=Paper=
{{Paper
|id=Vol-2090/AIC17_paper6
|storemode=property
|title=Dialogical Scaffolding for Human and Artificial Agent Reasoning
|pdfUrl=https://ceur-ws.org/Vol-2090/paper6.pdf
|volume=Vol-2090
|authors=Sanjay Modgil
|dblpUrl=https://dblp.org/rec/conf/aic/Modgil17
}}
==Dialogical Scaffolding for Human and Artificial Agent Reasoning==
<pdf width="1500px">https://ceur-ws.org/Vol-2090/paper6.pdf</pdf>
<pre>
     Dialogical Scaffolding for Human and Artificial Agent
                           Reasoning

                           Sanjay Modgil (sanjay.modgil@kcl.ac.uk)

                         Department of Informatics, King’s College London



          Abstract. This paper proposes use of computational models of argumentation
          based dialogue for enhancing the quality and scope (‘scaffolding’) of both hu-
          man and artificial agent reasoning. In support of this proposal I draw on work
          in cognitive psychology that justifies such a role for human reasoning. I also re-
          fer to recent concerns about the potential dangers of artificial intelligence (AI),
          and the consequent need to ensure that AI actions are aligned with human val-
          ues. I will advocate that argumentation based models of dialogue can contribute
          to value alignment by enabling joint human and AI reasoning that may indeed
          be better purposed to resolve challenging ethical issues. This paper also reviews
          research in formals models of argumentation based reasoning and dialogue that
          will underpin applications for scaffolding human and artificial agent reasoning.


1      Introduction
This position paper argues that computational models of argumentation based dialogue
can play a key role in enhancing the quality and scope (henceforth referred to as ‘scaf-
folding’1 ) of both human and artificial agent reasoning. In developing the argument I
will draw on ground-breaking work in cognitive psychology – Sperber and Mercier’s
‘argumentative theory of reasoning’ [25] – to support the scaffolding role of argumen-
tation based dialogue for human reasoning. I will also refer to work by N. Bostrom [7]
(and others), who argue the need for aligning the values of artificial intelligence (AI)
and humans, so as to avert the potential threats that AI poses to humans. I will propose
that argumentation based models of dialogue can contribute to solving the so called
‘value alignment problem’, through enabling joint human and AI reasoning that may
indeed be better purposed to resolve challenging moral and ethical issues, as compared
with such deliberations being exclusively within the purview of humans or AI.
    The remainder of this paper is structured as follows. In Section 2 I review work
on provision of argumentative characterisations of non-monotonic inference over given
static belief bases. I then describe how these characterisations can be generalised to dia-
logical models in which interlocutors effectively reason non-monotonically over sets of
beliefs that are incrementally defined by the contents of locutions communicated during
the course of such dialogues. Section 3 then reviews Sperber and Mercier’s theory of
the evolutionary impetus for human acquisition of explicit (system 2) reasoning capaci-
ties, and the theory’s empirically supported implication that reasoning delivers superior
 1
     I here use the connoting term ‘scaffolding’ in view of its pedagogical use to describe instruc-
     tional techniques for inculcating interpretative and reasoning skills.
outcomes when human reasoners engage in dialogue. This implication in turn suggests
benefits accruing from deployment of computational models of argumentation based
dialogue for scaffolding human reasoning. I then propose deployment of such models
in education, deliberative democracy, and, more speculatively, the puncturing of belief
bubbles erected by the filtering algorithms of social media. Section 4 then reviews argu-
ments to the effect that future AI systems may pose serious threats to humankind, due
to their single minded pursuit of operators’ goals. This has led researchers to focus on
the problem of how to ensure that the reasoning of AI systems account for human val-
ues. I argue that dialogical models will contribute to solving this problem, by enabling
joint human-AI reasoning, so that human values may inform AI reasoning tasks that
have an ethical dimension. In Section 5 I review current work that can contribute to the
development of dialogical models for the applications envisaged in Sections 3 and 4,
and point to future research challenges. Finally, Section 6 concludes the paper.


2   From Non-monotonic Inference to Distributing Non-monotonic
    Reasoning through Dialogue

AI research in the 80s and early 90s saw a proliferation of non-monotonic logics tackle
classical logic’s failure to formalise our common-sense ability to reason in the pres-
ence of incomplete and uncertain information. In the classical paradigm, the inferences
from a set of formulae grows monotonically, as the set of formulae grow. However in
practice, conclusions that we previously obtain may be withdrawn because new infor-
mation conflicts with what we concluded previously or with the assumptions made in
drawing previous conclusions. Essentially then, a key concern of non-monotonic rea-
soning is how to arbitrate amongst conflicting information; a concern that is central
to the argumentative enterprise. It is this insight that is substantiated by argumentative
characterisations of non-monotonic inference. Most notably, in Dung’s seminal theory
of argumentation [15] and subsequent developments of the theory, one constructs the
arguments A from a given set of formulae ∆ (essentially each argument being a self-
contained proof of a conclusion derived from the supporting formulae). Arguments are
then related to each other in an argument framework (AF ) hA, →i where the binary
attack relation →⊆ A × A denotes that one argument is a counter-argument to (at-
tacks) another; for example when the conclusion, or claim, of one argument negates
a formula in the support of the attacked argument. In this way the formulae ∆ are
said to ‘instantiate’ the AF , as henceforth indicated by AF∆ . Of particular relevance
here, is developments of Dung’s theory to account for preferences over arguments [1, 5,
32]. For example, preferences may be based on the relative reliability of the sources of
the arguments, the epistemic certainty attached to the arguments’ constituent formulae,
principles of precedence (such as when rules in legal arguments encoding more recent
legislation are given higher priority), or orderings of values associated with the decision
options supported by arguments in practical reasoning. Preferences can then be used to
distinguish those attacks that can be deployed dialectically; that is, even though X’s
claim negates a formula in the support of Y , we have that (X, Y ) ∈→ only if X ⊀ Y
(Y is not strictly preferred to X).
    Conflict free sets (i.e. sets that contain no attacking arguments) of acceptable ar-
guments (extensions) of an AF hA, →i are then identified under different ‘semantics’.
The fundamental principle of ‘defense’ licenses membership of an argument X in any
such extension E ⊆ A: X ∈ E iff (Y, X) ∈→ implies ∃Z ∈ E, (Z, Y ) ∈→ (E is
said to defend X). An admissible extension E is one that defends all its contained argu-
ments. E is a complete extension if all arguments defended by E are in E. Then E is a
preferred, respectively the grounded, extension, if E is a maximal (under set inclusion),
respectively the minimal (under set inclusion) complete extension. E is stable if all ar-
guments outside of E are attacked by some argument in E. The claims of sceptically or
credulously justified arguments (those arguments that are in all extensions, or, respec-
tively, at least one extension) identify a semantics-parameterised family of inference
relations over ∆:

