1

Thirteen Challenges in Formal and Com-

putational Argumentation
Liuwen Yu, Leendert van der Torre, and Réka Markovich

abstract. In this chapter, we present thirteen challenges in formal

and computational argumentation. They are organized around Dung’s
attack-defense paradigm shift. First, we describe four challenges pertain-
ing to the diversity of argumentation. Then we discuss five challenges
regarding the attack-defense paradigm shift. Finally, we discuss four chal-
lenges for computational AI argumentation arising after the paradigm
shift. We illustrate these challenges using examples from machine ethics,
AI & Law, decision-making, linguistics, philosophy, and other disciplines
to illustrate the breadth of argumentation research. We end each chal-
lenge by presenting several open questions for further research.

1 Introduction
Argumentation means different things to different people. Even in the two
volumes of the Handbook of Formal Argumentation, one can find a range of
definitions, some focusing more on the formal aspects, others focusing more on
the computational aspects. We believe that the thirteen challenges discussed
in this chapter pertain to all these definitions or can be rephrased to adhere to
all these definitions. Nevertheless, to clarify some of the issues we discuss, we
present in Table 1 the definitions we will use in this chapter.
Moreover, as its title suggests, this chapter is particularly concerned with
formal and computational argumentation as discussed in the Handbook of For-
mal Argumentation, the proceedings of the Computational Models of Argument
(COMMA) conferences, the Argument and Computation journal, and the wider
artificial intelligence (AI) literature on argumentation. In particular, whereas
formal argumentation has developed as a branch of knowledge representation
and reasoning, an essential part of AI, it now intersects with numerous disci-
plines, including natural language processing (NLP), and multiagent systems.
Therefore, when we refer to argumentation without further clarification, we
mean formal and computational argumentation. When we specifically discuss
formal argumentation only or computational argumentation only, we will make
this explicit. Similarly, when referring to natural argumentation, we will do so
explicitly.
To structure our discussion of these challenges, we use the attack-defense
paradigm shift brought about by Dung [1995] as a pivotal point. In particular,
Table 2 distinguishes between three groups of challenges. The first group is
concerned with the background to this paradigm shift, the second group is
2 Liuwen Yu, Leendert van der Torre, and Réka Markovich

Term Definition
Refers to the way humans naturally reason and communicate
Natural
in everyday language, combining elements of linguistics, phi-
argumentation
losophy, and rhetoric.
Formal A process of representing, managing and (sometimes) resolv-
argumentation ing conflicts.
Algorithmic A step-by-step procedure or set of rules designed to perform
argumentation a specific task or solve a particular argumentation problem.
Refers to the study and implementation of argumentation pro-
Computational cesses using computational methods. Involves theoretical and
argumentation practical aspects of how argumentation can be modeled and
executed by computers.
A computational approach incorporating argumentation rea-
Argumentation
soning mechanisms with other technologies, e.g., NLP, large
technology
language models (LLMs), distributed ledger technologies, etc.

Table 1: Five types of argumentation discussed in this chapter

concerned with the paradigm shift itself, and the third group is concerned with
the consequences of this paradigm shift for computational argumentation.
For each challenge, we begin by presenting an “observation”. Here, we mean
an observation about the above-mentioned literature on argumentation, i.e., the
Handbook on Formal Argumentation, the COMMA proceedings, the Argument
and Computation journal, and the wider AI literature on argumentation.
Given the wide range of topics discussed in this literature, and the changes
taking place due to technological developments such as LLMs, the observations
we chose to focus on have had a big influence on the contents of this chapter.
Other researchers in argumentation might make different observations and, as
a result, would approach this chapter differently. Thus, this chapter reflects
our personal interpretation of the literature on argumentation.
We use a diverse array of examples for illustrative purposes from areas such
as decision-making, ethical and legal reasoning, and practical reasoning, and
these are listed in Table 3. The selected examples cover a wide range of disci-
plines and issues, illustrating also the breadth of potential application domains
for the techniques discussed in this chapter. We reuse examples across different
challenges so that we can look at these examples from different angles. We end
each challenge by presenting several open questions for further research.
We selected the topics judiciously, leaving out many topics we would have
liked to cover but would have made this chapter too long. To provide the reader
with some additional information, we complement these challenges with open
research questions.
This chapter follows the structure of Table 2 and is organized as follows. Sec-
tion 2 discusses natural argumentation approaches that were prevalent before
the paradigm shift, identifying four key challenges. Section 3 focuses on the
paradigm shift itself, acknowledging its contributions while also highlighting
 Thirteen Challenges in Formal and Computational Argumentation 3

Sections Observations Challenges

1. Diversity in reasoning 1.Connecting individual and
across disciplines collective reasoning
2. Diversity of arguments in 2. Understanding and generating
Section 2.
natural language arguments
Context of the
3. Diversity in modeling the 3. Conceptualizing
attack-defense
process of argumentation argumentation
paradigm shift
4. Diversity of formal methods
4. Formalizing argumentation
used in formal argumentation
5. Creating argumentation
5. Universality of attack
frameworks and semantics
6. Variety of nonmonotonic
6. Representing nonmonotonic
logics and game theory
logics and solution concepts
solution concepts
7. Inconsistent knowledge 7. Rationality postulates for
Section 3.
bases defining a new logic
Attack-defense
8. Generalizing Dung’s
paradigm shift 8. Dialogue is based on agents,
attack-defense paradigm for
strategies, and games
dialogue
9. Generalizing Dung’s
9. Balancing is based on
attack-defense paradigm for
support
balancing
10. Conducting a principle-based
10. Diversity of argumentation
analysis of argumentation
11. Compositional nature of 11. Designing efficient algorithms
Section 4.
argumentation for argumentation semantics
Computational
12. Human-level and
turn after the 12. Explaining argumentation
human-centered argumentation
paradigm shift
13. Integrating argumentation
13. Argumentation technology
with technologies

Table 2: Thirteen observations and challenges discussed in this chapter

the challenges it presents to our community. Section 4 examines the conse-

quences of the paradigm shift, including the transition towards computational
AI argumentation, and provides a broader view of the community.

2 Context of the attack-defense paradigm shift

In this section, we describe the diversity of argumentation in the literature
to position and appreciate the attack-defense paradigm shift. In Section 2.1,
we introduce the diversity of reasoning across disciplines, in Section 2.2 we
discuss the diversity in natural argumentation, in Section 2.3 we discuss three
argumentation conceptualizations, and in Section 2.4, we discuss the diversity
of formal methods used in this area. We discuss the challenges of: connecting
individual and collective reasoning; understanding, analyzing, evaluating, and
generating arguments; and conceptualizing and formalizing the argumentation
4 Liuwen Yu, Leendert van der Torre, and Réka Markovich

Challenge
Example Discipline
Number(s)
Jiminy ethical governor Machine ethics 1, 2
Fatio dialogue protocol Speech act theory 2
Dialogue between au-
Speech act theory, argumentation
tonomous robot NS-4 and 2
schemes
Spooner
Child custody AI & law 2, 3, 9
Scottish fitness lover, snoring Knowledge representation and
4, 6
professor reasoning
Untidy room Neuro-symbolic AI 4
Knowledge representation and
Bachelor vs. married 7
reasoning
Dialogue between accuser and
AI & law 8
suspect
Murder at Facility C AI & law 12
Intelligent Human-Input-
Computer science, financial
Based Blockchain Oracle 13
markets, AI & law
(IHiBO)

Table 3: Examples, disciplines, and relevant challenges (see Table 2)

process.
2.1 Individual and collective reasoning
In this section, we discuss reasoning, its philosophical and mathematical foun-
dations, and its use across many disciplines. From this perspective, we illustrate
in Table 1 our definition of formal argumentation — representing, managing
and (sometimes) resolving conflicts —using as an example the six layers of
conflict addressed by the Jiminy ethical governor.
Russell and Norvig [2010] identify four schools of thought on AI — machines
that: think like humans, act like humans, think rationally, and act rationally.
We can interpret these four schools of thought from various perspectives:
Practical reasoning vs. theoretical reasoning Practical reasoning is ori-
ented towards choosing a course of action on the basis of goals and knowl-
edge of one’s own situation, while theoretical reasoning is oriented to-
wards finding reasons for determining that a proposition about the world
is true or false [Walton, 1990].
Descriptive reasoning vs. prescriptive reasoning Descriptive reasoning
aims to replicate human intelligence and behavior, while prescriptive rea-
soning aims to simulate decision-making and prescribe actions that align
with ethics and laws.
In all these different kinds of reasoning, there could be individual reasoning
and collective reasoning. This brings us to the distinction between intelli-
 Thirteen Challenges in Formal and Computational Argumentation 5

gent systems and multiagent systems across various disciplines. In the social
sciences, the distinction between individual and collective reasoning is called
the micro-macro dichotomy [Coleman, 1984]. Another prototypical example is
the distinction between classical decision theory based on the expected utility
paradigm and classical game theory [Savage, 1972]. Whereas classical deci-
sion theory is a kind of optimization problem (concerned with maximizing the
agent’s expected utility), classical game theory is a kind of equilibrium analy-
sis [Nash Jr, 1950]. This leads to the following challenge.
Challenge 1. Connecting individual and collective reasoning.
One of the central topics in reasoning is how to handle conflict, whether
it arises among beliefs (logical inconsistency) or choices of actions (practical
conflicts). This is relevant both to the reasoning process of an individual agent
and to interactions among multiple agents.
Example 2.1 illustrates a conflict from the perspective of a single agent.
Example 2.1 (Car accident dilemma in I, Robot). In the film I,
Robot, Detective Del Spooner is driving when he has an accident, plunging
his own car and another vehicle carrying a child into a river. An autonomous
general-purpose humanoid robot, NS-4, is passing by and is faced with a con-
flict because it cannot save all the humans involved in the accident, i.e., the
drivers and the child. NS-4 must make a descriptive analysis of the situation
and follow prescriptive actions guided by ethical and legal considerations.
Example 2.2 illustrates a conflict from a multiagent perspective.
Example 2.2 (Continued from Example 2.1). Instead of seeing the con-
flict in terms of saving Spooner or saving the girl, a conflict that is faced by
NS-4 only, we can consider it as a disagreement between two stakeholders: NS-
4 and Spooner. NS-4 wants to save Spooner, while Spooner wants to save the
girl.
NS-4 and Spooner might reach a consensus through a process known as
judgment aggregation where they combine their individual judgments to arrive
at a collective decision [Caminada and Pigozzi, 2011]. However, in the context
of game theory, the goal is not always to resolve all conflicts but rather to
understand the dynamics at play and sometimes agree to disagree [Aumann,
2016]. This concept, known as equilibrium analysis, allows for a situation
where Spooner and NS-4 recognize their differing priorities and accept the
disagreement without necessarily resolving the conflict.
In a multi-stakeholder setting, conflicts can be conceptualized and managed
at various layers. The Jiminy architecture [Liao et al., 2023] is an ethical
governor that uses techniques from normative systems and formal argumen-
tation to get moral agreements from multiple stakeholders. Each stakeholder
has their own normative system. The Jiminy architecture combines norms into
arguments, identifies their conflicts as moral dilemmas, and evaluates the argu-
ments to resolve each dilemma whenever possible. In particular cases, Jiminy
6 Liuwen Yu, Leendert van der Torre, and Réka Markovich

decides which of the stakeholders’ normative systems takes preference over the
others.
Example 2.3 (The six layers of conflict of I, Robot in the Jiminy
architecture).
Layer 1: conclusions only The conflict is based on the conclusions drawn by
each stakeholder. NS-4 concludes it should save Spooner, while Spooner
concludes that the girl should be saved. This layer represents a straight-
forward conflict of decisions without going into the underlying reasoning,
possibly making it difficult to resolve.
Layer 2: assumptions and reasons Agents present their conclusions along
with their assumptions and reasons. We refer to these conclusions to-
gether with the assumptions and reasons as arguments. Conflict reso-
lution may involve formal argumentation techniques such as assigning
attacks among arguments. The reason Spooner wants to save the girl is
that she has a longer potential lifespan. That reason could be attacked by
an argument from NS-4 that the girl has incurable cancer and therefore
has a short lifetime.
Layer 3: combining normative systems This layer involves combining mul-
tiple normative systems into a single normative system. As a conse-
quence, there may be new arguments built from the norms of distinct
stakeholders, and the combined knowledge may be sufficient to reach
a moral agreement. For example, NS-4 has information unknown to
Spooner — the child’s illness. By aligning their knowledge bases, they
may reach an agreement to save Spooner instead.
Layer 4: context-sensitive meta-reasoning as ethical governors Jiminy
considers the agents’ norm preferences. These meta-norms are context-
dependent norms that select the one stakeholder who has the most rel-
evant expertise. Jiminy may decide that NS-4’s preference takes prece-
dence over Spooner’s because NS-4 can get a more accurate evaluation of
the imminent accident, leading to a more reasonable decision. This mech-
anism is comparable to those used in private international law [Markovich,
2019].
Layer 5: suspend decisions and observation Traditional conflict resolu-
tion often assumes that dilemmas must be addressed immediately. How-
ever, suspending a decision to allow for additional information to emerge
can be beneficial in certain situations.
Layer 6: dialogue In this layer, stakeholders engage in a dialogue, attempt-
ing to persuade one other. Through structured communication, NS-4
and Spooner present their arguments, counterarguments, and justifica-
tions. The dialogue helps them explore each other’s perspectives and can
lead to a more informed and mutually agreeable resolution.
 Thirteen Challenges in Formal and Computational Argumentation 7

There are many other examples similar to the one in I, Robot. For instance,
the tunnel dilemma and the trolley dilemma [Awad et al., 2018] involve ethical
decisions by autonomous vehicles and the question of who should decide how
they respond in life-and-death situations. Another example is a smart speaker
that passively listens in and stores voice recordings, acting like a surveillance
device [Liao et al., 2023]. Should it assist in the prevention of or investigation
into crimes? This presents a moral dilemma involving household members,
law enforcement agencies, and the manufacturer of the smart speaker. Which
stakeholder should be alerted in such cases?
Jiminy explains the general problem of connecting individual and collective
reasoning, and its relation to practical reasoning. Different mechanisms could
be implemented in the Jiminy architecture to connect the two kinds of reason-
ing. For example, philosophical concepts such as the veil of ignorance [Rawls,
2001] and Kantian imperative [Kant, 1993] offer valuable perspectives. The veil
of ignorance principle requires individuals to make decisions without knowledge
of their own personal characteristics or societal position, thus promoting im-
partiality and fairness in collective decision-making. This aligns closely with
the AI challenge of designing a system that makes unbiased decisions. Simi-
larly, Kant’s categorical imperative suggests that one should act only according
to maxims that can be universally applied to build universally ethical and ra-
tional guidelines for behavior. Both principles emphasize the importance of
considering the broader implications of actions, and they encourage integrat-
ing individual actions into collective norms and ethics.
In this section, we discussed the general challenge of connecting individual
and collective behavior from the perspective of argumentation. We end this
section by raising a number questions for further research.

1. In this section, we considered argumentation as a kind of reasoning, which

raises the question: what kinds of reasoning count as argumentation, and
what kinds of argumentation do not count as reasoning? More generally,
what is the scope of argumentation?

2. Which kinds of reasoning cannot be handled by argumentation? For ex-

ample, can causal or case-based reasoning be cast as a kind of argumenta-
tion [Bengel et al., 2024; Roth and Verheij, 2004]? Can decision-making
be regarded as a kind of argumentation [Amgoud, 2009]?

3. How does argumentation relate to other kinds of reasoning? For example,

What are the distinctions and connections between reasoning as a cog-
nitive activity and argumentation as a communicative practice [Walton,
1990]?

4. How should argumentation be used in a general theory of reasoning? For

example, some articles refer to argumentation and negotiation [Sycara,
1990; Rahwan et al., 2003; Van Laar and Krabbe, 2018], examining how
8 Liuwen Yu, Leendert van der Torre, and Réka Markovich

argumentation can be used as part of a negotiation, or how a negotiation

can be seen as a kind of argumentation process.

5. How should argumentation be applied to legal and ethical reasoning? For

example, how can argumentation facilitate structured discourse among
agents to negotiate conflicting norms, particularly in multiagent legal pro-
ceedings where stakeholders argue for or against specific outcomes [Prakken
and Sartor, 2015]?

2.2 Natural arguments

In this section, we discuss the diversity of natural argumentation, psychological
analysis of natural (and formal) arguments, transitioning from NLP to foun-
dation models and chatbots, understanding and generating arguments using
foundation models, and the central role of questions in natural argumentation.
We illustrate the latter using critical questions in argument schemes to find
weaknesses and by asking why questions to obtain justifications according go
the so-called Fatio dialogue protocol.
Natural argumentation refers to the way humans reason and communicate
naturally in everyday language. It combines elements of linguistics, philosophy,
and rhetoric. It is characterised by considerable diversity, with arguments
taking various forms, styles, and contexts. Our aim is to avoid developing a
separate technical understanding and generating arguments and argumentation
with only a weak connection to how these concepts are understood in the
humanities and related fields by both scholars and practitioners [Gordon, 2018].
In linguistics, researchers evaluate the diversity of natural language argu-
ments and their role in human interaction with experimental methodologies
[Gillioz and Zufferey, 2020]. These methodologies often include human-based
techniques such as psychological experiments to verify linguistic intuitions
[Schindler et al., 2020] and determine sound arguments and standards of justi-
fication [Weinstock, 2006]. In AI, the focus shifts to modeling, formalizing, and
automating the argumentation process, generating arguments in both natural
and formal languages. However, evaluation in AI also relies heavily on human
assessment. For instance, empirical cognitive experiments have been conducted
by Cramer and Guillaume [2018a; 2018b; 2019] and Cerutti et al. [2021] to eval-
uate the connection between natural and abstract argumentation. Two main
research questions often guide this evaluation: Do the features shared by major
argumentation semantics (e.g., simple reinstatement) correspond to genuine
cognitive aspects of human reasoning? Which argumentation semantics are the
best predictors of human evaluation of arguments? Human evaluations of auto-
matically generated text are also conducted to assess their performance [van der
Lee et al., 2021].
In recent years, the transition from NLP [Budzynska et al., 2018] to
argumentation-based chatbots [Black et al., 2021] has been accelerated by ad-
vancements in foundation models, such as OpenAI’s Generative Pre-trained
Transformer (GPT) series. This leads us to the following challenge.
 Thirteen Challenges in Formal and Computational Argumentation 9

Challenge 2. Understanding and generating natural arguments.

Chatbots are conversational software that seek to understand input from
human users and generate human-like responses [Black et al., 2021]. In chat-
bot development, questions play a crucial role in enhancing the effectiveness
of argumentation-based chatbots and building engaging conversations [McBur-
ney et al., 2021]. The use of questions allows chatbots to guide dialogues,
challenge assertions, support critical thinking, and provide justifications. Be-
low, we discuss two question mechanisms that could be embedded into chat-
bots — argumentation schemes with their associated critical questions, and
justification-seeking questions defined by speech act theory.
Argumentation schemes are investigated by an approach developed in phi-
losophy and rhetoric, representing stereotypical patterns of reasoning that
are often informal or semi-formal, rarely fully formalized. Initially developed
for teaching critical thinking, these schemes were systematized by Walton et
al. [2008], who identified sixty-five basic schemes. Argumentation schemes in-
volve the activity of critically evaluating viewpoints and the reasons given to
support them. For this reason, every scheme has a corresponding set of crit-
ical questions to identify possible weak points, challenge the arguments, and
evaluate their strengths. The mechanism of argumentation schemes and criti-
cal questions could potentially improve chatbots. Chatbots, or conversational
agents like ChatGPT, are good at crafting human-like sentences [Alkaissi and
McFarlane, 2023]. But they often present falsehoods as facts and exhibit in-
consistent logic, and these can be difficult to detect. Users tend to follow the
chatbot’s logic when given ready-made answers. However, when chatbots pose
questions, they prompt users to engage in deeper critical thinking and question
their responses [Danry et al., 2023], which fosters more realistic and reliable
interactions.
For example, consider below the argumentation scheme for practical reason-
ing.

Major premise: I have a goal G.

Minor premise: Carrying out action A is a means to realize G.
Therefore, I ought (practically speaking) to carry out this
Conclusion:
action A.

There are five critical questions:

CQ1 What other goals do I have that should also be considered even though
they might conflict with G?

CQ2 Other than me bringing about A, what alternative actions should be

considered that would also bring about G?

CQ3 From the solutions of me bringing about A and these alternative actions,
which can be argued to be the most efficient?
10 Liuwen Yu, Leendert van der Torre, and Réka Markovich

CQ4 What grounds are there for arguing that it is possible for me to bring
about A in practice?
CQ5 What other consequences of me bringing about A should be taken into
account?
Example 2.4 illustrates the dialogue between NS-4 and Spooner based on
the argumentation schemes of practical reasoning.
Example 2.4 (Dialogue between NS-4 and Spooner).

Spooner: Save the girl! That is the moral and ethical choice. She deserves
the chance to live her life fully.
NS-4: What other goals do you have that might conflict with this one?
Spooner: My goal is to save the most vulnerable lives. There is no conflict.
NS-4: What alternative actions should be considered?
Spooner: Saving the girl should be the only course of action. It should have
the highest priority.
NS-4: What is the most efficient choice?
Spooner: Saving the girl. She is lighter, so this course of action is more likely
to succeed.
NS-4: What grounds are there for arguing that it is practically possible to
save the girl?
Spooner: The girl’s lighter weight makes her rescue quicker and less risky.
NS-4: What consequences should be considered?
Spooner: Saving the girl aligns with the duty to protect the vulnerable.
NS-4: Your argument is sound and aligns with ethical and practical consider-
ations. I will save the girl.

Speech act theory, a subfield of pragmatics, studies how words are used not
only to present information but also to carry out actions [Austin, 1975]. This
theory has been formalized in the Foundation for Intelligent Physical Agents
(FIPA) standards, which are widely used in computer science to facilitate com-
munication between autonomous agents [FIPA, 2002]. The scheme allows mul-
tiple labels to be applied to one utterance since a single utterance can perform
multiple actions in a dialogue [Kissine, 2013]. Such labels range from a few
basic types such as assertions, questions and commands to more complex ones
like promises and declarations [Searle, 1979]. One of the key features of speech
acts, as opposed to physical actions, is that their main effects are on the mental
 Thirteen Challenges in Formal and Computational Argumentation 11

and interactional states of agents [Traum, 1999]. The attitudes of belief, desire
and intention are familiar to agency theories [Georgeff et al., 1999]. In the
context of human-like chatbots, speech acts can be used to design interactions
between the chatbot and the user [Hakim et al., 2019]. Specifically, questions
that seek justification are crucial as they prompt the chatbot to provide reasons
and explanations, which not only enriches the interaction but also drives the
conversation towards deeper engagement and understanding.
McBurney and Parsons [2004] proposed an interaction protocol called Fatio
comprising of five locutions for argumentation which can be considered as a set
of speech acts.

F1: assert (Pi , ϕ): A speaker Pi asserts a statement ϕ. In doing so, Pi creates
a dialectical obligation within the dialogue to provide a justification for
ϕ if required subsequently by another participant.

F2: question (Pj , Pi , ϕ): A speaker Pj questions a prior utterance of as-

sert(Pi , ϕ) by another participant Pi and seeks a justification for ϕ. The
questioner Pj creates no dialectical obligations.

F3: challenge (Pj , Pi , ϕ): A speaker Pj challenges a prior utterance of as-

sert(Pi , ϕ) by another participant Pi and seeks a justification for ϕ. Pj
not only asks a question but also creates for himself a dialectical obli-
gation to provide a justification for not asserting ϕ. For example, he
must provide an argument against ϕ if questioned or challenged. Thus,
challenge (Pj , Pi , ϕ) is stronger than question (Pj , Pi , ϕ).

F4: justify (Pi , Φ ⊢ ϕ): A speaker Pi , who had uttered assert(Pi , ϕ) and was
then questioned or challenged by another speaker, is able to provide a
justification Φ ∈ A for the initial statement ϕ by means of this locution.

F5: retract (Pi , ϕ): A speaker Pi , who had uttered assert(Pi , ϕ) or jus-
tify(Pi , Φ ⊢ ϕ), can withdraw this statement with the utterance of re-
tract(Pi , ϕ) or the utterance of retract(Pi , Φ ⊢ ϕ) respectively. This
removes the earlier dialectical obligation on Pi to justify ϕ or Φ ⊢ ϕ if
questioned or challenged.

Example 2.5 illustrates the dialogue between NS-4 and Spooner following
the speech act.
Example 2.5 (A dialogue between NS-4 and Del Spooner). xx

Spooner: Saving the girl is the right choice. (assert)

NS-4: Why? (question)

Spooner: Because the girl is young and has a much longer lifespan. (justify)

NS-4: I disagree that she has a much longer lifespan. (challenge)

12 Liuwen Yu, Leendert van der Torre, and Réka Markovich

Spooner: Why? (question)

NS-4: I conducted a health evaluation and found that she has a terminal
disease. (justify)

In this section, we discussed the common understanding of natural argu-

ments and how they are generated. We end with some open research questions.

1. In natural argumentation, we often encounter fallacies [Hamblin, 1970],

and on the internet we encounter fake news [Visser et al., 2020]. How
should fallacies and fake news be handled in argumentation-based chat-
bots? For example, how can argumentation schemes be used in a chatbot
to evaluate if an argument is fallacious [Walton, 2013]?
2. Programming has been replaced by prompt engineering for interacting
with LLMs [Ross et al., 2023]. How can argument schemes and speech act
theory be used in prompt engineering? For example, argument schemes
can potentially help structure for the generation of well-formed argu-
ments [Musi and Palmieri, 2022].
3. An increasing number of arguments on the internet are generated by
AI [Hinton and Wagemans, 2023]. How should AI-generated arguments
be evaluated? The evaluation could focus on criteria like logical coher-
ence, the relevance and sufficiency of the evidence, adherence to ethical
principles, and the impact of the argument on the intended audience.
Additionally, metrics could be developed to assess how well AI argu-
ments handle counterarguments and whether they respect the norms of
constructive and respectful discourse.
4. Argumentation has been discussed as a key component of chat-
bots [Castagna et al., 2024a]. Which domain applications are
argumentation-based chatbots suitable for? For example, how can ar-
guments be used in AI therapy?
5. In the previous section, we introduced the Jiminy architecture. How can
foundation models, argument schemes, and speech act theory be used to
improve or enrich the Jiminy architecture?

