Credibility and its Attacks

Full text

Credibility and its Attacks
Antoine Venant, Nicholas Asher, C´edric D´egremont
To cite this version:
Antoine Venant, Nicholas Asher, C´edric D´egremont. Credibility and its Attacks. Workshop
on the Semantics and Pragmatics of Dialogue - SemDIAL 2014, Sep 2014, Edinburgh, United
Kingdom. pp. 154-162, 2014. <hal-01145832>
HAL Id: hal-01145832
https://hal.archives-ouvertes.fr/hal-01145832
Submitted on 27 Apr 2015
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-
entific research documents, whether they are pub-
lished or not. The documents may come from
teaching and research institutions in France or
abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est
destin´ee au d´epˆot et `a la diffusion de documents
scientifiques de niveau recherche, publi´es ou non,
´emanant des ´etablissements d’enseignement et de
recherche fran¸cais ou ´etrangers, des laboratoires
publics ou priv´es.

Open Archive TOULOUSE Archive Ouverte (OATAO)
OATAO is an open access repository that collects the work of Toulouse researchers and
makes it freely available over the web where possible.
This is an author-deposited version published in : http://oatao.univ-toulouse.fr/
Eprints ID : 13213
To cite this version : Venant, Antoine and Asher, Nicholas and
Dégremont, Cédric Credibility and its Attacks. (2014) In: Workshop on
the Semantics and Pragmatics of Dialogue - SemDIAL 2014, 1
September 2014 - 3 September 2014 (Edinburgh, United Kingdom).
Any correspondance concerning this service should be sent to the repository
administrator: staff-oatao@listes-diff.inp-toulouse.fr

Credibility and its Attacks
Antoine Venant1, Nicholas Asher2and Cedric D´
egremont1∗
1IRIT, Universit´
e Paul Sabatier and 2CNRS, IRIT
1 Introduction
Dialogues occurring in non-cooperative settings
often exhibit attempts at deception, such as misdi-
rections or lies. In many contexts, such as, e.g., tri-
als or political debates, the objectives of a conver-
sation’s participant cannot be expressed in terms
of her and her opponent’s beliefs toward the con-
tent of the different dialog moves. In such con-
texts, dialogues moves come with semantic com-
mitments of their own and challenges based on
other players’ semantic commitments.
In a political debate, an agent Amight ask a
question to another agent B, even though Aknows
the answer to the question. In such a case Ais
just seeking for B’s commitment to an answer. If
Bcomplies and provides an answer, it can be in
A’s interest to further challenge this answer, even
knowing it is correct. What is crucial here, are the
objective semantic commitments that agents can
force out of each other, rather than the subjective
beliefs of these agents about whether the content
of these commitments actually occurs or not.
Addressing the above requires us: 1. to have a
semantic theory of commitments in dialogues, 2.
to determine semantically what constitutes an at-
tack and 3. to distinguish between attacks from a
semantic perspective.
In the next section, we define credibility more
precisely and attacks on it, linking these to linguis-
tic commitments. In section 3, we give some ex-
amples of attacks on credibility, while sections 4
and 5 flesh out the analysis. Section 6 describes
related work. We conclude with some directions
for the future in section 7.
2 Credibility and commitments
An attack on credibility can be thought of as ex-
posing deceitful intention. But determining inten-
∗
We thank ERC Research Grant 269427 for research sup-
port
tions behind speech acts is a tricky business [14]
we will not be getting into. The notions of cred-
ibility and attacks we are considering depends on
overt and public linguistic commitments by speak-
ers.
Using commitments, we now precise our notion
of credibility: a dialogue agent iis not credible iff
(i) it is shown for some ϕthat ihas committed to
ϕthat is absurd or clearly refutable (shown to be
inconsistent with a prior claim of the agent or a
background common assumption), and that it was
plausibly in i’s interest to commit to ϕif ϕis not
attacked. An attack by player jon the credibility
of ioccurs iff jcommits to the following: ihas
committed to ϕ,ϕ|=ψ,ψis absurd or refutable,
and it is in o’s interest to commit to ψ, if ψis not
attacked. A move aby player imakes possible an
attack on credibility iff it is discourse coherent for
jto attach an attack on i’s credibility to a.
Our notion of credibility differs considerably
from that employed in the signaling games liter-
ature where credibility is defined in terms of be-
liefs, typically in equilibrium [10, 11]. Our no-
tion of credibility is defined in terms of commit-
ments, agent’s interests and logical consequence,
none of which depend on how the message affects
the agents’ beliefs.
