Interactive Explanations by Conflict Resolution via Argumentative Exchanges

Antonio Rago , Hengzhi Li and Francesca Toni

Department of Computing, Imperial College London, UK
{a.rago, hengzhi.li21, ft}@imperial.ac.uk

Abstract put deemed responsible (in different ways, according to the

arXiv:2303.15022v2 [cs.AI] 9 Jun 2023

method used) for the outputs of the explained AI model, and

As the field of explainable AI (XAI) is maturing, calls for
interactive explanations for (the outputs of) AI models are
offer little opportunity for interaction. For illustration, con-
growing, but the state-of-the-art predominantly focuses on sider a recommender system providing positive and negative
static explanations. In this paper, we focus instead on in- evidence drawn from input features as an explanation for a
teractive explanations framed as conflict resolution between movie recommendation to a user: this form of explanation
agents (i.e. AI models and/or humans) by leveraging on com- is static in that it does not support interactions between the
putational argumentation. Specifically, we define Argumen- system and the user, e.g. if the latter disagrees with the role
tative eXchanges (AXs) for dynamically sharing, in multi- of the input features in the explanation towards the recom-
agent systems, information harboured in individual agents’ mendation, or with the system’s recommendation itself.
quantitative bipolar argumentation frameworks towards re-
solving conflicts amongst the agents. We then deploy AXs A parallel research direction focuses on argumentative ex-
in the XAI setting in which a machine and a human inter- planations for AI models of various types (see (Cyras et
act about the machine’s predictions. We identify and assess al. 2021; Vassiliades, Bassiliades, and Patkos 2021) for re-
several theoretical properties characterising AXs that are suit- cent overviews), often motivated by the appeal of argumen-
able for XAI. Finally, we instantiate AXs for XAI by defining tation in explanations amongst humans, e.g. as in (Antaki
various agent behaviours, e.g. capturing counterfactual pat- and Leudar 1992), within the broader view that XAI should
terns of reasoning in machines and highlighting the effects of take findings from the social sciences into account (Miller
cognitive biases in humans. We show experimentally (in a 2019). Argumentative explanations in XAI employ compu-
simulated environment) the comparative advantages of these tational argumentation (see (Atkinson et al. 2017; Baroni et
behaviours in terms of conflict resolution, and show that the
al. 2018) for overviews), leveraging upon (existing or novel)
strongest argument may not always be the most effective.
argumentation frameworks, semantics and properties.
Argumentative explanations seem well suited to support
1 Introduction interactivity when the mechanics of AI models can be ab-
The need for interactivity in explanations of the outputs of stracted away argumentatively (e.g. as for some recom-
AI models has long been called for (Cawsey 1991), and the mender systems (Rago, Cocarascu, and Toni 2018) or neural
recent wave of explainable AI (XAI) has given rise to re- networks (Albini et al. 2020; Potyka 2021)). For illustration,
newed urgency in the matter. In (Miller 2019), it is stated consider the case of a movie review aggregation system, as
that explanations need to be social, and thus for machines in (Cocarascu, Rago, and Toni 2019), and assume that its
to truly explain themselves, they must be interactive, so that recommendation of a movie x and its reasoning therefor
XAI is not just “more AI”, but a human-machine interac- can be represented by the bipolar argumentation framework
tion problem. Some have started exploring explanations as (BAF) (Cayrol and Lagasquie-Schiex 2005) ⟨X , A, S⟩ with
dialogues (Lakkaraju et al. 2022) , while several are explor- arguments X = {e, m1 , m2 }, attacks A = ∅ and supports S =
ing forms of interactive machine learning for model debug- {(m1 , e), (m2 , m1 )} (see left of Figure 1 for a graphical vi-
ging (Teso et al. 2023). It has also been claimed that it is our sualisation). Then, by supporting e, m1 (statically) conveys
responsibility to create machines which can argue with hu- shallow evidence for the output (i.e. movie x being recom-
mans (Hirsch et al. 2018). However, despite the widespread mended). Argumentative explanations may go beyond the
acknowledgement of the need for interactivity, typical ap- shallow nature of state-of-the-art explanations by facilitating
proaches to XAI deliver “static” explanations, whether they dynamic, interactive explanations, e.g. by allowing a human
be based on feature attribution (e.g. as in (Lundberg and explainee who does not agree with the machine’s output or
Lee 2017)), counterfactuals (e.g. as in (Wachter, Mittel- the evidence it provides (in other words, there is a conflict
stadt, and Russell 2017)) or other factors such as prime between the machine and the human) to provide feedback (in
implicants (e.g. as in (Shih, Choi, and Darwiche 2018; Figure 1, by introducing attacks (h1 , e) or (h2 , m1 )), while
Ignatiev, Narodytska, and Marques-Silva 2019)). These ex- also allowing for the system to provide additional informa-
planations typically focus exclusively on aspects of the in- tion (in Figure 1, by introducing the support (m2 , m1 )). The
 deconfliction and based on abstract argumentation (Dung
1995); (Panisson, McBurney, and Bordini 2021) develop a
multi-agent frameworks whereby agents can exchange in-
formation to jointly reason with argument schemes and crit-
ical questions; and (de Tarlé, Bonzon, and Maudet 2022) let
agents debate using a shared abstract argumentation frame-
work. These works mostly focus on narrow settings us-
ing structured and abstract argumentation under extension-
based semantics, and mostly ignore the XAI angle ((Ray-
Figure 1: An argumentative explanation for a review aggregation mond, Gunes, and Prorok 2020; Calegari et al. 2022) are
system, amounting to the interactions between a machine and a exceptions). Instead, with XAI as our core drive, we fo-
human sharing their reasoning following a recommendation for x. cus on (quantitative) bipolar argumentation under gradual
semantics, motivated by their usefulness in several XAI
approaches (e.g. in (Cocarascu, Rago, and Toni 2019;
resulting interactive explanations can be seen as a conflict Albini et al. 2020; Potyka 2021; Rago, Baroni, and Toni
resolution process, e.g. as in (Raymond, Gunes, and Pro- 2022)). Other works consider (Q)BAFs in multi-agent ar-
rok 2020). Existing approaches focus on specific settings. gumentation, e.g. (Kontarinis and Toni 2015), but not for
Also, although the need for studying properties of explana- XAI. We adapt some aspects of these works on multi-agent
tions is well-acknowledged (e.g. see (Sokol and Flach 2020; argumentation approaches, specifically the idea of agents
Amgoud and Ben-Naim 2022)), to the best of our knowl- contributing attacks or supports (rather than arguments) to
edge properties of interactive explanations, e.g. relating to debates (Kontarinis and Toni 2015) and the restriction to
how well they represent and resolve any conflicts, have been trees rooted at explananda under gradual semantics from (de
neglected to date. We fill these gaps by providing a gen- Tarlé, Bonzon, and Maudet 2022). We leave other interest-
eral argumentative framework for interactive explanations as ing aspects they cover to future work, notably handling ma-
conflict resolution, as well as properties and instantiations liciousness (Kontarinis and Toni 2015), regulatory compli-
thereof, backed by simulated experiments. Specifically: ance (Raymond, Gunes, and Prorok 2020), and defining suit-
• We define Argumentative eXchanges (AXs, §4), in which able utterances (Panisson, McBurney, and Bordini 2021).
agents, whose reasoning is represented as quantitative Several approaches to obtain argumentative explanations
bipolar argumentation frameworks (QBAFs) under grad- for AI models exist (see (Cyras et al. 2021; Vassiliades,
ual semantics (Baroni, Rago, and Toni 2018), contribute Bassiliades, and Patkos 2021) for overviews), often rely-
attacks/supports between arguments, to interactively ob- ing upon argumentative abstractions of the models. Our ap-
tain BAFs as in Figure 1 towards resolving conflicts on the proach is orthogonal, as we assume that suitable QBAF ab-
agents’ stances on explananda. We use QBAFs, which stractions of models and humans exist, focusing instead on
are BAFs where arguments are equipped with intrinsic formalising and validating interactive explanations.
strengths, as they are well suited to modelling private Our AXs and agent behaviours are designed to resolve
viewpoints, public conflicts, and resolutions, as well as conflicts and are thus related to works on conflict res-
cognitive biases, which are important in XAI (Bertrand et olution, e.g. (Black and Atkinson 2011; Fan and Toni
al. 2022). We use gradual semantics to capture individual 2012a), or centered around conflicts, e.g. (Pisano et al.
evaluations of stance, taking biases into account. 2022), but these works have different purposes to inter-
• We identify and assess several properties (§5) which AXs active XAI and use forms of argumentation other than
may satisfy to be rendered suitable in an XAI setting. (Q)BAFs under gradual semantics. Our agent behaviours
These properties concern, amongst others, the representa- can also be seen as attempts at persuasion in that they
tion and possible resolution of conflicts within interactive aim at selecting most efficacious arguments for changing
explanations drawn from AXs. the mind of the other agents, as e.g. in (Fan and Toni
2012b; Hunter 2018; Calegari, Riveret, and Sartor 2021;
• We instantiate AXs to the standard XAI setting of two
Donadello et al. 2022). Further, our AXs can be seen
agents, a machine and a human, and define a catalogue
as supporting forms of information-seeking and inquiry, as
of agent behaviours for this setting (§6). We experiment
they allow agents to share information, and are thus re-
in a simulated environment (§7) with the behaviours, ex-
lated to work in this spectrum (e.g. (Black and Hunter
ploring five hypotheses about conflict resolution and the
2007; Fan and Toni 2015a)). Our framework however dif-
accuracy of contributed arguments towards it, noting that
fers from general-purpose forms of argumentation-based
the strongest argument is not always the most effective.
persuasion/information-seeking/inquiry in its focus on inter-
active XAI supported by (Q)BAFs under gradual semantics.
The importance of machine handling of information from
2 Related Work humans when explaining outputs, rather than the humans
There is a vast literature on multi-agent argumentation, exclusively receiving information, has been highlighted e.g.
e.g. recently, (Raymond, Gunes, and Prorok 2020) define for recommender systems (Balog, Radlinski, and Arakelyan
an argumentation-based human-agent architecture integrat- 2019; Rago et al. 2020) and debugging (Lertvittayakumjorn,
ing regulatory compliance, suitable for human-agent path Specia, and Toni 2020) or other human-in-the-loop methods
(see (Wu et al. 2022) for a survey). Differently from these ∀a ∈ X ∖ {e}, there is a path from a to e; and iii) ∄a ∈ X
works, we capture two-way interactions. with a path from a to a.