     ∆ |∼(a,s) α iff α is the claim of an a ∈ {sceptically, credulously} justified
                                                                                               (1)
          argument under semantics s ∈ {grounded , preferred , stable} in AF∆
    Argumentation thus provides for the definition of novel non-monotonic inference
relations. Moreover, Dung and others [2, 15, 32, 50] have shown that for various estab-
lished non-monotonic logics L 2 and their associated inference relations |∼L , that:
                         f or some a, s : ∆ |∼L α iff ∆ |∼(a,s) α                              (2)
    Given an AF hA, →i, argument game proof theories (e.g., [10, 30, 47]) establish
whether a given argument X ∈ A is justified. The essential idea is that a proponent wins
a game iff she successfully counter-attacks (defends) against all attacking arguments
moved by an opponent, where all attacks moved are licensed by reference to those in
the given AF . Players can backtrack to attack previous moves of their interlocutors,
so defining a tree of moves, with X as the root node, and Y a child node of Z iff
(Y, X) ∈→. A game is won in respect of showing that X is justified, iff X is justified
in the sense that it belongs to an extension of the framework under some semantics,
with rules on the allowable moves in the game varying according to the semantics3 .
    Argumentation based dialogues in which agents communicate to persuade one an-
other of the truth of a proposition, or decide amongst alternative action options (e.g.,
[17, 28, 36, 45]), can be seen as generalising the above argument games in two impor-
tant respects. Firstly, consider proponent and opponent agents attacking each others’
arguments, as in the above described games, where these attacks are not licensed by
reference to a given AF ; rather, the arguments are constructed from the agents’ private
belief bases, and the contents of these arguments incrementally define a public com-
mitment store Bp . At any point in the dialogue, an agent can then construct and move
arguments constructed from their own belief bases and the contents of Bp thus far de-
fined. An agent can at any point in the dialogue be said to have successfully established
the ‘topic’ α (a belief or decision option) iff α is the clam of a justified argument (under
some semantics implemented by the rules licensing allowable moves) in AFBp [17, 36,
 2
   Including Logic Programming, Reiter’s Default Logic, Pollock’s Inductive Defeasible Logic,
   Brewka’s Preferred Subtheories and Brewka’s Prioritised Default Logic.
 3
   E.g., in [30], variations in rules licensing allowable (legal) moves, yield games for membership
   of extensions under grounded, preferred and stable semantics.
28]. Dialogues also generalise games by allowing for agents to submit locutions that
not only consist of arguments, but locutions of other types (in the tradition of agent
communication languages that build on speech act theory [41]). For example, an agent
may simply make an individual claim rather than move an argument, or question why
a claim or premise of a moved argument is the case, or retract the contents of previ-
ous locutions, or concede that an interlocutor’s assertion is the case. Thus locutions
more typical of real world dialogues are defined, and dialogue protocols specify when
locutions are legal replies to other locutions. In such dialogues, only the contents of
assertional locutions (i.e., claims and arguments) define the contents of Bp .
    Now, let a dialogue D be defined by a sequence of moves (locutions) m1 , . . . , mn ,
where each mj (j 6= 1) replies to a single move mi<j . Then D can be represented as a
dialogue tree DT with root m1 , and mj a child of mi<j iff mj replies to mi . Let Bpn be
defined by the contents of assertional locutions in D = m1 , . . . , mn . Then:
                             D = m1 , . . . , mn |∼ α iff Bpn |∼(a,s) α                          (3)
     The above dialogical inference relation |∼D is defined in terms of the claims of jus-
tified arguments in AFBpn . However, one can also directly define dialogical inferences
by reference to the dialectical status of locutions in a dialogue tree DT , exploiting the
fact that some locutions can be said to attack the locutions they reply to [17, 28, 36].
Clearly, if an argument X is moved (in mj ) as an attack on Y (in mi<j ), then mj is an
‘attack reply’ to mi<j . Moreover if some premise or claim α in mi<j is replied to by
a locution mj questioning why α holds, then mj is an attack reply to mi<j , since the
burden of proof is on the agent (ag) moving mi<j to provide some argument justifying
α. If ag replies with such an argument in a reply mk to mi , then in fulfilling the burden
of proof, mk attack replies mi . Other locutions such as concede or retract do not affect
the dialectical status of their targets. The dialectical status of any locution m that claims
α, or moves an argument claiming α, is then said to be winning if all attack replies to
m are losing, else m is losing. The dialectical status is computed by initially assigning
winning status to all the leaves of DT and then propagating up to the root node. It can
then be shown that if interlocutors ‘play logically perfectly’ (i.e, any argument that can
be constructed from the contents of Bpn , and that can be legally moved according to the
protocol rules, is moved), then:

           ∃mi such that claim(mi )4 = α and mi is winning iff Bpn |∼(a,s) α                     (4)

    Hence under the assumption of logically perfect play, Equation 4 states correspon-
dence results for dialogical formalisations of distributed non-monotonic reasoning. The
proof theoretic realisation of D |∼ α consists in the evaluation of the dialectical status
of a locution claiming α as winning. This dialogical proof theory is shown to be sound
and complete w.r.t. the ‘public’ semantics whereby there exists a justified argument
claiming α in the AFBp instantiated by the contents of the locutions exchanged in D 5 .
 4
     Where mi is a locution asserting α or an argument that claims α.
 5
     That the semantics is ’public’, is in recognition of the fact that no reference is made to the
     contents of the agents’s belief bases. It should also be acknowledged that the work of Prakken
     [36] was instrumental in initiating this still somewhat nascent line of research into dialogical
     formalisations of non-monotonic inference.
    The above provides an overview of how argumentative characterisations of non-
monotonic inference can be generalised to dialogical formalisations of non-monotonic
reasoning in which agents communicate in order to engage in joint epistemic and prac-
tical reasoning. In the following sections we propose applications of such models for
scaffolding human and AI reasoning. We also briefly review current research aiming
at theoretical underpinnings for such applications, as well as suggested directions for
longer term future research, if such applications are to be realised.


3     The Role of Dialogue in Scaffolding Human Reasoning

We begin by reviewing recent influential research in cognitive psychology that gives
credence to the claim that deployment of computational models of argumentation based
dialogue can enhance the quality and scope of human epistemic and practical reasoning.


3.1   The argumentative theory of reasoning

Sperber and Mercier’s recent 2017 book: The Enigma of Reason: A New Theory of Hu-
man Understanding, summarises a programme of research building on their 2011 pa-
per: ‘Why do humans reason? Arguments for an argumentative theory’ [25]. The theory
proposes that reasoning evolved to produce and evaluate arguments when communicat-
ing. It is this understanding of the evolutionary function of reasoning that underpins an
empirically validated explanatory framework for the wealth of psychological evidence
suggesting that reasoning often leads us astray; evidence that contradicts the ‘Cartesian
view of reasoning’ which maintains the view that reasoning typically leads us to more
reliable beliefs and better decisions. However, while their theory explains the wayward-
ness of the lone reasoner, it also explains why and how reasoning keeps us on the path to
better beliefs and decisions when we reason together, in groups and through dialogue.
It is with this account of the origins of reasoning in mind, that we can appreciate the
value of research into formal models of logic based argument and dialogue.
     In a nutshell, their theory argues that reasoning evolved to support communication.
To avoid being misled and possibly manipulated into acting against one’s self-interest,
it is advantageous for an addressee to exercise epistemic vigilance (especially when in
receipt of information from sources that do not warrant a high degree of trust and in-
formation that does not cohere with what she believes). Vigilance is exercised by eval-
uating reasons (i.e, arguments) for the received information, and looking for counter-
arguments, before accepting the received information. In turn, it is to the advantage of
the sender that he produce arguments supporting the information being communicated,
in order that it be accepted. Reasoning thus increases the quantity and epistemic quality
of the information humans are able to share, by allowing communicators to argue for
what they claim, and by allowing addressees to assess these arguments.
     This evolutionary function of reasoning implies that a lone reasoner is disposed to
seek reasons in support of his beliefs, and overlook reasons that argue to the contrary,
especially when such beliefs are contentious and the reasoner anticipates that they will
be challenged. This, for example, manifests in the classic confirmation bias, which the
argumentative theory suggests is a normal feature of reasoning. Moreover, individual
decision makers are disposed to harness reasoning to the extent that they anticipate com-
municating their decisions to others; hence the evidence from experimental psychology
showing that we favour decision options that can be easily justified and are less at risk
of being criticised, rather than because they satisfy some criterion of rationality.
    However, the argumentative theory also implies that reasoning serves us better when
performed in groups and in particular when reasoning jointly through dialogue, under
the assumption that interlocutors are motivated to have a common interest in the truth or
the right decision. In these contexts, the dispositions of speakers and receivers to respec-
tively seek arguments for claims, and evaluate and seek counter-arguments, yields better
outcomes. That this is so is supported by evidence reviewed by Sperber and Mercier.
These benefits of dialogical reasoning have important implications for research into
logical and computational models of argument and dialogue, since implementations of
such models may potentially be used to enhance the quality and scope of human rea-
soning, as outlined in the following section.