To flesh out our picture of credibility and at-
tack, we need to explain our notions of conse-
quence and interest or preference. We have two
notions of consequence: ordinary, logical conse-
quence and defeasible consequence. We will as-
sume that our agents are logically (though not fac-
tually) omniscient and so if icommits to ϕhe pub-
licly also commits to ψif ψis a logical conse-
quence of ϕ(notation ϕ|=ψ). Agents also com-
mit to implicatures that are defeasible but what
we shall term normal consequences that interlocu-
tors would draw upon learning that icommits to
ϕ. Finally, implicatures may be more tentative, as
when idraws attention to an alternative to some-

thing to which he is explicitly committed. We’ll
assume that implicatures are modeled in a defeasi-
ble logic using a space of preferred models of the
conversation. We also allow that some weak im-
plicatures may exist only in some of the preferred
models while stronger ones are true in all preferred
models. We thus distinguish between the follow-
ing three levels of commitment. -Non-defeasible
commitment by ito ϕ:ϕis a logical consequence
of every possible interpretation of i’s contribution.
-Implicit defeasible commitment by ito ϕ: the
“prefered” interpretations of i’s contribution entail
ϕ. -Weak implicit defeasible commitment by i
to ϕ:some interpretations of i’s contribution im-
ply ϕ. Section 4 will provide more formal defini-
tions.
We take preferences to be tied to a conception
of rationality. In our framework, we will assume
two conversational partners 0,1and a third party
who observes and judges but does not participate
in linguistic exchanges. We will refer to this third
party as the jury. The jury should be thought off
as an abstract procedural entity with an objective
look on the conversation that serves in the pro-
cess of modeling rationality. The jury’s mecan-
isms also depend on some contextual parameters
which are common knowledge among the players:
for instance, at court, the jury should know that an
“honnest” expert witness must not share interest
with the defendent lawyer. To give another exam-
ple, he should know also when some facts are ir-
refutable and known as such. With this notion of
a jury, our players prefer moves which make them
look good in the eyes of the jury and make the
other look bad, or at least worse. An attack by i
on a player j’s credibility is a way to make jlook
less good. Part of i’s looking good is to not make
mistakes, to not invite attacks on her credibility,
but to make herself look good a player must pro-
vide positive reasons for the position she favors.
Mutatis mutandis for the preferences of player j.
More generally: (i) our players must play moves
that make them look good; (ii) if player iis ra-
tional, she will prefer moves that make possible
moves that jcannot attack; (iii) between 2 moves
that make ilook good but make possible attacks,
she will prefer the one with the more indirect or
weaker damaging context, since a more indirect
damaging consequence is one that has a rebuttal
move that’s not what I meant to say.
3 Linguistic examples and intuitions
In this section we offer some linguistic examples
featuring different sorts of commitments and at-
tacks on credibility. These examples involve not
only commitments to propositions expressed by
assertoric clauses but also to propositions involv-
ing rhetorical relations that link clauses, sentences
and larger units together into a coherent whole.
That is, players commit to a particular content and
to its relations with what has been said before. In
so doing a player may also commit to contents pro-
ferred by his conversational partner as in [1].
Consider a case in which speaker A takes C’s
initial moves to be ambiguous.
(1) a. C: N. isn’t coming to the meeting. It’s
been cancelled.
b. A: Did you mean that N. isn’t com-
ing because the meeting’s cancelled
or that the meeting is cancelled as a
result?
c. C: As a result.
A’s clarification question in (1)b presupposes that
C’s initial contribution was ambiguous between a
result and an explanation move [7, 16]. We take
this to imply at least a weak implicature for both
readings, either of which a conversational partici-
pant could have exploited. This is something we
want to model, and we’ll see in the next example
how such implicatures are exploited by an inter-
locutor.
Now consider the following example:
(2) a. C: N. isn’t coming to the meeting. It’s
been cancelled.
b. A: That’s not why N. isn’t coming.
He’s sick.
c. C: I didn’t say that N. wasn’t coming
because the meeting was cancelled.
The meeting is cancelled because N.
isn’t coming.
This example illustrates how commitments em-
bed. In (2)b A commits to the fact that C com-
mitted in (2)a to providing an explanation for why
N isn’t coming, even though (2)a is ambiguous.
Only such a commitment explains why A attacks
that commitment in the way that he does by giv-
ing an alternative explanation. But in fact, C takes
that commitment by A to have misinterpreted him;
C commits in (2)c that he committed in (2)a to of-

fering a consequence or result of N’s not coming
to the meeting.
Note that while A attacks a move of C’s in (2),
he does not attack C’s credibility in our sense. But
neither does (2) provide a case of misleading im-
plicature. However, the following example from
[3] does. During the Dan Quayle-Lloyd Bentsen
Vice-Presidential debate of 1988, Quayle was re-
peatedly questioned about his experience and his
qualifications to be President. Quayle’s attempted
to compare his experience to the young John
Kennedy’s (referred to below as Jack Kennedy to
convey familiarity) in his answer.