Some works advocate interactivity in XAI (Paulino- Here e plays the role of an explanandum.3 When inter-
Passos and Toni 2022), but do not make concrete sugges- preting the BAF/QBAF as a graph (with arguments as nodes
tions on how to support it. Other works advocate dialogues and attacks/supports as edges), i) amounts to sanctioning
for XAI (Lakkaraju et al. 2022), but it is unclear how these that e admits no outgoing edges, ii) that e is reachable from
can be generated. We contribute to grounding the problem any other node, and iii) that there are no cycles in the graph
of generating interactive explanations by a computational (and thus, when combining the three requirements, the graph
framework implemented in a simulated environment. is a multi-tree rooted at e). The restrictions in Definition 1
impose that every argument in a BAF/QBAF for e are “re-
3 Preliminaries lated” to e, in the spirit of (Fan and Toni 2015b).
In all illustrations (and in some of the experiments in §7)
A BAF (Cayrol and Lagasquie-Schiex 2005) is a triple
we use the DF-QuAD gradual semantics (Rago et al. 2016)
⟨X , A, S⟩ such that X is a finite set (whose elements are ar-
for QBAFs for explananda. This uses I = [0, 1] and:
guments), A ⊆ X ×X (called the attack relation) and S ⊆ X ×
X (called the support relation), where A and S are disjoint. • a strength aggregation function Σ such that Σ(())=0 and,
A QBAF (Baroni et al. 2015) is a quadruple ⟨X , A, S, τ ⟩ for v1 , . . . , vn ∈ [0, 1] (n ≥ 1), if n = 1 then Σ((v1 )) = v1 ,
such that ⟨X , A, S⟩ is a BAF and τ ∶ X → I ascribes base if n = 2 then Σ((v1 , v2 )) = v1 + v2 − v1 ⋅ v2 , and if n > 2
scores to arguments; these are values in some given I rep- then Σ((v1 , . . . , vn )) = Σ(Σ((v1 , . . . , vn−1 )), vn );
resenting the arguments’ intrinsic strengths. Given BAF • a combination function c such that, for v 0 , v − , v + ∈ [0, 1]:
⟨X , A, S⟩ or QBAF ⟨X , A, S, τ ⟩, for any a ∈ X , we call if v − ≥ v + then c(v 0 , v − , v + ) = v 0 − v 0 ⋅ ∣ v + − v − ∣ and if
{b ∈ X ∣(b, a) ∈ A} the attackers of a and {b ∈ X ∣(b, a) ∈ S} v − < v + , then c(v 0 , v − , v + ) = v 0 + (1 − v 0 )⋅ ∣ v + − v − ∣.
the supporters of a. Then, for F = ⟨X , A, S, τ ⟩ and any a ∈ X , given
We make use of the following notation: given BAFs B = A(a) = {b ∈ X ∣(b, a) ∈ A} and S(a) = {b ∈ X ∣(b, a) ∈
⟨X , A, S⟩, B ′ = ⟨X ′ , A′ , S ′ ⟩, we say that B ⊑ B ′ iff X ⊆ X ′ , S}, σ(F, a) = c(τ (a), Σ(σ(F, A(a))), Σ(σ(F, S(a))))
A ⊆ A′ and S ⊆ S ′ ; also, we use B ′ ∖ B to denote ⟨X ′ ∖ where, for any S ⊆ X , σ(F, S) = (σ(F, a1 ), . . . , σ(F, ak ))
X , A′ ∖A, S ′ ∖S⟩. Similarly, given QBAFs Q=⟨X , A, S, τ ⟩, for (a1 , . . . , ak ), an arbitrary permutation of S.
Q′ =⟨X ′ , A′ , S ′ , τ ′ ⟩, we say that Q ⊑ Q′ iff X ⊆ X ′ , A ⊆ A′ ,
S ⊆ S ′ and ∀a ∈ X ∩ X ′ (which, by the other conditions, is 4 Argumentative Exchanges (AXs)
exactly X ), it holds that τ ′ (a) = τ (a). Also, we use Q′∖Q to
denote ⟨X ′ ∖X , A′ ∖A, S ′ ∖S, τ ′′ ⟩, where τ ′′ is τ ′ restricted We define AXs as a general framework in which agents ar-
gue with the goal of conflict resolution. The conflicts may
to the arguments in X ′ ∖ X .1 Given a BAF B and a QBAF
arise when agents hold different stances on explananda. To
Q = ⟨X , A, S, τ ⟩, with an abuse of notation we use B ⊑ Q to
model these settings, we rely upon QBAFs for explananda
stand for B ⊑ ⟨X , A, S⟩ and Q ⊑ B to stand for ⟨X , A, S⟩ ⊑
as abstractions of agents’ internals. Specifically, we as-
B. For any BAFs or QBAFs F, F ′ , we say that F = F ′ iff
sume that each agent α is equipped with a QBAF and a
F ⊑F ′ and F ′ ⊑F, and F ⊏F ′ iff F ⊑F ′ but F ≠F ′ .
gradual semantics (σ): the former provides an abstraction
Both BAFs and QBAFs may be equipped with a gradual of the agent’s knowledge/reasoning, with the base score
semantics σ, e.g. as in (Baroni et al. 2017) for BAFs and as (τ ) representing biases over arguments; the latter can be
in (Potyka 2018) for QBAFs (see (Baroni, Rago, and Toni seen as an evaluation method for arguments. To reflect the
2019) for an overview), ascribing to arguments a dialectical use of QBAFs in our multi-agent explanatory setting, we
strength from within some given I (which, in the case of adopt this terminology (of biases and evaluation methods)
QBAFs, is typically the same as for base scores): thus, for a in the remainder. Intuitively, biases and evaluations repre-
given BAF or QBAF F and argument a, σ(F, a) ∈ I. sent agents’ views on the quality of arguments before and
Inspired by (de Tarlé, Bonzon, and Maudet 2022)’s use after, resp., other arguments are considered. For illustra-
of (abstract) argumentation frameworks (Dung 1995) of a tion, in the setting of Figure 1, biases may result from ag-
restricted kind (amounting to trees rooted with a single ar- gregations of votes from reviews for the machine and from
gument of focus), we use restricted BAFs and QBAFs: personal views for the human, and evaluation methods allow
Definition 1. Let F be a BAF ⟨X , A, S⟩ or QBAF the computation of the machine/human stance on the recom-
⟨X , A, S, τ ⟩. For any arguments a, b ∈ X , let a path from mendation during the interaction (as in (Cocarascu, Rago,
a to b be defined as (c0 , c1 ), . . . , (cn−1 , cn ) for some n > 0 and Toni 2019)). Agents may choose their own evaluation
(referred to as the length of the path) where c0 = a, cn = b range for measuring biases/evaluating arguments.
and, for any 1 ≤ i ≤ n, (ci−1 , ci ) ∈ A ∪ S.2 Then, for e ∈ X , Definition 2. An evaluation range I is a set equipped with a
F is a BAF/QBAF (resp.) for e iff i) ∄(e, a) ∈ A ∪ S; ii) pre-order ≤ (where, as usual x < y denotes x ≤ y and y ≰ x)
such that I = I+ ∪ I0 ∪ I− where I+ , I0 and I− are disjoint and
for any i ∈ I+ , j ∈ I0 and k ∈ I− , k < j < i. We refer to I+ , I0
1
Note that B′∖B, Q′∖Q may not be BAFs, QBAFs, resp., as they
may include no arguments but non-empty attack/support relations.
2
and I− , resp., as positive, neutral and negative evaluations.
Later, we will use paths(a, b) to indicate the set of all paths
3
between arguments a and b, leaving the (Q)BAF implicit, and use Other terms to denote the “focal point” of BAFs/QBAFs could
∣p∣ for the length of path p. Also, we may see paths as sets of pairs. be used. We use explanandum given our focus on the XAI setting.
 and (Iη , Qη , ση ), with Qµ = ⟨Xµ , Aµ , Sµ , τµ ⟩, Qη =
⟨Xη , Aη , Sη , τη ⟩ QBAFs for the same e and:
• I−µ = I−η = [0, 0.5), I0µ = I0η = {0.5} and I+µ = I+η = (0.5, 1];
• Xµ = {e, a, b, c}, Aµ = {(a, e)}, Sµ = {(b, e), (c, a)}
(represented graphically on the top left of Figure 2) and
τµ (e) = 0.7, τµ (a) = 0.8, τµ (b) = 0.4, and τµ (c) = 0.6;
• Xη ={e, a, b, d, f },Aη ={(a, e), (d, a)},Sη ={(b, e), (f, b)}
(represented on the top right of Figure 2) and τη (e) = 0.6,
τη (a) = 0.8, τη (b) = 0.2, τη (d) = 0.6 and τη (f ) = 0.5.
• σµ is the DF-QuAD semantics, giving σµ (Qµ , e) = 0.336,
σµ (Qµ , a)=0.92, σµ (Qµ , b)=0.4, and σµ (Qµ , c)=0.6;
• ση is also DF-QuAD, giving ση(Qη , e)=0.712,ση (Qη , a)=
0.32, ση (Qη , b)=0.6, ση (Qη , d)=0.6, ση (Qη , f )=0.5.
Thus, the machine and human agents hold entirely different
views on the arguments (based on their private QBAFs and
their evaluations) and Σµ (Qµ , e) = − while Ση (Qη , e) = +.
Thus, there is a conflict between the agents’ stances on e.
We define AXs so that they can provide the ground to
Figure 2: AX for explanandum e amongst agents AG = {µ, η}, identify and resolve conflicts in stance amongst agents.
with the exchange BAF representing an interactive explanation.
