Trichotomic Argumentation Representation

by   Merlin Göttlinger, et al.

The Aristotelian trichotomy distinguishes three aspects of argumentation: Logos, Ethos, and Pathos. Even rich argumentation representations like the Argument Interchange Format (AIF) are only concerned with capturing the Logos aspect. Inference Anchoring Theory (IAT) adds the possibility to represent ethical requirements on the illocutionary force edges linking locutions to illocutions, thereby allowing to capture some aspects of ethos. With the recent extensions AIF+ and Social Argument Interchange Format (S-AIF), which embed dialogue and speakers into the AIF argumentation representation, the basis for representing all three aspects identified by Aristotle was formed. In the present work, we develop the Trichotomic Argument Interchange Format (T-AIF), building on the idea from S-AIF of adding the speakers to the argumentation graph. We capture Logos in the usual known from AIF+, Ethos in form of weighted edges between actors representing trust, and Pathos via weighted edges from actors to illocutions representing their level of commitment to the propositions. This extended structured argumentation representation opens up new possibilities of defining semantic properties on this rich graph in order to characterize and profile the reasoning patterns of the participating actors.



There are no comments yet.


page 1

page 2

page 3

page 4


Using Semantic Wikis for Structured Argument in Medical Domain

This research applies ideas from argumentation theory in the context of ...

Enacting Social Argumentative Machines in Semantic Wikipedia

This research advocates the idea of combining argumentation theory with ...

An Approach to Characterize Graded Entailment of Arguments through a Label-based Framework

Argumentation theory is a powerful paradigm that formalizes a type of co...

Online Handbook of Argumentation for AI: Volume 1

This volume contains revised versions of the papers selected for the fir...

A Matrix Approach for Weighted Argumentation Frameworks: a Preliminary Report

The assignment of weights to attacks in a classical Argumentation Framew...

Representing Pure Nash Equilibria in Argumentation

In this paper we describe an argumentation-based representation of norma...

Discrete Argument Representation Learning for Interactive Argument Pair Identification

In this paper, we focus on extracting interactive argument pairs from tw...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Argumentation plays a central role in society for forming rational opinions about controversial topics by providing a way to resolve conflicting views. For understanding complex argumentative discussions it is helpful to draw diagrams called argument maps that visualize how the propositions and arguments interact with one another. Having such a formal representation for argumentation also allows for having computational semantics for them to aid humans in evaluating argumentation.

Over the years there have been many developments of such argumentation frameworks with varying information content and complexity. One of the simplest forms of argumentation frameworks are the so called Abstract Argumentation Frameworks that treat arguments as abstract objects without having attached internal structure.

Abstract Argumentation Frameworks

consist of a graph having arguments as nodes without any additional structure or information apart from the relations connecting them to other arguments. Dung [19] popularized the simplest form of these frameworks (see 1) with only a single relation representing attack between arguments.

Definition 1 (Dung Aaf [20])

An abstract argumentation framework is composed of a set of arguments and an attack relation . An argument is called an attacker of if . A set of arguments attacks if there exists s.t. . defends if attacks each attacker of . is conflict-free if it does not attack its own arguments. is admissible if it is conflict-free and defends each of its arguments. The characteristic function of is defined by . is:

  • a stable extension if it is conflict-free and attacks each argument ;

  • a preferred extension if it is a maximal (w.r.t. set inclusion) admissible set of arguments;

  • a complete extension if it is admissible and contains each argument it defends (or equivalently a conflict-free fixed point of );

  • a grounded extension if it is the least complete extension (or equivalently the least fixed point of ).

While being algorithmically and conceptionally simple, having only an attack relation limits the expressivity of the framework. In argumentation, it is often the case that one argument supports the truth of another argument, and it is not always easy to express this as an attack. Bipolar Argumentation Frameworks [15] add a support relation between arguments to capture those situations. The semantics of BAFs is a priori somewhat less clear-cut than that of AAF, however, as one needs to determine how support and attach interact.