(3) a. Quayle: ... the question you’re ask-
ing is, ”What kind of qualifications
does Dan Quayle have to be presi-
dent,” [...] I have as much experience
in the Congress as Jack Kennedy did
when he sought the presidency.
b. Bensten: Senator, I served with Jack
Kennedy. I knew Jack Kennedy. Jack
Kennedy was a friend of mine. Sena-
tor, you’re no Jack Kennedy.
Implicatures play a key role in his example.
Quayle argues, against the thesis that his little gov-
ernmental experience would make him unsuitable
for the presidency, that Kennedy before him, with
as much experience as he have, was able to han-
dle the presidency. But this answer to the question
suggests an implicit comparison between the two
politicians (both junior senators from a state, each
with little governmental experience) and gives rise
to the possibility of interpreting Quayle’s move
as a stronger commitment that he would likely
be able to handle the presidency in the same way
that John Kennedy handled his, which, if not chal-
lenged would serve Quayle’s claim better. Bentsen
seized upon this weak implicature of Quayle’s
contribution and refuted it, indirectly exposing to
the audience the self-serving nature of the compar-
ison.
Here’s an attested example from [17], in which
a prosecutor (P) wants Bronston (B) to say
whether he had a bank account in Switzerland or
not, and Bronston does not want to make such a
commitment for strategic reasons. But he defeasi-
bly commits to an answer with (4)d in an attempt
to avoid further questioning [2].
(4) a. P: Do you have any bank accounts in
Swiss banks, Mr. Bronston?
b. B: No, sir.
c. P: Have you ever?
d. B: The company had an account there
for about six months, in Zurich.
It is interesting to consider a continuation of this in
which the prosecutor would indirectly attack this
response in (4)d.
(5) Prosecutor: I would like to know whether
you personally ever had an account there?
If Bronston is forced on the threat of perjury to an-
swer affirmatively, his response in (4)d now looks
pretty deceiving to the Jury. The natural thought
arises: Bronston was trying to deceive us into
thinking that he didn’t have an account. Though
the prosecutor didn’t proceed as in (5), had he
done so he would have successfully attacked Bron-
ston’s credibility.
For our final example, consider the following
excerpt from a voir dire examination in [12]. As
background, the plaintiff lawyer (LP) has been re-
peatedly coming back to questions about the divi-
sion of a nerve during a surgery with the objective
of getting the witness (D) to characterize the surgi-
cal operation as incompetent and mishandled. Re-
peatedly coming back to the topic wore D down,
and the defense attorney (LD) was no help:
(6) a. LP: And we know in addition to that,
that Dr. Tzeng tore apart this medial
antebrachial cutaneous nerve?
b. D: Correct.
c. LD: Objection.
d. THE COURT: Overruled.
e. D: Correct. There was a division of
that nerve. I’m not sure I would say
tore apart would be the word that I
would use.
f. LP: Oh, there you go. You’re getting a
hint from your lawyer over here, so do
you want to retract what you’re say-
ing?
The defendant was resisting LP’s line of attack rel-
atively well, but then made an error by agreeing
to LP’s loaded question, in which LP makes the
proposition that is really at issue, that Dr. Tzeng
was negligent, a presupposition by embedding it
under a factive verb. This makes it difficult to an-
swer for D the question in a straightforward way.

Since D had already repeatedly been asked about
this issue, he wasn’t paying attention. LP suc-
cessfully attacks D’s credibility in (6)f when D at-
tempts to correct his mistake with (6)e, by seizing
on a weakly implicated discourse connection be-
tween (6)c and (6)e of Result* (the commitment
in (6)c caused the commitment in (6)e).
These examples suggest two general methods of
deception: moves that implicate propositions that
can’t be committed to explicitly for strategic rea-
sons, and moves that trap agents into making com-
mitments they should rationally refrain from.
Another feature of attacks is that generally they
work gradually in damaging an opponent’s credi-
bility. Perhaps no one move succeeds on its own
in convincing the jury that the opponent is duplici-
tous or incompetent; rather a series of moves grad-
ually move a jury to a skeptical view of the oppo-
nent over the course of a conversation. The victory
conditions for our players are to succeed in even-
tually moving the jury to a position in which the
opponent is no longer credible.
4 Dialogue model
We need a dialogue model in order to analyze our
examples and attacks on credibility in more de-
tail. We’ve already seen that we need to model as
part of a speaker’s contribution not only its com-
positional semantics but also its illocutionary ef-
fects, in particular the implicit discourse links be-
tween utterances, as these can trigger or convey
attacks on credibility. We will therefore build on
[15], as SDRT already offers a formal, logic-based
approach of dialogue content (semantics + illocu-
tionary effects).