2.3 Models of argument

In this section, we discuss three conceptualizations of argumentation — ar-
gumentation as inference, argumentation as dialogue, and argumentation as
balancing. Each conceptualization embodies: a unique perspective on the con-
struction and purpose of argumentation, a set of formal methods, and applica-
tion across different disciplines and contexts. As mentioned in the introduction,
we view argumentation as representing, managing and sometimes resolving
conflicts. We explain how this key idea becomes more concrete with the three
conceptualizations.
 Thirteen Challenges in Formal and Computational Argumentation 13

For argumentation as inference [Prakken, 2018], we consider: coherent po-

sitions in cases of conflict, what follows from each coherent position (or what
we can infer from all coherent positions), and what can be agreed upon in
cases of disagreement. For argumentation as dialogue [Prakken, 2018], we also
consider the stakeholders that may hold such coherent positions and how they
might interact, for example as proponents and opponents in a debate. This can
clarify the conflict that is being managed and sometimes even help to resolve
it. In such dialogues, the concerns or goals of the stakeholders can also be
made explicit. As in dispute resolution, the process becomes very important.
Finally, in argumentation as balancing [Gordon, 2018], we consider conflicts as
trade-offs involving taking into account various pros and cons. Here, the cen-
tral metaphor, referred to frequently in the law and ethical decision-making,
is a pair of scales. For fine-grained comparisons, it is not uncommon to use
quantitative metrics. In this section, we discuss the following challenge.
Challenge 3. Conceptualizing argumentation.
Table 4 provides a detailed comparison of the three conceptualizations. No-
tably, the applications of these conceptualizations are neither mutually exclu-
sive nor incompatible. The formal methods are discussed in Section 2.4.

Theories and
Conceptua-
Process Formal Application
lization
Approaches
Logical structure and Graph theory, Automated
reasoning to derive nonmonotonic logic, reasoning systems,
Argumentation
conclusions from computational logic, knowledge
as inference
incomplete and causal reasoning, representation,
inconsistent premises Bayesian reasoning expert systems
Dynamic verbal Speech act theory, Debating
interaction between game theory, technologies,
Argumentation
stakeholders to exchange axiomatic semantics, chatbots,
as dialogue
information or resolve operational recommender
conflicts of opinion semantics systems
Multi-criteria
Deliberative
Balancing pros and cons decision theory,
Argumentation decision-making in
to reach a justified machine ethics,
as balancing law, ethics, and
decision computational law,
economics
case-based reasoning

Table 4: Conceptualizations of argumentation

Argumentation as inference focuses on determining the conclusions that

can be derived from a given body of information, which may be incom-
plete, inconsistent, or uncertain. Relevant systems ultimately define a
nonmonotonic notion of logical consequence in terms of the intermediate
notions of argument construction, argument attack, and argument eval-
uation, and the arguments are seen as constellations of premises, con-
14 Liuwen Yu, Leendert van der Torre, and Réka Markovich

clusions, and inferences [Prakken, 2018]. These systems employ formal

methods like nonmonotonic logic for commonsense reasoning, graph the-
ory, computational logic, causal reasoning, and Bayesian reasoning. Ar-
gumentation as inference is primarily applied in knowledge representation
and reasoning.

Argumentation as dialogue conceptualizes argumentation as a form of ver-

bal interaction aimed at resolving conflicts of opinion [Prakken, 2018].
Relevant systems define argumentation protocols, which serve as the rules
of the argumentation game, and they address strategic aspects that guide
effective engagement in the game. The exploration of strategies involves
understanding how to engage in productive discourse and present argu-
ments effectively. Argumentation as dialogue utilizes speech act theory,
game theory, axiomatic semantics, and operational semantics. It is most
suitable for debates, chatbots, persuasion systems, negotiation systems,
etc.

Argumentation as balancing involves weighing the pros and cons of an is-

sue in order to reach a balanced decision or judgment. It is applicable not
only when resolving conflicts of opinion in persuasion dialogues but also,
e.g., when deciding courses of action in deliberation dialogues [Gordon,
2018]. In such a system, pro and con arguments for alternative resolutions
of the issues (options or positions) are put forward, evaluated, resolved,
and balanced. The formal methods used are multi-criteria decision the-
ory, machine ethics, computational law, and case-based reasoning, and
they are applied in the realms of law, ethics, and economics.

Table 4 might give the impression that the three approaches are distinct and
that they have distinct application areas. We would like to point out that this
is not the case. The approaches (or types) of argumentation are not mutually
exclusive or even incompatible. You can switch from one to another if you want
to look at the same problem or situation from different angles, highlight differ-
ent aspects, or select a modeling approach that is more suitable for a particular
purpose. This also means that complex application areas like the legal domain
can make very good use of each approach. Indeed, legal reasoning often en-
gages with each of the three conceptualizations — argumentation as inference,
dialogue, and balancing — across different contexts and legal roles. Judges
and attorneys may rely on one form of argumentation more than another, de-
pending on the nature of the case and their specific role in the legal process.
For instance, inference is commonly used by judges, attorneys, actually any
type of lawyer, when applying legal rules to facts or deriving conclusions from
incomplete or inconsistent premises. Dialogue plays a central role in courtroom
exchanges between opposing parties. The structure of a trial often resembles
a dialogue: each party presents their arguments and responds to those of the
other while the judge oversees the process to ensure it follows legal procedures.
 Thirteen Challenges in Formal and Computational Argumentation 15

Balancing is typically the domain of judges as they weigh multiple factors, con-
flicting interests, or values, to determine the most appropriate outcome. This is
particularly important in discretionary decision-making where the law, instead
of trying to provide detailed rules, assigns special power to judges so that they
can make decisions based on their own evaluations. In such cases, judges ex-
ercise their judicial discretion by carefully balancing competing considerations
within the framework of legal principles to reach a fair and just decision.
Hence, these different modes of reasoning can correspond to and interact
with one another, creating a comprehensive tool set for legal reasoning and
decision-making. Below, we shall illustrate each of the three conceptualizations
using the legal example of child custody in a divorce case.
Research on argumentation-based dialogue (see the overview of Black et
al. [2021]) is often carried out against the background of the six types of dia-
logue and in accordance with their respective goals [Walton and Krabbe, 1995],
as shown in Table 5. When argumentation is viewed as a kind of dialogue be-
tween multiple agents (whether human or artificial), new issues arise. One issue
is the distributed nature of information (among the agents). Another issue is
the dynamic nature of information — agents do not reveal everything they be-
lieve initially, and they can learn from one other. There are also strategic issues
— agents will have their own internal preferences, desires and goals [Prakken,
2018]. In Section 2.2, we described the speech act theory on dialogue forma-
tion [McBurney and Parsons, 2002]. For better comparison, we use a legal child
custody case [Yu et al., 2020] to illustrate the three conceptualizations.

Type of
Initial situation Participant goal Dialogue goal
dialogue
Persuasion Conflict of opinions Persuade other party Resolve or clarify issue
Find and verify Prove (or disprove)
Inquiry Need to have proof
evidence hypothesis
Get what they most Reasonable settlement
Negotiation Conflict of interests
want they can both live with
Information- Acquire or give
Need information Exchange information
seeking information
Dilemma or Co-ordinate goals and Decide best available
Deliberation
practical choice actions course of action
Verbally hit out at Reveal deeper basis of
Eristic Personal conflict
opponent conflict

Table 5: Types of dialogue [Walton and Krabbe, 1995]

Example 2.6 (Child custody dialogue). Alice and Lucy are talking about
a divorce case, specifically whether it is in the child’s best interest to live with
her mother or with her father. They have the following dialogue.

Alice: It is in the ten-year-old child’s best interest that she lives with her
mother. (assert)
16 Liuwen Yu, Leendert van der Torre, and Réka Markovich

Lucy: Why? (question)

Alice: Because the child wants to live with her mother and the civil code states
that the judge must take the child’s opinion into account. (justify)
Lucy: A ten-year-old child does not know what she wants. (challenge)
Alice: Why? (question)
Lucy: Public opinion says that ten-year-old children do not know what they
want. (justify)
Alice: Most ten-year-old children do know what they want. (assert)
Lucy: Why do you say that? (question)
Alice: Peter is a child psychologist, and Peter says that most ten-year-old
children know what they want. (justify)

Most of the literature in this area is concerned with argumentation as infer-

ence. Some formal work had already had been carried out on argumentation-
based inference before the publication of Dung’s 1995 paper, notably the exten-
sive research by Pollock [1987; 1992; 1994; 1995; 2001; 2009; 2010] on argument
structure, the nature of defeasible reasons, the interplay between deductive and
defeasible reasons, rebutting versus undercutting defeat, argument strength, ar-
gument labeling, self-defeating arguments, etc. Pollock identified reasoning as
a process of constructing arguments where reasons provide the atomic links in
arguments [Pollock, 1992]. He distinguished between two kinds of reasons: de-
feasible (prima facie) reasons and nondefeasible (conclusive) reasons [Pollock,
1987]. Nondefeasible reasons are those reasons that logically entail their con-
clusions while defeasible reasons may be destroyed with additional information.
There are two kinds of defeaters that can defeat defeasible reasons. Rebutting
defeaters deny the conclusion. Undercutting defeaters attack the connection
between the reason and the conclusion. Pollock [1992; 1994; 1995] used so-called
inference graphs to represent arguments and the nodes represented the steps
of inference. There are three kinds of arrows in the inference graph, and they
represent defeasible inferences, deductive inferences, and defeat links [Pollock,
1994].
Example 2.7 (Child custody in an inference graph). The dialogue be-
tween Alice and Lucy can also be illustrated in the format of Pollock’s inference
graph, as shown in Figure 1. Figure 1 illustrates two arguments rebutting the
two opposite conclusions “It is in the child’s best interest that she lives with
her father”, and “It is not in the child’s best interest that she lives with her
father”. An undercutting argument, “Public opinion is not reliable”, defeats
the argument “Most ten-year-old children do not know what they want”. In this
figure, nondefeasible and defeasible inferences are visualized respectively with
solid and dotted lines (without arrowheads). The arrows are defeat relations.
 Thirteen Challenges in Formal and Computational Argumentation 17

(1) Peter says that most (11) Public opinion says that
ten-year-old children (2) Peter is a child psychologist. most ten-year-old children
know what they want. do not know what they want.
r5
r1
(4) The child says (3) Most ten-year-old (12) Most ten-year-old
rebut
she wants to live children know what children do not
with her mother. they want. know what they want.

undercut

(6) The Civil Code states that r2

(5) The child indeed wants (13) Public opinion is
the judge must take into consideration
to live with her mother. not reliable.
the child’s opinion.

r3
(7) It is in the child’s best interest that
(9) The father is wealthy.
she lives with her mother.
r4 r6

(8) It is not in the child’s best interest rebut (10) It is in the child’s best interest
that she lives with her father. that she lives with her father.

Figure 1: A dialogue about a child custody case represented as a Pollock infer-

ence graph. The solid and dotted lines (without arrowheads) are nondefeasible
and defeasible inferences respectively. The arrows are defeat relations.

One model of argumentation as balancing is the Carneades Argumentation

System [Gordon et al., 2007]. The conception of argument graphs in Carneades
is similar to Pollock’s conception of an inference graph. There are nodes in the
graph representing statements (propositions) and links that indicate inference
relations between statements. In particular, the system distinguishes between
pro and con arguments. Semantically, con arguments are instances of pre-
sumptive inference rules for negating the conclusion. If the premises of a con
argument hold, this justifies rejecting the conclusion or, equivalently, accepting
its logical complement. With pro and con arguments, some statements need
to be ordered or otherwise aggregated to resolve the conflict. Then there are
several proof standards used to balance the pros and cons. Here are three
examples:
SE (Scintilla of Evidence): A statement meets this standard iff it is sup-
ported by at least one defensible pro argument.
BA (Best Argument): A statement meets this standard iff it is supported
by some defensible pro argument with priority over all defensible con
arguments.
DV (Dialectical Validity): A statement meets this standard iff it is sup-
ported by at least one defensible pro argument and none of its con argu-
ments are defensible.
18 Liuwen Yu, Leendert van der Torre, and Réka Markovich

Example 2.8 (Child custody in Carneades). We represent part of the

child custody example in Carnedes, as visualized in Figure 2. Statements are
depicted as boxes and arguments as circles. For the purpose of this discussion,
we assume that all the premises are ordinary without distinguishing between
different types of premises. Premises are shown as edges without arrowheads.
Pro arguments are indicated by circle arrowheads while con arguments are
shown with standard arrowheads. Argument a1 asserts that the child knows
what she wants and she wants to live with her mother, making it a pro argument
for the statement “It is in the child’s best interest that she lives with her
mother”. In contrast, argument a2 argues that the mother is less wealthy than
the father, serving as a con argument against that statement. In this scenario,
a1 is given priority over a2 . Consequently, according to the BA proof standard,
the statement “It is in the child’s best interest that she lives with her mother”
is accepted.

It is in the child’s
best interest that she
lives with her mother.
PRO CON

a1 a2

The mother is
The child knows The child wants to
less wealthy
what she wants. live with her mother.
than the father.

Figure 2: Child custody case represented in Carneades argument graphs

Argumentation as inference, argumentation as dialogue, and argumentation

as balancing are distinct conceptualizations but they are not incompatible.
A crucial question is how to move between individual reasoning (argumen-
tation as inference and argumentation as balancing) and collective reasoning
(argumentation as dialogue). Argumentation as inference can occur within an
individual’s mind, drawing upon a single knowledge base. However, in mul-
tiagent dialogues, each agent operates with a distinct and dynamic knowl-
edge base. Agents employ strategic moves requiring them to learn about
and understand other participants and to select or generate arguments from
their knowledge base to achieve the goal of the dialogue. Some works at-
tempt to integrate these different conceptualizations. For instance, discussion
games serve as a proof procedure for abstract argumentation semantics [Cam-
inada, 2018a], and multiagent argumentation considers agents with varying
attitudes and knowledge [Arisaka et al., 2022]. Moreover, several new agent
argumentation semantics are inspired by social choice theory [Yu et al., 2021;
Baumeister et al., 2021], and Carneades models legal dialogue by using critical
questions from argumentation schemes while incorporating balancing to model
 Thirteen Challenges in Formal and Computational Argumentation 19

judgments [Gordon et al., 2007].

In this section, we discussed argument models. We end again with some
open questions:
1. Are there any other conceptualizations of argumentation that should be
considered in the argumentation literature?
2. How should it be decided which conceptualization of argumentation to
use for an application? In particular, when should argumentation as
inference, dialogue, or balancing be used?
3. How are these conceptualizations related and should they be combined?
For example, how should agent interaction and dialogue [Arisaka et al.,
2022] be introduced into Pollock’s theory?
4. How should a general framework for argumentation as dialogue be de-
signed? The formal study of argumentation-based dialogue is less sub-
stantial than the formal study of argumentation-based inference. It
largely consists of a variety of different approaches and individual sys-
tems, with few unifying accounts or general frameworks [Prakken, 2018].
5. How should argumentation as balancing be represented formally? Com-
pared with argumentation as inference and dialogue, there is little for-
mal work on argumentation as balancing. Some examples are bipolar
argumentation [Cayrol and Lagasquie-Schiex, 2013; Cayrol et al., 2021;
Yu et al., 2020; Yu et al., 2023], and an investigation into balancing op-
erations [Knoks and van der Torre, 2023; Knocks et al., 2024].
2.4 Formalizing argumentation
In this section, we discuss the large variety of formal methods in formal argu-
mentation and their combination in applications and case studies. We show
the use of nonmonotonic logic for commonsense reasoning, the integration of
axiomatic and operational semantics in the Fatio dialogue system, and a com-
bination of different reasoning methods with an example of a mother reasoning
about her daughter.
Each conceptualization of argumentation comes with its own set of formal
methods, as shown in Table 4 in Section 2.3. Basically, argumentation as
inference uses most of the methods from graph theory (e.g., abstract argumen-
tation [Dung, 1995]) or methods from nonmonotonic and computational logic
(e.g., [Pollock, 1987; Nute, 1994; Reiter, 1980]) or causal reasoning [Giunchiglia
et al., 2004; Turner, 2004]. Argumentation as dialogue involves speech act the-
ory (e.g., [FIPA, 2002]), axiomatic semantics, operational semantics, and deno-
tational semantics, as well as game theory methods (e.g., [Moore, 1993]). In ar-
gumentation as balancing, there are methods from multi-criteria decision theory
(e.g., [Amgoud, 2005; Amgoud and Vesic, 2012; Knoks and van der Torre, 2023;
Knocks et al., 2024]), ethical theory, legal theory, and case-based reasoning
(e.g., [Yu and Gabbay, 2022]). In this section, we discuss that challenge.
20 Liuwen Yu, Leendert van der Torre, and Réka Markovich

Challenge 4. Formalizing argumentation.

Nonmonotonic logic was motivated by the fact that commonsense reason-
ing often involves incomplete or inconsistent information, in which case logical
deduction is not a particularly useful reasoning model [Toulmin, 1958]. Clas-
sical logic is characterized by its monotonic nature. It asserts that if a set
of statements S entails a proposition ϕ (denoted S ⊢ ϕ), any superset S ′ of S
also entails ϕ. This principle underpins traditional logical proofs where lemmas
remain valid while new premises are added. However, commonsense reasoning
often allows conclusions to be retracted in the light of new information. For
instance, the inference “Tweety, a bird, flies” (symbolized as |∼) may be valid
until additional contextual information like “Tweety is a penguin” necessitates
the retraction of the initial conclusion. This approach mirrors our everyday rea-
soning processes, which frequently involve default rules with exceptions (e.g.,
a → x for strict rules and a ⇒ x for default rules). Default rules apply unless
there is evidence to the contrary requiring us to revoke our conclusions upon
encountering such exceptions.
Here are some examples of the sources of nonmonotonicity [Prakken, 2010b]:

Empirical generalizations: e.g., adults are usually employed, birds can typ-
ically fly, etc.

Exceptions to legal rules: e.g., when a father dies, the child inherits, except
when the child killed the father.

Exceptions to moral principles: e.g., one should not lie, except when a lie
can save lives.

Conflicting information sources: experts who disagree, witnesses who con-

tradict each other, conflicting sensory input, etc.

Alternative explanations: e.g., the grass is wet so it must have rained, but
the sprinkler was on.

Conflicting reasons for actions: if we have a reason to do something, we

should do it, unless we have good reasons for not doing it.

Prioritized default logic (PDL) [Brewka, 1994] is one formalization of non-

monotonic reasoning. A knowledge base in PDL contains prioritized defaults
n n
a ⇒ b and facts, including tautologies. The notation a ⇒ b means “if a, then
normally b”, with n indicating the priority level; a higher n implies a higher
priority for the default rule a ⇒ b. PDL operates by selecting sets of defaults
and bringing their conclusions into extensions. At each step, the default rule
with the highest priority among the unapplied default rules is applied, while
consistency is maintained.
Example 2.9 (Fitness lover Scot). Assume we have the following defaults
and facts:
 Thirteen Challenges in Formal and Computational Argumentation 21

Defaults Facts
1
 
 BornInScotland ⇒ Scottish
 

3

Scottish ⇒ LikesWhisky BornInScotland, FitnessLover
2
 
FitnessLover ⇒ ¬LikesWhisky
 

We can obtain the extension iteratively.

E1 = {BornInScotland, FitnessLover}
E2 = {BornInScotland, FitnessLover, ¬LikesWhisky}
E3 = {BornInScotland, FitnessLover, ¬LikesWhisky, Scottish}
E4 = E3

We introduced the speech acts of Fatio [McBurney and Parsons, 2002] in

Section 2.2, and we showed how reasons are used in a dialogue by ‘question’
and ‘justify’ moves in Section 2.3. We now reference Fatio to show how a
dialogue system can make use of various formal methods: in this case, axiomatic
semantics and operational semantics.
An axiomatic semantics for a programming language defines a set of ax-
ioms that the language obeys such as the pre-conditions and post-conditions
for each command [Tennent, 1991]. It defines pre-conditions and post-
conditions for the locutions. In Fatio, the axiomatic semantics concerns the
beliefs and desires of the participating agents, which are written as Bi φ :
“Agent i believes that φ is true”, and Di φ : “Agent i desires that φ be true”.
Central to the axiomatic semantics is a publicly viewable store to record the di-
alectical obligations of the participants, which is called a dialectical obligations
store (DOS). The triple (Pi , φ, +) ∈ DOS(Pi ) denotes that participant Pi has
a dialectical obligation to provide a justification or an argument in support of
proposition φ, while the triple (Pi , φ, −) ∈ DOS(Pi ) denotes that participant
Pi has a dialectical obligation to provide a justification or an argument against
proposition φ.
For illustration, we list the pre- and post-conditions for assert and question,
and we refer the rest of the axiomatic semantics to the original paper [McBurney
and Parsons, 2004].

assert(Pi , φ) Pre-condition: Speaker Pi wants each participant Pj (j ̸= i) to

believe that Pi believes the proposition φ ∈ C.

((Pi , φ, +) ∈
/ DOS(Pi )) ∧ (∀j ̸= i)(Di Bj Bi φ).

Post-condition: Each participant Pk (k ̸= i) believes that participant Pi

wants each participant Pj (j ̸= i) to believe that Pi believes φ.

((Pi , φ, +) ∈ DOS(Pi )) ∧ (∀k ̸= i)(∀j ̸= i)(Bk Di Bj Bi φ).

Dialectical obligation: (Pi , φ, +) is added to DOS(Pi ), the dialectical obli-

gations store of speaker Pi .
22 Liuwen Yu, Leendert van der Torre, and Réka Markovich

question(Pj , Pi , φ) Pre-condition: One participant Pi (i ̸= j) has a dialectical

obligation to support φ and participant Pj wants every other participant
Pk (k ̸= j) to believe that Pj wants Pi to utters a justify(Pi , φ, .) locution.
Post-condition: Participant Pi must utter a justify locution.
Dialectical obligation: No effect.

Operational semantics in Fatio is defined from a traditional computer science

perspective. That means that the state of the system changes as a result of
executions of commands in a programming language. To ensure automated
generation of agent dialogues, participants need mechanisms to invoke specific
utterances at appropriate points in the dialogue, and these mechanisms are
called agent decision mechanisms. In this case, the commands in question are
the locutions in an argumentation dialogue conducted according to the rules
of the protocol. In Fatio, for example, an agent can decide whether to Claim
or Not, whether to React or Not, whether to Fold or Not, whether to
Defend or Not and, as a meta-level decision mechanism, whether to Listen
or Do. There are transition rules defined for Fatio’s operational semantics,
and they assume that agents are equipped with decision mechanisms to initiate
and respond to utterances. This enables the system to initiate utterances and
respond to utterances in the dialogue, and so the states we will take to be the
inputs and outputs of these decision mechanisms reflect that process.
One possible extension to the Fatio protocol is an additional semantics called
denotational semantics, described but not explicitly defined by McBurney and
Parsons [2004]. It would link the utterances made under the protocol to the
nodes and edges of a graph representing the arguments created by the partici-
pants in the course of a dialogue. This kind of graph would be similar to the
argumentation graph constructed in Thomas Gordon’s [1993] Pleadings Game,
which is a formal structure capturing the flow and relationships between differ-
ent arguments in a dialogue or argumentation context. It would thus provide
a mathematical structure to the dialogues, mapping the linguistic constructs
(utterances) to a formal representation (graph) that captures the logical re-
lationships and the dynamics of the argumentation process. This could be a
way to visualize and analyze the structure of the arguments and the interplay
between different participants’ statements and responses in a dialogue.
Lastly, we give another example where various formal methods are com-
bined. Neurosymbolic AI [Garcez et al., 2008] combines neural and symbolic
AI architectures to address the weaknesses of each, providing a robust AI sys-
tem capable of reasoning, learning, and cognitive modeling. This diversity of
formal methods brings many challenges to the area of formal argumentation.
Consider Example 2.10 from Gabbay and Rivlin [2017].
Example 2.10 (Untidy room [Gabbay and Rivlin, 2017]). A mother
goes into her teenage daughter’s bedroom. Her instant impression is that it is
a big mess. There is stuff scattered everywhere. The mother’s feeling is that it
is not like her daughter to be like this. What happened?
 Thirteen Challenges in Formal and Computational Argumentation 23

Conjecture: The girl may be experiencing boyfriend issues.

Further Analysis: The mother notices a collapsed shelf and realizes that
the disarray is due to the shelf collapsing under excessive weight which, upon
reflection, follows a logical (gravitational) pattern.
Several types of reasoning are illustrated through this scenario:

Neural network reasoning: The mess is perceived instantly, similarly to fa-

cial recognition by neural networks.

Nonmonotonic deduction: The mother deduces from the context and her
knowledge that her daughter does not typically live in disarray. Thus,
something extraordinary must have happened.

Abductive reasoning: She hypothesizes a plausible explanation that her

daughter has social-emotional issues, which is common among teenagers.

Database AI deduction: A reevaluation leads to the understanding that the

mess is due to gravitational effects rather than disorganization on the part
of her daughter.

Pattern recognition: Someone accustomed to similar patterns may identify

the cause as easily as they might recognize a face.

In practical reasoning, it is crucial to combine various formal methods. To

deal with scenarios similar to Example 2.10, D’Avila Garcez et al. [2005] pro-
posed a hybrid model of computation that allows for deduction and learning
with argumentative reasoning. The model manages to combine value-based
argumentation frameworks and neural-symbolic learning systems by providing
a translation from argumentation networks to neural networks. Another ex-
ample is the general argumentation framework presented by Williamson and
Gabbay [2005]. The framework incorporates the idea of recursive causality and
extends the Bayesian network formalism to cope with recursive causality. The
authors discussed how support relations behave analogously to causal relations
and how arguments are recursive structures; these two observations motivate
the use of recursive Bayesian networks for modeling arguments.
In this section, we discussed the formalization of argumentation. As usual,
we end with some open questions.