White (grey) boxes represent contributions (learnt relations, resp.). Definition 5. An Argumentative eXchange (AX) for an ex-
planandum e amongst agents AG (where ∣AG∣ ≥ 2) is a tuple
⟨Bx0 , . . . , Bxn , AG0 , . . . , AGn , C⟩ where n > 0 and:
Thus, an evaluation range discretises the space of possible
• for every timestep 0 ≤ t ≤ n:
evaluations into three categories.4
– Bxt = ⟨Xxt , Atx , Sxt ⟩ is a BAF for e, called the exchange
Definition 3. A private triple for an agent α and an ex- BAF at t, such that Xx0 = {e}, A0x = Sx0 = ∅ and for
planandum e is (Iα , Qα , σα ) where: t > 0, Bxt−1 ⊑ Bxt ;
• Iα = I+α ∪ I−α ∪ I0α is an evaluation range, referred to as α’s – AGt is a set of private triples (Itα , Qtα , σαt ) for e, one
private evaluation range; for each agent α ∈ AG, where, for t > 0, It−1 = Itα ,
α
• Qα = ⟨Xα , Aα , Sα , τα ⟩ is a QBAF for e, referred to as t−1 t t−1 t t t−1
σα = σα , Qα ⊑ Qα and Qα ∖ Qα ⊑ Bx ∖ Bx ; t t−1
α’s private QBAF, such that ∀a ∈ Xα , τα (a) ∈ Iα ; • C, referred to as the contributor mapping, is a mapping
• σα is an evaluation method, referred to as α’s pri- such that, for every (a, b) ∈ Anx ∪ Sxn : C((a, b)) = (α, t)
vate evaluation method, such that, for any QBAF Q = with 0 < t ≤ n and α ∈ AG.
⟨X , A, S, τ ⟩ (τ ∶X→Iα ) and, for any a∈X , σα (Q, a)∈Iα .
Agents’ private triples thus change over time during AXs,
Agents’ stances on explananda are determined by their with several restrictions, in particular that agents do not
private biases and evaluation methods. change their evaluation ranges and methods, and that their
Definition 4. Let (Iα , Qα , σα ) be a private triple for agent biases on known arguments propagate across timesteps (but
α (for some e), with Qα = ⟨Xα , Aα , Sα , τα ⟩. Then, for note that Definition 5 does not impose any restriction on
a ∈ Xα , α’s stance on a is defined, for ∗ ∈ {−, 0, +}, as the agents’ private triples at timestep 0, other than they
Σα (Qα , a) = ∗ iff σα (Qα , a) ∈ I∗α . are all for e). The restriction that all BAFs/QBAFs in
Note that a may be the explanandum or any other argu- exchanges are for the explanandum, means that all con-
ment (namely, an agent may hold a stance on any arguments tributed attacks and supports (and underlying arguments)
in its private QBAF). Also, abusing notation, we will lift the are “relevant” to the explanandum. Implicitly, while we
pre-order over elements of I to stances, whereby − < 0 < +. do not assume that agents share arguments, we assume
In general, agents may hold different evaluation ranges, that they agree on an underpinning ‘lingua franca’, so that,
biases, QBAFs and evaluation methods, but the discretisa- in particular, if two agents are both aware of two argu-
tion of the agents’ evaluation ranges to obtain their stances ments, they must agree on any attack or support between
allows for direct comparison across agents. them, e.g. it cannot be that an argument attacks another
argument for one agent but not for another (in line with
Example 1. Consider a machine agent µ and a human other works, e.g. (de Tarlé, Bonzon, and Maudet 2022;
agent η equipped resp. with private triples (Iµ , Qµ , σµ ) Raymond, Gunes, and Prorok 2020)). We leave to future
4
We choose three discrete values only for simplicity. This may work the study of the impact of this assumption in practice
mean that very close values, e.g. 0.49 and 0.51, belong to different when AXs take place between machines and humans.
categories. We leave to future work the analysis of further value During AXs, agents contribute elements of the at-
categorisations, e.g. a distinction between strongly and mildly pos- tack/support relations, thus “arguing” with one another.
itive values or comfort zones (de Tarlé, Bonzon, and Maudet 2022). These elements cannot be withdrawn once contributed, in
line with human practices, and, by definition of C, each ele- the contributing agents or the intrinsic quality of the argu-
ment is said once by exactly one agent, thus avoiding repe- ments. Depending on these biases, learnt attacks and sup-
titions that may occur in human exchanges. Note that we ports may influence the agents’ stances on the explanan-
do not require that all agents contribute something to an dum differently. For illustration, in Example 3, η opted
AX, namely it may be that {α∣C((a, b)) = (α, t), (a, b) ∈ for a low bias (0.2) on the learnt argument c, resulting in
Anx ∪ Sxn } ⊂ AG. Also, we do not force agents to contribute ση2 (Q2η , e) = 0.648, ση2 (Q2η , a) = 0.32 and ση2 (Q2η , c) = 0.2,
something at every timestep (i.e. it may be the case that and thus Σ2η (e) = + still, as in Examples 1, 2. If, instead, η
Bxt−1 = Bxt at some timestep t). Further, while the definition had chosen a high bias on the new argument, e.g. τη2 (c) = 1,
of AX does not impose that agents are truthful, from now on
we will focus on truthful agents only and thus assume that if this would have given ση2 (Q2η , e) = 0.432, ση2 (Q2η , a) = 0.88
(a, b) ∈ Anx or Sxn and C((a, b)) = (α, t) (with 0 < t ≤ n), and ση2 (Q2η , c) = 1, leading to Σ2η (e) = −, thus resolving the
then, resp., (a, b) ∈ Aα t−1
or Sαt−1 . conflict. This illustration shows that learnt attacks, supports
In the remainder, we may denote the private triple and arguments may fill gaps, change agents’ stances on ex-
(Itα , Qtα , σαt ) as αt and the stance Σα (Qtα , a) as Σtα (a). plananda and pave the way to the resolution of conflicts.
Example 2. An AX amongst {µ, η} from Example 1 may be Definition 7. Let E = ⟨Bx0 , . . . , Bxn , AG0 , . . . , AGn , C⟩ be
⟨Bx0 , Bx1 , AG0 , AG1 , C⟩ such that (see top row of Figure 2): an AX for explanandum e amongst agents AG such that
Σ0α (e) ≠ Σ0β (e) for some α, β ∈ AG. Then:
• Bx0 = ⟨{e}, ∅, ∅⟩, Bx1 = ⟨{e, a, b}, {(a, e)}, {(b, e)}⟩;
• E is resolved at timestep t, for some 0 < t ≤ n, iff ∀α, β ∈
• µ0 = µ1 and η 0 = η 1 are as in Example 1; AG, Σtα (e) = Σtβ (e), and is unresolved at t otherwise;
• C((a, e)) = (µ, 1) and C((b, e)) = (η, 1), i.e. µ and η • E is resolved iff it is resolved at timestep n and it is unre-
contribute, resp., attack (a, e) and support (b, e) at 1. solved at every timestep 0 ≤ t < n;
Here, each agent contributes a single attack or support jus- • E is unresolved iff it is unresolved at every 0 < t ≤ n.
tifying their stances (negative for µ and positive for η), but, Thus, a resolved AX starts with a conflict between at least
in general, multiple agents may contribute multiple relations two agents and ends when no conflicts amongst any of the
at single timesteps, or no relations at all. agents exist or when the agents give up on trying to find
When contributed attacks/supports are new to agents, they a resolution. Practically, AXs may be governed by a turn-
may (rote) learn them, with the arguments they introduce. making function π ∶ Z+ → 2AG determining which agents
should contribute at any timestep. Then, an AX may be
Definition 6. Let ⟨Bx0 , . . . , Bxn , AG0 , . . . , AGn , C⟩ be an AX deemed to be unresolved if, for example, all agents decide,
amongst agents AG. Then, for any α ∈ AG, with private when their turn comes, against contributing.
tuples (I0α , Q0α , σα0 ), . . . , (Inα , Qnα , σαn ): Note that, while agents’ biases and evaluations are kept
• for any 0 < t ≤ n, for ⟨Xα , Aα , Sα , τα ⟩ = Qtα ∖ Qt−1α , private during AXs, we assume that agents share their
Xα , Aα , and Sα are, resp., the learnt arguments, attacks, stances on the explanandum, so that they are aware of
and supports by α at timestep t; whether the underpinning conflicts are resolved. Agents’
• for ⟨Xα , Aα , Sα , τα ⟩ = Qnα ∖ Q0α , Xα , Aα , and Sα are, stances, when ascertaining whether an AX is resolved, are
resp., the learnt arguments, attacks, and supports by α. evaluated internally by the agents, without any shared evalu-
ation of the exchange BAF, unlike, e.g. in (de Tarlé, Bonzon,
Note that, by definition of AXs, all learnt arguments, at- and Maudet 2022) and other works we reviewed in §2.
tacks and supports are from the (corresponding) exchange Finally, note that our definition of AX is neutral as to the
BAFs. Note also that in Example 2 neither agent learns any- role of agents therein, allowing in particular that agents have
thing, as indeed each contributed an attack/support already symmetrical roles (which is natural, e.g., for inquiry) as well
present in the other agent’s private QBAF. as asymmetrical roles (which is natural, e.g., when machines
Example 3. Let us extend the AX from Example 2 to obtain explain to humans: this will be our focus from §5).
⟨Bx0 , Bx1 , Bx2 , AG0 , AG1 , AG2 , C⟩ such that (see the top two
rows of Figure 2): 5 Explanatory Properties of AXs
• Bx2 = ⟨{e, a, b, c}, {(a, e)}, {(b, e), (c, a)}⟩ Here we focus on singling out desirable properties that AXs
may need satisfy to support interactive XAI. Let us assume
• µ = µ1 = µ0 ; η 2 is such that Q2η ⊐ Q1η where Xη2 =
2
as given an AX E = ⟨Bx0 , . . . , Bxn , AG0 , . . . , AGn , C⟩ for e
Xη1 ∪ {c}, A2η = A1η , Sη2 = Sη1 ∪ {(c, a)} and τη2 (c) = 0.2; as in Definition 5. The first three properties impose basic re-
• C((c, a)) = (µ, 2), namely µ contributes the support quirements on AXs so that they result in fitting explanations.