[15] models the semantics of dialogue by as-
signing to each conversational agents a commit-
ment slate. Each commitment slate contains a
list of propositions that an agent is committed to,
which involve rhetorical relations as well as ele-
mentary propositions. [15] model explicit and im-
plicit agreements and denials of one agent about
another agent’s commitments. However, the anal-
ysis of credibility threats requires that we go a step
further. Conversational agents explicitly or im-
plicitly refer to, and dispute, others’ commitments.
They attack their opponent’s credibility by expos-
ing inconsistencies in something they claim the
opponent committed to or implicated, and defend
against such attacks by denying a commitment to
content that the opponent claims they committed
to or implicated. We need to represent the com-
mitments of all speakers from their own and their
interlocutors’ points of view, as in [18]. Moreover,
we need to represent arbitrary nesting of commit-
ments explicitly. Recall example (2). In (2)b A
corrects C’s prior utterance, and thereby commits
that C is committed to a false proposition p(N.
is not coming because the meeting is cancelled).
C rejects A’s correction. But what C rejects is
not the proposition that corrects p, but A’s com-
mitment that C commited to p. Therefore, C also
commits that A commits that C commits that p.
Further, we need to distinguish between weak and
strong commitments: when an agent tries to mis-
direct another, he might for instance give a weak
commitment the look of a stronger one. Thus our
dialogue model will add three things to [15]: ex-
plicit nested commitments, the commitments of
each agent from every agents’ point of view and
explicit strong and weak commitments.
Conversations proceed as follows in our model:
speakers alternate turns, each performing a se-
quence of discourse moves. Because we are in-
terested in commitments and attacks, we will not
import the full machinery of SDRT here. We will
symbolize clausal contents within a propositional
language, but incorporate labels for speech acts
and discourse relations so that we can roughly
express discourse-structures following [1]. Cru-
cially, however, our language allows us to embed
discourse structures under 3 modal operators [ ],
h i and N. A discourse move for an agent iis de-
fined as a discourse-level proposition labelled by
a speech act identifier. A discourse-level propo-
sition is either a base-level proposition, a formula
expressing commitment over a discourse structure
(i.e. icommits that a label have some particular
content), or a complex formula R(π1, π2)where
Ris a coherence-relation symbol and π1and π2
are speech act labels. A complex formula recur-
sively involves previously introduced speech acts
labels. The modalities make the language more
expressive, since we can express commitments of
different agents to different contents for a single
speech-act. The formula [π:γ]istates that agent
icommits that the content of the speech act πis
γ. Hence, she also commits that the speaker of π
commits to the discourse proposition γ. Its dual,
hπ:γii, expresses the proposition that it is pos-
sible for ithat the content of πis γ.Niϕmeans
that idefeasibly commits to the contents of the for-

mula ϕ. These modal operators express commit-
ments over discourse structures. From this we re-
trieve commitments over informational content by
looking at the content assigned to labels which are
maximal for a given speaker. (labels that are not
in the scope of another label of the same speaker):
a speaker is committed to a content ϕiff she com-
mits the content of one of his maximal labels to be
a proposition that entails ϕ.
Assume a set Φof base-level propositions, a
countably-infinite set of labels Π, a finite set of
relation symbols Rand a set of conversational
agents I. In order to keep track of which agent
xperform which speech act π, we assume Πpar-
titioned in |I|disjoint subsets (Πi)i∈I. We define
spk(π) = the unique i∈Xsuch that π∈Πi.
Γ(Φ) := ϕ|R(π1, π2)|[δ]i|Ni(δ)| hδii| ¬γ
∆ := π:γ|π:?(γ)|δ1∧δ2
Where the γiand δirespectively range over Γ(Φ)
and ∆, the πiand ϕirespectively range over Φand
Π, and iand Rrespectively range over Iand R.
Definition 1 (Model).A model Mis a tuple
hW, v, (⊲x)x∈X, <i, where Wdenotes set of pos-
sible worlds, v: Φ 7→ ℘(W)a coloration, <:
W→W2a function from worlds to partial order-
ings over W, and for each agent x,⊲x⊆W×W
is a transitive and euclidean accessibility relation.