1. We have observed that every conceptualization comes with their own

formal methods. What else do they depend on? For example, which
formal method(s) should be chosen for a case study or application?

2. Example 2.10 also illustrated that we often need to combine reasoning

and formal methods. How can various formal methods be combined in
a case study? For example, how can symbolic logic be combined with
network (neural and argumentation) reasoning?
24 Liuwen Yu, Leendert van der Torre, and Réka Markovich

3. Formal argumentation is often presented as a general way to deal with

nonmonotonicity. But how should arguments be conducted when there
are various sources of nonmonotonicity? For example, how should we
argue in legal or ethical contexts?

4. Concepts relevant for argumentation currently include, among others:

time, action, knowledge, belief, revision, deduction, learning, context,
neural networks, probabilistic networks, argumentation networks, consis-
tency, etc. How can these concepts be incorporated into existing formal
models of argumentation?

3 The paradigm shift: the attack-defense perspective

In this section, we discuss and critically reflect upon the attack-defense
paradigm shift. In section 3.1, we discuss the universality of attack and defense.
In section 3.2, we consider the variety of nonmonotonic logics and game theo-
retic solution concepts. In section 3.3, we discuss reasoning with inconsistent
knowledge bases. In section 3.4, we consider argumentation as dialogue that
is based on other concepts besides attack, like agents, strategies, and games.
And in section 3.5, we discuss Dung’s attack-defense paradigm shift for bal-
ancing that is based on both attack and support. We discuss the challenges
of: creating argumentation frameworks and semantics, representing nonmono-
tonic logics and game theoretic concepts, defining rationality postulates for
new logics, generalizing Dung’s attack-defense paradigm shift for dialogue, and
generalizing Dung’s attack-defense paradigm shift for balancing.
3.1 Universality of attack
In this section, we introduce the attack-defense paradigm shift initiated by
Dung’s [1995] paper, we discuss the requirement that every utterance can be
attacked including claims, arguments, and attacks, and we describe the flat-
tening of diverse extended argumentation frameworks into basic ones.
The attack-defense paradigm shift was a turning point in modern formal
argumentation, marked by Dung’s theory of abstract argumentation [Dung,
1995]. In this theory, the acceptability of arguments depends on the attack
relations between them, not their internal structure. An argument is accepted
if it is not attacked or is successfully defended — meaning all its attackers
are attacked. Pre-existing ideas and methods, such as Pollock’s defeasible
reasoning, dialogue theories, and balancing techniques, continue to persist and
influence contemporary research. Rather than being rendered obsolete, these
traditional theories are reinterpreted within the context of this new paradigm.
While the central notion of Dung’s theory is the acceptability or non-
acceptability of arguments based on attack and defense, Dung shows that
nonmonotonic logic is a special form of argumentation (more details in Sec-
tion 3.2). It can be visualized as the commutative diagram in Figure 3. There
are two approaches to deriving conclusions from a knowledge base. The first
is a direct approach where a given logic selects a set of rules with conclusions.
 Thirteen Challenges in Formal and Computational Argumentation 25

The other is an indirect approach through argumentation, as shown in Figure 3

(2—4). Structured argumentation studies the process that adds the structure
that turns collections of rules into arguments and assigns attack relations (2)
among arguments. This gives us abstract argumentation frameworks — di-
rected graphs where nodes represent arguments, and arrows represent attack
relations. Then argumentation semantics (3) determine the acceptance status
of arguments and their conclusions. To represent a given logic by structured
argumentation, eventually the conclusions from both approaches must be the
same.

(1) logic conclusion

knowledge base
extension
(2 ) attack (4 ) conclusion
assignment extraction
argumentation (3) arg. semantics accepted
framework arguments

Figure 3: Commutative diagram: two approaches to nonmonotonic inference:

(1) logic systems; (2)–(4) argumentation systems. With appropriate choices on
elements (2) and (3), one can obtain exactly the same conclusions as for the
given logic (1).

We now present informally the construction of arguments and attack rela-

tions from a knowledge base in the ASPIC+ structured argumentation frame-
works [Prakken, 2010a]. A knowledge base typically consists of a set of strict
rules (with a simple arrow →) and a finite set of defeasible rules (using a
double-lined arrow ⇒), where each defeasible rule is assigned a priority num-
n
ber, denoted as a ⇒ b. The knowledge base also includes a base of evidence
(BE ). An argument can be constructed as follows:

1. For each element α ∈ BE, the expression [α] constitutes an argument

having the conclusion α.

2. Let r be a rule of the form α1 , . . . , αn → / ⇒ α, where A1 , . . . , An are

arguments with conclusions αi (for 1 ≤ i ≤ n). In this case, the expression
[A1 , . . . , An → / ⇒ α] is regarded as an argument with conclusion α.

Each argument is derived by applying the steps above (1 and 2) finitely many
times to ensure a structured process for argumentation within the framework.
We now use Example 3.1 to illustrate the commutative diagram, and we
explain the technical details later in section 3.2.
Example 3.1 (Two approaches to nonmonotonic reasoning). Consider
n
a knowledge base containing three defeasible rules ⇒ as well as facts ({⊤}),
as in Figure 4(a). Logical approaches to defeasible reasoning select a subset of
26 Liuwen Yu, Leendert van der Torre, and Réka Markovich

rules whose conclusions are maximally consistent. For example, PDL [Brewka
and Eiter, 1999], discussed in Section 2.4, selects the strongest applicable rules,
i.e., the order (i) → (ii) in Figure 4(a), with output {a, ¬b}. While (iii) is now
made applicable by a, its consequent b conflicts with ¬b and cannot be selected.
Argumentation approaches, in turn, build explicit arguments (Figure 4(b))
and represent these conflicts (b, ¬b) as attacks between arguments (B and C).
Observe how the arguments in Figure 4(b) activate the rules in Figure 4(a). To
specifically capture PDL, one needs a selection of attacks (discussed in Section
3.2), such as the attack induced by the weakest link in Figure 4(c), which
defines that the strength of an argument reflects its weakest rules. Intuitively,
the jointly acceptable arguments here are {A, C}, which corresponds to the
PDL extension {a, ¬b}. But in Figure 4(c), the last link, which defines that the
strength of an argument is that of its last rule, selects the arguments {A, B}
with output {a, b}.

C
{⊤} z }| {
⊤ ⇒2 ¬b C • 2
C • 2
< >
1
 
 ⊤⇒a A 3 1
 
 (ii) z }| {
2
⊤ ⇒ ¬b (i) ⊤ 1
a ⇒3 }b A• •B A• • B
| ⇒ {z
3 (iii)
 
a ⇒b
 
B last link weakest link
(a) knowledge base (b) arguments (c) attack assignments; extensions

Figure 4: The PDL approach to nonmonotonic reasoning. (a) PDL iteratively

selects the strongest active rule up to the point of inconsistency. (b) Arguments
build upon facts (A, C) or other arguments (B). (c) Argumentation semantics
abstract from a logical structure. Attacks depend on argument strength.

While nonmonotonic logic provides a structured framework for managing

conflicts and reasoning through premises, inferences, and conclusions, it be-
comes difficult to capture complex conflicts among a large number of arguments.
Here, abstract argumentation frameworks can provide landscapes of how these
arguments relate to one another via attack relations. One of the main goals of
abstract argumentation theory is to answer the question: which sets of argu-
ments can be reasonably accepted in a discussion based on a given argumenta-
tion framework? In the simple argumentation frameworks of Figure 4(c), sets
of arguments could be selected by intuition. But a formal method is needed for
a more complex graph, for instance where attacks among arguments can form
even or odd cycles, which may be part of more complex structures like strongly
connected components (SCCs). In abstract argumentation, argumentation se-
mantics provides a way to deal with these complications.
We use graph labeling based on so-called gunfight rules [Caminada, 2006;
Caminada and Gabbay, 2009] to determine which arguments can survive in
 Thirteen Challenges in Formal and Computational Argumentation 27

conflicts. The concept is straightforward: in a gunfight, one stays alive iff all
attackers are dead, and one dies iff at least one attacker is still alive. Under-
standing this analogy essentially captures the core idea of abstract argumenta-
tion:

1. An argument is labeled in iff all its attackers are labeled out;

2. An argument is labeled out iff it has at least one attacker that is labeled
in;

Example 3.2 (Argumentation framework with two cycles). Consider

the argumentation framework in Figure 5, which has a set of arguments:
{a, b, c, d, e}. We follow the direction of the graph. On the left, we have an even
cycle: a and b attack each other. On the right, we have an odd cycle: c attacks
d, d attacks e, and e attacks c. The two cycles are connected by the attack
from b to c, thus the status of the arguments in the even cycle will influence
the status of the arguments in the odd cycle. In the left cycle, there are two
possibilities. In the first case, a is labeled in, then b is labeled out. However,
there is no way to label the arguments in the odd cycle on the right. Thus, we
need a third label called undec (undecided), indicating that one abstains from
an explicit judgment whether the argument is in or out. It means that not all
the attackers are labeled out and no attackers are labeled in. Therefore, c, d, e
will be labeled undec. In the second case, b is in, a is out, and c is out. Then d
is reinstated as an in argument because its only attacker c is out; we can also
say b defends d. It follows that e is out because d is now in. We can label a
and b with the third label undec, and all the arguments in the odd cycle are
also labeled undec.

a b c

Figure 5: Argumentation framework with two cycles

Now the question is: which labelings in Dung’s theory are called extensions?
We illustrate the extensions to the framework in Example 3.2 below. We use the
thick nodes for in, normal nodes for out, and dotted nodes for undec, to obtain
a visualization similar to a colored graph. We say that if a labeling is three-
valued, then it is a complete extension, as in the first item below. A complete
extension generalizes a stable extension and there is no argument labeled undec,
i.e., there is no dotted node, as in the second item below. The unique grounded
28 Liuwen Yu, Leendert van der Torre, and Réka Markovich

extension is the most skeptical complete extension; only arguments that cannot
avoid being accepted are labeled in, as in the third item. For some frameworks,
there are no stable extensions. Then we can use preferred extensions, which
are the maximal complete extensions, as in the fourth item.

• Three complete extensions (3-valued)

• One stable extension (2-valued)

• One grounded extension (minimal complete)

• Two preferred extensions (maximal complete)

While there have been many transformations of nonmonotonic reasoning

formalisms into Dung’s theory, direct usage is limited when modeling the ar-
gumentation of some realistic examples [Boella et al., 2009] such as multiagent
argumentation and dialogues [Yu et al., 2021; Arisaka et al., 2022], decision-
making [Kakas and Moraitis, 2003], coalition formation [Amgoud, 2005], com-
bining micro arguments [Toulmin, 1958], normative reasoning [Atkinson and
Bench-Capon, 2005], or meta-argumentation [Boella et al., 2009]. That leads
to the following challenge.
Challenge 5. Creating argumentation frameworks and semantics.
Several extensions to abstract argumentation frameworks are discussed in
the second volume of the Handbook of Formal Argumentation. Figure 6 visu-
alizes six examples of extended argumentation frameworks. Preference-based
argumentation [Kaci et al., 2021] introduces a preference relation between the
arguments, as shown in Figure 6(a), where a defeats b, and b is preferred over a.
Bipolar argumentation [Cayrol and Lagasquie-Schiex, 2005] defines support and
attack independently. There are arguments in favor of other arguments, i.e.,
with a support relation, and also arguments against other arguments, i.e., with
an attack relation, as shown in Figure 6(b). Weighted argumentation [Bistarelli
 Thirteen Challenges in Formal and Computational Argumentation 29

et al., 2021] specifies a numeric value that indicates the relative strength of an
attack, as shown in Figure 6(c). Abstract agent argumentation [Yu et al.,
2021] extends Dung’s framework with a set of agents and a relation associat-
ing arguments with agents, as shown in Figure 6(d). Value-based argumenta-
tion [Atkinson and Bench-Capon, 2021], as shown in Figure 6(e), defines values
that are associated with an argument. The preference ordering of the values
may depend on a specific audience. To model defeat for a specific audience: an
argument A attacks an argument B for audience a if A attacks B and the value
associated with B is not preferred to the value associated with A for audience
a. Higher (second)-order argumentation [Cayrol et al., 2021] introduces a new
kind of attack which is a binary relation from arguments to attack relations,
as shown in Figure 6(f).
One technique that has already proven to be useful in the past for studying
such extensions is a meta-argumentation methodology involving the notion of
flattening [Boella et al., 2009]. Flattening is a function that maps some ex-
tended argumentation frameworks into Dung frameworks. There are two main
flattening techniques. One is that we keep the arguments the same while re-
moving attacks or introducing auxiliary attacks (this is also called reductions
sometimes). This technique is used in preference-based argumentation, ab-
stract agent argumentation, bipolar argumentation, etc. The other technique
is to use not only auxiliary attacks but also auxiliary arguments in higher-order
argumentation.
Example 3.3 (Four reductions of preference-based argumentation).
Figure 7 illustrates the differences between the four reductions from a
preference-based argumentation framework to abstract argumentation frame-
works [Kaci et al., 2021]. The basic idea of Reduction 1 is that an attack
succeeds only when the attacked argument is not preferred to the attacker.
Reduction 2 enforces that one argument defeats another when the former is
preferred but attacked by the latter. The idea of Reduction 3 is that if an
argument is attacked by a less preferred argument, then the former should de-
fend itself against its attacker. Reduction 4 mixes the second and the third
reductions.
Flattening by adding auxiliary arguments is a way of implementing the
methodology of meta-argumentation [Boella et al., 2009]. Meta-argumentation
generally involves taking into account the arguments of, e.g., lawyers, com-
mentators, citizens, teachers, or parents (in accordance with the level of their
expertise) but it can also go beyond this — the arguers and the meta-arguers
can be represented by the same reasoners. For example, a lawyer may debate
whether a suspect’s argument attacks another argument, and she may also
argue in a similar way about her own arguments. To give another example,
people may be in the middle of an argument, but then start questioning the
rules of the dialogue game, and argue about that. A further example is that of
a child arguing that the argument I was ill attacks the argument I have to do
my homework but then finds that the argument I have a nice tan attacks the
30 Liuwen Yu, Leendert van der Torre, and Réka Markovich

a b
a

b b≻a c

(a) Preference-based argumentation (b) Bipolar argumentation

⟨Ar, →, ⪰⟩, where Ar is a set of argu- < Ar, →, 99K>, where Ar is a set of ar-
ments, → is an attack over arguments, guments, → is an attack over arguments,
and ⪰ is a preference relation over Ar. and 99K is support for arguments.
8 α β γ
8 5 6
a b c d
a b c
9
(c) Weighted argumentation (d) Abstract agent argumentation
⟨Ar, →, w⟩, where Ar is a set of argu- ⟨Ar, →, S, <⟩, where Ar is a set of argu-
ments, → is an attack over arguments, ments, → is an attack over arguments, S
and w is a function from Ar to [0, 1]. is a set of agents, and < is a binary rela-
tion associating arguments with agents.

v1 |v2 v2 >v1

A B
v1 v2

a b

v1 >v2
c

(e) Value-based extended argumen- (f) Higher-order argumentation

tation ⟨Ar, →, ↠⟩, where Ar is a set of argu-
⟨Ar, →, V, val, P ⟩, where Ar is a set of ar- ments, → is an attack over arguments,
guments, → is an attack over arguments, and ↠ is an attack relation over argu-
V is a set of values, val is a function map- ments that can be attacked.
ping arguments to values, and P is a set
of audiences.

Figure 6: Six extensions to the argumentation framework

 Thirteen Challenges in Formal and Computational Argumentation 31

a a a a a

b b b b b
b≻a

Figure 7: From left to right: the original argumentation framework, and the
results after applying the four reductions respectively. The defeat relation
is visualized with thick lines, and arguments that are accepted in grounded
semantics also have thick lines.

argument I was ill.

When we flatten the extended framework, if an argument a of the extended
argumentation framework also occurs in the flattened abstract argumentation
framework, then we no longer refer to it as argument a but as the meta-
argument “argument a is accepted”, denoted as accept(a). We use Example 3.4
to illustrate instantiating Dung’s abstract argumentation framework by intro-
ducing meta-arguments that use flattening.
Example 3.4 (Flattening with auxiliary arguments [Boella et al.,
2009]). Given the higher-order argumentation framework in Figure 6(f),
the flattened framework is as illustrated in Figure 8. We introduce the
meta-arguments Ya,b , which means that a is capable of attacking b, and
Xa,b . which means that a does not have the capability of attacking b. We
use the meta-arguments in the following way. Each a → b is replaced by
accept(a) −→ Xa,b −→ Ya,b −→ accept(b). The accepted arguments are
{accept(a), accept(c), Yc,Ya,b , accept(b)}.

accept(a) Xa,b Y a, b accept(b)

Yc,Ya,b Xc,Ya,b accept(c)

Figure 8: Flattened argumentation framework for Figure 6(f)

These examples illustrate how diverse extended argumentation frameworks

build upon Dung’s foundational concepts by introducing additional elements.
However, the key reasoning ontology fundamentally relies on just two elements:
arguments and attacks. Every utterance, be it a claim, an argument, or even an
32 Liuwen Yu, Leendert van der Torre, and Réka Markovich

attack, can be modeled as an argument within these frameworks. Consequently,

many of these extended frameworks can be flattened to basic ones, reinforcing
the idea that attack graphs serve as a universal model of reasoning — much
like how Turing machines serve as a universal model of computation.
In this section, we discussed the attack-defense paradigm shift introduced
by Dung, emphasizing that every type of utterance (claim, argument, or at-
tack) can be attacked. We also presented how diverse extended argumentation
frameworks can be flattened into basic ones, demonstrating the universality of
attack. We end this section with the following questions:
1. What is an argument? What is an attack? What is the interplay between
an argument and an attack?
2. Should attack always be the first-class citizen in formal argumentation?
For example, a novel notion of attack-defense is adopted as a first-class
citizen by Liao and van der Torre [2024]. It can represent some knowl-
edge that cannot be represented in Dung-style argumentation, e.g., some
context-sensitive knowledge in a dialogue.
3. How should the new diversity created by the attack-defense perspective
be handled?
4. What does the attack-defense paradigm shift mean for argumentation as
dialogue? What does the attack-defense paradigm shift mean for argu-
mentation as balancing?
5. If we flatten an extended argumentation framework we introduce auxil-
iary arguments. How can we then recognize these auxiliary arguments in
the instantiated Dung abstract argumentation framework? How can we
deal with the original arguments and the arguments introduced later?
3.2 Representing nonmonotonic logics
In this section, we discuss: structured argumentation as a bridge from classi-
cal to nonmonotonic logic, the variety of nonmonotonic logics available, and
the represention of nonmonotonic logic. We again refer to the commutative
diagram in Figure 3 that we used in section 3.1 to illustrate how the same
conclusions can be reached with two different approaches: the direct approach
using logic and the indirect approach through the construction of argumen-
tation frameworks, and semantics. Here, we illustrate this using the weakest
versus last link principle and we continue with PDL and the weakest link.
Despite the uniqueness of classical logic, a wide variety of nonmonotonic
logics are employed in different contexts. However, engaging in nonmonotonic
logics means the aim is to extend classical logic rather than replace it tout
court [Makinson, 2005]. Structured argumentation is used to classify exist-
ing nonmonotonic logics as a way to define a new nonmonotonic logic and
create a bridge from classical to nonmonotonic logic. Dung [1995] provides
semantics of attacks for structured argumentation. This has been used in the
 Thirteen Challenges in Formal and Computational Argumentation 33

ASPIC+ system by Modgil and Prakken [2013; 2014], and it has also been
used to reconstruct and compare a variety of nonmonotonic logics, namely de-
fault logic [Reiter, 1980], Pollock’s [1987] argumentation system, and several
logic-programming semantics. More representations have been developed —
for details, please refer to the work of Heyninck [2019]. However, as discussed
in Section 2, there is also a diversity of natural argumentation, conceptual-
izations, and formal methods. Notwithstanding the initial appeal of Dung’s
abstract argumentation theory, there are many different kinds of argumenta-
tion frameworks and semantics. That leads to the following challenge:
Challenge 6. Representing nonmonotonic logics and solution concepts.
Before we get into approximating PDL with argumentation, let us first talk
about methodologies employed to compare different nonmonotonic logics and,
in particular, their use of examples. There are different approaches to the use
of examples in different disciplines. In law, ethics, and linguistics, examples
are central to the development and validation of theories because they help
ground abstract concepts in real-world scenarios, which helps to align logical
frameworks with intuitive understanding. In contrast, knowledge representa-
tion (KR) and other areas of computer science often use examples as a practical
tool to test, demonstrate, and communicate the effectiveness of a formal theory
rather than using them as foundational elements in theory construction.

NLP task: translating natural language into formal language. Consider the
aforementioned example of the fitness-lover Scot and an additional ex-
ample about a snoring professor:
The fitness-lover Scot: It is commonly assumed that if a man was
born in Scotland, then he is Scottish. And if he is Scottish, we
can normally deduce that he likes whiskey. However, fitness lovers
normally avoid alcohol for health reasons. Stewart was born in Scot-
land, and he is also a fitness lover. Does he like whiskey?
The snoring professor A library has a general rule that misbehavior,
such as snoring loudly, leads to denial of access. However, there
is another rule that professors are normally allowed access. Bob is
a professor and he is snoring loudly in the library. Should he be
allowed access to the library?
NLP could be used to identify three rules for each example, and
then further abstract them into these formal (default) rules with
1 3 2
priorities: {⊤ ⇒ a, a ⇒ b, ⊤ ⇒ ¬b}. We then have:

Fitness-lover Scot Snoring professor in the library

1 1
   
 BornInScotland ⇒ Scottish
 
  snores ⇒ misbehaves
 

3 3
Scottish ⇒ LikesWhisky misbehaves ⇒ accessDenied
2 2
   
FitnessLover ⇒ ¬LikesWhisky professor ⇒ ¬accessDenied
   
34 Liuwen Yu, Leendert van der Torre, and Réka Markovich

KR task: after inputting some requirements, i.e., the goal of the reasoning,
the system asks what you want to derive from what you have. Although
the above two examples share a similar structure, there could be differ-
ent reasoning requirements that lead to the selection of different rules
and ultimately different conclusions. In the fitness-lover example, one
2
might prioritize the rule ⊤ ⇒ ¬b and conclude that Stewart does not
like whiskey. In contrast, in the snoring professor example, one might
3
prioritize the rule a ⇒ b and conclude that Bob should be denied access.

Logic design task: According to these requirements, the system asks what is
the best logic for your application. These two examples have been used to
illustrate the difference between prescriptive and descriptive reasoning in
nonmonotonic reasoning and between the weakest link and the last link,
which are two principles regarding how an argument draws strength from
its defaults.

In Section 2.4 and Section 3.1, we briefly mentioned Brewka and

Eiter’s [1999] PDL. Pardo et al. [2024] compared structured argumentation
based on the weakest link variant with that of the PDL variant. Let us start
with a reminder that PDL can be understood as a greedy approach, i.e., PDL
iteratively adds the strongest applicable and consistent default. Initially, peo-
ple thought that using the weakest link principle to construct argumentation
frameworks would capture this kind of greedy procedure. However, over time,
analysis of the weakest-link-related attack assignment reveals that it is more
complicated and ambiguous than it appears at first sight.
The history of the weakest link revolves around three key examples from
the literature, visualized in Figure 9 and described in Examples 3.5–3.7. Note
that these examples illustrate the role of formal argumentation in the context
of PDL. We refer to the work of Pardo et al. [2024] for the formal definitions.
Here, we discuss Examples 3.5–3.7 informally.
The following example illustrates the use of priorities. What does a stronger
priority mean? Under the prescriptive reading, it means priority in the order
of application: PDL always selects the strongest default (among those that
are applicable and consistent). Under the descriptive reading, the priority of a
default is its contribution to the overall status of any extension containing this
default [Delgrande et al., 2004]. The two readings clash in the most discussed
example of defeasible reasoning with prioritized rules.
1 3 2
Example 3.5 (Weakest vs. Last link). Let {⊤ ⇒ a, a ⇒ b, ⊤ ⇒ ¬b}
be again our defaults (Figure 9, top). The two readings of priorities give the
following outputs:
2 1
(Prescriptive.) Based on application order, one must select {⊤ ⇒ ¬b, ⊤ ⇒ a}
thereby obtaining the output {a, ¬b}, as in PDL. In fact, PDL is an implemen-
tation of the prescriptive reading. Let us call simple weakest link (swl) the
strength defined by the lowest priority of an argument:
 Thirteen Challenges in Formal and Computational Argumentation 35

swl−attack dwl−attack pdl-attack PDL

1 1 3
⊤⇒a ⊤⇒a⇒b
Ex. 3.5 {a, ¬b}
2
⊤ ⇒ ¬b
1 1 3 1 1 3
⊤⇒a ⊤⇒a⇒b ⊤⇒a ⊤⇒a⇒b
Ex. 3.6 {a, b}

1 1 2 1 1 2
⊤⇒a ⊤ ⇒ a ⇒ ¬b ⊤⇒a ⊤ ⇒ a ⇒ ¬b
1 1 2 1 1 2
⊤⇒a ⊤ ⇒ a ⇒ ¬b ⊤⇒a ⊤ ⇒ a ⇒ ¬b
{a, ¬b}
Ex. 3.7
{b, ¬a}
1 1 2 1 1 2
⊤⇒b ⊤ ⇒ b ⇒ ¬a ⊤⇒b ⊤ ⇒ b ⇒ ¬a

Figure 9: Approximating PDL in structured argumentation: a comparison of

three attacks (columns) for three examples (rows). Columns are not marked
when adjacent notions of attack agree on the induced attack relation at a
given row. Dotted rectangles are argument extensions. The rightmost attacks
approximate PDL better.