3.2   Applications for scaffolding human reasoning

The first area of application I propose for computational models of argumentation based
dialogue, is in education. Sperber and Mercier themselves note studies showing that
the teaching of critical reasoning skills has not yielded very good results. They argue
that we need institutional interventions that engender in people a readiness to engage
in argument with others (especially with those who disagree with them), and increase
people’s exposure to arguments. I suggest that this should begin in educational institu-
tions, through the use of applications that support student–student and student–teacher
dialogue and debate. In such applications, models of dialogue can provide dialectical
feedback to students as to who is currently winning a given discussion, the relevant lo-
cutions that need to be (attack) replied to in order to change the dialectical status of a
dialogue topic (as described in [36]), and the arguments that students may pose given
the contents of commitment stores. Moreover, with suitable advances in argument min-
ing [23] one can envisage such applications enriching the content and scope of such
discussions, by searching the web for arguments in support of claims. Indeed, the latter
technologies are actively being developed by researchers working on the next gener-
ation IBM Watson (see: Computers that can argue will be satnav for the moral maze
New Scientist, September 2016). The latter functionality also suggests computer agents
that play an active role as interlocutors, challenging human interlocutors, and sourcing
the web and other electronic data repositories for arguments. For example, consider an
envisaged E-learning tool – E-Clinic – for use by medical students. Studies show that
over 50% of junior doctors’ acquisition of clinical reasoning skills to decide amongst al-
ternative (i.e., conflicting) diagnoses and treatment options, occurs on ward rounds [11].
In these ward rounds teaching clinicians engage students in dialogue, challenging their
assumptions, suggesting alternative diagnoses and treatment options, and suggesting
hypothetical scenarios that may, for example, prompt the student to propose additional
interventions to ameliorate hazardous side effects of proposed treatments. However, de-
mands on teaching clinicians’ time present a significant barrier to such learning [11].
Furthermore, the range of medical scenarios is limited by the presenting patients, and
students on wards do not benefit from access to electronic data repositories. The envis-
aged E-Clinic will simulate clinicians on ward rounds engaging students in deliberating
over medical treatments, thus increasing students’ exposure to, and scope of coverage
of, a key methodology in medical education.
    A key challenge for the short to medium term implementation of the kinds of appli-
cations described above, is the inadequacy of current computational support for natural
language processing and understanding. However, in addition to the above mentioned
work on argument mining, the use of schemes and critical questions (SchCQ) devel-
oped by the informal logic community (most notably by D. Walton [49]), provides
for an enabling methodology to help address this challenge. Schemes provide generic
templates for construction of arguments that can be instantiated by computational and
natural languages, enabling computational and human agents to interact in comprehen-
sively reasoning about some issue. For example, variables in the argument for action
scheme (AS)[3] – ‘In circumstances S, doing action A will have effect E, achieving
goal G and so promoting desirable value V ’ – can be instantiated to define an argument
X for action A. The argument can then be challenged in a dialogue, by addressing AS’s
associated critical questions (CQs). For example, the CQ ‘Is S true ?’ can be posed as
a challenge, and responded to by instantiating a scheme to provide an argument X 0 for
why S is true. CQs can also be addressed by instantiating schemes that yield attack-
ing arguments. For example ‘does A have a side-effect that is undesirable ?’ can be
addressed by instantiating a scheme that argues against performing an action given its
undesirable effect. These arguments can then themselves be challenged by their CQs.
One can thus envisage a computational dialogue manager using SchCQ to guide human
and computational agents in challenging, and constructing attacking arguments that are
then evaluated to yield dialectical feedback on the state of the dialogue (as described
above). Indeed, such a dialogue manager has been implemented to support dialogue
amongst clinicians in their rational exploration of the arguments for and against as-
signment of transplant organs to potential recipients [45]. A key lesson learnt in this
work, is that the SchCQ developed by Walton and others are often too generic for prac-
tically guiding human authoring of arguments. Hence domain specific specialisations
of SchCQ are required; for example, SchCQ specialised (in [45]) for reasoning about
medical actions and in particular the safety of such actions.
    Another arena in which dialogical scaffolding can address the limitations and biases
of human reasoning so as to have significant societal impact, is in deliberative democ-
racy (as explicitly advocated by Mercier and Landemore in [24]). Indeed, current pro-
totype web applications (see http://cgi.csc.liv.ac.uk/˜parmenides/)
make use of the AS scheme to structure policy proposals, and then invite citizens to
agree or disagree with challenges to the assumptions in the policy proposal that are
posed by the associated critical questions [4]. However, one can go further, in support-
ing political deliberations amongst humans who are additionally challenged by compu-
tational agents with their vastly superior access to supporting evidence and arguments.
    The above use contexts for computational dialogue assume an intent on the part of
interlocutors to get to the truth of the matter or make better decisions. A more chal-
lenging speculative use of argumentation technologies, is in the dismantling of the echo
chambers erected by social media. These belief bubbles arise due to filtering algorithms
feeding news and opinions that entrench people’s ideological positions, and, as in clas-
sic examples of groupthink, even make those positions more extreme. Such algorithms
are digital incarnations of our dispositions to seek arguments that confirm what we be-
lieve, but now unbounded by the limitations of a human reasoner reasoning alone. One
might envisage argument mining technologies trawling the web to curate and present
arguments, opinions and news that challenge the beliefs of bubble dwellers. But is this
what a ‘typical’ user would really want ? What might also be required is an attitudinal
shift in the way humans engage with information, so that exposure to arguments and
counter-arguments is the default; a shift that may in part be brought about by the kinds
of educational interventions alluded to at the beginning of this section.