Our language has a dynamic semantics: the in-
terpretation of a formula is context-change poten-
tial i.e. a relation between world-assigment pairs
(w, σ). To account for polar question in our exam-
ples, we adopt a simplistic version of [13] and take
propositions to semantically denote a set of set of
worlds (a proposition denotes a set of possibili-
ties which is partitioned into equivalence classes
raised by questions). For intance, the question
whether p?partitions a set of world in two,
those worlds at which pon the one hand, and
those at which ¬pon the other. An assigment
σ: Π ×W7→ ℘(℘(W)) is a function that as-
signs a proposition as a set of set of worlds to a
speech act label at a particular world. σ(w, π)is
roughly the (partitioned) set of worlds in which
the interpretation of πat world wis true. Given
a model M, the function J·KMmaps each for-
mula δof the language to a binary relation JδKM
over world-assignment pairs. Discourse-level as-
sertoric propositions in Γ(ϕ)always leave the as-
signment component unchanged and act as fil-
ters that let through only the worlds at which the
proposition is true. Discourse moves in ∆on the
other end modify the assignment. Another bit of
needed machinery is for interpreting discourse re-
lations. In our semantics each relation affects the
contents assigned to its terms. Veridical relations
like Explanation or Result will simply update the
contextually given values to its terms with the se-
mantic effects of the relation on those terms [1].
Non veridical relations like Correction or alterna-
tion place constraints on the truth of the contents
associated with the terms at worlds verifying the
relation in question. We need some notation first:
assume a model M=hW, v, s, (⊲x)x∈X, <i.
Let pdenote a dynamic proposition (i.e a relation
between world/assigments pairs). Define |p|σ
M
as {w∈W|(σ, w)p(σ, w)}and |?p|σ
Mas
{|p|σ
M, W \ |p|σ
M}. Define Acc(w)as the set of set
of world containing a single element which is the
set of all worlds accessible from w:Accx(w) =
{{w′|w ⊲xw′}}. Finally define the update oper-
ation ⋆:℘(℘(W)) ×℘(℘(W)) 7→ ℘(℘(W)) as
a ⋆ b ={x∩y:x∈a∧y∈b}.
Definition 2 (Semantics).Discourse proposi-
tions:
(w, σ)JϕKM(w′, σ ′)iff (σ, w) = (σ′, w′)
w∈v(ϕ)
(w, σ)JR(π1, π2)KM(w′, σ ′)iff (σ, w) = (σ′, w′)
and w∈IR(σ(π1, w), σ(π2, w))
(σ, w)J[δ]xKM(σ′, w′)iff w=w′
and ∀w′′ w ⊲xw′′ →(σ, w′′)JδKM(σ′, w′′)
(σ, w)JhδixKM(σ′, w′)iff w=w′
and ∃w′′ w ⊲xw′′ ∧(σ, w′′)JδKM(σ′, w′′)
(σ, w)JNxδKM(σ′, w′)iff w=w′and
∀u(w ⊲xu∧ ∀v(w ⊲xv→u≥wv))
→(σ, u)JδKM(σ′, u)
Discourse moves:
(σ, w)Jπ:γKM(σ′, w′)iff w=w′and
σ′(π, w) = σ(π , w)⋆|γ|σ
M⋆ Accspk(π)(w)
(σ, w)Jδ1∧δ2KM(σ′, w′)iff w=w′and
(σ, w)Jδ1KM◦Jδ1KM(σ′, w′)

Armed with this semantics for formulas, we can
now define the commitments of each agent iat ev-
ery initial prefix (sequence of turns) in the conver-
sation. Because commitments will depend on dis-
course structure, we define commitments at maxi-
mal labels in the logical forms for the turns (those
that are not within the scope of any other label).
Given a logical form for n conversational turns (or
the whole conversation), we can define the com-
mitment of the players:
Definition 3. (σ, w)JCiϕKM(σ, w)iff
∃π(spk(π, i)∧maximal(π)
∧ ∀u(w ⊲xu→σ(π, v)⊆ |ϕ|))
We thus have a dynamic picture of how speakers’
commitments evolve throughout a conversation.
Examples revisited. We start with example (2).
[15] would analyse the two first turns as in table 1
turn C’s SDRS A’s SDRS
(2-a)
π1:¬N
π2:ccl meeting
π3:Res(π1, π2)
(2-b) π4:¬Exp(π1, π2)
π5:Corr(π3, π4)
Table 1: Analyisis of (2) following [15].
This is problematic, since Ais committed to an ab-
surdity. The semantic conditions of Corr(π3, π4)
require that the content of π3implies the nega-
tion of π4, but Res(π1, π2)does not imply
Exp(π1, π2)(the two are even contradictory).