1 3 2
⊤ ⇒ a ⇒ b 7−→ 1 = min{1, 3} ⊤ ⇒ ¬b 7−→ 2 = min{2}
A comparison of the strengths in this conflict produces the attack shown in
Figure 9 (top). The semantics then gives the argument selection also shown.
Our three attack relations (swl, dwl, pdl) do in fact agree on the verdict for this
example.1
1 3
(Descriptive.) This reading favours the set {⊤ ⇒ a, a ⇒ b} as its priorities
{1, 3} are more desirable than the rival ones {1, 2}. Last link can be seen as an
implementation of this reading: the contribution of a new default to a selection
1
or argument, say {⊤ ⇒ a}, is defined by the desirability of this default (2 vs. 3
in the example). Last link thus agrees on the above preference but arrives at
it through argumentative means. First, one computes argument strength:
1 3 2
⊤ ⇒ a ⇒ b 7−→ 3 = last(1, 3) ⊤ ⇒ ¬b 7−→ 2 = last(2)
1 3 2
Based on this, argument ⊤ ⇒ a ⇒ b attacks ⊤ ⇒ ¬b. Using a standard
argumentation semantics, one obtains the output {a, b}, not shown in Figure 9
(top).
1 This example represents the Tweety scenario {penguin → bird, bird ⇒ flies,
penguin ⇒ ¬flies} with priorities instead of the strict rule (→). Without priorities, the
solution {penguin, bird, ¬flies} obtains from specificity (of penguin over bird): birds fly is
overruled by the more specific penguins do not fly. Without specificity the solution obtains
from appropriate priorities using PDL or swl.
36 Liuwen Yu, Leendert van der Torre, and Réka Markovich

The simple weakest link, though, does not always capture the prescriptive
reading. In response to this, a more intuitive disjoint variant of the weakest
link has been considered [Young et al., 2016]. This variant assumes a rela-
tional measure of argument strength. It ignores all the shared defaults before
searching for the weakest link between two arguments.
1 3
Example 3.6 (Simple vs. disjoint weakest link). Let {⊤ ⇒ a, a ⇒ b,
2 1 3
a ⇒ ¬b} define our knowledge base. Note that the two arguments ⊤ ⇒ a ⇒ b
1 2 1
and ⊤ ⇒ a ⇒ ¬b share a default ⊤ ⇒ a with the lowest priority. See the
middle row in Figure 9.

(Simple weakest link) Pollock’s definition assigns the same strength of 1 to

these two arguments. This gives the mutual swl-attack in Figure 9 (mid,
1 3
left). Now, one argument selection ⊤ ⇒ a ⇒ b matches the PDL exten-
1 2
sion {a, b}; the other ⊤ ⇒ a ⇒ ¬b, though, gives us a non-PDL extension,
{a, ¬b}.

(Disjoint weakest link) The attack relation defined by disjoint weakest link
(dwl) assigns strengths 3 > 2 to the above arguments, after excluding
the default they share. This generates the tie-breaking dwl-attack shown
in Figure 9 (mid, right). This figure also shows the set of arguments se-
lected by our semantics. The selected arguments’ conclusions match the
PDL output {a, b}.

Pollock’s definition of weakest link swl [Pollock, 2001] was adopted and
studied for ASPIC+ by Modgil and Prakken [2013; 2014]. Then, Young et
al. [2016; 2017] introduced dwl and proved that argument extensions under the
dwl-attack relation correspond to PDL extensions under total orders; see also
the results presented by Liao et al. [2019] and Pardo and Straßer [2022]. Under
total preorders, a new attack relation is needed for more intuitive outputs and
a better approximation of PDL — that is, better than dwl.
1 1 2 2
Example 3.7 (Beyond dwl). Let {⊤ ⇒ a, ⊤ ⇒ b, a ⇒ ¬b, b ⇒ ¬a} be the
defaults.

(swl, dwl) Weakest link attacks, depicted in Figure 9 (bottom, left), admit
1 1
the selection of arguments {⊤ ⇒ a, ⊤ ⇒ b}. This selection fits neither
the prescriptive interpretation nor PDL. Selecting either default ought to
2
be followed by the selection of a stronger default, namely a ⇒ ¬b and
2
b ⇒ ¬a respectively.

(PDL) As PDL selects the strongest default one at a time, this excludes by
1 1
construction the concurrent selection of {⊤ ⇒ a, ⊤ ⇒ b}. The PDL-
inspired attack relation in Figure 9 (bottom, right) also excludes this
selection.
 Thirteen Challenges in Formal and Computational Argumentation 37

An important research question is then how to characterize, or at least ap-

proximate, the PDL extensions of a prioritized default theory. For total orders,
an attack that characterizes PDL extensions already exists: attdwl [Young et
al., 2016].
But for total preorders, how to characterize PDL extensions using an attack
relation assignment is an open problem. Certainly, such a characterization can
no longer be based on the disjoint weakest link, as shown in Example 3.8.
Example 3.8 (Disjoint weakest link vs. PDL). Example 3.7 shows a
stable belief set {a, b} under attdwl (K) that is not a PDL extension of K.
1 2 1
Example 3.9 (PDL vs. Disjoint weakest link). Let {⊤ ⇒ a, a ⇒ b, ⊤ ⇒
2
c, a, c ⇒ ¬b} define our knowledge base. Figure 10 shows that the shared rule
1
⊤ ⇒ a produces only one stable extension E under the disjoint weakest link,
and so we have a unique stable belief set of (ARK , attdwl (K)):

E = {A, C, [A ⇒ b]} 7−→ S = {a, b, c}

In contrast, two PDL constructions exist for K, and so do two PDL extensions:
1 2 1
(⊤ ⇒ a, a ⇒ b, ⊤ ⇒ c) 7 → {a, b, c}
−
1 1 2
(⊤ ⇒ c, ⊤ ⇒ a, a, c ⇒ ¬b) −7 → {a, ¬b, c}
As a consequence, disjoint weakest link cannot characterize PDL under stable
semantics. Observe that attswl here coincides with PDL.

A 1 1 2
⊤⇒a ⊤⇒a⇒b
1
C ⊤⇒c

2
A, C ⇒ ¬b

Figure 10: The stable belief set {a, c, b} under attdwl for Example 3.9. Two
extensions, {a, c, b} and {a, c, ¬b}, exist under PDL.

Attack relations have become a major subject of study in logic-based ar-

gumentation. Dung [2014; 2016] recently proposed an axiomatic method that
supersedes all argumentation systems with defeasible rules. Pardo et al. [2024]
attempted to identify an attack relation that captures PDL extensions, and
they compared it with attacks based on the simple and disjoint weakest link
using the eight principles advanced by Dung. They proved an impossibility
theorem: representing PDL in formal argumentation should preserve a princi-
ple (attack closure), but this is incompatible with another principle (context
independence).
38 Liuwen Yu, Leendert van der Torre, and Réka Markovich

1 1 2
⊤⇒a ⊤ ⇒ a ⇒ ¬b

1 1 2
⊤⇒b ⊤ ⇒ b ⇒ ¬a

Figure 11: pPDL differs from PDL. PDL has two extensions {a, ¬b} and
{b, ¬a}. pPDL has an additional extension {a, b}. Arrows denote logical con-
flicts.

As seen in Examples 3.7 and 3.9, disjoint weakest link and PDL are incom-
parable under total preorders. As a first step towards their convergence, one
can slightly modify PDL to make it closer to the disjoint weakest link. To this
end, Pardo et al. [2024] propose parallel PDL (pPDL), a concurrent variant of
PDL. The main novelty of pPDL is that each inductive step can concurrently
select a set of defaults, rather than just one, for the technical details, we refer
to the paper of Pardo et al. [2024].
Example 3.10 (pPDL, DWL vs. PDL). Let us use Example 3.7 to show
that the default logic PDL differs from pPDL. Figure 11 illustrates the three
pPDL extensions {a, ¬b}, {b, ¬a}, {a, b}, of which {a, b} is not a PDL extension.
Although pPDL and attdwl agree in this and other examples, pPDL does
not always match the disjoint weakest link.
Example 3.11 (pPDL vs. DWL). Example 3.9 showed a unique stable
belief set, {a, b, c}, under attdwl . But there are two pPDL extensions: {a, b, c}
and {a, ¬b, c}.
To sum up, the first goal of Pardo et al. [2024] was to identify an attack
relation that captures PDL extensions and compare it with attacks based
on the simple and disjoint weakest link using the eight principles advanced
by Dung [2016; 2018]. They proved which principles for attack relations are
satisfied by weakest link, disjoint weakest link and PDL based attacks. Their
principle-based analysis presented the difference between several kinds of attack
relation assignment. They identified and explained the nature of the weakest
link principle and revealed that there is still the potential to improve the weak-
est link attack. On this last question, they proposed pPDL (parallel PDL), a
concurrent variant of PDL, and they showed by way of examples that it falls
closer to the disjoint weakest link than PDL does. While the pPDL variant
still does not match the disjoint weakest link, one might conjecture that some
further refinement might do.
In addition to presenting the argumentation framework, Dung [1995] also
investigated two examples of problems from microeconomics — cooperative
game theory and matching theory. In each case, Dung showed how an appro-
 Thirteen Challenges in Formal and Computational Argumentation 39

priate framework can represent a given cooperative game or a given instance of

the stable marriage problem, and that the sets of winning arguments in such
argumentation frameworks correspond to meaningful solutions in both these
domains.
Cooperative game theory studies how rational agents cooperate to form
coalitions to maximize their payoffs. A coalition’s payoff is measured by its
value, and agents ideally cooperate to form a coalition. The Von Neumann-
Morgenstern (vNM) stable set [Von Neumann and Morgenstern, 1947] is the
solution concept for distributing the grand coalition’s payoff and ensuring that
no agent defects. Dung showed that stable extensions of an argumentation
framework correspond to vNM stable sets [Dung, 1995, Theorem 37]. How-
ever, just like the stable extensions of an argumentation framework may not
exist, vNM stable sets also do not always exist. Dung proposed that sets of
payoff distributions that form preferred extensions could serve as an alternative
solution concept because preferred extensions always exist, and therefore this
is well defined for all cooperative games.
In this section, we discussed the representation of nonmonotonic logics using
the attack-defense paradigm. We end with several questions concerning such
representations:

1. We showed that PDL and the weakest link definitions are similar but
not exactly the same. How can PDL be changed to make it fit one of
the weakest link definitions? How can the weakest link be changed to fit
PDL?

2. We discussed the logic of the weakest link. What is the logic that corre-
sponds to the last link?

3. We showed various alternative formalizations of the weakest link princi-

ple. Likewise, are there variants of the last link principle?

4. PDL is only one of many logics for prioritized rules. How can all the
other systems for prioritized rules be represented?

5. We discussed representation of nonmonotonic logics, but Dung also talked

about logic programming and game theory. What is the relation between
different solution concepts in game theory and (extended) abstract argu-
mentation frameworks semantics?

3.3 Postulates from paraconsistent reasoning

In this section, we continue our discussion of formal argumentation as a log-
ical framework for nonmonotonic reasoning. We consider inconsistent knowl-
edge bases and so-called rationality postulates from paraconsistent reasoning,
which is used to define new nonmonotonic logics in ASPIC+. We illustrate the
postulates using the example of three persons on a two-person tandem taken
from Caminada and Wu [2011].
40 Liuwen Yu, Leendert van der Torre, and Réka Markovich

In previous sections, we discussed universality of attack and one resultant

challenge — representing existing nonmonotonic logic. In particular, we dis-
cussed representing PDL with structured argumentation, comparing attack as-
signments using variants of the weakest link with principles. What we showed
is just the tip of the iceberg. There are numerous options based on different
knowledge bases containing various types of information such as strict and/or
defeasible rules. There are different methods for constructing argumentation
frameworks. There are applications of distinct semantics. The combination
of all these factors defines different argumentation-based logics that can be
adopted or rejected, depending on their applicability in different contexts.
Rationality postulates are a list of desiderata that structured argumentation
systems should satisfy in order to be logically well-behaved [Caminada and
Amgoud, 2007]. In this section, we address the following challenge:
Challenge 7. Rationality postulates for defining a new logic.
Various rationality postulates are inspired by paraconsistent reasoning.
Paraconsistent logic [Da Costa, 1974; Priest, 2002] is a non-classical logical
system designed to handle contradictions without leading to the collapse (or
“explosion”) of the entire systems (as would occur in classical logic). These
logics have inspired the development of modal and nonmonotonic logics as well
as various rationality postulates [Da Costa et al., 2007]. Such postulates ensure
that logic can handle inconsistencies without leading to the kind of trivializa-
tion where any and every conclusion becomes derivable from a set of contra-
dictory premises. One key postulate of paraconsistent logic is noninterference,
i.e., independent knowledge bases do not influence each other’s outcomes. An-
other is avoiding contamination, i.e., the outcome of a set of formulas remains
unchanged when merged with an unrelated set [Caminada et al., 2012].
A side note regarding terminology: we use terms such as postulates, axioms,
requirements and desiderata in a rather interchangeable manner, and they dif-
fer slightly from principles and properties. All six terms refer to the behavior
of logic, the construction of an argumentation framework, and the semantics
of argumentation frameworks. Abstract properties are formally specified, and
in this section, postulates are treated as desiderata, akin to formal require-
ments in computer science. In Section 4.1, where we discuss the principle-based
methodology in detail, postulates are regarded as more general properties, with
some being desirable and others not.
There are three fundamental rationality postulates [Caminada and Amgoud,
2007]. Direct Consistency means that any extension should be consistent ac-
cording to certain semantics. Indirect Consistency means that the set of the
conclusions of arguments in a given extension is consistent when closed under
the strict rule. Closure means that arguments with conclusions derived from
arguments in an extension using strict rules should also be in the extension.
Given these postulates, the question is under what conditions do structured
argumentation satisfy them. When assigning attack relations among argu-
ments from a knowledge base, there are so-called rebuts when the conclusions
 Thirteen Challenges in Formal and Computational Argumentation 41

of two arguments conflict with one other. Two kinds of rebuts have been
discussed in the literature: restricted rebuts and unrestricted rebuts. The in-
tuition behind restricted rebut is: if an argument is built up with only strict
rules, then the conclusion should also be strict, and the argument cannot be
attacked. The intuition behind unrestricted rebut is that a conclusion is de-
feasible, i.e., it can be attacked iff it is built up with at least one defeasible
rule. Different choices on rebuts influence how to define the argumentation for-
malism that derives reasonable conclusions. This exists in the ASPIC family
of argumentation frameworks, including ASPIC+ [Modgil and Prakken, 2013;
Modgil and Prakken, 2014], ASPIC- [Caminada et al., 2014] and ASPIC-
END [Dauphin and Cramer, 2018].
Example 3.12 illustrates rationality postulates, comparing unrestricted rebut
and restricted rebut, and it shows the solutions to restricted rebut required to
satisfy the rational postulates for this example.
Example 3.12 (Married John [Caminada and Amgoud, 2007]). Con-
sider an argumentation system consisting of the strict rules {→ r, → n, m →
hs, b → ¬hs} and the two defeasible rules {r ⇒ m, n ⇒ b}. An intuitive in-
terpretation of this example is the following: “John wears a ring (r) on his
finger. John is also a regular nightclubber (n). Someone who wears a ring
on his finger is usually married (m). Someone who is a regular nightclubber is
usually a bachelor (b). Someone who is married has a spouse (hs) by definition.
Someone who is bachelor does not have a spouse (¬hs) by definition.” We can
construct the following arguments:

A1 :→ r A3 : A1 ⇒ m A5 : A3 → hs
A2 :→ n A4 : A2 ⇒ b : A6 : A4 → ¬hs

If we apply unrestricted rebut, we have A5 and A6 attacking each other, and

we obtain the grounded extension of {A1 , A2 , A3 , A4 } with the conclusion ex-
tension {r, n, m, b}, which does not satisfy the direct consistency property. If
we apply restricted rebut, the situation is even worse. Because we do not have
any attack relations, we have the extension {A1 , A2 , A3 , A4 , A5 ,A6 } and the
conclusion extension {r, n, m, b, hs, ¬hs}, which are not consistent.
Two solutions for argumentation systems applying restricted rebut to satisfy
the rationality postulates are closure of transposition and closure of contrapo-
sition, as adopted in ASPIC+ [Modgil and Prakken, 2014].
Example 3.13 (Example 3.12 continued). Given that we have m → hs
and b → ¬hs in the knowledge base, we add their “contraposed” versions:
¬hs → ¬m and hs → ¬b. We can construct additional arguments: A7 : A5 →
¬b and A8 : A6 → ¬m. We have that A7 restrictively rebuts A4 , and that
A8 restrictively rebuts A3 . As a result, each set of conclusions yielded under
grounded or preferred semantics satisfies the postulates of direct consistency,
closure, and indirect consistency.
42 Liuwen Yu, Leendert van der Torre, and Réka Markovich

We use the example of three persons on a two-person tandem [Caminada

and Amgoud, 2007] to take a closer look at unrestricted rebut and restricted
rebut — the latter is applied where unrestricted rebut can lead to undesired
behavior.
Example 3.14 (Restricted rebut vs. unrestricted rebut). Consider a
knowledge base consisting of three defeasible rules, {⊤ ⇒ p, ⊤ ⇒ q, ⊤ ⇒ r},
and three strict rules, {p, q → ¬r, p, r → ¬q, q, r → ¬p}. We can construct six
arguments as shown below. If we apply unrestricted rebut, we can obtain the
abstract argumentation framework on the left hand side of Figure 12. One of
the complete extensions is {A1 , A2 , A3 }, yielding conclusion extension {p, q, r}.
If we close this extension under strict rules, we have {p, q, r, ¬p, ¬q, ¬r}, which
is not consistent. If we apply restricted rebut, we obtain the framework at
the right hand side of Figure 12, where we have the complete extensions of
{A1 , A2 , A6 }, {A1 , A3 , A5 } and {A2 , A3 , A4 }. They are also consistent under
the closure of strict rules.

A1 : ⊤ ⇒ p A4 : q, r → ¬p
A2 : ⊤ ⇒ q A5 : p, r → ¬q
A3 : ⊤ ⇒ r A6 : p, q → ¬r

A1 A1

A4 A4

A5 A6 A5 A6

A2 A3 A2 A3

Figure 12: Resticted rebut vs. unrestricted rebut

There are more postulates. For example, noninterference and crash resis-
tance are particularly relevant when the strict rules are derived from classical
logic, and again we examine various ways of satisfying these properties. How-
ever, there have been comparatively fewer results that would establish them in
systems of the ASPIC family.
 Thirteen Challenges in Formal and Computational Argumentation 43

Noninterference: no set of formulas can influence the entailment of an unre-

lated set of formulas when they are merged with a completely unrelated
(syntactically disjoint) defeasible theory.

Crash resistance: no set of formulas can make an unrelated set of formulas

completely irrelevant when they are merged with a completely unrelated
(syntactically disjoint) defeasible theory.

A violation of non-interference means that a defeasible theory somehow in-

fluences the entailment of a completely unrelated (syntactically disjoint) de-
feasible theory when being merged with it. A violation of the crash resistance
property is more severe, as this means that a defeasible theory influences the
entailment of a completely unrelated (syntactically disjoint) defeasible theory
to such an extent that the actual content of this other defeasible theory becomes
irrelevant.
In this section, we discussed the use of postulates from paraconsistent rea-
soning in argumentation. We end with some open questions.

1. In structured argumentation, arguments can be attacked by either de-

feasible premises, defeasible inference rules, or the conclusion of defeasi-
ble rules. In assumption-based argumentation, there are only defeasible
premises, while ASPIC+ allows all defeasibilities. How should we decide
upon and clarify the various defeasibilities in structured argumentation?

2. Incorporating formal argumentation and social concepts has attracted

much interest. One example is the use of an argumentative approach to
normative reasoning [Dong et al., 2019; Pigozzi and van der Torre, 2018;
Dong et al., 2021; Straßer and Arieli, 2019]. The question then is how to
construct and evaluate deontic arguments.

3. Dialectical concepts like multiple agents, communication steps, or com-

mitment stores (like those of the Fatio dialogue system) do not play a
role in ASPIC+, which is more monolithic. If we want to add dialecti-
cal aspects to structured argumentation [Prakken, 2024a], how should we
design an argumentation system that behaves logically?

3.4 Extensions of the attack-defense paradigm for dialogue

In this and the following sections, we discuss extensions to abstract argumen-
tation. There are various approaches to extracting more information from
frameworks, and there is a variety of qualitative and quantitative enrichments
of frameworks. Semantics can be defined by reductions, selections, or adapta-
tions of defense. In argumentation as inference, only preference is clearly linked
to structured argumentation. In this section, we focus on extensions inspired
by dialogue.
There are two kinds of extensions to abstract argumentation in the commuta-
tive diagram of Figure 3. The first extends the argumentation framework with
44 Liuwen Yu, Leendert van der Torre, and Réka Markovich

qualitative and quantitative components from the knowledge base. In section

3.1, we mentioned various examples of such extensions. The second pertains to
step (3), where the argumentation semantics contains more information rather
than the acceptance of arguments.
For the second type of extension, Villata et al. [2011], for instance, generalize
the argument semantics by selecting from the graph not only a set of nodes but
also a set of edges. This represents intuitively that attacks can be successful
or unsuccessful. A similar kind of intuition is formalized in extended argumen-
tation frameworks with second or higher-order attacks [Barringer et al., 2005;
Cayrol et al., 2021]. Attacks are treated as arguments that can be attacked, and
thus can be accepted and rejected too. Consider, for example, the two-three
cycle framework shown in Figure 5. One possible output is a subframework
where we retain only the attack from b to a in the cycle on the left, effectively
reducing the complexity of the argumentation structure while maintaining spe-
cific attack relations.
Extended abstract argumentation frameworks enhance the expressive ca-
pacity of frameworks. However, it is not clear how these extensions can be
constructed directly from a knowledge base while incorporating additional com-
ponents such as agents, supports, numerical values or weights. One exception
is preference, which is clearly linked to structured argumentation as inference.
In structured argumentation, preferences play a central role in determining
formal outcomes. For example, in ASPIC+ [Modgil and Prakken, 2013], the
defeat relation between arguments is governed by a preference order, typically
derived through mechanisms such as the weakest link or last link principles.
Specifically, an attack from one argument to another only succeeds as a de-
feat if the attacked argument is not stronger than or strictly preferred to the
attacking argument, according to the given preference relation. In abstract
preference-based argumentation [Kaci et al., 2021], the first reduction in Fig-
ure 7 corresponds to this type of attack assignment.
In section 2.3, we introduced argumentation as dialogue. In section 2.4, we
discussed its formal methods, e.g., speech acts, game theory, axiomatic seman-
tics, and operational semantics. Inspired by dialogue, we have the following
challenge:
Challenge 8. Generalizing Dung’s attack-defense paradigm for dialogue.
At the structured level, it is natural to have the role of agents. One exam-
ple is Jiminy architecture [Liao et al., 2023], discussed in section 2.1. Jiminy
involves multiple stakeholders, each with their own knowledge base. When
dilemmas and conflicts arise, the argumentation engine considers the combina-
tion of all the arguments constructed by each stakeholder. Either there is a
large framework consisting of all the stakeholders’ arguments and the attack
relations, or all the knowledge bases are combined first, and the argumentation
frameworks are constructed afterward.
At the abstract level, agent-based extensions typically introduce various as-
pects such as agents, coalitions, knowledge, uncertainty, support, and so on.
 Thirteen Challenges in Formal and Computational Argumentation 45

As a result, there are various ways to define the semantics. Below we discuss
abstract agent argumentation [Yu et al., 2021], which uses a minimal extension
of Dung’s framework as a common core. This work only introduces an abstract
set of agents and arguments are associated with agents. There are four types
of semantics, defined by adaptations of defense, reductions, aggregations, and
selections:
Agent defense approaches adapt Dung’s notion of defense to argumenta-
tion semantics.
Social approaches are based on counting the number of agents [Leite and
Martins, 2011] and a reduction to preference-based argumentation [Am-
goud and Cayrol, 2002].
Agent reductions take the perspective of individual agents and aggregate
their individual perspectives [Giacomin, 2017].
Filtering methods are inspired by agents’ knowledge or trust [Arisaka et al.,
2022]. They leave out some arguments or attacks because they do not
belong to any agent.
Yu et al. [2021] have defined individual agent defense and collective agent
defense. Roughly, in individual agent defense, if an agent puts forward an
argument, it can only be defended by arguments from that same agent, i.e.,
a set of arguments E from individual agent α defends an argument c iff there
exists an agent α who has argument c such that for all arguments b attacking
c, there exists an argument a in E from α attacking b. Whereas with collective
agent defense, a set of agents α can do that, i.e., a set of arguments E defends
c collectively iff for all arguments b attacking c, there exists an agent α who
has c and an argument a in E from a set of agents α such that a attacks b.
Example 3.15 illustrates these two agent defenses.
Example 3.15 (Individual agent defense vs. collective agent defense).
In Figure 6(d), argument c defends argument a, but it does not individually
agent-defend it because c and a come from different agents. Consider another
abstract agent framework visualized in Figure 13. Here, {c1 , c2 } collectively
agent-defend argument a, but they do not individually agent-defend it.
Social semantics is based on a reduction to preference-based argumentation
for each argument, by counting the number of agents that have those argu-
ments. It thus interprets agent argumentation as a kind of voting procedure.
Example 3.16 illustrates social reduction.
Example 3.16 (Social reduction). Consider the agent argumentation
framework visualized in Figure 14. Arguments a and b both belong to agent α,
b also belongs to agent β, and a attacks b. In that situation, argument b is pre-
ferred to argument a because it is held by more agents. We can then apply the
four reductions from preference-based argumentation framework to abstract
argumentation framework, followed by application of Dung’s semantics.
46 Liuwen Yu, Leendert van der Torre, and Réka Markovich

b1 b2 γ

α c1 c2 β

Figure 13: {c1 , c2 } collectively agent-defending a

a a a a a
a b

α β b b b b b
b≻a
Figure 14: Social reduction: the left hand is an abstract agent argumentation
framework, the middle is the corresponding preference-based framework, and
the right hand are four corresponding abstract argumentation frameworks (as
discussed in Example 3.3)

Agent reductions take the perspective of each agent and obtain the semantics
accordingly. Intuitively, agents prefer their own arguments over the arguments
of others. Thus, for each agent, there is a preference-based argumentation
framework. In social reduction semantics, there is a unique abstract argumen-
tation framework after each reduction. However, in agent reduction semantics,
we obtain a set of abstract argumentation frameworks — one for each agent.
The final step is to take the union of all these frameworks to form a combined
abstract argumentation framework. Example 3.17 illustrates agent reductions.
Example 3.17 (Agent reduction). Reconsider the abstract argumentation
framework in Figure 14. Agent β prefers argument b over argument a. Thus,
we get the same preference-based framework as depicted in Figure 14, but for
a very different reason to that of social reduction. For agent α, argument a
and b are equivalent. To compute the agent extensions, we take the union of
the reductions for each agent.
Agent filtering semantics remove arguments that do not belong to an agent,
or they remove attacks that do not belong to an agent. An attack belongs to
an agent if both the attacking and attacked arguments belong to that agent.
Example 3.18 illustrates agent filtering semantics.
Example 3.18 (Agent filtering). Consider the two abstract agent frame-
works visualized in Figure 15. For the framework on the left, we might say
that argument a is not known because it doesn’t belong to any agent. And for
 Thirteen Challenges in Formal and Computational Argumentation 47

the framework on the right, we might say that the attack is unknown because
no agent holds both arguments a and b. The filtering methods remove such
unknown arguments and unknown attacks. This is followed by the application
of Dung’s semantics.

a b a b

β α β

Figure 15: Unknown argument and unknown attack

There are several aspects of dialogue beyond associating arguments with

agents that can be represented in abstract argumentation. One significant
aspect is to make time explicit: unlike inference, dialogue inherently unfolds
over time, with the dynamic argumentation framework evolving as the dialogue
progresses. Dialogue can also be strategic — sometimes it is advantageous for
an agent to not reveal certain arguments (this is discussed further in Section
4.3). A prototypical example of this is the content of the Miranda warning:
“Anything you say can and will be used against you in a court of law”. Ex-
ample 3.19 illustrates how a suspect’s argument can be strategically turned
against him/her in a dialogue.
Example 3.19 (Dialogue between accuser and suspect [Okuno and
Takahashi, 2009]). Let Pr and Op be the players involved in the following
argumentation dialogue (Pr and Op denote, respectively, a proponent and an
opponent):

Pr0 : “You killed the victim.”