4    The Role of Dialogue in Scaffolding Artificial Agent Reasoning

The previous section considered the potential for enhancing the quality and scope of hu-
man reasoning via dialogical models harness the signature strengths of AI interlocutors.
In this section I want to argue for the other ‘direction of travel’, suggesting that such
models may facilitate the influence of human concerns and values on AI reasoning, and
thus contribute to solving what has been termed the “value loading problem”.
     Recent years have witnessed a dramatic increase in reports of AI successes, which
have in large part been due to advances in machine learning. These successes have been
accompanied by leading researchers warning of the possible dangers of AI [40]. It is
argued that the evolution of artificial general intelligence and accompanying benefits,
will license the development of, and trust in, machines that are increasingly more pow-
erful (with cognitive powers far outstripping those of humans), autonomous and capable
of acting in diverse and open environments 6 . However, such machines may formally
achieve their operator’s goals in ways that not only diverge from their operators inten-
tions, but may actually be contrary to the interests and values of their operators [7, 40,
44]. This concern recalls arguments to the effect that adhering to any rule based ethical
system may result in unintended, harmful consequences (as exemplified by Asimov’s
fictional accounts of robots adhering to laws intended to guarantee ethical behaviour,
but that in so doing cause unforseen harm [12, 13]). However, this problem has acquired
renewed urgency given that it is a feature of learning systems that they find unforseen
ways of achieving goals, and that achievement of any operator’s goal will incentivise
‘instrumental goals’ that thwart corrective measures to prevent harm [7, 42–44]7 .
 6
   See [7] for a rigorously argued chartering of possible trajectories to artificial general intelli-
   gence (i.e,. intelligence exhibited across a wide range of tasks, as opposed to narrowly defined
   tasks) and its widespread deployment, and the subsequent evolution of superintelligent ma-
   chines, given recursive self-improvement and massive data access and processing power.
 7
   Instrumental goals include self-preservation (preserving ones self increases the probability of
   achieving a current goal), goal maintanance (maintaining any current goal into the future,
   increases the probability of achieving a current goal), increasing intelligence, technological
   perfection and resource acquisition. In [7], Bostrom gives examples of how seemingly benign
   goals may lead to unforseen harmful courses of actions which are maintained by virtue of in-
   strumental convergence. I suggest that a conceivable scenario is the utilitarian tasking of future
   AIs to maximise human happiness. Given that future AIs will share our current understand-
    The need to ensure AI acts in accordance with human values has prompted consider-
able intellectual investment into what in a machine learning context has been termed the
‘value loading (alignment) problem’ [7, 44], more broadly understood as the problem
of how to design ‘ethical agents’. Whether the envisaged agents’ ethical behaviour is
implemented through use of machine learning techniques via the maximising of utility
functions encoding human preferences, and/or through the use of ‘top down’ symbolic
logic based reasoning adhering to explicitly encoded ethical theories [34, 48], two key
research problems need to be addressed [7, 39, 44]. Firstly, there is the problem of how
to specify objective utility functions (deontic axiomatisations) that are perfectly aligned
with human values and applicable in changing environments and to novel situations (in
particular ethical dilemmas with no precedent that most saliently expose the Humean
is/ought gap). Secondly, there is the above described problem of unintended behaviours
misaligned with human values. Run time learning of values has been proposed to ad-
dress these problems [42], for example through the use of inverse reinforcement learn-
ing [35], in which AI systems are incentivised to observe and query humans [39]; the
assumption being that actions reveal preferences and hence values, and that humans are
sufficiently informed and have the requisite capacity to definitively arbitrate on matters
of ethical importance. However, humans clearly do not always behave ethically, and
moreover are often uncertain about how to resolve ethical issues; in particular those
arising from the use of novel technologies whose use lack precedent (e.g., see foot-
notes 7 and 8). I argue that we therefore require that AI systems and humans engage
in comprehensive, rational exchange of arguments purposed to decide ethical issues, as
do humans when faced with difficult ethical choices8 . Indeed, such dialogues, by har-
nessing AI’s vastly superior access to information and the capacity to look further into
the future, may be better purposed to decide ethical issues, as compared with humans
reasoning without the support of AI. In turn, human input regarding values will inform
such deliberations. In effect, one can view such dialogues as contributing to the ongoing
inculturation of AI, mirroring the gradual ongoing acquisition of norms and values that
we witness in human societies.
    To meet requirements for what I refer to as ‘value deliberation’ dialogues (cf. ‘value
learning’) will require building on research into models of argument and dialogue re-
viewed in Section 2. In the following section I review recent work relevant to the fur-
ther development of logic based models of argument and dialogue that will facilitate
the scaffolding of human and artificial agent reasoning.


   ing of happiness as subjective well being mediated by brain biochemistry, and that changes
   in external conditions have been shown to have relatively minimal impact on happiness, one
   might envisage AI development of virtual reality technologies, and the leveraging of existing
   societal trends towards online experience, dating, friendship etc, in order to manipulate hu-
   mans into living ever more increasingly virtual lives (for our own benefit !). Such scenarios
   raise morally challenging questions upon which human values bear; hence the need for value
   loading/alignment in order to ensure AI actions in compliance with human concerns.
 8
   For example, consider the UK government’s appointment of the moral philosopher Baroness
   Mary Warnock to chair the Committee of Inquiry into Human Fertilisation and Embryology.
   This gave rise to the Human Fertilisation and Embryology Act 1990, which governs human
   fertility treatment and experimentation using human embryos.
5      Towards Formal Models for Applications of Dialogical Models