One approach, extensively discussed by Cohen et al. [17], is to derive complex attacks from joining support and attack relations and thus reduce to Dung semantics. Another way is to move from crisp notions of accept and reject towards a weighted acceptance where supporters and attackers are accumulated into the acceptability of an argument [1, 29, 2]. The weighted approach however comes at the cost of requiring acyclicity of the graph as evident in the overview of current bipolar semantics by Amgoud and Ben-Naim [1].

Structured Argumentation Frameworks

go beyond this abstract view of argumentation and attach structure to the argument map. There are multiple points where one can add information to obtain a richer argument representation. One way is to add structure to the nodes, e.g. in the form of logical formulas representing the propositions and reasoning patterns used in an argument [5, 13, 3, 28]. This allows using logical reasoning to find conflicts and inferences between the arguments and generating new arguments from existing facts and inference rules. Structure can also be added to the edges in the form of identifying different kinds of inferences, conflicts, or exceptions. Argument structure and the classification of different types of arguments have been studied widely [39, 33, 42, 45, 41].

Toulmin [39] proposed that arguments generally obey a basic structure, commonly referred to as the Toulmin Scheme, consisting of identifiable parts that justify why a listener should believe the qualified conclusion. Perelman and Olbrechts-Tyteca [33] informally categorized different types of arguments and analysed what makes them convincing. Walton [42] combined these efforts and created a compendium of different Argumentation Schemes with a common structure of identifying multiple premises necessary to reasonably derive the truth of an argument’s claim. Additionally, he identified so called critical questions specific to each scheme, which a critically thinking audience can put forth to undermine arguments employing that scheme, as in the following example.

Example 1 (Scheme: Argument from Position to Know [45])
Major Premise:

Source is in a position to know about things in a certain subject domain containing proposition

Minor Premise:

asserts that is true (false)


is true (false)

Critical Questions:
  1. Is in a position to know whether is true (false)?

  2. Is an honest (trustworthy, reliable) source?

  3. Did assert that is true (false)?

For an in-depth review of this style of argumentation schemes as well as a historical overview of classifying different patterns of argumentation see the book by Walton et al. 

[45]. Due to their nature of clearly identifying premises and claim of an argument, Walton-style schemes are a good candidate for complex edge labels in an argumentation map.

The Aif

is a formal representation of argumentation general enough to include many existing representations in order to establish a standard for storing and exchanging argumentation [16]. Inferences in AIF can be labelled with instances of rule schemes which can be Walton–style argumentation schemes but are not limited to those. Some of these schemes are implemented in the AIF ontology, which specifies the overall format and accompanying concepts intended as labels for the participating propositions.

To represent argumentative discourse between multiple actors, the AIF was later extended to AIF+ by Modgil and McGinnis [27] and by Reed et al. [36], who introduce new types of nodes representing the dialogue structure through dialogue moves relating locutions.

IAT introduced the idea to incorporate illocutionary force into argument maps for relating locutions and illocutions as well as transitions and inferences/conflicts [12, 10, 11]. Having such links between statements and their content also allows specifying requirements on force edges, e.g. that the speaker is trustworthy. This can be seen as capturing ethical requirements similar to premises of Walton-style argumentation schemes, expressing that a witness is trustworthy in an Argument from Witness Testimony. In the extension by Reed et al. [35], AIF+ already has the capabilities necessary to represent IAT annotated argumentation maps.

Recently, Lawrence et al. [22] formalized an adjunct ontology to AIF+ called the S-AIF, which incorporates the authors of locutions into the model. They use this to calculate various statistics about the discussion, such as participation and agreement-based author clustering. We build on this idea in Section 2 where we propose the Trichotomic T-AIF, which formally captures even more aspects of argumentation. S-AIF distinguishes between the roles of Argument Web participants and is tailored towards augmented argument construction and statistics, while our aim is to represent existing arguments in more detail.

Argumentation in AIF+ is represented through different types of connected nodes with corresponding semantic meaning. The nodes in the actual argumentation graph are instances of the concepts from the upper ontology with related instances of concepts from the forms ontology specifying the schemes for the S-nodes as well as concepts to classify the I-nodes into the various forms (see the appendix).