Op1 : “I did not commit murder! There is no evidence!”

Pr1 : “There is evidence. We found your ID card near the scene.”

Op2 : “That is not evidence! I had my ID card stolen!”

Pr2 : “It is you who killed the victim. Only you were near the scene at the
time of the murder.”

Op3 : “I did not go there. I was at facility A at that time.”

Pr3 : “At facility A? Then, it is impossible that you had your ID card stolen
because facility A does not allow any person to enter without an ID card.”

In this example, the opponent tries to defend himself with the claim “I had
my ID card stolen!” (Op2 ). However, the proponent strategically uses this very
claim against the opponent (Pr3 ), arguing that if the opponent was at facility
48 Liuwen Yu, Leendert van der Torre, and Réka Markovich

A, it would have been impossible for his ID card to have been stolen because
the facility does not permit entry without an ID card. This demonstrates how
an argument can backfire in a strategic dialogue.
In this section, we discussed extending the attack-defense paradigm for di-
alogue, particularly with agents. We end the section with the following ques-
tions:

1. In this section, we discussed abstract agent argumentation, and we pro-

vided various semantics. How should a theory of structured agent argu-
mentation be designed?

2. Strategic dialogue goes beyond argumentation as inference by incorpo-

rating agency. How should dialogue strategies be designed? And should
they be evaluated?

3. What is the next step required to bridge the gap between 1) Dung’s
attack-defense paradigm and 2) strategic argumentation and dialogue?

4. There are many kinds of dialogues. What are the main components of a
dialogue? For example, what are the components of persuasion dialogue
systems like Fatio?

5. For all these kinds of dialogue, what makes a good dialogue? For example,
what is a successful Fatio dialogue? Does a successful dialogue happen
when someone is convinced of an argument they did not hold previously
or does it happen when the parties agree about where they disagree?

3.5 Extensions of the attack-defense paradigm for balancing

In this section, we continue our discussion of extensions to abstract argumen-
tation, focusing on extensions inspired by argumentation as balancing.
Argumentation as balancing brings to mind a double pan scale. The pros go
on one pan and the cons go on the other. The pros and cons may have relative
weights, and one needs to balance them from a utilitarian lens to determine
the status of the issues, e.g., what action to take. Balancing finds applications
in ethics and the law. In the legal domain, balancing is a metaphoric term that
is generally used to describe an important conceptual operation [Aleinikoff,
1986]. In many conflicts, there is something to be said in favor of two or more
outcomes. Whatever result is chosen, someone will be advantaged and someone
will be disadvantaged; some policies will be promoted at the expense of some
others. Hence it is often said that a “balancing operation” must be undertaken,
with the “correct” decision seen as the one yielding the greatest net benefit. In
medical ethics, for example, there are models of clinical ethics case consultation
that often refer to ‘balancing’ or ‘weighing’ moral considerations [McDougall
et al., 2020].
Challenge 9. Generalizing Dung’s attack-defense paradigm for balancing.
 Thirteen Challenges in Formal and Computational Argumentation 49

At an abstract level, it seems that pro and con arguments and the rela-
tions between them can be represented intuitively in bipolar argumentation
frameworks discussed by Cayrol and Lagasquie-Schiex [2005; 2009; 2010; 2013],
extending abstract argumentation framework with support relation that is in-
dependent of attack. Figure 16 illustrates three bipolar argumentation frame-
works, where attack relations are depicted by solid arrows, and support rela-
tions are depicted by dashed arrows. Similarly to abstract agent argumentation
semantics, there are also three types of bipolar argumentation semantics de-
fined by Yu et al. [2023].

The defense-based approach defines new notions of defense using both

support and attack.

The selection-based approach utilizes support only for selecting some of

the extensions provided in Dung’s semantics.

The reduction-based approach introduces indirect attacks based on inter-

pretations of support.

There are three new defense based on both attack and support relations,
called defense1 , defense2 , and defense3 , all of which have additional require-
ments for Dung’s defense. Defended1 requires that the argument defends (in
Dung’s theory) another argument also supports it. Defended2 requires that a
defender is supported. Moreover, defended3 requires not only that the attack-
ers are attacked, but also that all supporters of the attackers are attacked as
well. We illustrate the three defense in Figure 16.

a b a b a b

c c d c d e

Figure 16: Three bipolar argumentation frameworks (BAFs) illustrating the

three defense notions. In the left hand framework, {c} defends1 a. In the middle
framework, {c, d} defends2 a. In the right hand framework, {c, e} defends3 a.

The selection-based approach uses support during the post-processing step

for Dung’s theory of abstract argumentation [Gargouri et al., 2020], i.e., first
Dung’s semantics are obtained, then support can be used to select extensions
from Dung’s semantics. One way selects the extensions that have the largest
number of internal supports. This reflects the idea that for a coalition, the more
internal supports they have, the more cohesive they are. The other way is to
select the extensions that receive the most support from outside. This reflecting
the idea that the more support a coalition receives, the stronger it is. It thus
interprets support as a kind of voting procedure. We say that argument b in E
50 Liuwen Yu, Leendert van der Torre, and Réka Markovich

is internally supported if b receives support from arguments in E. Argument

b in E is externally supported if b receives support from arguments that are
outside E.
Example 3.20 (Selection-based approach to bipolar argumentation).
Consider the bipolar argumentation framework on the right hand side of Fig-
ure 17. There are four extensions to Dung’s stable semantics: {{a, d}, {a, c},
{b, d}, {b, c}}. By following the selection based on internal supports, {{a, c}}
is the stable semantics, while by following selection based on external supports,
{{a, d}} is the stable semantics.

a b

c d

Figure 17: A bipolar argumentation framework

The reduction-based approach has been studied extensively by Cayrol and

Lagasquie-Schiex [2005; 2009; 2013], and support is used as pre-processing for
Dung semantics. The corresponding abstract argumentation frameworks are
reduced by adding indirect attacks from the interaction between attacks and
supports with different interpretations, i.e., deductive support and necessary
support. Based on these two interpretations, four reductions have been dis-
cussed introducing additional attacks.
Deductive support [Boella et al., 2010] captures the intuition that if a sup-
ports b, then the acceptance of a implies the acceptance of b. Based on deduc-
tive interpretation, there are two kinds of additional attacks:
Supported attack and mediated attack. For example, in Figure 18(a), a
supports c, and c attacks b. Acceptance of a implies acceptance of c,
and acceptance of c implies non-acceptance of b. So, acceptance of a
implies non-acceptance of b. Thus, the supported attack from a to b is
introduced, depicted as a double-headed arrow. Similarly, the mediated
attack is visualized in Figure 18(b).
Necessary support [Nouioua and Risch, 2010] captures the intuition that if
a supports b, then the acceptance of a is necessary to obtain the acceptance of
b, or equivalently, the acceptance of b implies the acceptance of a.
Secondary attack and extended attack For example, in Figure 18.(c), a
attacks c, c supports b. The acceptance of a implies the non-acceptance
of c and the non-acceptance of c implies the non-acceptance of b; so, the
acceptance of a implies the non-acceptance of b. Thus, the secondary
attack from a to b is introduced, depicted as a double-headed arrow.
Similarly, the extended attack is visualized in Figure 18.(d).
 Thirteen Challenges in Formal and Computational Argumentation 51

a c b a c b

(a) Supported attack (b) Mediated attack

a c b a c b

(c) Secondary attack (d) Extended attack

Figure 18: Four kinds of reductions of bipolar argumentation frameworks

We now use the child custody example to illustrate reduction-based seman-

tics.
Example 3.21 (Child custody in bipolar argumentation). Consider the
bipolar framework visualized below. The figure should be read as follows.
Normal arrows are attack relations, dashed arrows are support relations, a
double box represents a prima facie argument which is self-supporting, and a
single box represents a standard argument that does not support itself. Our
focus is on how to interpret the support from (OP): “Child wants to live with
her mother” to (M): “Child’s best interest is that she lives with her mother.”
For a comprehensive analysis, we refer to the work of Yu et al. [2020]. The
supporting argument (OP) might have a special status because of the rules of
the Civil Code: the judge has to take the child’s opinion into consideration when
deciding about custody. Analysis of this rule shows how various interpretations
of the support interpretations relate to legal interpretations. We assume that
the child wants to live with her mother (OP). What does this mean? One can
say that the obligation to take argument (OP) into consideration means that
(OP) is a prima facie argument and thus has to be accepted. If it is a prima
facie argument, (M) receives the evidentiary support it needs. But this in
itself doesn’t decide how argument (OP) affects the extension. The extension
depends on how we interpret the support relation between (OP) and (M):
deductive or necessary. It seems very intuitive to interpret the support relation
as deductive: the obligation to take the child’s opinion into consideration is
apparently very much in line with what deductive support means: if we accept
the child’s opinion (which is prima facie) then we have to accept (M) too. We
assume the support from W: “Father is wealthy because he inherited” to F:
“Child’s best interest is that she lives with her father”, is not deductive.
Such an analysis contributes to the discussion on the formalization of legal
interpretation in the following way. The role of interpretation is crucial in the
law, but it is also a source of criticism of the use of logic-based methods for mod-
eling legal reasoning. For example, Leith warns that the knowledge engineer’s
interpretation when formalizing norms is necessarily premature because the
authority for interpreting the law has been assigned to the judiciary [Prakken,
52 Liuwen Yu, Leendert van der Torre, and Réka Markovich

Figure 19: A child custody deliberation with possible arguments and their
relations in a divorce case [Yu et al., 2020].

2013]. Addressing this criticism, the literature on legal interpretation has dis-
cussed the possibility that legal knowledge-based systems contain alternative
syntactic formalizations. It has been observed that while, on the syntactic
level, formalization commits us to a given interpretation, on the conceptual
level, classification of factual situations as legal concepts is not an issue of
logical form [Prakken, 2013]. Alternatively, we can restrict the investigation
by saying that “the only aspects of legal reasoning that can be formalized are
those aspects that concern the following problem: given a particular interpre-
tation of a body of legal knowledge, and given a particular description of some
legal problem, what are the general rational patterns of reasoning with which a
solution to the problem can be obtained?” [Prakken, 2013]. If a formal frame-
work offers the different interpretations itself, though, then using it might be
directly exploitable to the comparison of the different possibilities and routes
of reasoning given each interpretation.
It has been argued that for the validation of a bipolar argumentation theory,
so-called theory-based validation is preferable to empirical validation [Prakken,
2020; Polberg and Hunter, 2018], which itself is preferable to intuition-based
validation. Nevertheless, in this context, the principle-based analysis discussed
in Section 4.1 complements these validation methods. The theory of formal
argumentation needs to be complemented with examples and case studies con-
cerning the use of the theory.
In this section, we discussed extending the attack-defense paradigm for bal-
ancing. We end this section with the following questions:

1. The attack-defense paradigm introduced the distinction between struc-

tured and abstract argumentation. For every extension, we need to decide
whether to represent it at the structured or abstract level. For example,
should justification (as in Fatio) be expressed as in structured argumen-
tation (within the argument) or as a support (among arguments)?
 Thirteen Challenges in Formal and Computational Argumentation 53

2. When we introduce a new concept like support at the abstract level, it can
be interpreted in various ways at the structured level. For example, what
does support mean other than inferential relation (e.g., in ASPIC+)?
3. How can we better represent argumentation as balancing (e.g., in law
and ethics)? For example, how should the pros and cons be aggregated?
What is the role of weights?
4. How should other aspects of balancing be incorporated? For example,
how should argument strength be represented and evaluated? It is im-
portant to distinguish between different kinds of argument strength, in
particular logical, dialectical and rhetorical argument strength [Prakken,
2024a].
5. In the previous and present sections, we discussed extended abstract ar-
gumentation inspired by dialogue and balancing. Which other inspira-
tions can be utilized to design extensions of abstract argumentation? For
example, how can natural argumentation inspire extensions of abstract
argumentation?

4 After the paradigm shift: the computational turn

In this section, we transition from the paradigm shift in formal argumentation
to computational argumentation. Section 4.1 introduces principle-based anal-
ysis as a methodology for handling the diversity of argumentation models at a
higher level of abstraction, providing a systematic approach to designing and
choosing methods for different computational contexts. In Section 4.2, we focus
on compositionality principles such as locality, modularity, and transparency,
which play a central role from the attack-defense perspective and are exploited
in algorithmic strategies like divide and conquer. Section 4.3 discusses the re-
lationship between explanation and argumentation, highlighting, for example,
strategic argumentation, discussion games and reason-based models for under-
standing argumentation as dialogue, inference, and balancing, respectively. Fi-
nally, Section 4.4 addresses integrating argumentation techniques with existing
and emerging technologies, showcasing the potential of distributed argumenta-
tion systems and their applications in diverse technological contexts.
4.1 A principle-based analysis
In this section, we turn towards computational argumentation by discussing
methodology. Principle-based analysis is a methodology for managing the di-
versity of argumentation, such as when selecting among existing methods or
designing new ones. Principles describe formal argumentation at a higher level
of abstraction, and a wide range of principles exists across all models of argu-
mentation.
The principle-based approach is also called the axiomatic approach and the
postulate-based approach. Principles are properties, while postulates are nor-
mally desirable properties or requirements. This approach is particularly useful
54 Liuwen Yu, Leendert van der Torre, and Réka Markovich

when there is a diversity of alternative methods. It has been successfully ap-

plied in various areas. For example, Alchourrón-Gärdenfors-Makinson (AGM)
postulates [Alchourrón et al., 1985] have been applied in belief revision opera-
tions to ensure rationality, and axiomatic principles are applied when searching
for and choosing suitable voting rules for various contexts.
The principle-based approach has also been used to describe formal argu-
mentation at a higher level of abstraction. Abstraction in mathematics is the
process of extracting the underlying structures, patterns or properties of a
mathematical concept. In software engineering and computer science, abstrac-
tion is the process of generalizing concrete details. In formal argumentation,
one form of abstraction is to focus on the attack and defense relations between
arguments rather than their internal structures. The attack-defense relation is
used to define the functions of semantics. Principles can be defined as sets of
such functions and are represented as constraints on such functions. Principles
can be used to compare or define new functions.
The challenge addressed in this section is:
Challenge 10. Conducting a principle-based analysis of argumentation.
There is a diversity of principles and postulates in all models of argumen-
tation, particularly in the context of argumentation as inference, less so in
argumentation as dialogue and balancing. To illustrate, let us reuse the com-
mutative diagram in Figure 20 featuring examples of principles that are used
for different purposes. For step (1), there are Kraus-Lehmann-Magidor (KLM)
principles [Kraus et al., 1990] that a logical inference relation ought to sat-
isfy. For structured argumentation in steps (2-4), there are axiomatic anal-
yses of various attack relation assignments among arguments [Dung, 2016;
Dung and Thang, 2018; Pardo et al., 2024], as discussed in Section 3.2. For
the whole commutative diagram, rationality postulates are used to ensure
that the conclusions drawn at the end of the overall process have desirable
properties [Caminada and Amgoud, 2005; Caminada and Ben-Naim, 2007;
Caminada, 2018b], as discussed in Section 3.3. For step (3), there is a di-
versity of semantics available for the abstract argumentation framework. Ba-
roni and Giacomin [2007] classified argumentation semantics based on a set of
principles, which was extended by van der Torre and Vesic [2018]. For diverse
extended argumentation frameworks, with even more semantics. There are the
principles-based analysis of ranking-based semantics, multiagent argumenta-
tion [Yu et al., 2021], and bipolar argumentation [Yu et al., 2023].
There are three steps in a principle-based approach [van der Torre and Vesic,
2018].
Define the function that will be the object of the study. For instance, ab-
stract argumentation semantics are functions that map graphs to sets of
sets of graph nodes.
Define the principles — examine the relations between functions and prin-
ciples to see if the semantics satisfy the principles.
 Thirteen Challenges in Formal and Computational Argumentation 55

Postulates
1. Reflexivity
2. Right weakening
3. Left logical equivalence
4. Cautious monotonicity
5. Cut
6. ...

Axioms knowledge (1) logic conclusion Rationality postulates

1. Direct consistency base extension 1. Direct consistency
2. Indirect consistency 2. Indirect consistency
(2) attack (4) conclusion
3. Attack monotonicity 3. Closure
assignment extraction
4. Attack closure 4. Non-interference
5. Subargument structure argumentation (3) arg. semantics accepted 5. Crash resistance
6. ... framework arguments 6. ...

Principles
1. Conflict-free
2. Naivety
3. Admissibility
4. Reinstatement
5. Directionality
6. SCC-recursivenss
7. ...

Figure 20: Principle-based analysis in commutative diagram

Classify and study the sets of principles — study the relations between
principles. For example, a set of principles may imply another principle.
Or there may be incompatibilities among principles. Or there may be a
set of principles that characterizes a semantics.

There are three main branches of abstract argumentation semantics.

Admissibility-based semantics (AB) uses gunfight rules requiring that an

extension E defends itself against all attackers [Dung, 1995], i.e., when-
ever each argument attacking E from the outside is itself attacked by
some element of E.

Weak admissibility-based semantics (WA) only requires that an exten-

sion E defends itself against reasonable arguments [Baumann et al., 2020].

Non-admissibility based semantics (NA) requires an extension E to be a

maximal conflict-free set of arguments. The most prominent example of
non-admissibility based semantics is CF2 semantics [Baroni et al., 2005].

We illustrate the above three branches of semantics with Example 4.1 be-
low.
Example 4.1 (Three branches of abstract argumentation semantics).
Consider the three frameworks in Figure 21. For framework (a), the only
extension in AB semantics is the empty set whereas under the CF2 semantics,
56 Liuwen Yu, Leendert van der Torre, and Réka Markovich

b is accepted. To get the desirable properties of directionality and strongly-

connected-component (SCC) recursiveness (discussed further below), CF2 is
defined in terms of a local function that computes the maximal conflict-free
subsets for each strongly connected component of a framework. Under WA
semantics, the set of weakly preferred extensions of framework (a) is {{b}}
because its only attacker a is self-attacking. It is like a “zombie”, it is there
but it can do no harm [Baumann et al., 2020]. For framework (b), the set
{d} is not admissible because it does not defend itself from b. Nevertheless,
under WA semantics, d is acceptable because b is part of an odd-length cycle
of arguments that are never accepted, and so it does not pose an actual threat.
These extensions are listed in Table 6.

c a
c d
a b c
a b d b

(a) (b) (c)

Figure 21: Three argumentation frameworks

Semantics (a) (b) (c)

AB {∅} {∅} {{a, d}, {b, d}}
NA(CF2) {{b}} {{a, d}, {c, d}, {b}} {{a, d}, {b, d}}
WA {{b}} {{d}} {{a, d}, {b, d}}

Table 6: Three semantics applied to the frameworks in Figure 21. AB =

admissibility-based; NA = non-admissibility based; WA = weak-admissibility
based.

To compare the diverse semantics, we can categorize formal argumentation

principles into three types: traditional principles (the most discussed), variants
of traditional principles, and new principles specifically designed for emerging
semantics. Below, we provide examples of these principles to illustrate how
they are used. This will enable us to compare the different branches of ar-
gumentation semantics as well as the different agent argumentation semantics
described in Section 3.4.
We list some of the traditional principles that have been used to compare
these semantics.

Conflict-freeness states that every extension under semantics is a conflict-

free set, i.e., there is no attack relation among the arguments in the
extension.
 Thirteen Challenges in Formal and Computational Argumentation 57

Admissibility is satisfied by a semantics if and only if every extension under

that semantics is admissible.

Naivety states that every extension under the semantics is a maximal conflict-
free set.

Directionality states that an argument a should be affected only by a’s at-

tacker. The arguments that only receive an attack from a should not
have any effect on the state of a.

SCC (strongly-connected-component) recursion states that extension

construction carried out in an initial SCC do not depend on those con-
cerning the other ones, while they directly affect the choices about the
subsequent SCCs and so on.

Modularity states that the semantics of a framework can be obtained by the

semantics of the smaller parts of that framework.

Example 4.2. Given the framework shown in Figure 22, the complete and
weak complete semantics are the same: {∅}, while ∅ is not a maximal conflict-
free set in this framework, e.g., {a} is also conflict-free.

a b

Figure 22: Complete and weak complete semantics do not satisfy naivety

In Section 3.4, we discussed abstract agent argumentation [Yu et al., 2021].

The analysis focused on the four traditional Dung semantics (complete, pre-
ferred, stable, and grounded) denoted as T R. Two new concepts, individual
defense and collective defense, have been introduced. By applying these con-
cepts to each of the four traditional semantics, there are two new variants for
each: one based on individual defense and the other on collective defense. This
results in a total of eight distinct defense-based semantics, denoted as Sem1
and Sem2 . Additionally, social agent semantics, which prioritizes arguments
supported by more agents, produce sixteen semantics through four reductions
(denoted as SR1 − SR4 ) from preference-based argumentation frameworks to
abstract argumentation frameworks. Agent reduction semantics, which consid-
ers the perspective of individual agents, also yields sixteen semantics through
the four reductions, denoted as AR1 − AR4 . Lastly, agent filtering semantics,
inspired by limited knowledge, introduces eight additional semantics, denoted
as OR and N BR. Altogether, this results in a total of fifty-two semantics.
The paper provides a full analysis of fifty-two agent semantics, includ-
ing Dung’s traditional semantics, with seventeen principles. The results of
58 Liuwen Yu, Leendert van der Torre, and Réka Markovich

principle-based analysis are typically summarized in tables, as seen in Ta-

ble 7, 8, 9. These principles are categorized into three groups: five traditional
principles (Table 7), four variations of these traditional principles (Table 8),
and eight new principles specifically designed for agent-based argumentation
(Table 9).

Semantics P1 P2 P3 P4 P5
TR CGPS CGPS CGP CGPS CGPS
Sem1 CGPS CGPS CGP S S
Sem2 CGPS CGPS CGP S S
SR1 × × CGP × ×
SR2 CGPS × × × ×
SR3 × CGPS CGP CGPS CGPS
SR4 CGPS G × × G
AR1 × × CGP × S
AR2 CGPS × × × ×
AR3 CGPS CGPS CGP CGPS CGPS
AR4 CGPS G × × G
OR CGPS × CGP CGPS CGPS
NBR × × CGP × S

Table 7: Comparison of abstract agent argumentation semantics and traditional

principles. When a principle is never satisfied by a certain reduction for all
semantics, we use the × symbol, and we use a question mark to represent an
open problem. P1 refers to Principle 1, and the same convention holds for all
the others. P1 = conflict-free, P2 = admissibility, P3 = directionality, P4 =
SCC-recursiveness, P5 = modularity.

For example, the agent admissibility principle is a variation of the tradi-

tional admissibility principle, with agent defense replacing the standard notion
of defense. Since admissibility can be applied to either individual defense or
collective defense, this gives rise to two distinct agent admissibility principles.
Similarly, the agent SCC-recursiveness principles are adapted to reflect the
concepts of individual and collective defense, resulting in two corresponding
principles. Additionally, eight entirely new agent principles have been intro-
duced to address the unique aspects of agent-based argumentation, and these
are shown in Table 9.
Below we list the eight new principles. Some of them are expected to be
satisfied by all of the semantics. Some can be used to distinguish between
different semantics, since only certain semantics satisfy or do not satisfy certain
principles.

Principle 10: AgentAdditionPersistence states that if more agents adopt

an argument that is already accepted, this does not affect the extension.
 Thirteen Challenges in Formal and Computational Argumentation 59

Semantics. P6 P7 P8 P9
TR × × × ×
Sem1 CGP CGP S S
Sem2 × CGP S S
SR1 × × × ×
SR2 × × × ×
SR3 × × × ×
SR4 × × × ×
AR1 × × × ×
AR2 × × × ×
AR3 × × × ×
AR4 × × × ×
OR × × × ×
NBR × × × ×

Table 8: Comparison of the reductions and agent admissibility principles, and

agent SCC-recursion. P6 = agent admissibility1 , P7 = agent admissibility2 , P8
= agent SCC-recursiveness1 , P9 = agent SCC-recursiveness2 .

Principle 11: AgentAdditionUniversalPersistence reflects the same

idea as principle 10 but is based on the assumption that a is accepted in
all extensions.