Keeping with the same kind of tabular represen-
tation as [15] our proposal amounts to further di-
vide each cell of the table above in two, introduc-
ing A′interpretation of C′smoves, and repeating
this process potentially infinitely to express arbi-
trary nestings as in table 2. For readability, we
simplify the table by recopying at each step only
the moves whose interpretation is controversial in
the nested cells. In our language (2) is analysed
as:
[π1:¬N]c∧[π2:ccl meeting]c
∧[π3:Res(π1, π2)]c
∧[π4:¬Exp(π1, π3)]a∧[π5:hπ3:Exp(π1, π2)ic
∧π5:Corr(π3, π4)]a
∧[[π5:hπ3:Exp(π1, π2)ic]a
∧π6:¬Cx(Exp(π1, π2)) ∧π7:C orr(π5, π6)]c
Correcting move like π5triggers presupositions:
here, a presuposition that c’s move π3possibly
commits him to the negation of π4’s content, ac-
comodated as part of π5’s content. In the tabular
representation, C’s final move is:
C’s SDRS
C A
π6:
¬Cx(Ex(π1, π2))
π7:Cor(π5, π6)
C A
π3:
Exp(π1, π2)
In (2), we have only encountered explicit com-
mitments [ϕ]x. But in (1)b, Atakes C’s commit-
ments to involve two possibilities, and he does not
know which Chas in fact committed to. Thus, in
(1)b, Arepresents C’s commitments as
[π1:¬N]c∧[π2:ccl meeting]c
∧[π3:Res(π1, π2)]c
∧[π5:Clar-q(π4, π3)]a∧[π5:hπ3:Exp(π1, π2)ic
∧π5:hπ3:Res(π1, π2)ic]a
In (3), Bentsen (B) seizes on a weak implica-
ture of Quayle’s (Q). Q explicitly commits to a
direct comparison between his experience in gov-
ernment and that of the young JFK, but B corrects
a more general equivalence between the presiden-
tial promise of JFK and his own. If we symbol-
ize the latter with JFK, we take B’s turn to yield
[π′:¬JFK ∧ hπ:JFKiq∧π2:Cor(π, π′)]b.
We see that even a weak implicature is sufficient
to warrant B’s corrective move in π2. The success
of this attack relies on the jury’s decision on the
admissibility of hπ:JFKiq,i.e. the possibility
of a commitent of qto JFK. Finally, in (6), we
see that LP commits that D is committed to a dis-
course link between the defense attorney and his
own self-correction:
[[π1:?p]lp ∧π2:p∧π3:QAP (π1, π2)]d
∧[π5:Obj(π4, π3)∧π9:hπ7:Res(π5, π6)id
∧π9: [π8:Cor(π3, π6)id]lp
C’s SDRS A’s SDRS
π1:¬N
π2:ccl meeting
π3:Res(π1, π2)
C A
π3:
Exp(π1, π2)
π4:
¬Exp(π1, π2)
π5:
Corr(π3, π4)
Table 2: Adding nested commitments

5 The strategic model
Speakers choose the sequences of discourse moves
they do because they want to convey commitments
that will make them look good in the eyes of the
Jury; they also want to make an opponent look bad
if possible by attacking her weak points. We call
this their winning condition. We will assume as in
[3] that speakers may have incomplete knowledge
of the other players’ moves, leading to nasty sur-
prises as in (3), where Quayle clearly didn’t antici-
pate Bentsen’s move in (3)b. A final desirable fea-
ture of the strategic model is that the moves open
to a participant that lead her to her winning condi-
tion may decrease or even vanish if her credibility
is repeatedly attacked. Thus the underlying frame-
work of a sequential game is essential for analyz-
ing conversation.
During play, a player has to weigh whether to
make a move that makes her look good but that is
risky in that it can be attacked; if the attack has
no grounded rebuttal [9], the move could be disas-
trous. Further, when an opponent jmakes a move
involving an implicature it is up to the player ito
decide whether it can be taken as a safe commit-
ment in the sense of [2], and to exploit it in subse-
quent conversational moves, as the prosecutor of
(4) does; and conversely jhas to weigh whether
the player will take the implicature on board or
not, as one of i’s commitments. If not, the decep-
tive move may fail if an opponent makes a request
for an explicit version of an implicated commit-
ment as in (5).
All of these calculations depend on the effect
of play on the Jury, who ultimately decides the
winner according to positive points and lack of
bad moves (inconsistencies or deceptions) on the
part of iand other players. Our Jury entertains a
space of possibilities concerning player types for
the players and a probability distribution Pover
them. Our model is simple; we assume just two
types for each player GOOD and BAD. At the start
of the conversation the Jury entertains only the
possibility that all players are GOOD; that is the
probability distribution is such that P(BADi) =
1−P(GO ODi) = 0 for any player i. As the con-
versation proceeds, P(BADi)is successively up-
dated given what has happened over the last turn;
i.e. Pn(BADi)=Pn−1(BADi/tn). As long as the
opponent does not convincingly refute the argu-
ments of iat n,Pn(BADi)=Pn−1(BADi). How-
ever, a successful attack on, say, iby jat turn tn,
which results in a refutation of an argument by i
with no convincing rebuttal gets the Jury to up-
date Pvia Bayesian conditionalisation such that
Pn(BADi)>Pn−1(BADi).