(c, a) in Bx2 at timestep 2. Property 1. E satisfies connectedness iff for any 0 ≤ t ≤ n,
We will impose that any attack/support which is added if ∣Xxt ∣ > 1 then ∀a ∈ Xxt , ∃b ∈ Xxt such that (a, b) ∈ Atx ∪ Sxt
to the exchange BAF by an agent is learnt by the other or (b, a) ∈ Atx ∪ Sxt .
agents, alongside any new arguments introduced by those at- Basically, connectedness imposes that there should be no
tacks/supports. Thus, for any α∈AG and t>0, Bxt ∖Bxt−1 ⊑Qtα . floating arguments and no “detours” in the exchange BAFs,
However, agents have a choice on their biases on the learnt at any stage during the AX. It is linked to directional con-
arguments. These biases could reflect, e.g., their trust on nectedness in (Cyras, Kampik, and Weng 2022). A violation
of this property would lead to counter-intuitive (interactive) of pro or con arguments (depending on how stances have
explanations, with agents seemingly “off-topic”. changed). For example, in Figure 2, b, d, f are pro argu-
Property 2. E satisfies acyclicity iff for any 0 ≤ t ≤ n, ments which could justify an increase in stance for e, while
∄a ∈ Xxt such that paths(a, a) ≠ ∅. a, c are con arguments which could justify its decrease.
Note that this property does not hold in general, e.g., given
Acyclicity ensures that all reasoning is directed towards an agent which (admittedly counter-intuitively) increases its
the explanandum in AXs. A violation of this property may evaluation of arguments when they are attacked. However, it
lead to seemingly non-sensical (interactive) explanations. holds for some evaluation models, notably DF-QuAD again:
Property 3. E satisfies contributor irrelevance iff for any AX Proposition 3. If E is resolved and ∀α ∈ AG, σα is DF-
′ ′ ′ ′
for e ⟨Bx0 , . . . , Bxn ′ , AG0 , . . . , AGn , C ′ ⟩, if Bx0 = Bx0 , Bxn ′ = QuAD, then E satisfies resolution representation.
′
Bxn , AG0 =AG0 , then ∀α ∈ AG: Σnα (Qnα , e)=Σnα (Qnα ′ , e). The final property we consider concerns unresolved AXs,
Contributor irrelevance ensures that the same final ex- in the same spirit as resolution representation.
change BAF results in the same stances for all agents, re- Property 5. E satisfies conflict representation iff E is unre-
gardless of the contributors of its attacks and supports or the solved, pro(Bxn ) ≠ ∅ and con(Bxn ) ≠ ∅.
order in which they were contributed.
These three properties are basically about the exchange This property thus requires that the conflict in an unre-
BAFs in AXs, and take the viewpoint of an external “judge” solved AX is apparent in the exchange BAF, namely it in-
for the explanatory nature of AXs. These basic properties cludes both pro and con arguments (representing the con-
are all satisfied, by design, by AXs:5 flicting stances). For example, if the AX in Figure 2 con-
cluded unresolved at t = 2, this property requires that Bx2
Proposition 1. Every AX satisfies Properties 1 to 3. contains both pro arguments for e (e.g. a or c) and con ar-
We now introduce properties which AXs may not always guments against it (e.g. b). This property does not hold in
satisfy, but which, nonetheless, may be desirable if AXs are general, e.g. for an agent who rejects all arguments by im-
to generate meaningful (interactive) explanations. First, we posing on them minimum biases and contributes no attack
define notions of pro and con arguments in AXs, amounting or support. Proving that this property holds requires consid-
to positive and negative reasoning towards the explanandum. eration of the agents’ behaviour, which we examine next.
Definition 8. Let B = ⟨X , A, S⟩ be any BAF for e. Then,
the pro arguments and con arguments for B are, resp.: 6 Agent Behaviour in AXs for XAI
●pro(B)={a∈X ∣∃p∈paths(a, e), where ∣p ∩ A∣ is even}; All our examples so far have illustrated how AXs may sup-
●con(B)={a∈X ∣∃p∈paths(a, e), where ∣p ∩ A∣ is odd}. port explanatory interactions amongst a machine µ and a hu-
Note that the intersection of pro and con arguments may man η. This specific XAI setting is our focus in the remain-
be non-empty as multiple paths to explananda may exist, so der, where we assume AG = {µ, η}. Also, for simplicity, we
an argument may bring both positive and negative reasoning. impose (as in all illustrations) that Iµ = Iη = [0, 1], I−µ = I−η =
Pro/con arguments with an even/odd, resp., number of [0, 0.5), I0µ = I0η = {0.5} and I+µ = I+η = (0.5, 1]. We also
attacks in their path to e are related to chains of sup- restrict attention to AXs governed by a turn-making func-
ports (supported/indirect defeats, resp.) in (Cayrol and tion π imposing a strict interleaving such that π(i) = {µ} if
Lagasquie-Schiex 2005) (we leave the study of formal links i is odd, and π(i) = {η} otherwise (thus, in particular, the
to future work). Pro/con arguments are responsible for in- machine starts the interactive explanation process).
creases/decreases, resp., in e’s strength using DF-QuAD: In line with standard argumentative XAI, the machine
Proposition 2. For any α ∈ AG, let σα indicate the evalua- may draw the QBAF in its initial private triple (at t = 0)
tion method by DF-QuAD. Then, for any 0 < t ≤ n: from the model it is explaining. This QBAF may be obtained
● if σα (Qtα , e) > σα (Qt−1 t t−1 by virtue of some abstraction methodology or may be the ba-
α , e), then pro(Bx ) ⊃ pro(Bx );
● if σα (Qα , e)<σα (Qα , e), then con(Bx ) ⊃ con(Bxt−1 ).
t t−1 t sis of the model itself (see (Cyras et al. 2021)). The humans,
instead, may draw the QBAF in their initial private triple, for
We conjecture (but leave to future work) that this result example, from their own knowledge, biases, and/or regula-
(and more later) holds for other gradual semantics satisfying tions on the expected machine’s behaviour. The decision on
monotonicity (Baroni, Rago, and Toni 2019) or bi-variate the evaluation method, for machines and humans, may be
monotony/reinforcement (Amgoud and Ben-Naim 2018). dictated by specific settings and desirable agent properties
Property 4. E satisfies resolution representation iff E is re- therein. Here we focus on how to formalise and evaluate in-
solved and ∀α ∈ AG: if Σnα (e) > Σ0α (e), then pro(Bxn ) ≠ ∅; teractive explanations between a machine and a human using
and if Σnα (e) < Σ0α (e), then con(Bxn ) ≠ ∅. AXs, and ignore how their initial private triples are obtained.
This property also takes the viewpoint of an external Below we define various behaviours dictating how ma-
“judge”, by imposing that the final exchange BAF convinc- chines and humans can engage in AXs for XAI, focusing on
ingly represents a resolution of the conflicts between agents’ ways to i) determine their biases and ii) decide their contri-
stances, thus showing why stances were changed. Specif- butions (attacks/supports) to (unresolved) AXs.
ically, it imposes that a changed stance must be the result Biases. As seen in §4, the degree to which learnt at-
tacks/supports impact the stances of agents on explananda
5
Proofs for all propositions are in the supplementary material. is determined by the agents’ biases on the learnt arguments.
In XAI different considerations regarding this learning ap- agent, without accessing their private QBAFs, thus differ-
ply to machines and humans. Firstly, not all machines may ing from other works, e.g. (de Tarlé, Bonzon, and Maudet
be capable of learning: simple AI systems which provide ex- 2022), which rely on shared evaluations: in our case, rea-
planations but do not have the functionality for understand- soning is shared but it is not evaluated in a shared manner.
ing any input from humans are common in AI. Secondly, The shallow behaviour selects a (bounded by max) max-
machines capable of learning may assign different biases to imum number of supports for/attacks against the explanan-
the learnt arguments: a low bias indicates scepticism while dum if the agent is arguing for/against, resp., it, as follows:
a high bias indicates credulity. Machines may be designed
to give low biases to arguments from sources which cannot Definition 11. Let max ∈ N. Agent α ∈ AG exhibits shallow
be trusted, e.g. when the expertise of a human is deemed behaviour (wrt max) iff, at any 0 ≤ t < n where π(t) = {α},
insufficient, or high biases to arguments when the human C = {(a, b)∣C((a, b)) = (α, t)} is a maximal (wrt cardinal-
is deemed competent, e.g. in debugging. Here, we refrain ity) set {(a1 , e), . . . , (ap , e)} with p ≤ max such that:
from accommodating such challenges and focus instead on • if α is arguing for e then C⊆Sαt−1 ∖Sxt−1 where ∀i∈{1, . . . ,
the restrictive (but sensible, as a starting point) case where p},∄(b, e)∈Sαt−1∖(Sxt−1 ∪ C) with σαt−1 (b)>σαt−1 (ai );
machines assign constant biases to arguments from humans.
• if α is arguing against e then C⊆At−1 t−1
α ∖Ax where ∀i∈{1,
Definition 9. Let c ∈ [0, 1] be a chosen constant. For any . . . , p},∄(b, e)∈Aα ∖(Ax ∪C) with σα (b)>σαt−1 (ai ).
t−1 t−1 t−1

learnt argument a ∈ Xµt ∖ Xµt−1 at timestep t, τµt (a) = c.