Principle 12: PermutationPersistence reflects a principle we expect to

hold for all agent semantics — anonymity. If we permute the agents,
it does not affect the extensions. This principle is analogous to language
independence for arguments, as defined by Baroni and Giacomin [2007].

Principle 13: MergeAgent states that if the arguments of two agents do

not attack each other, we can merge these agents into one single agent.
The principle does not hold for agent defense semantics because new
agent defenses may be created.

Principle 14: RemovalAgentPersistence states that if two agents have

the same arguments, we can remove one of these agents without changing
the extensions. This represents the opposite of social semantics, where
the number of agents does make a difference.

Principle 15: AgentNumberEquivalence is inspired by social agent se-

mantics. It states that where there are two argumentation frameworks
with the same arguments and attacks, if for every argument the number
of agents holding that argument is the same, then the extensions are the
same.
60 Liuwen Yu, Leendert van der Torre, and Réka Markovich

Principle 16: ConflictfreeInvolvement is inspired by agent reduction se-

mantics. It states that if the set of an agent’s arguments is conflict-free,
then there is an extension containing those arguments.
Principle 17: RemovalArgumentPersistence is inspired by OrphanRe-
duction semantics. It states that if we have arguments that do not belong
to any agents, then they can be removed from the framework without af-
fecting the extensions.
In the resulting Table 9, all the agent semantics satisfy P12. Perhaps sur-
prisingly, neither social agent semantics nor agent reduction semantics satisfy
P10 while trivial reduction semantics, social agent semantics, and agent filtering
semantics satisfy P13. Moreover, all agent semantics except social agent seman-
tics satisfy P14. No semantics satisfy P16. As expected, only OrphanRemoval
satisfies P17. The only semantics that are not distinguished yet concern the use
of different preference reductions, or different Dung semantics. To distinguish
between these, the principles proposed in preference-based argumentation [Kaci
et al., 2021] and in Dung’s semantics can be used [Baroni and Giacomin, 2007;
van der Torre and Vesic, 2018]. In that sense, the principle-based analyses can
complement one other.

Sem. P10 P11 P12 P13 P14 P15 P16 P17

TR CGPS CGPS CGPS CGPS CGPS CGPS × ×
Sem1 S S CGPS × CGPS × × ×
Sem2 S S CGPS × CGPS × × ×
SR1 × CGPS CGPS CGPS × CGPS × ×
SR2 × CGPS CGPS CGPS × CGPS × ×
SR3 × CGPS CGPS CGPS × CGPS × ×
SR4 × CGPS CGPS CGPS × CGPS × ×
AR1 × CGPS CGPS × CGPS × × ×
AR2 × CGPS CGPS × CGPS × × ×
AR3 × CGPS CGPS × CGPS × × ×
AR4 × CGPS CGPS × CGPS × × ×
OR CGPS CGPS CGPS CGPS CGPS CGPS × CGPS
NBR CGPS CGPS CGPS CGPS CGPS × × ×

Table 9: Comparison between the reductions and new agent principles. P10 =
AgentAdditionPersistence, P11 = AgentAdditionUniversalPersistence, P12 =
PermutationPersistence, P13 = PermutationPersistence, P14 = RemovalAgent-
Persistence, P15 = AgentNumberEquivalence, P16 = ConflictfreeInvolvement,
Principle 17 = RemovalArgumentPersistence.

Finally, the principle-based approach to formal argumentation may lead to

the study of impossibility and possibility results. For instance, Arrow’s impos-
sibility theorem in voting and social choice theory demonstrates that no voting
 Thirteen Challenges in Formal and Computational Argumentation 61

system can simultaneously satisfy the whole set of seemingly reasonable criteria
— non-dictatorship, unrestricted domain, Pareto efficiency, and independence
of irrelevant alternatives — when there are three or more options available.
This kind of result highlights the inherent trade-offs involved in designing sys-
tems that attempt to balance competing principles. Similarly, as discussed in
Section 3.2, any attempt to realize PDL in ASPIC+ should preserve the def-
initional principle of attack closure. The impossibility theorem explains how
this is incompatible with context independence [Pardo et al., 2024]. These im-
possibility results are crucial because they reveal the boundaries of what can
be achieved within a given formal system. Additionally, they also guide re-
searchers to either accept certain trade-offs or seek alternative approaches that
might circumvent these limitations.
In this section, we discussed principle-based analysis and, as usual, we list
several research questions about that topic.

1. How can we provide guidance to users who are not experts in formal
or computational argumentation on how to use the theory of argumen-
tation for their needs? Can we develop a user guide for the theory of
argumentation?

2. How do we decide which conceptualization of formal argumentation to use

for an application (argumentation as inference, dialogue or balancing),
and how do we connect or combine these conceptualizations?

3. What needs to be changed to move from constructing comparison tables

(as shown in Tables 7-9) to characterization, or proving possibility and
impossibility results? For example, how to characterize last vs. weakest
link in structured argumentation, or characterize various kinds of abstract
argumentation semantics?

4. Which methodology can be developed for formal and computational argu-

mentation to guide the search for and design of principles? For example,
the principle of resolution was defined by Baroni and Giacomin [2007],
then the resolution-based semantics were defined and studied by Baroni
et al. [2011].

5. How can we combine principles from various extended argumentation

frameworks? For example, reductions in preference-based argumentation
often remove attacks whereas reductions in bipolar argumentation often
add attacks.

4.2 Algorithmic argumentation

In this section, we consider the role of principles in algorithmic argumentation.
Algorithmic argumentation, as illustrated in Table 1, refers to a step-by-step
procedure or set of rules designed to perform a specific task or solve a particular
argumentation problem. We focus mainly on the calculation of argumentation
62 Liuwen Yu, Leendert van der Torre, and Réka Markovich

semantics. On the one hand, compositionality principles play a central role in

the attack-defense perspective. On the other hand, algorithms and other com-
putational approaches exploit these principles. We illustrate this by discussing
locality, modularity, and transparency principles on one side, and “divide and
conquer” and robustness principles on the other.
Traditionally, Dung’s abstract argumentation frameworks are viewed as
monolithic entities where various semantics are applied globally to determine
which arguments are acceptable. While this unified approach preserves gener-
ality, it has been shown to be computationally intractable. That complexity
presents the following challenge:
Challenge 11. Designing efficient algorithms for argumentation semantics.
The idea of compositionality is that an abstract argumentation framework
is broken down into interacting smaller subframeworks. This motivates a lo-
cal focus and further investigation into locality and modularity principles in
abstract argumentation. Related principles are, for example, directionality,
SCC-recursiveness, and decomposability.
The directionality property corresponds to the idea that the attack relation
encodes a form of dependency and that arguments can affect one other only by
following the direction of the attacks. This consideration can be extended from
individual arguments to sets of arguments. If a set of arguments is unattacked
(i.e., it does not receive attacks from arguments outside the set) it should
not be affected by the rest of the argumentation framework. In more formal
terms, given an argumentation semantics, projecting the semantics of the global
framework onto an unattacked set should result in the semantics producing an
evaluation of an argumentation framework consisting of only that unattacked
set.
Example 4.3 illustrates directionality and why stable semantics does not
satisfy this property.
Example 4.3 (Directionality [Baroni and Giacomin, 2007]). Consider
the stable semantics of the argumentation framework in Figure 5. For the
subframework consisting of {a, b}, the stable semantics is {{a}, {b}}. The
stable semantics of the whole framework is {{b, d}} whose projection — the
unattacked set {a, b} — is {b}. However, {{b}} does not coincide with the sta-
ble semantics of the unattacked framework. This is a counterexample proving
that stable semantics does not satisfy the directionality.
The property of SCC-recursiveness [Baroni et al., 2005] is based on decom-
position along the SCCs of the argumentation framework. Different from di-
rectionality, SCC-recursiveness has a direct constructive interpretation. The
structure of the argumentation frameworks drives the incremental definition of
extensions step by step. Without complex technicalities, we illustrate SCC-
recursiveness with Example 4.4.
Example 4.4 (SCC-recursiveness [Baroni et al., 2005]). Consider again
the framework in Figure 5.
 Thirteen Challenges in Formal and Computational Argumentation 63

Step 1 Partition the argumentation framework into SCCs. There are two
SCCs in the framework: S1 = {a, b} and S2 = {c, d, e}. Here, S1 is
identified as the initial SCC as it does not depend on S2 .

Step 2 Construct extensions incrementally using a base function. Determine

the possible extensions within each initial SCC using a semantic-specific
base function. This function returns the extensions for argumentation
frameworks consisting of a single SCC. For the SCC of S1 = {a, b}, the
base function provides two possible partial candidate extensions: {a} and
{b}.

Step 3 For each possible extension determined in Step 2, apply the reinstate-
ment principle. This involves suppressing the nodes directly attacked
within subsequent SCCs and considering the distinction between defended
and undefended nodes. Let’s take candidate extension {b}. Here, argu-
ment c in S2 cannot be included in the extension because it is attacked
by argument b. Therefore, only {d, e} can be taken into consideration.

Step 4 Recursively apply the steps on restricted frameworks. Consider the

subframework ({d, e}, {(d, e)}). This subframework is again partitioned
into SCCs, resulting in S1′ = {d} and S2′ = {e}. Considering S1′ = {d},
since e is attacked by d, e is excluded. Hence, the only extension left is
∅. Thus, the final extension is {b} ∪ {d} ∪ ∅ = {b, d}.

The idea of decomposability is to break down an abstract argumentation

framework arbitrarily into interacting smaller subframeworks called modules.
Input/Output frameworks have been defined on this basis [Baroni et al., 2014].
Each module can be described as a black box whose Input/Output behavior —
specifically referring to the labeling — fully determines its role in the system’s
global behavior. That makes it possible to describe and analyze the frame-
work’s global behavior in terms of the combination of the local behaviors of
its constituent modules. Each local behavior can be characterized individually.
This characterization is independent of the internal details of other modules.
Instead, it focuses only on the connections and mutual interactions between
the module and the other modules. Additionally, if two modules have the same
input and output behavior, they are interchangeable in a way that does not
influence the global behavior.
Example 4.5 illustrates the interfaces of subframeworks.
Example 4.5 (Interface [Baroni et al., 2014]). Given the abstract ar-
gumentation framework (AF) visualized in Figure 23 and the subframeworks
induced by the sets {a, b, c} and {d}, these subframeworks are denoted as
AF↓{a,b,c} and AF↓{d} . The subframeworks interact with one other through
the attacks a → d and d → a respectively. For AF↓{a,b,c} , the interface, or
input argument, is argument d, while for AF↓{d} , the input argument, or in-
terface, is argument a.
64 Liuwen Yu, Leendert van der Torre, and Réka Markovich

b
a d
c

Figure 23: A partition of an abstract argumentation framework

A local function determines the labeling of a subframework based on the

labeling of external input arguments, ensuring that the internal labeling of a
subframework is consistent with its external influences. Example 4.6 illustrates
how external input arguments enforce the internal labelings.
Example 4.6 (External input arguments enforcement [Baroni et al.,
2014]). Consider the argumentation framework in Example 4.5. If we apply
the local function to subframework AF↓{a,b,c} with the external argument d,
if d is labeled as in, the resulting labeling of the subframework is {(a, out),
(b, in), (c, out)}. If d is labeled as out, the resulting labeling of the sub-
framework is {{(a, undec), (b, undec), (c, undec)}}. If d is labeled as undec, the
resulting labeling of the subframework is {{(a, undec), (b, undec), (c, undec)}}.
We can apply the same analysis for the subframework AF↓{a,b,c} with external
argument a.
The property of decomposability states that given an arbitrary partition of
an argumentation framework into a set of subframeworks, the outcomes pro-
duced by a given semantics can be obtained as a combination of the outcomes
produced by a local counterpart applied separately on each subframework and
vice versa.
Example 4.7 (Decomposability [Baroni et al., 2014]). Considering again
the argumentation framework in Example 4.5 and the partition {{a, b, c}, {d}},
the decomposability of the complete semantics requires a local function such
that the labelings of AF are exactly those obtained by the union of the com-
patible labelings of AF↓{a,b,c} and AF↓{d} given by the local function itself.
The labeling {(a, out), (b, in), (c, out)} is compatible with {(d, in)}. The first is
obtained with d labeled in, and the latter is obtained with a labeled out. On
the other hand, the labeling {(a, out), (b, in), (c, out)} is not compatible with,
e.g., {(d, out)}. Overall, exactly two global labelings arise from combining
the compatible outcomes — {(a, undec), (b, undec), (c, undec), (d, undec)} and
{(a, out), (b, in), (c, out), (d, in)}, which corresponds to the complete labelings
of the whole AF.
The property of transaparency states that if two modules have the same
Input/Output behavior, then we can replace one with the other without influ-
encing the framework’s global behavior. This ensures that the invariant part of
the framework is unaffected. Example 4.8 illustrates the transparency property.
 Thirteen Challenges in Formal and Computational Argumentation 65

Example 4.8 (Transparency [Baroni et al., 2014]). Consider the argu-

mentation frameworks AF1 and AF2 shown in Figure 24. M1 and M2 have the
same Input/Output behavior, i.e., they are equivalent under preferred seman-
tics. The invariant set of the replacement is the set {e1 , e2 }. However, after the
replacement of M1 by M2 in AF1 , the preferred extension changes. In fact, the
preferred labelings of AF1 are {(a1 , in), (a2 , out), (o, out), (e2 , in), (e1 , out)} and
{(a1 , out), (a2 , in), (o, undec), (e2 , undec), (e1 , undec)}, while {(b, in), (c, out),
(a1 , in), (a2 , out), (o, out), (e2 , in), (e1 , out)} is the only preferred labeling of
AF2 .

M2

e1 b c
M1

e1 a1 a2 a1 a2

e2 o e2 o

(a) AF1 (b) AF2

Figure 24: Preferred semantics does not satisfy transparency

Now, we illustrate how to exploit modularity, which leads us to the concept of

summarization [Baroni et al., 2014]. Summarization allows a complex part of an
argumentation process, such as the analysis and discussion of factual evidence
in a legal case, to be replaced by a more synthetic representation. This process
focuses on the facts that have an actual impact on the decision, leaving out
unnecessary details. The main concern with summarization is ensuring that,
as the argumentation framework is simplified, the overall outcome remains
consistent and preserved.
We use Example 4.9 to illustrate how summarization works.
Example 4.9 (Summarization [Baroni et al., 2014]). Consider the ar-
gumentation frameworks AF1 and AF2 shown in Figure 25. AF2 can be ob-
tained from AF1 by “summarizing” the component M1 , including the argu-
ments a1 , a2 , a3 and a4 , with the component M2 , including the arguments a1
and a2 . Then, e1 and e2 are the same in the two frameworks, i.e. e1 is labeled in
and e2 is labeled out. More generally, consider a finite sequence of n arguments
a1 , . . . , an such that each argument attacks the subsequent one, i.e. ai attacks
ai+1 with 1 ≤ i < n, and suppose that only a1 can receive further attacks
from other arguments and that only an can attack other arguments. Then, it
66 Liuwen Yu, Leendert van der Torre, and Réka Markovich

is intuitive to see that the “black-box behavior” of a sequence of arguments of

this kind, whose external “terminals” are a1 and an , only depends on whether
n is even or odd. In fact, the behavior of any even-length sequence is the same
as where n = 2 (if a1 is in then an is out, if a1 is out then an is in, if a1 is
undec then an is undec), while for any odd-length sequence, the behavior is the
same as for a1 alone (if n is an odd number, an necessarily gets the same label
as a1 ).

M1 M2

e1 a1 a2 a3 a4 e2 e1 a1 a2 e2

(a) (b)

Figure 25: Summarizing a chain of arguments

The first area where locality and modularity principles find application is in
the development of algorithms for efficiently computing argumentation seman-
tics, particularly through divide-and-conquer strategies. By leveraging locality,
one can focus on specific parts of an argumentation framework and thus reduce
the computational burden. There are three locality- and modularity-based
approaches that demonstrate how these principles can enhance the efficiency
of computing semantics in dynamic, static, and partial argumentation frame-
works. For instance, when only partial semantics are required — such as in sce-
narios where the status of certain arguments needs to be queried — algorithms
can be designed to focus solely on the relevant arguments, disregarding those
that do not impact the outcome. Similarly, in dialogues where new arguments
are introduced, the computation can be streamlined by ignoring the effects of
irrelevant arguments. For a comprehensive overview of these approaches, we
refer to the work of Baroni et al. [2018].
Other principles used in the design of algorithms include robustness princi-
ples [Rienstra et al., 2020], which deal with the behavior of a semantics when
the argumentation framework changes due to the addition or removal of an
attack between two arguments. Robustness principles have been classified into
two kinds: persistence principles and monotonicity principles. The former deal
with the question of whether a labeling in an argumentation framework under
a given semantics persists after a change. The latter deal with the question of
whether new labelings are created after a change. They are listed as follows:

XY addition persistence: a labeling of an argumentation framework in

which x is labeled X and y is labeled Y is still a labeling of F after
adding an attack from x to y.

XY removal persistence: a labeling of an argumentation framework in

 Thirteen Challenges in Formal and Computational Argumentation 67

which x is labeled X and y is labeled Y is still a labeling of F if re-

moving the attack from x to y.

XY addition monotonicity: if in all labelings of an argumentation frame-

work, x is labeled X and y is labeled Y , then adding an attack from x to
y does not lead to new labelings.

XY removal monotonicity: if in all labelings of an argumentation frame-

work, x is labeled X and y is labeled Y , then removing an attack from x
to y does not lead to new labelings.

Persistence and monotonicity principles are also useful for addressing en-
forcement problems [Baumann, 2012] in abstract argumentation. This is about
the problem of determining minimal sets of changes to an argumentation frame-
work in order to enforce some result, such as the acceptance of a given set of
arguments. Because persistence and monotonicity principles can be used to de-
termine which changes to the attack relation of an argumentation framework
do or do not change its evaluation, these principles can be used to guide the
search for sets of changes that address the enforcement problem. This idea has
already been used for extension enforcement under grounded semantics [Niska-
nen et al., 2018].
In this section, we have discussed compositionality principles, algorithms and
other computational approaches that exploit principles. We end this section
with the following questions:

1. Most algorithms are developed for abstract argumentation and for argu-
mentation as inference. What are the computational tasks for structured
argumentation, for argumentation as dialogue, and for argumentation as
balancing?

2. Apart from algorithms, what other tools does computer science have to
offer, e.g., analysis of computational complexity, efficient implementation
of algorithms?

3. As discussed in this section, principle-based analysis is a bridge between

formal and algorithmic argumentation. Which principles are particularly
useful? And how can we use these principles to speed up computation?

4. What else can we learn from artificial intelligence? The rise of machine
learning and foundation models is changing the landscape of argumenta-
tion. How could these new approaches speed up computation?

5. Which topics need to be addressed first in computational argumentation?

Should we address algorithms for formal argumentation concepts or focus
our attention on the challenges of natural argumentation?
68 Liuwen Yu, Leendert van der Torre, and Réka Markovich

4.3 Explanation and argumentation

In this section, we discuss some relations between computational argumenta-
tion and explanation. Strategic argumentation explains argumentation as dia-
logue, discussion games explain argumentation as inference, and reason-based
explanation can be used for argumentation as balancing. We illustrate the
explanation for argumentation through the example of discussion games for
grounded semantics.
In recent years, the field of explainable artificial intelligence (XAI) [Longo
et al., 2024] has gained significant attention due to the increasing complexity
and opacity of AI systems, particularly when it comes to systems being po-
tentially deployed in critical decision-making areas such as healthcare, finance,
and the law. The main focus is usually on making the reasoning behind the
decisions or the predictions made by the AI system [Phillips et al., 2021] more
understandable and transparent.
The relationship between explanation and argumentation can be seen from
different perspectives. On the one hand, an explanation for argumentation
mostly concerns the question of whether a certain argument or claim can be
accepted (or not) and why. This has been studied not only at the abstract
level [Ulbricht and Wallner, 2021] but also at the structured level [Borg and
Bex, 2024]. On the other hand, explanation through argumentation is often
intuitively seen as reasonable [Sklar and Azhar, 2018]. For example, it can
clarify the decision-making process of an AI system through argumentation
procedures. This can be done in a static manner by illustrating the argument
inference process or showing the relations between arguments, or it can be done
through an interactive dialogue that explains the reasoning [Castagna et al.,
2024b]. In this section, we talk about the following challenge:
Challenge 12. Explaining argumentation.
Take the example of argumentation as inference, which is about how reason-
able decisions or conclusions can be reached by constructing for and against
arguments and then evaluating those arguments. It makes it possible to un-
derstand decisions by tracing exactly why a particular conclusion was reached.
It also makes it possible to see how certain decisions it relates to other poten-
tial conclusions. Explanations are often found to be argumentative. Mercier
and Sperber [2017] highlighted that the effectiveness of interactive argumenta-
tion in changing people’s minds, at least for simple arguments, stems from the
chance to address counterarguments during discussions. Participants can raise
and rebut counterarguments, which makes the exchange more dynamic. Con-
trast this with one-sided messaging campaigns, where counterarguments are
generated but remain unaddressed [Altay et al., 2022]. Interactive argumen-
tation can involve a form of dialogue where users interact with an AI system,
asking for clarifications or further information, and the system responds with
explanations. In this sense, explanation is intertwined with dialogue — a con-
versation where arguments are presented, challenged, and defended, as in the
Fatio design [McBurney and Parsons, 2004].
 Thirteen Challenges in Formal and Computational Argumentation 69

Strategic argumentation (see Governatori et al.’s [2021] overview) can be

used to explain argumentation as dialogue. By analyzing the strategies em-
ployed by an agent, it is possible to understand how and why that agent chooses
to disclose certain arguments or information during a debate in order to achieve
a specific objective and prevent the opposing party from gaining an advantage.
To give an example, Arisaka et al. [2022] propose an abstract agent argu-
mentation model that distinguishes the global argumentation of judges from
the local argumentation of accused persons, prosecutors, defense lawyers, wit-
nesses and experts. All the “local” agents have partial knowledge of the ar-
guments and attacks of the other “local” agents, on which basis they decide
autonomously whether to accept or reject their own arguments and whether to
bring their own arguments forward in court. The arguments accepted by the
judge are based on a game-theoretic equilibrium among the argumentation of
the other agents. The theory can be used to distinguish between the various
direct and indirect ways in which an agent’s arguments can be used against
his/her other arguments. The global abstract agent argumentation framework
is viewed differently by the different agents.
Example 4.10 (Murder at Facility C). There was a murder at Facility
C. Acc has been accused of the crime. There is a witness Wit and a prosecutor
Prc. Acc has two arguments in mind:

a1 He was at Facility A on the day of the murder [this is a fact Acc knows].

a2 He is innocent [this is Acc’s claim].

Prc entertains the following arguments:

a6 Only Acc could have killed the victim [this is Prc’s claim].

Meanwhile, Wit has certain beliefs on the basis of which he has three arguments:

a3 Acc stayed home on the day of the murder, having previously lost his ID
card [this Wit originally believes to be a fact].

a4 Acc could enter any facility provided he had his ID card on him [this is a
fact known to Wit].

a5 Acc could not have been at Facility C at the time of the murder [this is
Wit’s claim].

Further, the relationship between the three arguments is such that a3 attacks
a4 , which attacks a5 . Altogether, these arguments by the three agents form
the argumentation framework in Figure 26(a). Acc, Prc and Wit reveal their
own internal argumentation, partially or elaborately, for the judge to evaluate.
But since agents may come to learn the arguments of other agents if, say, they
are expressed before they present their own arguments, it is possible that they
take the additional information into account when deciding which arguments
70 Liuwen Yu, Leendert van der Torre, and Réka Markovich

to present. In this example, both Prc and Acc have the characteristic of being
unaware agents. Prc has no reason to drop argument a6 , and neither does Acc,
as he sees no benefit in admitting that a6 . However, how Wit responds to the
fact known to Acc (a1 ) could prove crucial to whether he is found innocent or
guilty.

Case A. Suppose Wit is unaware and open-minded. Wit presents what he

believes, i.e., his local argumentation framework (see Figure 26(a).) She
locally accepts a3 and a5 . The judge evaluates all the arguments, con-
cluding that a2 is not acceptable, i.e., Acc is guilty. The judge starts his
inference with Acc’s acceptable a1 and proceeds to reject a2 . The two
arguments a3 and a5 accepted by Wit are not accepted by the judge.
This illustrates indirect use of an argument against Acc.
Case B. Suppose Wit is unaware and closed-minded. Instead of presenting all
the reasoning he had developed in his local argumentation, Wit states the
following key points concisely: that Acc stayed home on the day of the
murder, and that Acc could not have been at Facility C (see Fi ). Omission
of a fact known to Wit (a4 ), which Wit perhaps considers irrelevant to
the criminal case, changes the judge’s decision completely. In Fi , a5 is
seen an argument that is globally acceptable. That argument rejects a6
in favor of a2 .
Case C. Suppose Wit learns a1 beforehand. Wit realizes that a3 , which she
thought was a fact, is not actually true. She no longer claims a5 in
her local argumentation, but she nevertheless discloses her entire original
argumentation (see Figure 26(a)). Her conclusion that a4 is acceptable
concurs with the judge’s view on the matter, and the judge ultimately
concludes that a2 shall be rejected.
Case D. Suppose again that Wit learns a1 beforehand but that she mentions
the key arguments concisely. She states that entry into any facility re-
quires an ID card (see Fj ). Here again, the judge has no objection to the
evidence that might have been provided by Proc. As such, a6 is accepted,
which proves that Acc is guilty. This illustrates direct use of an argument
by Wit against Acc.