The effect of a higher probability on BADiis
that the positive reasons advanced by iare given a
lower score; that is the effect of a bad reputation—
the good things you say get discounted. Thus, if
the positive arguments by iin her favor provide
some positive score σi, then the effect on the Jury
at turn nis:1
overall-scorei
n=Pn(GO ODi)(σi
n)(1)
Our model should also reflect the duplicitous
nature of weak implicatures that agents don’t dare
put out as full commitments. So the update of
the probability on Pn(BADi)will depend on (i)
the strength of the implicature, (ii) whether the at-
tack is successful in so far as there is no rebut-
tal that refutes it. For weak implicatures, there is
a rebuttal: you misinterpreted what I said; but it
renders the move useless for the player. The up-
shot of our model is that agents pay dearly if their
credibility is successfully attacked when they ad-
vance a weak implicature, as evidenced in the ex-
ample (3). What exactly was Quayle’s (DQ) mis-
take? It was that he weakly implicated that he was
of the same caliber as JFK, and it is this impli-
cature that Bentsen (B) seizes on and shows to
be ridiculous. He also implicates that DQ insin-
uated the direct comparison without directly say-
ing so, which is a deceptive move. B’s attack was
very successful, especially since DQ did not vigor-
ously rejoin with you misinterpreted me unfairly.
Our model considerably increases Pn(BADi)con-
ditional on B’s attack, rendering DQ’s successful
points much weaker. Not only did B refute DQ’s
argument and expose his deceptive move, but he
affected the overall outcome of the debate.
In example (4) on the other hand Bronston (B)
carries off his strongly implicated commitment to
an answer in (4)d without being challenged by the
prosecutor (P). However, it was somewhat dan-
gerous. Had P continued as in (5), B would
have had to contradict NB(¬bank account)with
[bank account]B, the two modal formulas being
inconsistent. Bronston could have claimed that
he had not understood the first prosecutor’s ques-
tion as being directly about him, but the response
1In future work, we plan to experiment with refinements
of this basic idea, such as using updated probabilities to score
continuations.

would have been weak and our intuition is that his
credibility would have suffered. Example (6) ex-
hibits a slightly different pattern: D has committed
to there being no negligence in the operation, but
in (6)b, he commits to the presupposition of the
question that entails negligence. Conditional on
such a contradiction, P(BADR)increases, but not
much because it was a trick question. But when D
attempts to retract his affirmative answer to LP’s
biased question, then pins on him reasoning that
attacks his credibility as an impartial witness via a
weakly implicated connection between LD’s and
D’s contributions. At this point, P(BAD D)in-
creases considerably, and weakens D’s testimony
in the eyes of the Jury.
Our model predicts that as a player’s credibility
is repeatedly attacked and duplicitous moves are
exposed, her credibility decreases monotonically.
As a consequence, after a certain point the player
may have no moves open to her that achieve her
winning condition—the probability of her being of
BAD type is now too high.
6Related work
Our work assumes a commitment based view of
conversation rather than one based on the internal,
mental states of the participants [14, 20, 19] and
builds on and complements the model proposed in
[18], which in turn extends [7]. They introduce
a dynamic Bayesian model for discourse actions
based on prior moves. Our paper is more limited
in scope but also goes into more detail: our model
details how attacks on credibility function with re-
spect to various types of commitments that come
from different kinds of discourse moves; we show
that even in simple conversations levels of embed-
ded commitments can be very complex (contrary
to a suggestion of [18]); and our Bayesian update
on player types details a part of the picture of [18].
Our model assumes a sequential game view of
conversation, differing from extended signaling
games [4], and uses a notion of credibility which
differs from the standard one in signaling games,
according to which a message is credible iff its
standardly accepted content understood as a set
of evaluation points is a superset of its meaning
in reflective equilibrium (roughly how the mes-
sage content affects belief). In strategic environ-
ments of the sort we have in mind, signaling games
have severe limitations: in our strategic contexts, a
player will send a message only if it benefits him,
but then that message will not benefit his oppo-
nent. In a signaling game, the opponent should
rationally ignore it [6]. However, in a debate, it
would be irrational to ignore the message of the
opponent. Our notion of credibility does not mix
belief and action in the way signaling games do,
and is immune to this problem. A further problem
with signaling games is that they assume common
knowledge of the preferences of each player over
moves. But if these are used to define or to guide
credibility, then there is no room for maneuver
or deception, which is manifest in our examples.
However, our model leaves a place for a signaling
game analysis between the Jury and the players,
which we will pursue in future work.
Related to our work are also recent attempts to
investigate argumentation in actual dialogue [5].