This behaviour thus focuses on reasoning for or against
If c = 0 then the machine is unable to learn, whereas the explanandum e exclusively. It selects supports or at-
0 < c < 1 gives partially sceptical machines and c = 1 tacks in line with the agent’s stance on e and with the
gives credulous machines. The choice of c thus depends highest evaluation in the contributing agent’s private QBAF.
on the specific setting of interest, and may have an impact This behaviour is inspired by static explanation methods in
on the conflict resolution desideratum for AXs. For exam- XAI, which deliver all information in a single contribution.
ple, let µ use DF-QuAD as its evaluation method: if c = 1 Clearly, if we let µ exhibit this shallow behaviour and η be
we can derive guarantees of rejection/weakening or accep- unresponsive, i.e. never contribute any attack/support, then
tance/strengthening of arguments which are attacked or sup- the AX cannot satisfy conflict representation.
ported, resp., by learnt arguments,6 demonstrating the po- The greedy behaviour allows an agent arguing for e to
tential (and dangers) of credulity in machines (see §7). support the pro or attack the con arguments, while that argu-
Humans, meanwhile, typically assign varying biases to ing against can support the con or attack the pro arguments.
arguments based on their own internal beliefs. These as-
Definition 12. Agent α ∈ AG exhibits greedy behaviour iff,
signments may reflect cognitive biases such as the confirma-
at any 0 ≤ t < n where π(t) = {α}, C = {(a, b)∣C((a, b)) =
tion bias (Nickerson 1998) – the tendency towards looking
(α, t)} is empty or amounts to a single attack or support
favourably at evidence which supports one’s prior views. In
(a, b) ∈ (At−1 t−1 t−1 t−1
α ∪ Sα ) ∖ (Ax ∪ Sx ) such that:
§7 we model humans so that they assign random biases to
learnt arguments, but explore confirmation bias by applying 1. if α is arguing for e then: (a, b) ∈ Sαt−1 and b ∈
a constant offset to reduce the bias assigned by the human. pro(Bxt−1 ) ∪ {e}; or (a, b) ∈ At−1 α and b ∈ con(Bxt−1 );
This differs, e.g., from the modelling of confirmation bias in if α is arguing against e then: (a, b) ∈ Sαt−1 and b ∈
(de Tarlé, Bonzon, and Maudet 2022), acting on the prob- con(Bxt−1 ); or (a, b) ∈ At−1 t−1
α and b ∈ pro(Bx ) ∪ {e};
ability of an argument being learned. We leave the explo- 2. ∄(a′ , b′ ) ∈ (At−1 t−1 t−1 t−1
α ∪ Sα ) ∖ (Ax ∪ Sx ) satisfying 1.
ration of alternatives for assigning biases to future work. t−1 ′
such that σα (a ) > σα (a); t−1
Attack/Support Contributions. We consider shallow,
greedy and counterfactual behaviours: intuitively, the first 3. ∄(a′′ , b′′ ) ∈ (At−1 t−1 t−1 t−1
α ∪ Sα ) ∖ (Ax ∪ Sx ) satisfying 1.
t−1 ′′ t−1
corresponds to the one-shot explanations in most XAI, such that σα (a ) = σα (a) and
the second contributes the (current) strongest argument in ∣argminP ′′ ∈paths(a′′ ,e) ∣P ′′ ∣ ∣ < ∣argminP ∈paths(a,e) ∣P ∣ ∣.
favour of the agent position, and the third considers how Intuitively, 1. requires that the attack or support, if any,
each attack/support may (currently) affect the exchange is in line with the agent’s views; 2. ensures that the attack-
BAF before it is contributed. All behaviours identify ar- ing or supporting argument has maximum strength; and 3.
gument pairs to be added to the exchange BAF as attacks ensures that it is “close” to the explanandum. We posit that
or supports reflecting their role in the private QBAFs from enforcing agents to contribute at most one argument per turn
which they are drawn. We use the following notion: will aid minimality without affecting conflict resolution neg-
Definition 10. For E resolved at timestep t, if Σtµ (e) > atively wrt the shallow behaviour (see §7). Minimality is a
common property of explanations in XAI, deemed benefi-
Σtη (e) then the states of µ and η at t are, resp., arguing for
cial both from a machine perspective, e.g. wrt computational
and arguing against e (else, the states are reversed).
aspects (see computational complexity in (Sokol and Flach
The agents’ states point to a “window for persuasion”, 2020)), and from a human perspective, e.g. wrt cognitive
whereby an agent arguing for (against) e may wish to at- load and privacy maintenance (see parsimony in (Sokol and
tempt to increase (decrease, resp.) the stance of the other Flach 2020)). Naturally, however, conflict resolution in AXs
should always take precedence over minimality, as prioritis-
6
Propositions on such effects are in the supplementary material. ing the latter would force AXs to remain empty.
Proposition 4. If E is unresolved and ∀α ∈ AG: α exhibits Identifying attacks and supports based on their effect on
greedy behaviour and {(a,b)∈Anx ∪ Sxn ∣C((a, b))=(α, t), t ∈ the explanandum is related to proponent and opponent argu-
{1, . . . , n}}≠∅, then E satisfies conflict representation. ments in (Cyras, Kampik, and Weng 2022), defined however
Proposition 5. If ∀α ∈ AG, for all 0 ≤ t < n and ∀a ∈ Xαt , in terms of quantitative dispute trees for BAFs.
paths((a, e)) = {(a, e)}, then the shallow (with max = 1) The counterfactual behaviour may better consider argu-
and greedy behaviours are aligned. mentative structure, towards resolved AXs, as shown next.
The greedy behaviour may not always lead to resolutions: Example 5. Consider the AX from Example 4 but where:
Example 4. Let us extend the AX from Example 3 to • Bx3 =⟨{e, a, b, c, f }, {(a, e)}, {(b, e), (c, a), (f, b)}⟩;
⟨Bx0 , . . . , Bx3 , AG0 , . . . , AG3 , C⟩ such that (see Figure 2): • µ3 is such that Q3µ ⊐ Q2µ where Xµ3 = Xµ2 ∪ {f }, A3µ =
• Bx3 =⟨{e, a, b, c, d}, {(a, e), (d, a)}, {(b, e), (c, a)}⟩; A2µ , Sµ3 = Sµ2 ∪ {(f, b)}, τµ3 (f ) = 0.5; then, the argument
• η 3 = η 2 ; µ3 is such that Q3µ ⊐ Q2µ where Xµ3 = Xµ2 ∪ {d}, evaluations are σµ3 (Q3µ , e) = 0.546, σµ3 (Q3µ , a) = 0.92,
A3µ = A2µ ∪ {(d, a)}, Sµ3 = Sµ2 , τµ3 (d) = 0.6; then, the σµ3 (Q3µ , b) = 0.7, σµ3 (Q3µ , c) = 0.6 and σµ3 (Q3µ , f ) = 0.5;
argument evaluations are σµ3 (Q3µ , e) = 0.42, σµ3 (Q3µ , a) = • C((f, b)) = (η, 3), i.e. η contributes support (f, b) to Bx3 .
0.8, σµ3 (Q3µ , b) = 0.4, σµ3 (Q3µ , c) = 0.6, σµ3 (Q3µ , d) = 0.6; Here, η contributes (f, b) in line with the counterfactual be-
• C((d, a)) = (η, 3), i.e. η contributes attack (d, a) at t = 3. haviour as ϵ((f, b), Q3η ) = 0.24 > ϵ((d, a), Q3η ) = 0.216.
This sufficiently modifies µ’s private QBAF such that Σ3µ =
Here, in line with the greedy behaviour, µ learns the attack
+, and the AX is now resolved: the counterfactual behaviour
(d, a) contributed by η at timestep 3. Then, even if µ assigns
succeeds where the greedy behaviour did not (Example 4).
the same bias to these learnt arguments as η (which is by no
means guaranteed), this is insufficient to change the stance, We end showing some conditions under which conflict
i.e. Σ3µ (e) = −, and so the AX remains unresolved. representation is satisfied by the counterfactual behaviour.
The final counterfactual behaviour takes greater consid- Proposition 6. If E is unresolved and is such that ∀α ∈
eration of the argumentative structure of the reasoning avail- AG: α exhibits counterfactual behaviour; σα is DF-QuAD;
able to the agents in order to maximise the chance of conflict {(a, b) ∈ Anx ∪ Sxn ∣C((a, b)) = (α, t), t ∈ {1, . . . , n}} ≠ ∅;
resolution with a limited number of arguments contributed. then E satisfies conflict representation.
This behaviour is defined in terms of the following notion.
Definition 13. Given an agent α ∈ AG, a private view 7 Evaluation
of the exchange BAF by α at timestep t is any Qtαv = We now evaluate sets of AXs obtained from the behaviours
⟨Xαvt
, Atαv , Sαv t t
, ταv ⟩ such that Bxt ⊑ Qtαv ⊑ Qtα . from §6 via simulations, using the following metrics:7
Resolution Rate (RR): the proportion of resolved AXs.
An agent’s private view of the exchange BAF thus
Contribution Rate (CR): the average number of argu-
projects their private biases onto the BAF, while also po-
ments contributed to the exchange BAFs in the resolved
tentially accommodating counterfactual reasoning with ad-
AXs, in effect measuring the total information exchanged.
ditional arguments. Based on arguments’ evaluations in an
Persuasion Rate (PR): for an agent, the proportion of
agent’s private view, the agent can then judge which attack
resolved AXs in which the agent’s initial stance is the other
or support it perceives will be the most effective.
agent’s final stance, measuring the agent’s persuasiveness.
Definition 14. Given an agent α ∈ AG, α’s perceived effect Contribution Accuracy (CA): for an agent, the propor-
t−1
on e at 0<t≤n of any (a, b) ∈ (Aα ∪Sαt−1 )∖(At−1 t−1
x ∪Sx ), tion of the contributions which, if the agent was arguing
t−1 t−1 t−1 t
where a ∈ Xα ∖ Xx and b ∈ Xx , is ϵ((a, b), Qα ) = for (against) e, would have maximally increased (decreased,
σα (Qtαv , e) − σα (Qαv t−1
, e) for Qtαv ⊐ Qt−1 αv a private view resp.) e’s strength in the other agent’s private QBAF.
t t−1
of the exchange BAF at t by α such that Xαv = Xαv ∪ {a}, We tested PR and CA for machines only. Let unrespon-
Atαv = (Xαv t
× Xαvt
) ∩ At−1
α and S t
αv = (X t
αv × X t
αv ) ∩ Sαt−1 . sive behaviour amount to contributing nothing (as in §6).
The counterfactual view underlying this notion of per- Then, our hypotheses were:
ceived effect relates to (Kampik and Cyras 2022), although H1: For a shallow machine and an unresponsive human,
we consider the effect of adding an attack or support, as the max constant increases, RR, CR and CA increase.
whereas they consider an argument’s contribution by remov- H2: For a shallow machine and an unresponsive human,
ing it. It also relates to the hypothetical value of (de Tarlé, as the human’s confirmation bias increases, RR decreases.
Bonzon, and Maudet 2022), which however amounts to the H3: For a greedy machine and a counterfactual human,
explanandum’s evaluation in the shared graph. RR increases relative to a shallow machine and an unrespon-
sive human.