We now move on to discussion games designed to explain argumentation as

inference. As discussed in Section 4.2, calculating semantics, or determining
whether a specific argument is present in some or all labelings, can be computa-
tionally expensive. Discussion games provide an alternative to that. Discussion
games [Caminada, 2017] take place between two parties, typically called the
“proponent” and the “opponent”, who argue about whether a particular argu-
ment within a formal argumentation framework should be accepted. Discussion
games can be seen as proof procedures for the argumentation semantics they
are associated with, e.g., grounded semantics, preferred semantics, or stable
 Thirteen Challenges in Formal and Computational Argumentation 71

With With
closed-minded unaware Wit closed-minded aware Wit

Fi Fj
a3 a3
a1 a1 a1
a4 a4
a6 a6 a6
a2 a2 a2
a5 a5

Acc Prc Wit Acc Prc Wit Acc Prc Wit

(a) {a1 , a4 , a6 } {a1 , a5 , a2 }

Figure 26: Left: Argumentation by an accused (Acc), by a witness (Wit),

and by a prosecutor (Prc). Middle: multiagent semantics with open-minded
unaware Wit (Fi ). Right: multiagent semantics with closed-minded aware Wit
(Fj ).

semantics. These games provide a local explanation, focussing on the admissi-

bility of arguments.
A discussion game for grounded semantics is won by one agent iff a particular
argument is labeled in. There are two players: Proponent (P) and Opponent
(O). There are four moves with respect to arguments A and B:
P: HTB(A): The labeling of A is in.
O: CB(A): Maybe the labeling of A is out in every complete labeling.
O: CONCEDE(A): Agree that the labeling of A is in in every complete
labeling.
O: RETRACT(A) The labeling of A is out in every complete labeling.
The following are the discussion rules on grounded semantics.
P: HTB(A): Either this is the first move, or:
• the previous move was CB(B), where A attacked B, and:
• no CONCEDE or RETRACT move is applicable.
O: CB(A): A is an attacker of the last HTB(B) statement, which has not yet
been conceded, and:
• the directly preceding move was not a CB statement,
• argument A has not yet been RETRACTed, and
• no CONCEDE or RETRACT move is applicable.
72 Liuwen Yu, Leendert van der Torre, and Réka Markovich

O: CONCEDE(A): There has been a HTB(A) statement in the past, and

• every argument attacking HTB(A) has been RETRACTed, and
• CONCEDE(A) has not yet been moved.
O: RETRACT(A): There has been a CB(A) statement in the past, and:
• there exists an argument attacking CB(A) that has been CON-
CEDEed, and
• RETRACT(A) has not yet been moved.
General rule: No “HTB or CB repeats” are allowed. HTB(A) is only
allowed once, CB(A) is only allowed once. For any A, not both ofHTB(A)
and CB(A) are allowed.
We use Example 4.11 to illustrate how grounded discussion games work.
Example 4.11 (Grounded discussion game). Given the abstract argu-
mentation framework visualized in Figure 27, the grounded discussion game
for argument C proceeds as follows:
(1) P : HT B(c) (3) P : HT B(a) (5) O : RET RACT (b)
(2) O : CB(b) (4) O : CON CEDE(a) (6) O : CON CEDE(c)
P wins the grounded discussion game for argument c, and c is labeled in in
the grounded labeling.

a b c

e f

g h d

Figure 27: An abstract argumentation framework where c is being discussed

Yet, despite its potential advantages, the use of argumentation as balanc-

ing for explanatory purposes has not been adequately explored. This method
employs reason-based [Knoks and van der Torre, 2023] and scale-based balanc-
ing as a decision model. Thus, the explanation could provide an overview of
the pros and cons concerning a decision, like judges do when explaining their
rulings.
Explanations are available to not just the experts who designed the system
(which is then called a white box [Vilone and Longo, 2021]), but also to lay peo-
ple — non-experts who may not understand all the intricacies of certain models.
This issue is about how to personalize explanations, which is particularly rele-
vant given the diversity of backgrounds, contexts, mental states, emotions, and
 Thirteen Challenges in Formal and Computational Argumentation 73

abilities of subjects receiving explanations generated by AI systems (humans

such as patients and healthcare professionals, or virtual intelligent autonomous
agents). To this end, new forms of knowledge representation should be envi-
sioned and synergistically integrated to enable argumentation reasoning over
that.
In this section, we discussed argumentation and explanation. We end this
section with the following open questions:

1. Argumentation and explanation can be related to each other in various

ways. What is the role of argumentation in explanation, and what is the
role of explanation in argumentation? For example, how can the Fatio
dialogue protocol be extended with explanation dialogues?

2. Dialogue is often cited as a distinctive feature of argumentation that

can be used for interactive explanation in human-computer interactions,
but other features may also be relevant. How about, for example, using
balancing or inference in explanation?

3. On the topic of dialogue, it is often observed that argumentation is per-

suasion and that there are many other kinds of dialogues types. How
can different dialogue types be integrated into dialogue systems? For
example, how can information seeking be included in argumentation?

4. Since explanations are often personal, expressed in natural language, and

use common sense reasoning, foundation models and chatbots have been
promoted as part of the explanation toolbox. How can we use LLMs in
argumentation to incorporate context, mental states and emotions?

5. Explanation techniques can be evaluated in terms of the degree to which

they improve a system’s goals, i.e., how does the combination of argu-
mentation and explanation techniques improve system behavior?

4.4 Argumentation technology

We conclude our discussion of computational argumentation by discussing the
integration of argumentation techniques with existing and emerging technolo-
gies in computer science like NLP, LLMs, distributed argumentation technol-
ogy, and dialogue technology. We illustrate the integration of these technologies
by using as an example the integration of argumentation with blockchain tech-
nology into the architecture of the IHiBO.
Recent years have seen remarkable advancements in deep learning, particu-
larly with the development and deployment of LLMs. This presents a significant
opportunity to integrate argumentation. Argumentation is inherently suitable
for enhancing the reasoning and conversational capabilities of LLMs [Bezou-
Vrakatseli, 2023; Castagna et al., 2024c]. Argumentation provides a formal
mechanism for capturing interactions between agents, and it manages the infor-
mation conflicts that arise during these interactions. This makes it an potent-
74 Liuwen Yu, Leendert van der Torre, and Réka Markovich

cial tool for improving the logical consistency and depth of responses generated
by LLMs.
Additionally, LLMs prompts to reevaluate the relationship between abstract
and structured argumentation. Traditionally, structured arguments were nec-
essary because the attack relations were defined based on the internal structure
of the arguments, as discussed in Section 3.2. However, with LLM capabili-
ties, it becomes possible to retain arguments in their natural language form
and use an LLM to extract the underlying argumentation framework. This
approach allows argumentation to be integrated more naturally with conversa-
tional models because LLMs provide the contextual understanding needed to
facilitate these processes.
Furthermore, continuous improvements in computational power, together
with the capabilities of foundational models like LLMs, have opened up new
avenues for integrating argumentation into more complex systems. In such
systems, argumentation can synergize with other technologies, enhancing the
overall functionality and enabling more sophisticated applications. Such inte-
gration will not only advance the field of computational argumentation but will
also push the boundaries of what can be achieved in AI-driven reasoning and
decision-making systems. In this section, we discuss the following challenge:
Challenge 13. Integrating argumentation with technologies.
Distributed argumentation technology [Yu, 2023] is a computational approach
that incorporates argumentation reasoning mechanisms within multiagent sys-
tems. An instantiation of distributed argumentation technology is Intelligent
Human-Input-Based Blockchain Oracle (IHiBO) [Yu et al., 2022]. The mo-
tivation for IHiBO comes from fund management for the securities market.
Figure 28 shows a toy fund management procedure. Investors first pool their
money together and then fund managers conduct investment research and pre-
pare the specific plan for the investment portfolio. Fund managers invest secu-
rities on behalf of their clients (investors) according to their research and the
final decision in the investment plan. The investment generates returns, and
the returns are passed down to investors. Fund managers play an important
role in the investment and financial world as they give investors peace of mind
that their money is in the hands of experts. However, reality is not always
as hoped for and investors are supposed to know (but do not actually know)
where their money is going, why, and how much is the true profit.
A significant aspect of IHiBO is its human-input-based oracle, which bridges
the gap between a blockchain and real-world data, allowing human experts
to input information into the blockchain. IHiBO was envisioned for use in
fund management, where managers provide their arguments in terms of the
investment plan for stocks. Specifically, IHiBO utilizes multiagent abstract
argumentation frameworks to model decision-making processes, which are then
implemented by smart contracts and stored on a blockchain. The integration
of argumentation reasoning and blockchain makes the decision-making process
more transparent and traceable.
 Thirteen Challenges in Formal and Computational Argumentation 75

Figure 28: A fund investment process

IHiBO leverages a two-layer distributed ledger technology (DLT), shown in

Figure 29, to ensure security and immutability of data while maintaining effi-
ciency and scalability. The secondary layer, a private permissioned blockchain,
accessible only to authorized users, facilitates the decision-making smart con-
tract. This layer maintains privacy and reduces transaction costs, providing a
balanced approach to data security and operational efficiency. The main layer is
a public permissionless blockchain such as Ethereum, where the smart contract
for executing stock transactions is invoked by the output of the decision-making
process. This two-layered design is particularly important in fund manage-
ment, where decisions may involve sensitive information. IHiBO’s architecture
supports not only multiagent abstract argumentation but also other kinds of
reasoning that can be encoded in smart contracts.
In this section, we have discussed he integration of argumentation techniques
with existing and emerging technologies, such as IHiBO. We end this section
with the following questions.

1. What can technologies do for argumentation, and what can argumenta-

tion do for technologies? For example, what can foundational models do
for (natural, formal, and computation) argumentation? How can LLMs
be used to develop technologies like IHiBO 2.0? Another discussed ex-
ample outlines the potential roles of computational models of legal ar-
gumentation [Prakken, 2024b]: as tools for guiding prompt engineering,
as benchmarks for evaluating the outputs of legal generative AI, and as
symbolic alternatives to legal generative AI, that could be integrated as
conversational interfaces.

2. As we emphasized in this chapter, conceptualizations of argumentation

can take the form of inference, dialogue or balancing, and these models
have their own formal methods. Do they also have their own technologies?
76 Liuwen Yu, Leendert van der Torre, and Réka Markovich

Figure 29: The architecture of the IHiBO framework. DLT = distributed ledger
technology.

3. In the attack-defense paradigm shift, commutative diagrams play a cen-

tral role. How can these technologies be integrated to make structured
argumentation diagrams commute? How should various technologies be
integrated with structured and abstract argumentation?

4. Another recurring discussion in this chapter is methodology, e.g., how can

we develop a user guide about the new formal methods? Likewise, we
may ask: how can we use new argumentation technologies like IHiBO?

5. This is just the beginning of the use of the attack-defense paradigm shift
in argumentation technology. For a start, how can we use algorithmic
argumentation methods in argumentation technologies?

5 Summary
This chapter has discussed the evolving landscape of argumentation, explor-
ing its natural forms, the paradigm shift initiated by Dung, and subsequent
advancements in computational approaches.
Natural argumentation is rooted in both theoretical and practical reasoning,
with formal argumentation grounded in philosophical and mathematical foun-
dations. This foundational approach is essential for representing, managing,
and resolving conflicts in various disciplines. For instance, the Jiminy ethical
governor, which operates across six layers of conflict, exemplifies the complexity
and depth of formal reasoning in ethical decision-making. However, natural ar-
gumentation is inherently diverse, reflecting the complexities of human thought
and communication. This diversity is evident in psychological evaluations of
arguments, where understanding and generating arguments involves intricate
 Thirteen Challenges in Formal and Computational Argumentation 77

cognitive processes. Foundation models are increasingly employed to construct

and decode arguments, highlighting the importance of questions, particularly
“why” questions, in uncovering weaknesses and justifying decisions. These
questions play a crucial role in frameworks like the Fatio dialogue protocol,
emphasizing the centrality of inquiry in argumentation. Argumentation can be
conceptualized in various ways such as inference, dialogue, or balancing, each
offering distinct perspectives and applications. For example, a divorce court
case can be modeled differently depending on the chosen conceptual framework,
which demonstrates the flexibility of argumentation theories. Additionally, for-
malizing argumentation through a variety of methods allows for a combination
of different reasoning techniques, for instance, in the representation of practical
scenarios like a mother reasoning with her daughter with mixed formal meth-
ods. This highlights the practical utility of formal argumentation for navigating
complex real-world situations.
The paradigm shift in argumentation was significantly influenced by the
attack-defense framework introduced by Dung [1995]. This framework marked
a turning point in argumentation theory by emphasizing that every utterance,
whether a claim, argument, or attack, can be contested. This led to a more
comprehensive and universal approach to analyzing arguments. Structured ar-
gumentation has served as a bridge between classical and nonmonotonic logic,
representing various logic and game theory concepts. This is particularly ev-
ident in the design of nonmonotonic logic, where rationality postulates from
paraconsistent reasoning are crucial. Examples like the weakest versus last
link principles and PDL illustrate the depth and versatility of formal argumen-
tation in designing and representing these logics. Furthermore, extensions to
abstract argumentation frameworks have been developed so that we can extract
more information and incorporate qualitative and quantitative elements such
as bipolarity, preferences, and so on. These extensions, inspired by dialogue
and balancing, have enriched Dung’s abstract argumentation frameworks and
allow for a more nuanced understanding and modeling of complex argumenta-
tive situations.
Computational argumentation has advanced significantly with the develop-
ment of the principle-based approach, which handles the diversity of argumen-
tation models by providing a higher level of abstraction. This approach is
essential for selecting appropriate methods and designing new ones. It ensures
that the diverse landscape of argumentation models can be navigated and ap-
plied effectively. Compositionality principles such as locality, modularity and
transparency are central to the attack-defense perspective in computational
argumentation. These principles are exploited by algorithms and computa-
tional techniques to enhance their efficiency, robustness, and scalability, as
seen in the divide and conquer approach which breaks down complex frame-
works into manageable components. The relationship between explanation and
argumentation was also discussed. Strategic argumentation explains dialogues,
discussion games clarify inference, and underdeveloped reason-based explana-
78 Liuwen Yu, Leendert van der Torre, and Réka Markovich

tions are used for balancing. These connections underscore the importance of
argumentation for making AI systems more transparent and understandable.
Additionally, the integration of argumentation techniques with emerging tech-
nologies such as distributed argumentation technology has expanded the poten-
tial applications of argumentation in areas like blockchain and AI. For instance,
the IHiBO architecture integrates argumentation with blockchain technology
to enhance transparency and trust in decision-making processes.
In summary, this chapter presents an overview of argumentation: past
achievements, the current state of the art, and future directions. We discussed
the three pillars in the context of natural argumentation before discussing the
attack-defense paradigm shift initiated by Dung and advancements in compu-
tational argumentation that are shaping the future of the field.

6 Acknowledgments
All authors acknowledge financial support from the Luxembourg National
Research Fund (FNR) — L. van der Torre through the project The Epis-
temology of AI Systems (EAI) (C22/SC/17111440), L. van der Torre and
R. Markovich through the projects Logical Methods for Deontic Explana-
tions (LODEX) (INTER/DFG/23/17415164/LoDEx) and Study of the Lim-
its, Problems and Risks Associated with Autonomous Technologies (INTE-
GRAUTO) (INTER/AUDACE/21/16695098) and all authors through the
project Deontic Logic for Epistemic Rights (DELIGHT) (O20/14776480). R.
Markovich and L.Yu are also supported by the University of Luxembourg’s
Marie Speyer Excellence Grant of 2024 Formal Analysis of Discretionary Rea-
soning (MSE-DISCREASON).

BIBLIOGRAPHY
[Alchourrón et al., 1985] Carlos E Alchourrón, Peter Gärdenfors, and David Makinson. On
the logic of theory change: partial meet contraction and revision functions. The journal
of symbolic logic, 50(2):510–530, 1985.
[Aleinikoff, 1986] T Alexander Aleinikoff. Constitutional law in the age of balancing. Yale
lj, 96:943, 1986.
[Alkaissi and McFarlane, 2023] Hussam Alkaissi and Samy I McFarlane. Artificial halluci-
nations in chatgpt: implications in scientific writing. Cureus, 15(2), 2023.
[Altay et al., 2022] Sacha Altay, Marlène Schwartz, Anne-Sophie Hacquin, Aurélien Allard,
Stefaan Blancke, and Hugo Mercier. Scaling up interactive argumentation by providing
counterarguments with a chatbot. Nature Human Behaviour, 6(4):579–592, 2022.
[Amgoud and Cayrol, 2002] Leila Amgoud and Claudette Cayrol. Inferring from inconsis-
tency in preference-based argumentation frameworks. International Journal of Approxi-
mate Reasoning, 29(2):125–169, 2002.
[Amgoud and Vesic, 2012] Leila Amgoud and Srdjan Vesic. On the use of argumentation for
multiple criteria decision making. In International Conference on Information Processing
and Management of Uncertainty in Knowledge-Based Systems, pages 480–489. Springer,
2012.
[Amgoud, 2005] Leila Amgoud. An argumentation-based model for reasoning about coali-
tion structures. In International Workshop on Argumentation in Multi-Agent Systems,
pages 217–228. Springer, 2005.
[Amgoud, 2009] Leila Amgoud. Argumentation for decision making. Argumentation in
artificial intelligence, pages 301–320, 2009.
 Thirteen Challenges in Formal and Computational Argumentation 79

[Arisaka et al., 2022] Ryuta Arisaka, Jérémie Dauphin, Ken Satoh, and Leendert van der
Torre. Multi-agent argumentation and dialogue. IfCoLog Journal of Logics and Their
Applications, 9(4):921–954, 2022.
[Atkinson and Bench-Capon, 2005] Katie Atkinson and Trevor Bench-Capon. Legal case-
based reasoning as practical reasoning. Artificial Intelligence and Law, 13:93–131, 2005.
[Atkinson and Bench-Capon, 2021] Katie Atkinson and Trevor Bench-Capon. Value-based
argumentation. Handbook of Formal Argumentation, 2:397–441, 2021.
[Aumann, 2016] Robert J Aumann. Agreeing to disagree. Springer, 2016.
[Austin, 1975] John Langshaw Austin. How to do things with words. Harvard University
Press, 1975.
[Awad et al., 2018] Edmond Awad, Sohan Dsouza, Richard Kim, Jonathan Schulz, Joseph
Henrich, Azim Shariff, Jean-François Bonnefon, and Iyad Rahwan. The moral machine
experiment. Nature, 563(7729):59–64, 2018.
[Baroni and Giacomin, 2007] Pietro Baroni and Massimiliano Giacomin. On principle-based
evaluation of extension-based argumentation semantics. Artificial Intelligence, 171(10-
15):675–700, 2007.
[Baroni et al., 2005] Pietro Baroni, Massimiliano Giacomin, and Giovanni Guida. Scc-
recursiveness: a general schema for argumentation semantics. Artificial Intelligence,
168(1-2):162–210, 2005.
[Baroni et al., 2011] Pietro Baroni, Paul E Dunne, and Massimiliano Giacomin. On the
resolution-based family of abstract argumentation semantics and its grounded instance.
Artificial Intelligence, 175(3-4):791–813, 2011.
[Baroni et al., 2014] Pietro Baroni, Guido Boella, Federico Cerutti, Massimiliano Giacomin,
Leendert van der Torre, and Serena Villata. On the input/output behavior of argumen-
tation frameworks. Artificial Intelligence, 217:144–197, 2014.
[Baroni et al., 2018] Pietro Baroni, Massimiliano Giacomin, and Beishui Liao. Locality and
modularity in abstract argumentation. Handbook of formal argumentation, pages 937–979,
2018.
[Barringer et al., 2005] Howard Barringer, Dov Gabbay, and John Woods. Temporal dy-
namics of support and attack networks: From argumentation to zoology. Mechanizing
Mathematical Reasoning: Essays in Honor of Jörg H. Siekmann on the Occasion of His
60th Birthday, pages 59–98, 2005.
[Baumann et al., 2020] Ringo Baumann, Gerhard Brewka, and Markus Ulbricht. Revis-
iting the foundations of abstract argumentation–semantics based on weak admissibility
and weak defense. In Proceedings of the AAAI Conference on Artificial Intelligence,
volume 34, pages 2742–2749, 2020.
[Baumann, 2012] Ringo Baumann. What does it take to enforce an argument? Minimal
change in abstract argumentation. In ECAI 2012, pages 127–132. IOS Press, 2012.
[Baumeister et al., 2021] Dorothea Baumeister, Daniel Neugebauer, and Jörg Rothe. Collec-
tive acceptability in abstract argumentation. Handbook of Formal Argumentation, Volume
2, 2021.
[Bengel et al., 2024] Lars Bengel, Lydia Blümel, Tjitze Rienstra, and Matthias Thimm.
Argumentation-based probabilistic causal reasoning. In Conference on Advances in Ro-
bust Argumentation Machines, pages 221–236. Springer, 2024.
[Bezou-Vrakatseli, 2023] Elfia Bezou-Vrakatseli. Evaluation of LLM reasoning via argument
schemes. Online Handbook of Argumentation for AI, 4, 2023.
[Bistarelli et al., 2021] Stefano Bistarelli, Francesco Santini, et al. Weighted argumentation.
Handbook of Formal Argumentation, Volume 2, 2021.
[Black et al., 2021] Elizabeth Black, Nicolas Maudet, and Simon Parsons. Argumentation-
based dialogue. Handbook of Formal Argumentation, Volume 2, 2021.
[Boella et al., 2009] Guido Boella, Dov M Gabbay, Leendert van der Torre, and Serena
Villata. Meta-argumentation modelling I: Methodology and techniques. Studia Logica,
93:297–355, 2009.
[Boella et al., 2010] Guido Boella, Dov M Gabbay, Leendert van der Torre, and Serena
Villata. Support in abstract argumentation. In Proceedings of the Third International
Conference on Computational Models of Argument (COMMA’10), pages 40–51. Frontiers
in Artificial Intelligence and Applications, IOS Press, 2010.
80 Liuwen Yu, Leendert van der Torre, and Réka Markovich

[Borg and Bex, 2024] AnneMarie Borg and Floris Bex. Minimality, necessity and sufficiency
for argumentation and explanation. International Journal of Approximate Reasoning,
page 109143, 2024.
[Brewka and Eiter, 1999] G. Brewka and T. Eiter. Preferred answer sets for extented logic
programs. Artif. Intell., 109:297–356, 1999.
[Brewka, 1994] Gerhard Brewka. Reasoning about priorities in default logic. In Barbara
Hayes-Roth and Richard E. Korf, editors, Proc. of the 12th National Conference on AI,
volume 2, pages 940–945. AAAI Press / The MIT Press, 1994.
[Budzynska et al., 2018] Katarzyna Budzynska, Serena Villata, et al. Processing natural
language argumentation. Handbook of formal argumentation, 1:577–627, 2018.
[Caminada and Amgoud, 2005] Martin Caminada and Leila Amgoud. An axiomatic account
of formal argumentation. In AAAI, volume 6, pages 608–613, 2005.
[Caminada and Amgoud, 2007] Martin Caminada and Leila Amgoud. On the evaluation of
argumentation formalisms. Artificial Intelligence, 171(5-6):286–310, 2007.
[Caminada and Ben-Naim, 2007] Martin Caminada and Jonathan Ben-Naim. Postulates for
paraconsistent reasoning and fault tolerant logic programming. PhD thesis, Department
of Information and Computing Sciences, Utrecht University, 2007.
[Caminada and Gabbay, 2009] Martin WA Caminada and Dov M Gabbay. A logical account
of formal argumentation. Studia Logica, 93:109–145, 2009.
[Caminada and Pigozzi, 2011] Martin Caminada and Gabriella Pigozzi. On judgment aggre-
gation in abstract argumentation. Autonomous Agents and Multi-Agent Systems, 22:64–
102, 2011.
[Caminada and Wu, 2011] Martin Caminada and Yining Wu. On the limitations of abstract
argumentation. In Proceedings of the 23rd Benelux Conference on Artificial Intelligence
(BNAIC 2011), pages 59–66, 2011.
[Caminada et al., 2012] Martin WA Caminada, Walter A Carnielli, and Paul E Dunne.
Semi-stable semantics. Journal of Logic and Computation, 22(5):1207–1254, 2012.
[Caminada et al., 2014] Martinus Wigbertus Antonius Caminada, Sanjay Modgil, and Nir
Oren. Preferences and unrestricted rebut. Computational Models of Argument, 2014.
[Caminada, 2006] Martin Caminada. Semi-stable semantics. COMMA, 144:121–130, 2006.
[Caminada, 2017] Martin Caminada. Argumentation semantics as formal discussion. Jour-
nal of Applied Logics, 4(8):2457–2492, 2017.
[Caminada, 2018a] Martin Caminada. Argumentation semantics as formal discussion. Hand-
book of Formal Argumentation, 1:487–518, 2018.
[Caminada, 2018b] Martin Caminada. Rationality postulates: Applying argumentation the-
ory for non-monotonic reasoning. Handbook of Formal Argumentation, Volume 1, pages
771–796, 2018.
[Castagna et al., 2024a] Federico Castagna, Nadin Kökciyan, Isabel Sassoon, Simon Par-
sons, and Elizabeth Sklar. Computational argumentation-based chatbots: a survey. Jour-
nal of Artificial Intelligence Research, 80:1271–1310, 2024.
[Castagna et al., 2024b] Federico Castagna, Peter McBurney, and Simon Parsons.
Explanation–question–response dialogue: An argumentative tool for explainable AI. Ar-
gument & Computation, (Preprint):1–23, 2024.
[Castagna et al., 2024c] Federico Castagna, Isabel Sassoon, and Simon Parsons. Can formal
argumentative reasoning enhance LLMs performances? arXiv preprint arXiv:2405.13036,
2024.
[Cayrol and Lagasquie-Schiex, 2005] Claudette Cayrol and Marie-Christine Lagasquie-
Schiex. On the acceptability of arguments in bipolar argumentation frameworks. In
European Conference on Symbolic and Quantitative Approaches to Reasoning and Un-
certainty, pages 378–389. Springer, 2005.
[Cayrol and Lagasquie-Schiex, 2009] Claudette Cayrol and Marie-Christine Lagasquie-
Schiex. Bipolar abstract argumentation systems. In Argumentation in Artificial Intelli-
gence, pages 65–84. Springer, 2009.
[Cayrol and Lagasquie-Schiex, 2010] Claudette Cayrol and Marie-Christine Lagasquie-
Schiex. Coalitions of arguments: A tool for handling bipolar argumentation frameworks.
International Journal of Intelligent Systems, 25(1):83–109, 2010.
[Cayrol and Lagasquie-Schiex, 2013] Claudette Cayrol and Marie-Christine Lagasquie-
Schiex. Bipolarity in argumentation graphs: Towards a better understanding. Inter-
national Journal of Approximate Reasoning, 54(7):876–899, 2013.
 Thirteen Challenges in Formal and Computational Argumentation 81

