air hypothesis is left to future work. 
7 Discussion 
It should be apparent from the account that the types 
of utterances examined here are not really redundant. 
The reason that many models of belief transfer in di- 
alogue would characterize them as redundant follows 
from a combination of facts: (1) The representation of 
belief in these models has been binary; (2) The effects 
of utterance actions are either assumed to always hold, 
or to hold as defaults unlcss the listener already be- 
lieved otherwise. This means that these accounts can- 
not represent the fact that a belief must be supported 
by some kind of evidence and that the evidence may be 
stronger or weaker. It also follows from (2) that these 
models assume that agents are not autonomous, or at 
least do not have control over their own mental states. 
But belief revision is surely an autonomous process; 
agents can choose whether to accept a new belief or 
revise old beliefs\[4, 8\]. 
The occurrence of IRU's in dialogue bas many ramifi- 
cations for a model of dialogue. Accounting for IRU's 
has two direct effects on a dialogue model. First it re- 
quires a model of nmtual beliefs that specifies how mu- 
tual beliefs are inferred and how some mutual beliefs 
can be as weak as mutual suppositions. One function 
of IRU's is to address the assumptions on which mutual 
beliefs are based. Second the assumption that propo- 
sitions representing beliefs and intentions get added to 
the context by the unilateral action of one conversant 
must be dropped. This account replaces that assump- 
tion with a model in which the evidence of the hearer 
must be considered to establish mutual beliefs. The 
claim here is that both understanding and acceptance 
are monitored. The model outlined here can be used 
for different types of dialogue, including dialogues in 
which agents are constructing mutual beliefs to sup- 
port future action by them jointly or alone. 
ltow and when agents decide to augment the strength 
of evidence for a belief has not been addressed in this 
work as yet. Future work includes analyzing the corpus 
with respect to whether the IRU plays a role in the final 
plan that is negotiated between the conversants. 
8 Acknowledgements 
Discussions with Aravind Joshi, Ellen Prince and Bon- 
nie Webber have been extremely helpful in the devel- 
oprnent of these ideas. In addition I would like to 
thank Herb Clark, Sharon Cote, Julia Galliers, Ellen 
Germain, Beth Ann Hockey, Megan Moser, Hideyuki 
Nakashima, Owen Rainbow, Craige Roberts, Phil Sten- 
ton, and Steve Whittaker for the influence of their ideas 
and for useful discussions. 
ACRES DE COLING-92, NARI~S. 23-28 nOt\]T 1992 3 5 1 PROC. oF COLING-92, NAntES, AUG. 23-28, 1992 

References 

\[1\] Jean C. Carletta. Risk Taking and Recovery in 
Task-Oriented Dialogue. PhD thesis, Edinburgh 
University, 1991. 

\[2\] Herbert H. Clark and Catherine R. Marshall. Def- 
inite reference and nmtual knowledge. In Joshi, 
Webber, and Sag, editors, Elements of Discourse 
Understanding, pages 10-63. CUP, 1981. 

\[3\] H H. Clark and Edward F. Schaefer. Contributing 
to discourse. Cognitive Science, 13, 1989. 

\[4\] Julia R. Galliers. Cooperative interaction as 
strategic belief revision. In M.S. Deen, editor, Co- 
operating Knowledge Based Systems, 1991. 

\[5\] H. P. Grice. William James Lectures. 1967. 

\[6\] Barbara J. Grosz, Aravind K. Josbi, and Scott 
Weinstein. Towards a computational theory of dis- 
course interpretation. Unpublished Manuscript, 
1986. 

\[7\] Barbara J. Grosz and Caudaee L. Sidner. Atten- 
tions, intentions and tile structure of discourse. 
Computational Linguistics, 12:pp. 175-204, 1986. 

\[8\] Barbara J. Grosz and Candace L. Sidner. Plans 
for discourse. In Cohen, Morgan and Pollack, eds. 
Intentions in Communication, MIT Press, 1990. 

\[9\] Julia Iiirschberg. A Theory of Scalar lmplicature. 
PhD thesis, University of Pennsylvania, Computer 
and Information Science, 1985. 

\[10\] David Lewis. Convention. llarvard University 
Press, 1969. 

\[11\] Diane Litman and James Allen. Recognizing 
and relating discourse intentions and task-oriented 
plans. In Cohen, Molyan and Pollack, eds. Inten- 
tions in Communication, MIT Press, 1990. 

\[12\] Martha Pollack, Julia Hirschberg, and Bonnie 
Webber. User participation in the reasoning pro- 
cess of expert systems. In AAAI, 1982. 

\[13\] Ellen F. Prince. On the function of existential 
presupposition in discourse. In Papers from l~th 
Regional Meeting. CLS, Chicago, IL, 1978. 

\[14\] Robert C. Stalnaker. Assertion. In Peter Cole, 
editor, Syntax and Semantics, Volume 9: Prag- 
mattes, pages 315-:\]32. Academic Press, 1978. 

\[15\] Marilyn A. Walker and Steve Whittaker. Mixed 
initiative in dialogue: An investigation into dis- 
course segmentation. In ACL, pages 70-79, 1990. 