Definition 15. Agent α ∈ AG exhibits counterfactual be- H4: For a greedy machine and a counterfactual human,
haviour iff, at any 0 ≤ t < n where π(t) = {α}, C = as the machine’s bias on learnt arguments increases, RR in-
{(a, b)∣C((a, b))=(α, t)} is empty or is {(a, b)} such that: creases while CR and PR decrease.
● if α is arguing for e then ϵ((a, b), Qtα ) > 0 and (a, b) is H5: For a counterfactual machine and a counterfactual
′ ′ t
argmax(a′ ,b′ )∈(At−1 α ∪Sα )∖(Ax ∪Sx )
t−1 t−1 t−1 ϵ((a , b ), Q );
α human, RR and CA increase relative to a greedy machine.
t
● if α is arguing against e then ϵ((a, b), Qα )<0 and (a,b)
′ ′ t 7
is argmin(a′ ,b′ )∈(At−1 α ∪Sα )∖(Ax ∪Sx )
t−1 t−1 t−1 ϵ((a , b ), Q ).
α See the supplementary material for exact formulations.
 Experimental Setup. For each AX for e (restricted as in Behaviour Learning
RR CR PRµ CAµ
§6), we created a “universal BAF”, i.e. a BAF for e of which µ η µ η
all argumentation frameworks are subgraphs. We populated S (1) - - 0 5.4 1 100 45.4
the universal BAFs with 30 arguments by first generating a S (2) - - 0 9.6 1.96 100 51.9
6-ary tree with e as the root. Then, any argument other than H1 S (3) - - 0 13.0 2.76 100 56.7
e had a 50% chance of having a directed edge towards a S (4) - - 0 13.9 3.22 100 58.1
random previous argument in the tree, to ensure that multi- S (5) - - 0 13.7 3.38 100 58.3
ple paths to the explanandum are present. 50% of the edges S (4) - - -0.1 11.2 3.26 100 57.6
in the universal BAF were randomly selected to be attacks, S (4) - - -0.2 8.6 3.27 100 58.0
H2
and the rest to be supports. We built agents’ private QBAFs S (4) - - -0.3 6.7 3.30 100 58.3
from the universal BAF by performing a random traversal S (4) - - -0.4 5.3 3.38 100 58.5
through the universal BAF and stopped when the QBAFs G (≤3) C 0 -0.2 9.8 3.15 83.7 38.8
reached 15 arguments, selecting a random argument from H3 G (≤4) C 0 -0.2 11.9 3.88 79.0 37.1
each set of children, as in (de Tarlé, Bonzon, and Maudet G C 0 -0.2 18.8 7.16 79.3 35.7
2022). We then assigned random biases to arguments in the G C 0.5 -0.2 42.2 6.73 31.5 37.5
H4
agents’ QBAFs, and (possibly different) random evaluation G C 1.0 -0.2 55.5 5.24 20.4 38.2
methods to agents amongst QuAD (Baroni et al. 2015), DF- H5 C C 0.5 -0.2 48.4 7.37 41.5 50.5
QuAD, REB (Amgoud and Ben-Naim 2017) and QEM (Po-
tyka 2018) (all with evaluation range [0, 1]). We used differ- Table 1: Results in the simulations for the five hypotheses for three
ent evaluation methods to simulate different ways to evalu- behaviours: Shallow (max constant given in parentheses); Greedy
ate arguments in real-world humans/machines. We repeated (where any limit on the number of contributed arguments by the
this process till agents held different stances on e. agent is in brackets); and Counterfactual. Learning amounts to c in
For each hypothesis, we ran 1000 experiments per config- Definition 9 for µ and to the confirmation bias offset for η (where
uration, making sure the experiments for different strategies appropriate). We report RR, PRµ and CAµ as percentages. We
are run with the same QBAFs. We ran the simulations on indicate in bold the chosen baseline for the next hypothesis.
the NetLogo platform using BehaviorSpace.8 We tested the
significance between testing conditions in a pairwise man-
ner using the chi-squared test for the discrete measures RR we restricted the greedy machine’s contributed arguments to
and PR, and Student’s t-test for the continuous measures CR 4 (p < 0.005), but not to 3 (p = 0.202).
and CA. We rejected the null hypotheses when p < 0.01. H4: RR increased significantly with the bias on learnt
Experimental Results. Table 1 reports the results of our arguments (p < 0.001 for both comparisons of learning con-
simulations: all hypotheses were (at least partially) verified. figurations: 0 vs 0.5 and 0.5 vs 1). However, the machine’s
H1: As expected, increasing max for shallow machines CR and PR fell significantly (p < 0.001 for similar pairwise
results in significantly higher RR, CR and CA up to max = comparisons, except for 0 vs 0.5 for CR, where p = 0.27).
3 (p < 0.005 for max values of 1 vs 2 and 2 vs 3 for all highlighting the naive nature of machines learning credu-
metrics). Above this limit (max values of 3 vs 4 and 4 vs 5), lously (i.e. assigning all learnt arguments the top bias).
this trend was no longer apparent, suggesting that there was H5: The counterfactual behaviour outperformed the
a limit to the effectiveness of contributing arguments at this greedy behaviour significantly in terms of both RR (p <
distance from e. Note that the machine’s PR is always 100% 0.01) and CA (p < 0.001), showing, even in this limited set-
here, since the (unresponsive) human does not contribute. ting, the advantages in taking a counterfactual view, given
H2: We fixed max = 4 (the value with the maximum that the strongest argument (as selected by the greedy be-
RR for H1) and found that increasing the confirmation bias haviour) may not always be the most effective in persuading.
in the human significantly decreased the machine’s RR ini-
tially (p < 0.01 for 0 vs −0.1 and −0.1 vs −0.2), before the 8 Conclusions
effect tailed off as RR became very low (p = 0.09 for −0.2 We defined the novel concept of AXs, and deployed AXs
vs −0.3 and p = 0.03 for −0.3 vs −0.4), demonstrating the in the XAI setting where a machine and a human engage
need for behaviours which consider deeper reasoning than in interactive explanations, powered by non-shallow reason-
the shallow behaviour to achieve higher resolution rates. ing, contributions from both agents and modelling of agents’
H3: From here onwards we tested with a counterfactual learning and explanatory behaviour. This work opens sev-
human9 and fixed the level of confirmation bias therein to eral avenues for future work, besides those already men-
−0.2. We compared shallow against greedy machines, also tioned. It would be interesting to experiment with any num-
limiting the number of arguments they contributed to max- ber of agents, besides the two that are standard in XAI,
ima of three and four to compare fairly with the shallow and to identify restricted cases where hypotheses H1-H5 are
machine with the fixed max constant. RR increased signifi- guaranteed to hold. It would also be interesting to accommo-
cantly with the greedy behaviour (p < 0.001), over the shal- date mechanisms for machines to model humans, e.g. as in
low machine which remained statistically significant when opponent modelling (Hadjinikolis et al. 2013). Also fruitful
could be an investigation of how closely AXs can represent
8
See github.com/CLArg-group/argumentative exchanges. machine and human behaviour. Further, while we used AXs
9
Experiments with greedy humans gave similar findings. in XAI, they may be usable in various multi-agent settings.
 Acknowledgements Calegari, R.; Riveret, R.; and Sartor, G. 2021. The burden
This research was partially funded by the ERC under the of persuasion in structured argumentation. In ICAIL 2021,
EU’s Horizon 2020 research and innovation programme 180–184.
(No. 101020934, ADIX) and by J.P. Morgan and by the Cawsey, A. 1991. Generating interactive explanations. In
Royal Academy of Engineering, UK. AAAI 1991, 86–91.
Cayrol, C., and Lagasquie-Schiex, M. 2005. On the accept-
References ability of arguments in bipolar argumentation frameworks.
Albini, E.; Lertvittayakumjorn, P.; Rago, A.; and Toni, F. In ECSQARU 2005, 378–389.
2020. DAX: deep argumentative explanation for neural net- Cocarascu, O.; Rago, A.; and Toni, F. 2019. Extracting
works. CoRR abs/2012.05766. dialogical explanations for review aggregations with argu-
Amgoud, L., and Ben-Naim, J. 2017. Evaluation of ar- mentative dialogical agents. In AAMAS 2019, 1261–1269.
guments in weighted bipolar graphs. In ECSQARU 2017, Cyras, K.; Rago, A.; Albini, E.; Baroni, P.; and Toni, F.
25–35. 2021. Argumentative XAI: A survey. In IJCAI 2021, 4392–
Amgoud, L., and Ben-Naim, J. 2018. Evaluation of argu- 4399.
ments in weighted bipolar graphs. Int. J. Approx. Reason. Cyras, K.; Kampik, T.; and Weng, Q. 2022. Dispute trees
99:39–55. as explanations in quantitative (bipolar) argumentation. In
Amgoud, L., and Ben-Naim, J. 2022. Axiomatic founda- ArgXAI 2022 co-located with COMMA 2022.
tions of explainability. In IJCAI 2022, 636–642. de Tarlé, L. D.; Bonzon, E.; and Maudet, N. 2022. Mul-
Antaki, C., and Leudar, I. 1992. Explaining in conversation: tiagent dynamics of gradual argumentation semantics. In
Towards an argument model. Europ. J. of Social Psychology AAMAS 2022, 363–371.
22:181–194. Donadello, I.; Hunter, A.; Teso, S.; and Dragoni, M. 2022.
Atkinson, K.; Baroni, P.; Giacomin, M.; Hunter, A.; Machine learning for utility prediction in argument-based
Prakken, H.; Reed, C.; Simari, G. R.; Thimm, M.; and Vil- computational persuasion. In AAAI 2022, 5592–5599.
lata, S. 2017. Towards artificial argumentation. AI Magazine Dung, P. M. 1995. On the Acceptability of Arguments and
38(3):25–36. its Fundamental Role in Nonmonotonic Reasoning, Logic
Balog, K.; Radlinski, F.; and Arakelyan, S. 2019. Transpar- Programming and n-Person Games. Artificial Intelligence
ent, scrutable and explainable user models for personalized 77(2):321–358.
recommendation. In SIGIR 2019, 265–274. Fan, X., and Toni, F. 2012a. Argumentation dialogues for
Baroni, P.; Romano, M.; Toni, F.; Aurisicchio, M.; and two-agent conflict resolution. In COMMA 2012, 249–260.