[Cayrol et al., 2021] Claudette Cayrol, Andrea Cohen, and Marie Christine La-
gasquie Schiex. Higher-order interactions (bipolar or not) in abstract argumentation:
a state of the art. 2021.
[Cerutti et al., 2021] Federico Cerutti, Marcos Cramer, Mathieu Guillaume, Emmanuel
Hadoux, Anthony Hunter, and Sylwia Polberg. Empirical cognitive studies about for-
mal argumentation. Handbook of Formal Argumentation, Volume 2, 2021.
[Coleman, 1984] James S Coleman. Micro foundations and macrosocial behavior. Ange-
wandte Sozialforschung anc AIAS Informationen Wien, 12(1-2):25–37, 1984.
[Cramer and Guillaume, 2018a] Marcos Cramer and Mathieu Guillaume. Directionality of
attacks in natural language argumentation. In CEUR Workshop Proceedings. RWTH
Aachen University, Aachen, Germany, 2018.
[Cramer and Guillaume, 2018b] Marcos Cramer and Mathieu Guillaume. Empirical cogni-
tive study on abstract argumentation semantics. In Computational Models of Argument,
pages 413–424. IOS Press, 2018.
[Cramer and Guillaume, 2019] Marcos Cramer and Mathieu Guillaume. Empirical study on
human evaluation of complex argumentation frameworks. In Logics in Artificial Intelli-
gence: 16th European Conference, JELIA 2019, Rende, Italy, May 7–11, 2019, Proceed-
ings 16, pages 102–115. Springer, 2019.
[Da Costa et al., 2007] Newton CA Da Costa, Décio Krause, and Otávio Bueno. Paracon-
sistent logics and paraconsistency. In Philosophy of logic, pages 791–911. Elsevier, 2007.
[Da Costa, 1974] Newton CA Da Costa. On the theory of inconsistent formal systems. Notre
dame journal of formal logic, 15(4):497–510, 1974.
[Danry et al., 2023] Valdemar Danry, Pat Pataranutaporn, Yaoli Mao, and Pattie Maes.
Don’t just tell me, ask me: AI systems that intelligently frame explanations as questions
improve human logical discernment accuracy over causal AI explanations. In Proceedings
of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1–13, 2023.
[Dauphin and Cramer, 2018] Jérémie Dauphin and Marcos Cramer. ASPIC-END: struc-
tured argumentation with explanations and natural deduction. In Theory and Applica-
tions of Formal Argumentation: 4th International Workshop, TAFA 2017, Melbourne,
VIC, Australia, August 19-20, 2017, Revised Selected Papers 4, pages 51–66. Springer,
2018.
[D’Avila Garcez et al., 2005] Artur S D’Avila Garcez, Dov M Gabbay, and Luis C Lamb.
Value-based argumentation frameworks as neural-symbolic learning systems. Journal of
Logic and Computation, 15(6):1041–1058, 2005.
[Delgrande et al., 2004] James Delgrande, Torsten Schaub, Hans Tompits, and Kewen
Wang. A classification and survey of preference handling approaches in nonmonotonic
reasoning. Computational Intelligence, 20(2):308–334, 2004.
[Dong et al., 2019] Huimin Dong, Beishui Liao, Reka Markovich, and Leendert van der
Torre. From classical to non-monotonic deontic logic using ASPIC+. In Logic, Ratio-
nality, and Interaction: 7th International Workshop, LORI 2019, Chongqing, China,
October 18–21, 2019, Proceedings 7, pages 71–85. Springer, 2019.
[Dong et al., 2021] Huimin Dong, Réka Markovich, and Leendert van der Torre. Towards
AI logic for social reasoning. arXiv preprint arXiv:2110.04452, 2021.
[Dung and Thang, 2018] Phan Minh Dung and Phan Minh Thang. Fundamental properties
of attack relations in structured argumentation with priorities. Artificial Intelligence,
255:1–42, 2018.
[Dung, 1995] Phan Minh Dung. On the acceptability of arguments and its fundamental role
in nonmonotonic reasoning, logic programming and n-person games. Artificial intelligence,
77(2):321–357, 1995.
[Dung, 2014] Phan Minh Dung. An axiomatic analysis of structured argumentation for
prioritized default reasoning. volume 263 of Frontiers in Artificial Intelligence and Ap-
plications, pages 267–272. IOS Press, 2014.
[Dung, 2016] Phan Minh Dung. An axiomatic analysis of structured argumentation with
priorities. Artif. Intell., 231:107–150, 2016.
[FIPA, 2002] FIPA. Communicative act library specification. http://www. fipa. org/specs/-
fipa00037, 2002.
[Gabbay and Rivlin, 2017] Dov M Gabbay and Lydia Rivlin. Heal2100: human effective
argumentation and logic for the 21st century. The next step in the evolution of logic.
IFCoLog Journal of Logics and Their Applications, 2017.
82 Liuwen Yu, Leendert van der Torre, and Réka Markovich

[Garcez et al., 2008] Artur SD’Avila Garcez, Luis C Lamb, and Dov M Gabbay. Neural-
symbolic cognitive reasoning. Springer Science & Business Media, 2008.
[Gargouri et al., 2020] Anis Gargouri, Sébastien Konieczny, Pierre Marquis, and Srdjan
Vesic. On a notion of monotonic support for bipolar argumentation frameworks. In
20th International Conference on Autonomous Agents and MultiAgent Systems, 2020.
[Georgeff et al., 1999] Michael Georgeff, Barney Pell, Martha Pollack, Milind Tambe, and
Michael Wooldridge. The belief-desire-intention model of agency. In Intelligent Agents V:
Agents Theories, Architectures, and Languages: 5th International Workshop, ATAL’98
Paris, France, July 4–7, 1998 Proceedings 5, pages 1–10. Springer, 1999.
[Giacomin, 2017] Massimiliano Giacomin. Handling heterogeneous disagreements through
abstract argumentation (extended abstract). In PRIMA 2017: Principles and Practice
of Multi-Agent Systems, pages 3–11, 2017.
[Gillioz and Zufferey, 2020] Christelle Gillioz and Sandrine Zufferey. Introduction to exper-
imental linguistics. John Wiley & Sons, 2020.
[Giunchiglia et al., 2004] Enrico Giunchiglia, Joohyung Lee, Vladimir Lifschitz, Norman
McCain, and Hudson Turner. Nonmonotonic causal theories. Artificial Intelligence, 153(1-
2):49–104, 2004.
[Gordon et al., 2007] Thomas F Gordon, Henry Prakken, and Douglas Walton. The
Carneades model of argument and burden of proof. Artificial Intelligence, 171(10-15):875–
896, 2007.
[Gordon, 1993] Thomas F Gordon. The Pleadings Game: An exercise in computational
dialectics. Artificial Intelligence and Law, 2:239–292, 1993.
[Gordon, 2018] Thomas F Gordon. Towards requirements analysis for formal argumentation.
In Pietro Baroni, Dov Gabbay, Massimiliano Giacomin, and Leendert van der Torre, edi-
tors, Handbook of formal argumentation, Volume 1, pages 145–156. College Publictions,
2018.
[Governatori et al., 2021] Guido Governatori, Michael J Maher, and Francesco Olivieri.
Strategic argumentation. Handbook of Formal Argumentation, Volume 2, 2021.
[Hakim et al., 2019] Fauzia Zahira Munirul Hakim, Lia Maulia Indrayani, and Rosaria Mita
Amalia. A dialogic analysis of compliment strategies employed by Replika chatbot. In
Third International conference of arts, language and culture (ICALC 2018), pages 266–
271. Atlantis Press, 2019.
[Hamblin, 1970] C. L. Hamblin. Fallacies. Tijdschrift Voor Filosofie, 33(1):183–188, 1970.
[Heyninck, 2019] Jesse Heyninck. Investigations into the logical foundations of defeasible
reasoning: an argumentative perspective. PhD thesis, Ruhr University Bochum, Germany,
2019.
[Hinton and Wagemans, 2023] Martin Hinton and Jean HM Wagemans. How persuasive
is AI-generated argumentation? An analysis of the quality of an argumentative text
produced by the GPT-3 AI text generator. Argument & Computation, 14(1):59–74, 2023.
[Kaci et al., 2021] Souhila Kaci, Leendert van Der Torre, Srdjan Vesic, and Serena Villata.
Preference in abstract argumentation. Handbook of Formal Argumentation, Volume 2,
2021.
[Kakas and Moraitis, 2003] Antonis Kakas and Pavlos Moraitis. Argumentation based de-
cision making for autonomous agents. In Proceedings of the second international joint
conference on Autonomous agents and multiagent systems, pages 883–890, 2003.
[Kant, 1993] Immanuel Kant. Groundwork of the Metaphysics of Morals. Hackett Publish-
ing Company, Indianapolis, 3rd edition, 1993. [1785].
[Kissine, 2013] Michail Kissine. Speech act classifications. Pragmatics of speech actions,
173:202, 2013.
[Knocks et al., 2024] Aleks Knocks, Muyun Shao, Leendert ver der Torre, Vincent De Wit,
and Liuwen Yu. A principle-based analysis for numerical balancing. In Logics for New-
Generation Artificial Intelligence (LNGAI2024). College Publications, United Kingdom,
2024.
[Knoks and van der Torre, 2023] Aleks Knoks and Leendert van der Torre. Reason-based
detachment. In Logics for New-Generation Artificial Intelligence (LNGAI2023). College
Publications, London, United Kingdom, 2023.
[Kraus et al., 1990] Sarit Kraus, Daniel Lehmann, and Menachem Magidor. Nonmonotonic
reasoning, preferential models and cumulative logics. Artificial intelligence, 44(1-2):167–
207, 1990.
 Thirteen Challenges in Formal and Computational Argumentation 83

[Leite and Martins, 2011] João Leite and João G. Martins. Social abstract argumentation.
In Toby Walsh, editor, IJCAI 2011, Proceedings of the 22nd International Joint Con-
ference on Artificial Intelligence, Barcelona, Catalonia, Spain, July 16-22, 2011, pages
2287–2292. IJCAI/AAAI, 2011.
[Liao and van der Torre, 2024] Beishui Liao and Leendert van der Torre. Attack-defense
semantics of argumentation. In Computational Models of Argument, pages 133–144. IOS
Press, 2024.
[Liao et al., 2019] Beishui Liao, Nir Oren, Leendert van der Torre, and Serena Villata. Prior-
itized norms in formal argumentation. Journal of Logic and Computation, 29(2):215–240,
2019.
[Liao et al., 2023] Beishui Liao, Pere Pardo, Marija Slavkovik, and Leendert van der Torre.
The Jiminy advisor: Moral agreements among stakeholders based on norms and argumen-
tation. Journal of Artificial Intelligence Research, 77:737–792, 2023.
[Longo et al., 2024] Luca Longo, Mario Brcic, Federico Cabitza, Jaesik Choi, Roberto Con-
falonieri, Javier Del Ser, Riccardo Guidotti, Yoichi Hayashi, Francisco Herrera, Andreas
Holzinger, et al. Explainable artificial intelligence (XAI) 2.0: A manifesto of open chal-
lenges and interdisciplinary research directions. Information Fusion, 106:102301, 2024.
[Makinson, 2005] David Makinson. Bridges from classical to nonmonotonic logic. King’s
College, 2005.
[Markovich, 2019] Réka Markovich. On the formal structure of rules in conflict of laws. In
JURIX, pages 199–204, 2019.
[McBurney and Parsons, 2002] Peter McBurney and Simon Parsons. Games that agents
play: A formal framework for dialogues between autonomous agents. Journal of Logic,
Language and Information, 11(3):315–334, 2002.
[McBurney and Parsons, 2004] Peter McBurney and Simon Parsons. Locutions for argu-
mentation in agent interaction protocols. In International Workshop on Agent Commu-
nication, pages 209–225. Springer, 2004.
[McBurney et al., 2021] Peter McBurney, Simon Parsons, et al. Argument schemes and
dialogue protocols: Doug walton’s legacy in artificial intelligence. FLAP, 8(1):263–290,
2021.
[McDougall et al., 2020] Rosalind McDougall, Cade Shadbolt, and Lynn Gillam. The prac-
tice of balancing in clinical ethics case consultation. Clinical Ethics, 15(1):49–55, 2020.
[Mercier and Sperber, 2017] Hugo Mercier and Dan Sperber. The enigma of reason. Harvard
University Press, 2017.
[Modgil and Prakken, 2013] Sanjay Modgil and Henry Prakken. A general account of argu-
mentation with preferences. Artif. Intell., 195:361–397, 2013.
[Modgil and Prakken, 2014] Sanjay Modgil and Henry Prakken. The ASPIC + framework
for structured argumentation: a tutorial. Argument Comput., 5(1):31–62, 2014.
[Moore, 1993] David John Moore. Dialogue game theory for intelligent tutoring systems.
PhD thesis, Leeds Metropolitan University, 1993.
[Musi and Palmieri, 2022] Elena Musi and Rudi Palmieri. The fallacy of explainable gener-
ative AI: evidence from argumentative prompting in two domains. 2022.
[Nash Jr, 1950] John F Nash Jr. Equilibrium points in n-person games. Proceedings of the
national academy of sciences, 36(1):48–49, 1950.
[Niskanen et al., 2018] Andreas Niskanen, Johannes P Wallner, and Matti Järvisalo. Exten-
sion enforcement under grounded semantics in abstract argumentation. In Sixteenth In-
ternational Conference on Principles of Knowledge Representation and Reasoning, 2018.
[Nouioua and Risch, 2010] Farid Nouioua and Vincent Risch. Bipolar argumentation frame-
works with specialized supports. In 2010 22nd IEEE International Conference on Tools
with Artificial Intelligence, volume 1, pages 215–218. IEEE, 2010.
[Nute, 1994] Donald Nute. Defeasible logic. Handbook of logic in artificial intelligence and
logic programming, 3:353–395, 1994.
[Okuno and Takahashi, 2009] Kenichi Okuno and Kazuko Takahashi. Argumentation sys-
tem with changes of an agent’s knowledge base. In Twenty-First International Joint
Conference on Artificial Intelligence. Citeseer, 2009.
[Pardo and Straßer, 2022] Pere Pardo and Christian Straßer. Modular orders on defaults in
formal argumentation. Journal of Logic and Computation, 2022.
84 Liuwen Yu, Leendert van der Torre, and Réka Markovich

[Pardo et al., 2024] Pere Pardo, Liuwen Yu, Chen Chen, and Leendert van der Torre. Weak-
est link, prioritised default logic and principles in argumentation. Journal of Logic and
Computation, 2024. Forthcoming.
[Phillips et al., 2021] P Jonathon Phillips, P Jonathon Phillips, Carina A Hahn, Peter C
Fontana, Amy N Yates, Kristen Greene, David A Broniatowski, and Mark A Przybocki.
Four principles of explainable artificial intelligence. 2021.
[Pigozzi and van der Torre, 2018] Gabriella Pigozzi and Leendert van der Torre. Arguing
about constitutive and regulative norms. Journal of Applied Non-Classical Logics, 28(2-
3):189–217, 2018.
[Polberg and Hunter, 2018] Sylwia Polberg and Anthony Hunter. Empirical evaluation of
abstract argumentation: Supporting the need for bipolar and probabilistic approaches.
Int. J. Approx. Reason., 93:487–543, 2018.
[Pollock, 1987] John L Pollock. Defeasible reasoning. Cognitive science, 11(4):481–518,
1987.
[Pollock, 1992] John L Pollock. How to reason defeasibly. Artificial Intelligence, 57(1):1–42,
1992.
[Pollock, 1994] John L Pollock. Justification and defeat. Artificial Intelligence, 67(2):377–
407, 1994.
[Pollock, 1995] John L Pollock. Cognitive carpentry: A blueprint for how to build a person.
Mit Press, 1995.
[Pollock, 2001] John L Pollock. Defeasible reasoning with variable degrees of justification.
Artificial intelligence, 133(1-2):233–282, 2001.
[Pollock, 2009] John L Pollock. A recursive semantics for defeasible reasoning. In Argumen-
tation in artificial intelligence, pages 173–197. Springer, 2009.
[Pollock, 2010] John L Pollock. Defeasible reasoning and degrees of justification. Argument
and Computation, 1(1):7–22, 2010.
[Prakken and Sartor, 2015] Henry Prakken and Giovanni Sartor. Law and logic: A review
from an argumentation perspective. Artificial intelligence, 227:214–245, 2015.
[Prakken, 2010a] Henry Prakken. An abstract framework for argumentation with structured
arguments. Argument & Computation, 1(2):93–124, 2010.
[Prakken, 2010b] Henry Prakken. Slides on nonmonotonic logic for commonsense reasoning,
2010.
[Prakken, 2013] Henry Prakken. Logical tools for modelling legal argument: a study of
defeasible reasoning in law, volume 32. Springer Science & Business Media, 2013.
[Prakken, 2018] Henry Prakken. Historical overview of formal argumentation. In Handbook
of formal argumentation, pages 73–141. College Publications, 2018.
[Prakken, 2020] Henry Prakken. On validating theories of abstract argumentation frame-
works: the case of bipolar argumentation frameworks. In Proceedings of the 8th Workshop
on Computational Models of Natural Argument (CMNA 2020), Perugia, Italy (and on-
line). CEUR-WS.org, 2020.
[Prakken, 2024a] Henry Prakken. An abstract and structured account of dialectical argu-
ment strength. Artificial Intelligence, 335:104193, 2024.
[Prakken, 2024b] Henry Prakken. On evaluating legal-reasoning capabilities of generative
ai. In Proceedings of the 24th Workshop on Computational Models of Natural Argument,
pages 100–112, 2024.
[Priest, 2002] Graham Priest. Paraconsistent logic. In Handbook of philosophical logic, pages
287–393. Springer, 2002.
[Rahwan et al., 2003] Iyad Rahwan, Sarvapali D Ramchurn, Nicholas R Jennings, Peter
McBurney, Simon Parsons, and Liz Sonenberg. Argumentation-based negotiation. The
Knowledge Engineering Review, 18(4):343–375, 2003.
[Rawls, 2001] John Rawls. Justice as fairness: A restatement. Harvard University Press,
2001.
[Reiter, 1980] Raymond Reiter. A logic for default reasoning. Artificial intelligence, 13(1-
2):81–132, 1980.
[Rienstra et al., 2020] Tjitze Rienstra, Chiaki Sakama, Leendert van der Torre, and Beishui
Liao. A principle-based robustness analysis of admissibility-based argumentation seman-
tics. Argument & Computation, 11(3):305–339, 2020.
 Thirteen Challenges in Formal and Computational Argumentation 85

[Ross et al., 2023] Steven I Ross, Fernando Martinez, Stephanie Houde, Michael Muller,
and Justin D Weisz. The programmer’s assistant: Conversational interaction with a
large language model for software development. In Proceedings of the 28th International
Conference on Intelligent User Interfaces, pages 491–514, 2023.
[Roth and Verheij, 2004] Bram Roth and Bart Verheij. Cases and dialectical arguments–an
approach to case-based reasoning. In On the Move to Meaningful Internet Systems 2004,
pages 634–651. Springer, 2004.
[Russell and Norvig, 2010] Stuart J Russell and Peter Norvig. Artificial intelligence: A
modern approach. London, 2010.
[Savage, 1972] Leonard J Savage. The foundations of statistics. Courier Corporation, 1972.
[Schindler et al., 2020] Samuel Schindler, Anna Drozdzowicz, and Karen Brøcker. Linguistic
intuitions: Evidence and method. Oxford University Press, 2020.
[Searle, 1979] John R Searle. Expression and meaning: Studies in the theory of speech acts.
Cambridge University Press, 1979.
[Sklar and Azhar, 2018] Elizabeth I Sklar and Mohammad Q Azhar. Explanation through
argumentation. In Proceedings of the 6th International Conference on Human-Agent
Interaction, pages 277–285, 2018.
[Straßer and Arieli, 2019] Christian Straßer and Ofer Arieli. Normative reasoning by
sequent-based argumentation. Journal of Logic and Computation, 29(3):387–415, 2019.
[Sycara, 1990] Katia P Sycara. Persuasive argumentation in negotiation. Theory and deci-
sion, 28:203–242, 1990.
[Tennent, 1991] Robert D Tennent. Semantics of programming languages. (No Title), 1991.
[Toulmin, 1958] Stephen E Toulmin. The uses of argument. Cambridge university press,
1958.
[Traum, 1999] David R Traum. Speech acts for dialogue agents. In Foundations of rational
agency, pages 169–201. Springer, 1999.
[Turner, 2004] Hudson Turner. Strong equivalence for causal theories. In Logic Program-
ming and Nonmonotonic Reasoning: 7th International Conference, LPNMR 2004 Fort
Lauderdale, FL, USA, January 6-8, 2004 Proceedings 7, pages 289–301. Springer, 2004.
[Ulbricht and Wallner, 2021] Markus Ulbricht and Johannes P Wallner. Strong explana-
tions in abstract argumentation. In Proceedings of the AAAI Conference on Artificial
Intelligence, volume 35, pages 6496–6504, 2021.
[van der Lee et al., 2021] Chris van der Lee, Albert Gatt, Emiel van Miltenburg, and Emiel
Krahmer. Human evaluation of automatically generated text: Current trends and best
practice guidelines. Computer Speech & Language, 67:101151, 2021.
[van der Torre and Vesic, 2018] Leendert van der Torre and Srdjan Vesic. The principle-
based approach to abstract argumentation semantics. In Pietro Baroni, Dov Gabbay,
Massimiliano Giacomin, and Leendert van der Torre, editors, Handbook of formal argu-
mentation, Volume 1, pages 797–838. College Publictions, 2018.
[Van Laar and Krabbe, 2018] Jan Albert Van Laar and Erik CW Krabbe. The role of ar-
gument in negotiation. Argumentation, 32(4):549–567, 2018.
[Villata et al., 2011] Serena Villata, Guido Boella, and Leendert van der Torre. Attack
semantics for abstract argumentation. In IJCAI. IJCAI/AAAI, 2011.
[Vilone and Longo, 2021] Giulia Vilone and Luca Longo. Classification of explainable artifi-
cial intelligence methods through their output formats. Machine Learning and Knowledge
Extraction, 3(3):615–661, 2021.
[Visser et al., 2020] Jacky Visser, John Lawrence, and Chris Reed. Reason-checking fake
news. Communications of the ACM, 63(11):38–40, 2020.
[Von Neumann and Morgenstern, 1947] John Von Neumann and Oskar Morgenstern. The-
ory of games and economic behavior, 2nd rev. 1947.
[Walton and Krabbe, 1995] Douglas Walton and Erik CW Krabbe. Commitment in dia-
logue: Basic concepts of interpersonal reasoning. State University of New York Press,
1995.
[Walton et al., 2008] Douglas Walton, Christopher Reed, and Fabrizio Macagno. Argumen-
tation schemes. Cambridge University Press, 2008.
[Walton, 1990] Douglas N Walton. What is reasoning? What is an argument? The Journal
of Philosophy, 87(8):399–419, 1990.
[Walton, 2013] Douglas Walton. Argumentation schemes for presumptive reasoning. Rout-
ledge, 2013.
86 Liuwen Yu, Leendert van der Torre, and Réka Markovich

[Weinstock, 2006] Michael P Weinstock. Psychological research and the epistemological

approach to argumentation. Informal Logic, 26(1):103–120, 2006.
[Williamson and Gabbay, 2005] Jon Williamson and Dov Gabbay. Recursive causality in
Bayesian networks and self-fibring networks. Laws and models in the sciences, pages
173–221, 2005.
[Young et al., 2016] Anthony P. Young, Sanjay Modgil, and Odinaldo Rodrigues. Prioritised
default logic as rational argumentation. In Catholijn M. Jonker, Stacy Marsella, John
Thangarajah, and Karl Tuyls, editors, Proceedings of the 2016 International Conference
on Autonomous Agents & Multiagent Systems, pages 626–634. ACM, 2016.
[Young et al., 2017] Anthony P. Young, Sanjay Modgil, and Odinaldo Rodrigues. On the
interaction between logic and preference in structured argumentation. In Elizabeth Black,
Sanjay Modgil, and Nir Oren, editors, Theory and Applications of Formal Argumentation
- 4th International Workshop, volume 10757, pages 35–50. Springer, 2017.
[Yu and Gabbay, 2022] Liuwen Yu and Dov Gabbay. Case-based reasoning via comparing
the strength order of features. In International Workshop on Explainable, Transparent
Autonomous Agents and Multi-Agent Systems, pages 143–151. Springer, 2022.
[Yu et al., 2020] Liuwen Yu, Réka Markovich, and Leendert Van Der Torre. Interpretations
of support among arguments. In Legal Knowledge and Information Systems, pages 194–
203. IOS Press, 2020.
[Yu et al., 2021] Liuwen Yu, Dongheng Chen, Lisha Qiao, Yiqi Shen, and Leendert van der
Torre. A Principle-based Analysis of Abstract Agent Argumentation Semantics. In Pro-
ceedings of the 18th International Conference on Principles of Knowledge Representation
and Reasoning, pages 629–639, 11 2021.
[Yu et al., 2022] Liuwen Yu, Mirko Zichichi, Markovich Réka, Najjar Amro, et al. Intel-
ligent human-input-based blockchain oracle (IHiBO). In Proceedings of the 14th Inter-
national Conference on Agents and Artificial Intelligence (ICAART 2022), pages 1–12.
SCITEPRESS, 2022.
[Yu et al., 2023] Liuwen Yu, Caren Al Anaissy, Srdjan Vesic, Xu Li, and Leendert van der
Torre. A principle-based analysis of bipolar argumentation semantics. In European Con-
ference on Logics in Artificial Intelligence, pages 209–224. Springer, 2023.
[Yu, 2023] Liuwen Yu. Distributed Argumentation Technology. PhD thesis, 2023.