A key development in research on argumentation based characterisations of non-monotonic
inference has been the formulation of rationality postulates identified as desiderata for
satisfaction by instantiations of argument frameworks [8, 9]. For example, the conclu-
sions of arguments in any extension should be mutually consistent. This in turn has led
to frameworks specifying guidelines for how to formulate instantiations of Dung AF s,
such that the postulates are satisfied. The ASPIC framework [8] considers guidelines
for instantiations that do not make use of preferences. The ASPIC+ framework [37, 32]
then generalised ASPIC so as to provide guidelines that account for preferences and a
broader range of instantiations. A notable feature of ASPIC+ is its provision of argu-
mentative characterisations of existing non-monotonic logics9 , as well as formalising
argumentation based on arguments and attacks obtained through use of the schemes and
critical questions methodology (see [33, 37]) described in Section 3. I therefore contend
that the ASPIC+ framework provides a suitable basis for distributing non-monotonic
reasoning through dialogue, by providing a basis for provably rational dialogues that
accommodate artificial and human agents.
     However, ASPIC+ and other approaches augmenting Dung’s theory to accommo-
date the use of preferences, assume that preferences, or value orderings used to derive
preferences, are fixed, consistent, and exogenous to the domain of reasoning. In prac-
tice this is clearly not the case: reasoning about preferences and values, and arguing
to resolve disagreements amongst preferences and values is the norm, and needs to be
accommodated in the kinds of applications envisaged in this paper (especially for the
value deliberation dialogues proposed in Section 4). Hence, Dung’s theory has been
extended to accommodate argumentation based reasoning about preferences and values
[26, 29]. The essential idea is that an argument X that justifies a preference for argument
A over argument B (B ≺ A), attacks the attack from B to A. Of course, an argument
Y claiming A ≺ B, attacks and is attacked by X, and one may then need to argue
about which preference is preferred. Preliminary work has then extended ASPIC+ so
as to provide guidelines for instantiations of these Extended AF s (EAF s), and such
that the outcomes yielded by evaluation of the extensions of EAF s are rational [31].
Moreover, [28] has recently adopted the methodology described in Section 2, to define
dialogical protocols that enable agents to reason about and possibly disagree about the
preferences and value orderings that are expressed over other arguments. Evaluation of
the outcomes of these dialogues is defined on the basis of ASPIC+ instantiations of the
EAF s instantiated by the contents of the locutions exchanged. [28] also defines the
dialectical status of moves in such dialogues, but it remains to show correspondences
of the type described in Equation 4 (which are stated as conjectures in [28]).
     Thus far I have reviewed the ASPIC+ framework that provides guidelines for di-
alectical characterisations of non-monotonic inference that are rational, and preliminary
work extending ASPIC+ so as to also accommodate arguments that express preferences
and reason about values, and the subsequent generalisation of this work to define for-
mal models of dialogues in which agents can jointly reason and engage in epistemic and
practical reasoning, while also resolving disagreements about the preferences and val-
 9
     Including Preferred Subtheories (see [32]), and Prioritised Default Logic (see [50])
ues that may bear on these kinds of reasoning. However, the ASPIC+ study of sufficient
conditions for satisfaction of rationality postulates (as well as other studies[8, 16, 19])
assume that agents have unbounded resources10 , and that the consistency of arguments’
premises is verified prior to submitting an argument. But real-world agents are resource
bounded, and the Socratic move of showing that an interlocutor contradicts himself by
committing to inconsistent premises (in one or a number of arguments) is a ubiquitous
feature of dialectical reasoning. We therefore need to address the challenge of develop-
ing rational resource bounded models of argumentation and dialogue that formalise real
world modes of dialectical reasoning. In this regard, [14] represents a first step toward
achieving this aim, in the context of classical logic argumentation [19]. The essential
insight in [14], is a refinement of the ontology of arguments so as to account for the
dialectical character of argumentative characterisations of non-monotonic inference. In
particular, an argument’s supporting formulae are distinguished according to whether
they are premises assumed true, or supposed true for the sake of argument, where the
latter may refer to the premises of an interlocutor.
     As well as the above mentioned Socratic move, other modes and features of real-
world dialectical reasoning need to be formally modelled in dialogues. For example
the use of enthymemes (arguments with some missing logical structure and/or internal
components) [6, 21] and relations other than binary attacks between arguments, such as
support relations and collective attacks from multiple arguments. Indeed, [27] argues
that these additional relation types are characteristic of human dialogue, and that they
are essentially abstractions of logical relationships between constituents of arguments.
The contents of these arguments when included in a commitment store, only then need
yield arguments related by binary attacks. For example, that X supports Y indicates that
X concludes some α that is a premise or intermediate conclusion in Y . Hence, when the
interlocutors’ assertional commitments instantiate an AF , it suffices that one construct
an argument Y 0 that is the argument Y ‘extended backwards’ with X, on what is now
an intermediate conclusion α in Y 0 . Moreover, [27] argues that in real world dialogues,
abstract relations amongst arguments typically encode declarative information that is
not explicitly communicated due to the use of enthymemes. For example, suppose Paul
argues that “Tony Blair is no longer a public figure (α), and the information about his
affair is not in the public interest (β), so the information should not be published (γ)”
(X). Trevor counter-argues with “but Blair is UN envoy for the Middle East” (Y ). The
latter is an example of an ‘indirect’ speech act [41], where it is inferred from the use of
“but”, that Y is an enthymeme that attacks X, and the missing structure in Y is a rule of
the form “if Blair is UN envoy for the Middle East then δ”, and δ negates either one of
the premises α or β, or the conclusion γ of X. The fact that there is some uncertainty as
to what δ denotes when instantiating the AF defined by the interlocutors commitments,
should then instigate dialectical feedback that promotes the rationality of the dialogue,
by prompting Paul to question Trevor as to what δ denotes.
     Other areas for future work include studying instantiations of argument frameworks
by deontic and temporal logics, so as to formalise non-monotonic normative and tem-
poral reasoning, and argumentative formalisations of hypothetical reasoning. Another
10
     Frameworks are assumed to be instantiated by all arguments defined by a set of formulae ∆,
     which may be infinite as in classical logic argumentation [19].
topic that has recently received attention is the modelling of interlocutors’ mental states
and the effects of these so called ‘opponent models’ on strategic choices made during
dialogues [20, 38]. Finally, the applications envisaged in this paper will require more
concerted interdisciplinary research efforts, in particular with areas such as argument
mining [23] and computational linguistics more generally, informal logic and philoso-
phy (e.g., integration of logic based argumentation with schemes and critical questions,
speech act theory, and the pragma-dialectic school of argumentation [46]), and integra-
tion with machine learning to address the symbol grounding problem [18].