I-nodes represent Illocutions and hence contain the propositions relevant for the arguments as premises, exceptions, and conclusions. The AIF defines concepts necessary to classify these nodes to participate in Walton-style argumentation schemes.

L-nodes represent Locutions and contain the raw utterances that make up the represented dialogue. They can be seen as illocutions in the sense that they carry the propositional content that their text was uttered in the conversation. In S-AIF, L-nodes are connected with nodes representing the authors of the locution.

S-nodes represent Schemes for interconnecting nodes. These schemes are associated with scheme application nodes for the various kinds of interconnections between nodes:


Instances of rule schemes, such as Walton-style schemes.


Instances of transition schemes representing speech acts in dialogue.


Instances of conflict schemes capture that truth of propositions is in conflict.


Instances of preference schemes, which resolve conflicts via preference rules.


Instances of illocutionary schemes modelling illocutionary force as a link between locutions, illocutions, inferences, and transitions.

The semantics of AIF graphs is defined by translating into ASPIC [6] and then evaluating using, e.g., the TOAST algorithm [38].


is a framework for building structured argumentation systems with a strong logical background following the ideas of Besnard and Hunter [5] where the nodes are sets of formulas representing a derivation tree [34, 6]. They distinguish between strict and defeasible inferences and require the syntax of the logic language of the nodes to contain each instance of a defesaible inference to allow attacks on inferences. Apart from this form of attack they also have attacks on premises and conclusions via a unidirectional relation specifying that one formula is contrary to another. ASPIC builds on an extended notion of abstract argumentation system [40] and was generalised to ASPIC+ [34] by partitioning inferences and facts into an undeniable and a defeasible part [26].

2 The Trichotomic T-Aif

Aristotle [4] noted that argument content is not the only relevant aspect for human argument evaluation. He distinguished between three means of persuasion forming the so called Aristotelian Trichotomy:


Appeal to logic. The structure of the argumentation both of individual reasoning steps and the overall interaction of multiple arguments, e.g. argumentation schemes used.


Appeal to authority. The properties of the speaker relevant for the evaluation of argument, e.g. credibility or moral values.


Appeal to emotion. The emotional aspects of an argument, e.g. enthusiasm of the speaker or intended emotional reaction from the audience.

Surprisingly, to our knowledge all formal representations for argumentation are mostly focused on the logos aspect of argumentation. The T-AIF aims to incorporate all three aspects into the formal representation to allow representing arguments in a richer way. It consists of three interconnected parts representing different aspects of the argumentation.

2.1 Trust Network

Building on ideas of Lawrence et al. [22], we include the speakers into the argumentation representation. Other entities such as politicians, news organisations, or other groups that might participate in the discussion through citation or quotation are also included. We call these E-nodes given that they represent Entities as unifying concept over active and passive participants of the conversation.

While Lawrence et al. use their speaker nodes mostly for statistics and grouping, we also use them to represent ethos and pathos through weighted edges in the graph. The ethos aspect is incorporated by weighted edges between E-nodes capturing the notion of trust of active participants towards other entities forming a trust network [32]. One could have multiple different trust relations between the entities to capture the different means of trust described in the literature [14, 18, 21, 31, 37] but in the following we are mostly concerned with trust as the confidence an actor has in another actor’s utterances.

The grade of commitment of the speakers towards their illocutions is a pathos aspect represented in the graph. Distinguishing between different levels of commitment or tracking the commitment of actors is mostly studied in the context of dialogue games [44, 24, 25, 46, 11] but we believe it has relevant interconnections with a speaker’s ethos. Strongly committing to propositions obligates an actor to defend them when challenged and should have a negative effect on her trustworthiness when this fails [30]. In contrast, an actor voicing arguments for the opposing view besides her own should receive increased trust due to seeming well-informed and not being afraid of attacks on her own stance.

2.2 Dialogue Structure

This part of the argumentation representation captures the raw locutions the actors have put forth. The Locutions are represented as textual labels in L-nodes. Connecting the L-nodes via the reply relation forms a lattice structure. Intuitively one would assume this to form a tree structure but an utterance may reference multiple other locutions resulting in multiple parents.