Bertanza, G. 2015. Automatic evaluation of design alter- Fan, X., and Toni, F. 2012b. Mechanism design for
natives with quantitative argumentation. Argument Comput. argumentation-based persuasion. In COMMA 2012, 322–
6(1):24–49. 333.
Baroni, P.; Comini, G.; Rago, A.; and Toni, F. 2017. Ab- Fan, X., and Toni, F. 2015a. Mechanism design for
stract games of argumentation strategy and game-theoretical argumentation-based information-seeking and inquiry. In
argument strength. In PRIMA 2017, 403–419. PRIMA 2015, 519–527.
Baroni, P.; Gabbay, D.; Giacomin, M.; and van der Torre, Fan, X., and Toni, F. 2015b. On computing explanations in
L., eds. 2018. Handbook of Formal Argumentation. College argumentation. In AAAI 2015, 1496–1502.
Publications.
Hadjinikolis, C.; Siantos, Y.; Modgil, S.; Black, E.; and
Baroni, P.; Rago, A.; and Toni, F. 2018. How many proper- McBurney, P. 2013. Opponent modelling in persuasion dia-
ties do we need for gradual argumentation? In AAAI 2018, logues. In IJCAI 2013, 164–170.
1736–1743.
Hirsch, T.; Soma, C. S.; Merced, K.; Kuo, P.; Dembe, A.;
Baroni, P.; Rago, A.; and Toni, F. 2019. From fine-grained Caperton, D. D.; Atkins, D. C.; and Imel, Z. E. 2018. “It’s
properties to broad principles for gradual argumentation: A hard to argue with a computer”: Investigating psychothera-
principled spectrum. Int. J. Approx. Reason. 105:252–286. pists’ attitudes towards automated evaluation. In DIS 2018,
Bertrand, A.; Belloum, R.; Eagan, J. R.; and Maxwell, W. 559–571.
2022. How cognitive biases affect XAI-assisted decision- Hunter, A. 2018. Towards a framework for computational
making: A systematic review. In AIES ’22, 78–91. persuasion with applications in behaviour change. Argument
Black, E., and Atkinson, K. 2011. Choosing persuasive Comput. 9(1):15–40.
arguments for action. In AAMAS 2011, 905–912. Ignatiev, A.; Narodytska, N.; and Marques-Silva, J. 2019.
Black, E., and Hunter, A. 2007. A generative inquiry dia- Abduction-based explanations for machine learning models.
logue system. In AAMAS 2007, 241. In AAAI 2019, 1511–1519.
Calegari, R.; Omicini, A.; Pisano, G.; and Sartor, G. 2022. Kampik, T., and Cyras, K. 2022. Explaining change in
Arg2P: an argumentation framework for explainable intelli- quantitative bipolar argumentation. In COMMA 2022, 188–
gent systems. J. Log. Comput. 32(2):369–401. 199.
Kontarinis, D., and Toni, F. 2015. Identifying malicious Vassiliades, A.; Bassiliades, N.; and Patkos, T. 2021. Ar-
behavior in multi-party bipolar argumentation debates. In gumentation and explainable artificial intelligence: a survey.
EUMAS/AT 2015, 267–278. The Knowledge Engineering Review 36:e5.
Lakkaraju, H.; Slack, D.; Chen, Y.; Tan, C.; and Singh, S. Wachter, S.; Mittelstadt, B. D.; and Russell, C. 2017. Coun-
2022. Rethinking explainability as a dialogue: A practi- terfactual explanations without opening the black box: Au-
tioner’s perspective. CoRR abs/2202.01875. tomated decisions and the GDPR. CoRR abs/1711.00399.
Lertvittayakumjorn, P.; Specia, L.; and Toni, F. 2020. Wu, X.; Xiao, L.; Sun, Y.; Zhang, J.; Ma, T.; and He, L.
FIND: human-in-the-loop debugging deep text classifiers. In 2022. A survey of human-in-the-loop for machine learning.
EMNLP 2020, 332–348. Future Gener. Comput. Syst. 135:364–381.
Lundberg, S. M., and Lee, S. 2017. A unified approach to
interpreting model predictions. In NIPS 2017, 4765–4774.
Miller, T. 2019. Explanation in artificial intelligence: In-
sights from the social sciences. Artif. Intell. 267:1–38.
Nickerson, R. S. 1998. Confirmation bias: A ubiquitous
phenomenon in many guises. Review of General Psychology
2:175 – 220.
Panisson, A. R.; McBurney, P.; and Bordini, R. H. 2021. A
computational model of argumentation schemes for multi-
agent systems. Argument Comput. 12(3):357–395.
Paulino-Passos, G., and Toni, F. 2022. On interactive ex-
planations as non-monotonic reasoning. In XAI 2022 co-
located with IJCAI 2022.
Pisano, G.; Calegari, R.; Prakken, H.; and Sartor, G. 2022.
Arguing about the existence of conflicts. In COMMA 2022,
284–295.
Potyka, N. 2018. Continuous dynamical systems for
weighted bipolar argumentation. In KR 2018, 148–157.
Potyka, N. 2021. Interpreting neural networks as quantita-
tive argumentation frameworks. In AAAI 2021, 6463–6470.
Rago, A.; Baroni, P.; and Toni, F. 2022. Explaining causal
models with argumentation: the case of bi-variate reinforce-
ment. In KR 2022, 505–509.
Rago, A.; Toni, F.; Aurisicchio, M.; and Baroni, P. 2016.
Discontinuity-free decision support with quantitative argu-
mentation debates. In KR 2016, 63–73.
Rago, A.; Cocarascu, O.; Bechlivanidis, C.; and Toni, F.
2020. Argumentation as a framework for interactive expla-
nations for recommendations. In KR 2020, 805–815.
Rago, A.; Cocarascu, O.; and Toni, F. 2018. Argumentation-
based recommendations: Fantastic explanations and how to
find them. In IJCAI 2018, 1949–1955.
Raymond, A.; Gunes, H.; and Prorok, A. 2020. Culture-
based explainable human-agent deconfliction. In AAMAS
2020, 1107–1115.
Shih, A.; Choi, A.; and Darwiche, A. 2018. A symbolic ap-
proach to explaining bayesian network classifiers. In IJCAI
2018, 5103–5111.
Sokol, K., and Flach, P. A. 2020. Explainability fact sheets:
a framework for systematic assessment of explainable ap-
proaches. In FAT* 2020, 56–67.
Teso, S.; Alkan, Ö.; Stammer, W.; and Daly, E. 2023.
Leveraging explanations in interactive machine learning: An
overview. Frontiers Artif. Intell. 6.
 A Supplementary Material contributed some (a, b) such that either: (a, b) ∈ Sαt−1 and
(Interactive Explanations by Conflict b ∈ con(Bxt−1 ); or (a, b) ∈ At−1 t−1
α and b ∈ pro(Bx ) ∪ {e}. By
Definition 8, it can be seen that in both cases, a ∈ con(Bxt )
Resolution via Argumentative Exchanges) and so, by Definition 5, con(Bxn ) ≠ ∅. Thus, conflict resolu-
A.1 Proofs of Propositions in the Main Body tion is satisfied.
Proof for Proposition 1:
Proof for Proposition 5:
Proof. Let E = ⟨Bx0 , . . . , Bxn , AG0 , . . . , AGn , C⟩
be an AX
for e. For Properties 1 and 2, E trivially satisfies connected- Proof. Let (a, e) be such that C((a, e)) = (α, t). If α
ness and acyclicity given that Bxn is a BAF for e. For Prop- is arguing for e, then the shallow behaviour, by Defini-
erty 3, consider that ∀α ∈ AG′ , Σnα (Qnα ′ , e) depends only tion 11, requires that (a, e) ∈ Sαt−1 and ∄(b, e) ∈ Sαt−1 ∖
on σαn (Qnα ′ , e) by Definition 4. By Definition 5, we can see (Sxt−1 ∪ {(a, e)}) such that σα (Qt−1 t−1
α , b) > σα (Qα , a).
′ ′ Meanwhile, the greedy behaviour, by Point 1 of Defini-
that Qnα ′ ⊐ Bxn ′ such that Xαn ′ = Xα0 ∪Xxn ′ , Anα ′ = A0α ∪Anx ′
′ tion 12, requires that (a, e) ∈ Sαt−1 , and by Point 2, re-
and Sαn ′ = Sα0 ∪ Sxn ′ , regardless of C ′ , so it can be seen that quires that ∄(b, e) ∈ Sαt−1 ∖ (Sxt−1 ∪ {(a, e)}) such that
n n′
σα (Qα , e) = σαn (Qnα , e) and Σnα (Qnα ′ , e) = Σnα (Qnα , e), σα (Qt−1 t−1
α , b) > σα (Qα , a). (Point 3 is not relevant here
thus E satisfies contributor irrelevance. since paths((c, e)) = {(c, e)} ∀c ∈ Xαt−1 at t > 0.) The
Proof for Proposition 2: greedy behaviour is thus aligned with the shallow behaviour
when α is arguing for e. Similarly for when α is arguing
Proof. Let Case 1 for e be that σα (Qtα , e) > σα (Qt−1 α , e) against e, the shallow behaviour, by Definition 11, requires
and Case 2 for e be that σα (Qtα , e) < σα (Qt−1 α , e). It can that (a, e) ∈ At−1 t−1 t−1
α and ∄(b, e) ∈ Aα ∖(Ax ∪{(a, e)}) such
be seen from Definition 5 that ταt (a) = ταt−1 (a) ∀a ∈ Xαt for t−1 t−1
that σα (Qα , b) > σα (Qα , a). Meanwhile, the greedy be-
t > 0. Then, by the definition of DF-QuAD (see §3), Case 1 haviour, by Point 1 of Definition 12, requires that (a, e) ∈
must be due to: 1a.) an added supporter of e, i.e. Sαt (e) ⊃ At−1 t−1
α , and by Point 2, requires that ∄(b, e) ∈ Aα ∖ (Ax ∪
t−1

Sαt−1 (e); 1b.) a strengthened supporter of e, i.e. ∃a ∈ Sαt (e) {(a, e)}) such that σα (Qt−1
α , b) > σ α (Q t−1
α , a). (Point 3 is
such that σα (Qtα , a) > σα (Qα t−1
, a); or 1c.) a weakened not relevant here since paths((c, e)) = {(c, e)} ∀c ∈ Xαt−1
t
attacker of e, i.e. ∃b ∈ Aα (e) such that σα (Qtα , b) < at t > 0.) The greedy behaviour is thus aligned with the
σα (Qt−1
α , b). Similarly, and also by the definition of DF- shallow behaviour when α is arguing against e. Thus, the
QuAD, Case 2 must be due to: 2a.) an added attacker of proposition holds.