Argumentation theory provides a framework for
analyzing attacks and counterattacks [9]. We have
given much more linguistic detail on how such at-
tacks are carried out and how this can affect ones’
strategy in conversation. On the other hand, we
have presented a general model for credibility in
strategic conversation. Different contexts may af-
fect the parameters of the model that we have
set up. For instance,2sometimes the Jury may
be a participant in the conversation in the sense
that it is allowed to ask questions, sometimes not.
Given a particular context, Jury might also func-
tion according to persuasion rules that are differ-
ent from the simple one we have used in section
5. We have chosen simple settings to illustrate
our model. Finally, we have not gone into the de-
tails of how particular conversational contexts may
dictate specific linguistic forms of attacks and de-
fense, e.g., [8]. Our model is general enough, we
believe, so that we can tune the parameters to fit
the particularities of specific contexts.
7 Conclusions
In this paper, we have presented new notion of
credibility and attacks on credibility that are rel-
evant to conversations in strategic settings where
interlocutor preferences may be opposed. We have
developed a dialogue model extending both [15]
and [18] with a semantics for dialogue turns and
commitments that allows for arbitrary nestings of
commitments. We have also shown that this com-
plexity is required to analyze many examples of
dialogue with attacks on credibility.
2Thanks to a Semdial reviewer for this point.

References
[1] N. Asher and A. Lascarides. Logics of
Conversation. Cambridge University Press,
2003.
[2] N. Asher and A. Lascarides. Strategic con-
versation. Semantics and Pragmatics, Vol
6.2:1–62, 2013.
[3] N. Asher and S. Paul. Infinite games with un-
certain moves. In F. Mogavero, A. Murano,
and M. Vardi, editors, Proceedings of the
First Workshop on Strategic Reasoning, ET-
PCS 2013, Rome, Italy, pages 25–32, Rome,
Italy, 2013. Springer.
[4] R. Aumann and S. Hart. Long cheap talk.
Econometrica, 71(6):1619–1660, 2003.
[5] Elena Cabrio, Sara Tonelli, and Serena Vil-
lata. A Natural Language Account for Argu-
mentation Schemes. In AI*IA - XIII Confer-
ence of the Italian Association for Artificial
Intelligence - 2013, Turin, Italie, December
2013. Springer.
[6] V. Crawford and J. Sobel. Strategic in-
formation transmission. Econometrica,
50(6):1431–1451, 1982.
[7] D. DeVault and M. Stone. Managing ambi-
guities across utterances in dialogue. In Pro-
ceedings from the International Workshop on
the Semantics and Pragmatics of Dialogue
(DECALOG 2007), Trento, Italy, 2007.
[8] Paul Drew. Contested evidence in courtroom
cross-examination: The case of a trial for
rape. Talk at work: Interaction in institu-
tional settings, pages 470–520, 1992.
[9] P.M. Dung. On the acceptability of ar-
guments and its fundamental role in non-
monotonic reasoning, logic programming
and n-person games. Artificial intelligence,
77(2):321–357, 1995.
[10] Joseph Farrell. Meaning and credibility in
cheap-talk games. Games and Economic Be-
havior, 5(4):514–531, 1993.
[11] M. Franke, T. de Jager, and R. van Rooij.
Relevance in cooperation and conflict. Jour-
nal of Logic and Language, 2009.
[12] R. Friedman and P. Malone. Rules of the
Road: A Plaintiff Lawyers Guide to Proving
Liability. Trial Guides, 2nd edition, 2010.
[13] J. Groenendijk and M. Stokhof. Studies on
the Semantics of Questions and the Pragmat-
ics of Answers. PhD thesis, Centrale Interfac-
ulteit, Amsterdam, 1984.
[14] C. Hamblin. Imperatives. Blackwells, 1987.
[15] A. Lascarides and N. Asher. Agreement, dis-
putes and commitment in dialogue. Journal
of Semantics, 26(2):109–158, 2009.
[16] M. Purver. The Theory and Use of Clarifica-
tion Requests in Dialogue. PhD thesis, De-
partment of Computer Science, King’s Col-
lege, London, 2004.
[17] L.M. Solan and P.M. Tiersma. Speaking of
Crime: The Language of Criminal Justice.
University of Chicago Press, Chicago, IL,
2005.
[18] M. Stone and A. Lascarides. Grounding as
implicature. In Proceedings of the 14th SEM-
DIAL Workshop on the Semantics and Prag-
matics of Dialogue, pages 51–58, Poznan,
2010.
[19] D. Traum. Computational models of non-
cooperative dialogue. In Proceedings of
the International Workshop on the Semantics
and Pragmatics of Dialogue (LONDIAL),
London, 2008.
[20] D. Traum and J. Allen. Discourse obliga-
tions in dialogue processing. In Proceedings
of the 32nd Annual Meeting of the Associa-
tion for Computational Linguistics (ACL94),
pages 1–8, Las Cruces, New Mexico, 1994.