Apart from the reply relation, locutions are connected to moves from a dialogue game giving meaning to their interconnections. Moreover, they are connected to illocution nodes (described in Section 2.3) via illocutionary force edges as seen widely in previous work [35, 12, 10, 11]. Note that in the interest of clarity, our representation uses typed labelled edges to represent illocutionary force and dialog relations in favour of having YA- and TA-nodes.

It is beneficial to keep the locutions of the dialogue in the representation of the argumentation to enable a broader applicability of the format as a source structure containing both argumentation and dialogue. This structure serves as witness for the temporal development of the argumentation, and thus e.g. allows penalizing unfavourable moves or decisions in the dialogue game as proposed by Walton [43]. Penalizing or rewarding based on logical content as briefly mentioned at the end of Section 2.1 also becomes expressible as a result of having the locutions as nodes in T-AIF graphs.

2.3 Argument Map

Due to unknown background knowledge of the discussion participants and enthymemes dominating real argumentative discussions, we cannot assume the represented argument maps to be complete without unknown parts. From the dialogue relations we can infer argumentative relations even if the precise reasoning patterns in use are not clear. We assume that the participants of the discussion voice complaints if the connections between propositions are unclear to a human participant with relevant background knowledge. Hence, we assume that enthymemes and skipped reasoning steps are not fallacious but covered by knowledge available at the time.

Our argument map is similar to the one defined in the AIF specification [16] but our I-nodes are labelled with logical formulas (without committing to a particular logic) representing the propositional content of the Illocution. Attacks and supports have schemes similar to the ones proposed by the AIF but restrict the shape of participating propositions according to their nature (e.g. as specified by Walton [42] or Parsons et al. [31]). The Application of these schemes is represented in the graph by SA-nodes and AA-nodes for Support and Attacks, respectively. Exceptions to these schemes are represented as propositions of specified shape connecting to the application nodes directly. We allow E-nodes to be connected via attack and support nodes to express ethical requirements on entities or infer ethical properties of these respectively.

We leave out the preference nodes from AIF in our formalism as preferences are inherently a per entity concept and our argument map represents all actors’ contributions. Preferences of the actors are not inexpressible this way but are rather expressed through trust and commitment in the sense that an actor likely prefers her own arguments or those made by highly truted entities over those made by untrusted entities.

Due to elided reasoning steps the uttered propositions in real argumentative conversations might not have the required shapes for the reasoning patterns they are involved in (see 2). Filling these reasoning steps from common sense and common knowledge is an easy task for humans (albeit not uniquely) but a very hard task for machines as analysed by Boltužić and Šnajder [8].

Example 2 (Elided Steps in Typical Reasoning)

‘Experts say that Brexit would hurt the economy, so we should vote against Brexit.’ Here the main argumentation scheme involved is Argument from Expert Opinion based on the fact that experts are quoted. The drawn conclusion by the definition of that scheme [45] would be that ‘Brexit would hurt the economy’ and not that ‘UK citizens should vote against Brexit’.

We resort to accepting this lack of information and capturing the unknown reasoning patterns by having default inferences connecting the perceived to the required proposition. So in the case of 2 we would have the conclusion of the argument from expert opinion be that ‘Brexit would hurt the economy’ and add a default inference from there to ‘UK citizens should vote against Brexit’. Similarly, we add required premises and exceptions to the scheme applications to complete enthymemes without committing entities to these.


EU harms UK

EU has bad intentions

EU harms UK fishermen

EU must harm EU fishermen

EU harms Norway fishermen

Norway is in the EU

EU protects employees

EU protects nature










Figure 1: Simplified extract from a Twitter discussion between A and B on Brexit. Red arrows represent attack while green arrows represent support. The weighted black edges represent the commitment of the actors to the propositions.

3 Semantics

For reasoning about T-AIF we define a reduced representing structure and define properties on it as fuzzy formulas to give an example of what is possible within this formalism. We use the standard Łukasiewicz semantics of fuzzy logic (see Lukasiewicz and Straccia [23] for an overview):

Definition 2 (Łukasiewicz Semantics [23])

In particular, is interpreted into in Łukasiewicz Logic, which matches the intuition we intend in our formulas. All sets mentioned in the following definitions are assumed to be finite unless mentioned otherwise.