6      Conclusions

The last two decades have witnessed an explosion of interest in formal argumentation
theory, with researches often locating the significance of the theory in its: 1) provision
of an integrative bridge between computational models of logic based reasoning and
human reasoning, and 2) provision of formal underpinnings for dialogue. This paper
has set out to further substantiate these claims. Firstly, while the work of Sperber and
Mercier has been cited to support the need for development of artificial cognitive sys-
tems whose reasoning processes are compatible with those of humans [22], this paper
has cited the argumentative theory of reasoning to support use of computational mod-
els of dialogue for enhancing the quality and scope of human reasoning. Moreover, I
have motivated the use of such models in supporting the reasoning of AI agents, so that
through joint reasoning with human interlocutors, AI decisions that have ethical im-
plications are aligned with human values. This paper also outlines how argumentative
characterisations of non-monotonic reasoning are generalised to dialogical models that
facilitate distributed reasoning, and current work11 that aims at further developing these
models for use in the kinds of applications envisaged in this paper.

References
 1. L. Amgoud and C. Cayrol. A reasoning model based on the production of acceptable argu-
    ments. Annals of Mathematics and Artificial Intelligence, 34(1-3):197–215, 2002.
 2. L. Amgoud and S. Vesic. Handling inconsistency with preference-based argumentation. In
    4th International Conference on Scalable Uncertainty Management, pages 56–69, 2010.
 3. K. Atkinson, T.J.M. Bench-capon, and P. Mcburney. Computational representation of prac-
    tical argument. Synthese, 152:2006, 2005.
 4. K. Atkinson, T.J.M. Bench-Capon, and P. McBurney. Parmenides: Facilitating deliberation
    in democracies. Artificial Intelligence and Law, 14:261–275, 2006.
 5. T. J. M. Bench-Capon. Persuasion in practical argument using value-based argumentation
    frameworks. Journal of Logic and Computation, 13(3):429–448, 2003.
 6. E. Black and A. Hunter. A relevance-theoretic framework for constructing and deconstruct-
    ing enthymemes. Journal of Logic and Computation, 22(1):55–78, 2012.
 7. M. Bostrom. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014.
11
     Note that the work reviewed is not exhaustive in its coverage of relevant research, but, given
     that this position paper sets out a personal vision of the potential role of argumentation based
     dialogue that has informed my research, I have focussed on research involving myself in col-
     laboration with others.
 8. M. Caminada and L. Amgoud. On the evaluation of argumentation formalisms. Artificial
    Intelligence, 171(5-6):286–310, 2007.
 9. M. Caminada, W. Carnielli, and P. Dunne. Semi-stable semantics. Logic and Computation,
    22(5):1207–1254, 2012.
10. C. Cayrol, S. Doutre, and J. Mengin. On decision problems related to the preferred semantics
    for argumentation frameworks. Journal of Logic and Computation, 13(3):377–403, 2003.
11. A. Claridge. What is the educational value of ward rounds? a learner and teacher perspective.
    Journal of Clinical Medicine, 11(6):558–62, 2011.
12. R. Clarke. Asimov’s laws of robotics: Implications for information technology - part 1.
    Computer, 26(12):53–61, 1993.
13. R. Clarke. Asimov’s laws of robotics: Implications for information technology - part 2.
    Computer, 27(1):57 – 66, 1994.
14. M. D’Agostino and S. Modgil. A rational account of classical logic argumentation for real-
    world agents. In European Conference on Artificial Intelligence, pages 141 – 149, 2016.
15. P. M. Dung. On the acceptability of arguments and its fundamental role in nonmonotonic rea-
    soning, logic programming and n-person games. Artificial Intelligence, 77:321–357, 1995.
16. P.M. Dung and P.M. Thang. Closure and consistency in logic-associated argumentation.
    Journal of Artificial Intelligence Research, 49(1):79–109, 2014.
17. X. Fan and F. Toni. A general framework for sound assumption-based argumentation dia-
    logues. Artificial Intelligence, 216:20 – 54, 2014.
18. M. Garnelo, K. Arulkumaran, and M. Shanahan. Towards deep symbolic reinforcement
    learning. CoRR, abs/1609.05518, 2016.
19. N. Gorogiannis and A. Hunter. Instantiating abstract argumentation with classical-logic ar-
    guments: postulates and properties. Artificial Intelligence, 175:1479–1497, 2011.
20. C. Hadjinikolis, Y. Siantos, S. Modgil, E. Black, and P. McBurney. Opponent modelling in
    persuasion dialogues. In Proc. 23rd International Joint Conference on Artificial Intelligence,
    pages 164–170, 2013.
21. S.A. Hosseini, S. Modgil, and O. Rodrigues. Enthymeme construction in dialogues using
    shared knowledge. In Proc. 5th Computational Models of Argument: COMMA 2014, pages
    325–332, 2014.