e, i.e. Atα (e) ⊃ Aα t−1
(e); 2b.) a weakened supporter of e,
i.e. ∃a ∈ Sα (e) such that σα (Qtα , a) < σα (Qt−1
t
α , a); or
Proof for Proposition 6:
2c.) a strengthened attacker of e, i.e. ∃b ∈ Atα (e) such that
σα (Qtα , b) > σα (Qα t−1
, b). In Cases 1a and 2a, it can be Proof. Since E is unresolved and by Definition 10, let α ∈
seen that pro(Bx ) ⊃ pro(Bxt−1 ) and con(Bxt ) ⊃ con(Bxt−1 ),
t AG be arguing for e and β ∈ AG be arguing against e. By
resp. In Cases 1b and 2b, we repeat Cases 1 and 2 for the Definition 15, we know that α contributed some (a, b) such
supporter a. In Cases 1c and 2c, we repeat Cases 1 and 2 that ϵ((a, b), Qtα ) > 0. Thus, by Proposition 2, pro(Bxt ) ⊃
for the attacker b, noting that the pro and con sets will be pro(Bxt−1 ) and, by Definition 5, pro(Bxn ) ≠ ∅. Similarly, by
inverted due to the extra attack in the path to e. Since Qtα Definition 15, we know that β contributed some (a, b) such
is a(n acyclic) QBAF for e, all paths through the multitree that ϵ((a, b), Qtβ ) < 0. Thus, by Proposition 2, con(Bxt ) ⊃
eventually reach leaves and Cases 1a and 2a apply, thus the con(Bxt−1 ) and, by Definition 5, con(Bxn ) ≠ ∅. Thus, con-
proposition holds. flict resolution is satisfied.
Proof for Proposition 3: A.2 Additional Propositions for §6
Proof. It can be seen, by Definition 4, that if > Σnα (e) We now give more detail on the effects of varying the biases
Σ0α (e), then σα (Qnα , e) > σα (Q0α , e). Then, by Defini- applied to learnt arguments by the machine. The first case
tion 5 and Proposition 2, pro(Bxn ) ≠ ∅. Analogously, it demonstrates the difficulty in achieving consensus between
can be seen, by Definition 4, that if Σnα (e) < Σ0α (e), then agents when one or all agents are incapable of learning.
σα (Qnα , e) < σα (Q0α , e). Then, by Definition 5 and Propo- Proposition 7. Let, ∀α ∈ AG, σα be the DF-QuAD seman-
sition 2, con(Bxn ) ≠ ∅. Thus, resolution representation is tics. Then, for any α ∈ AG and t > 1, if Xαt = Xαt−1 ∪ {a}
satisfied. where ταt (a) = 0 and Atα (a) ∪ Sαt (a) = ∅, then ∀b ∈
Proof for Proposition 4: Xαt ∖ {a}, σα (Qtα , b) = σα (Qt−1
α , b).

Proof. Since E is unresolved and by Definition 10, let α ∈ Proof. Since ταt (a) = 0 and Atα (a) ∪ Sαt (a) = ∅, by the
AG be arguing for e and β ∈ AG be arguing against e. definition of DF-QuAD (see §3), it must be the case that
By Definition 12, we know that α contributed some (a, b) σα (Qtα , a) = 0. Also by the definition of DF-QuAD, an
such that either: (a, b) ∈ Sαt−1 and b ∈ pro(Bxt−1 ) ∪ {e}; or argument with an evaluation of zero has no effect on the
(a, b) ∈ At−1 t−1
α and b ∈ con(Bx ). By Definition 8, it can be arguments it attacks or supports, and thus any argument in
seen that in both cases, a ∈ pro(Bxt ) and so, by Definition 5, α’s private QBAF, i.e. σα (Qtα , b) = σα (Qt−1 t
α , b) ∀b ∈ Xα ∖
pro(Bxn ) ≠ ∅. Similarly, by Definition 12, we know that β {a}, thus the proposition holds.
 We conjecture (but again leave to future work) that this Definition 16. Let E be a set of AXs for the same explanan-
behaviour is not limited to agents which evaluate arguments dum e between agents AG = {µ, η}, and let R ⊆ E be the
with DF-QuAD but also any semantics which ignore argu- set of all resolved exchanges in E. Then
ments with the minimum strength (e.g. as is described in ∣R∣
(Baroni, Rago, and Toni 2019) and by the property of neu- • the resolution rate (RR) of E is defined as RR(E) = ∣E∣
;
trality in (Amgoud and Ben-Naim 2017)). • the contribution rate (CR) of E is defined
The second case, meanwhile, guarantees that any learnt ∑E∈R E#
as CR(E) = ∣R∣
, where for E =
argument will have an effect all other downstream argu-
⟨Bx0 , . . . , Bxn , AG0 , . . . , AG , C⟩, E# = ∣Anx ∪ Sxn ∣;
n
ments’ (including the explanandum’s) strengths so long as
no arguments in the private QBAF are assigned the minim- • the persuasion rate (PR) of α ∈ AG over E is defined as
∣{E∈R∣∀β∈AG,Σβ (Qnβ ,e)=Σα (Qα ,e)}∣
0
ium or maximum biases. P R(α, E) = .
∣R∣
Proposition 8. Let, ∀α ∈ AG, σα be the DF-QuAD seman- Before giving the final measure we let the following
tics. Then, for any α ∈ AG and t > 1 where 0 < ταt (a) < 1 indicate the set of arguments which would have had the
∀a ∈ Xαt , if Xαt = Xαt−1 ∪ {b}, then ∀c ∈ Xαt such that biggest effect on the explanandum in the other agent’s pri-
∣paths(b, c)∣ = 1, σα (Qtα , c) ≠ σα (Qt−1
α , c). vate QBAF:
Proof. It can be seen from the definition of the DF-QuAD if α is arguing for e, then: Cmax (α, t) =
semantics (see §3) that if ∄a ∈ Xαt such that ταt (a) = 0 ′
or ταt (a) = 1, then ∄a ∈ Xαt such that σα (Qtα , a) = 0 or {(a,b)∈argmax(a′,b′ )∈(Atα∪Sαt )∖(Atx∪Sxt ) σβ (Qt+1 t+1 t
β , e)∣Xβ =Xβ ∪{a }}
σα (Qtα , a) = 1. Then, also by the definition of DF-QuAD, it if α is arguing against e, then: Cmax (α, t) =
must be the case that if Xαt = Xαt−1 ∪ {b}, then ∀c ∈ Xαt such {(a,b)∈argmin(a′,b′ )∈(Atα∪Sαt )∖(Atx∪Sxt ) σβ (Qt+1 t+1 t ′
β , e)∣Xβ =Xβ ∪{a }}
that (b, c) ∈ Atα ∪ Sαt , σα (Qtα , c) ≠ σα (Qt−1
α , c). This same
logic follows ∀d ∈ Xαt such that ∣paths(b, d)∣ = 1, and thus We can then define the contribution accuracy.
the proposition holds. Definition 17. Let E be a set of AXs for the same explanan-
Finally, the third case demonstrates the potential of in- dum e, each between two agents AG = {µ, η}. Then, the
corporating credulity in machines with guarantees of rejec- contribution accuracy (CA) of α ∈ AG over E is defined as
tion/weakening or acceptance/strengthening of arguments CA(α, E) = ∑E∈E acc(α,E)
∣E∣
, where for E = E, acc(α, E) = 0
which are attacked or supported, resp., by learnt arguments. n n
if {(a, b) ∈ Ax ∪ Sx ∣C((a, b)) = (α, t)} = ∅, otherwise:
Proposition 9. Let, ∀α ∈ AG, σα be the DF-QuAD seman- ∣{(a, b) ∈ Cmax (α, t)∣C((a, b)) = (α, t)}∣
tics. Then, for any α ∈ AG and t > 1, if Xαt = Xαt−1 ∪ {a} acc(α, E) =
where ταt (a) = 1 and Atα (a) ∪ Sαt (a) = ∅, then for any ∣{(a, b) ∈ Anx ∪ Sxn ∣C((a, b)) = (α, t)}∣
b ∈ Xαt such that a ∈ Atα (b) ∪ Sαt (b): .

• if a ∈ Atα (b):
– if Sαt (b) = ∅ then σα (Qtα , b) = 0;
– if {c ∈ Sαt (b)∣σα (Qtα , c) = 1} = ∅ then σα (Qtα , b) <
ταt (b);
• if a ∈ Sαt (b):
– if Atα (b) = ∅ then σα (Qtα , b) = 1;
– if {c ∈ Atα (b)∣σα (Qtα , c) = 1} = ∅ then σα (Qtα , b) >
ταt (b).

Proof. Since ταt (a) = 1 and Atα (a)∪Sαt (a) = ∅, by the Def-
inition of DF-QuAD (see §3), we can see that σα (Qtα , a) =
1. Let a ∈ Atα (b). If Sαt (b) = ∅ then, by the defini-
tion of DF-QuAD, σα (Qtα , b) = 0. Also by the definition
of DF-QuAD, if {c ∈ Sαt (b)∣σα (Qtα , c) = 1} = ∅, then
σα (Qtα , b) < ταt (b). Similarly, we let a ∈ Sαt (b). If Atα (b) =
∅ then, by the definition of DF-QuAD, σα (Qtα , b) = 1. Also
by the definition of DF-QuAD, if {c ∈ Atα (b)∣σα (Qtα , c) =
1} = ∅, then σα (Qtα , b) > ταt (b). Thus, the proposition
holds.

A.3 Evaluation Measures for §7

The following gives the formal definitions for the first three
of the evaluation measures used in the simulations. In all
measures, when the denominator is zero, the measure is also.