22. A. Kakas and M. Loizos. Cognitive systems: Argument and cognition. IEEE Intelligent
    Informatics Bulletin, 17(1):14–20, 2016.
23. M. Lippi and P. Torroni. Argumentation mining: State of the art and emerging trends. ACM
    Transactions on Internet Technology, 16(2):10:1–10:25, 2016.
24. H. Mercier and H. Landemore. Reasoning is for arguing: Understanding the successes and
    failures of deliberation. Political Psychology, 23(2):243?–258, 2012.
25. H. Mercier and D. sperber. Why do humans reason? arguments for an argumentative theory.
    Behavioral and Brain Sciences, 34(2):57–747, 2011.
26. S. Modgil. Reasoning about preferences in argumentation frameworks. Artificial Intelli-
    gence, 173(9-10):901–934, 2009.
27. S. Modgil. Revisiting abstract argumentation frameworks. In Second International Workshop
    on Theory and Applications of Formal Argumentation (TAFA 2013), pages 1–15, 2013.
28. S. Modgil. Towards a general framework for dialogues that accommodate reasoning about
    preferences. In Fourth International Workshop on Theory and Applications of Formal Argu-
    mentation, TAFA 2017 (to appear), 2017.
29. S. Modgil and T.J.M Bench-Capon. Metalevel argumentation. Journal of Logic and Com-
    putation, 21(6):959–1003, 2011.
30. S. Modgil and M. Caminada. Chapter 6 : Proof theories and algorithms for abstract argu-
    mentation frameworks. In I. Rahwan and G. Simari, editors, Argumentation in AI, pages
    105–129. Springer, 2009.
31. S. Modgil and H. Prakken. Reasoning about preferences in structured extended argumen-
    tation frameworks. In Proc. Computational Models of Argument: COMMA 2010, pages
    347–358, 2010.
32. S. Modgil and H. Prakken. A general account of argumentation with preferences. Artificial
    Intelligence, 195:361–397, 2013.
33. S. Modgil and H. Prakken. The aspic+ framework for structured argumentation: A tutorial.
    Argument and Computation, 5(1):31–62, 2014.
34. J. H Moor. The nature, importance, and difficulty of machine ethics. IEEE Intelligent Sys-
    tems, 21(4):18–21, 2006.
35. A. Y. Ng and S. Russell. Algorithms for inverse reinforcement learning. In Proc. 17th
    International Conference on Machine Learning, ICML ’00, pages 663–670, 2000.
36. H. Prakken. Coherence and flexibility in dialogue games for argumentation. Journal of Logic
    and Computation, 15:1009–1040, 2005.
37. H. Prakken. An abstract framework for argumentation with structured arguments. Argument
    and Computation, 1(2):93–124, 2010.
38. T. Rienstra, M. Thimm, and N. Oren. Opponent models with uncertainty for strategic argu-
    mentation. In Proc. 23rd International Joint Conference on Artificial Intelligence, (IJCAI
    ’13), pages 332–338, 2013.
39. S. Russell, D. Dewey, and M. Tegmark. Research priorities for robust and beneficial artificial
    intelligence. CoRR, abs/1602.03506, 2016.
40. S. Russell, T. Dietterich, E. Horvitz, B. Selman, F. Rossi, D. Hassabis, S. Legg, M. Su-
    leyman, D. George, and S. Phoenix. Research priorities for robust and beneficial artificial
    intelligence: An open letter. AI Magazine, 36(4):3–4, 2016.
41. J. R. Searle. Speech acts. In Cambridge University Press, Cambridge UK., 1962.
42. N. Soares. The value learning problem. In Ethics for Artificial Intelligence Workshop at 25th
    International Joint Conference on Artificial Intelligence (IJCAI-2016), 2016.
43. N. Soares, B. Fallenstein, S. Armstrong, and E. Yudkowsky. Corrigibility. In Artificial
    Intelligence and Ethics, Papers from the 2015 AAAI Workshop, 2015.
44. J. Taylor, E.Yudkowsky, P. LaVictoire, and A. Critch.                    Alignment for ad-
    vanced machine learning systems. https://intelligence.org/2016/07/27/
    alignment-machine-learning/, 2016.
45. P. Tolchinsky, S. Modgil, K. Atkinson, P. McBurney, and U. Cortes. Deliberation dialogues
    for reasoning about safety critical actions. Journal of Autonomous Agents and Multi-Agent
    Systems, 25:209–259, 2012.
46. F. H. van Eemeren, B. Garssen, E. C.W. Krabbe, A. Henkemans, S. Francisca, , B. Verhei,
    and J. H. M. Wagemans. The Pragma-Dialectical Theory of Argumentation, pages 517–613.
    Springer, Dordrecht, 2014.
47. G. A. W. Vreeswijk and H. Prakken. Credulous and sceptical argument games for preferred
    semantics. In Proc. 7th European Workshop on Logic for Artificial Intelligence, pages 239–
    253, 2000.
48. W. Wallach, C. Allen, and I. Smit. Machine morality: Bottom-up and top-down approaches
    for modelling human moral faculties. AI & Society - Special Issue: Ethics and artificial
    agents, 22(4):565–582, 2008.
49. D. N. Walton. Argument Schemes for Presumptive Reasoning. Lawrence Erlbaum Asso-
    ciates, 1996.
50. A.P. Young, S. Modgil, and O. Rodrigues. Prioritised default logic as rational argumenta-
    tion. In Proc. 15th International Conference on Autonomous Agents and Multiagent Systems
    (AAMAS’16), pages 626–634, 2016.

</pre>