Definition 3 ( Trichotomic Argumentation Framework (T-Af))

A Trichotomic Argumentation Framework is a structure over sets

  • of propositions formed from illocutions and entities ,

  • of atomic argumentation schemes formed from attack schemes and support schemes ,

with additional data

  • representing the arity of schemes,

  • representing the exception arity of schemes,

  • representing the interpretation of atomic schemes,

  • representing atomic argumentation actions, relating the givens or premises and the exceptions to the claim or conclusion, and

  • representing belief.

Remark 1 (Entities as Propositions)

Argumentation actions can include entities as premises and conclusions. An entity seen as a proposition is to be read as the trustworthiness of the entity. Therefore, an Ad Hominem argument against entity  would be an argumentation action with  as conclusion prescribing low trustworthiness. Trust establishment or propagation schemes as described by Parsons et al. [31] can be modelled as inferences perscribing high trustworthiness to an entity. Additionally, treating entities as propositions enables the trust relations decribed in Section 2.1 to be incorporated into the belief predicate .

Example 3 (Kinds of inferences in T-AFs)

Note that the signature of the scheme interpretation function has as additional argument the acceptability of the conclusion. It can therefore be seen as a fuzzy predicate over the combination of givens, exceptions, and claim where givens and exceptions are the propositions related to a claim by an argumentation action. This allows expressing different concepts, and combinations thereof, via this single notion:

  • Exceptions can produce a predicate that accepts all truth values of the claim as the exception defeats the argumentation action;

  • Weighted premises as seen in the specification of Walton-style schemes can be incorporated through linear combination of the givens;

  • Necessary conditions can produce a predicate specifying that the claim can not be accepted unless the givens are accepted: (read, e.g., as a fuzzy formula). This is what Boudhar et al. [9] call necessary support;

  • Sufficient conditions can produce a predicate specifying that the claim has to be accepted when the givens are accepted: . This is what Boella et al. [7] call deductive support;

  • Inhibiting conditions can produce a predicate specifying that the claim can not be accepted unless the givens are rejected: ;

  • Disjunctive conditions can produce a predicate specifying that the givens or the claim need to be accepted: ;

Similarly, belief is interpreted as a predicate rather than a single truth value, thus capturing belief and disbelief as well as uncertainty.

Definition 4 (Scheme Interpretation)

The scheme interpretation is required to adhere to the given type of argumentation scheme: An attack (support) is interpreted as a fuzzy predicate that allows increasing (decreasing) acceptance of the claim if the acceptance of the premises increases. This is made explicit in the following constraints for scheme interpretation:

Example 4 (Logically: Argument from Position to Know)

Consider a logical representation of Argument from Position to Know (1) using an epistemic modality , read “ believes that”, and a modality  for locution, read “ said that”:

Critical questions 1 and 3 are modelled as attacks against truth of the premises while question 2 would be an attack against an added honesty premise . As E-nodes are treated as propositions for their honesty the third premise would be represented simply by . Note that the inference is deductively valid and therefore can have no exceptions but only attacks on premises. This also leads to a uniform weighting of premises; hence the interpretation for the scheme would correspond to the formula involving illocutions , , and  as well as the entity .

Due to having diverse and joint notions of inference, forming complex attacks to derive sensible semantics becomes more challenging [1, 17].

Definition 5 (Composite Schemes)

Define the extended sets of complex support (attack) schemes () and their union of complex schemes inductively:

  • and

  • Given schemes and define the composite scheme . Depending on wether was an attack or a support, belongs to or , respectively.

The composite scheme is interpreted as follows:

Note that or may be infinite if there are cycles in the graph that allow infinite composition.

For classifying extensions in our T-AFs, we follow the approach of graded bipolar frameworks, which define a fuzzy labelling specifying the acceptability of an argument. Using graded acceptance, uncertainty and preferences can be incorporated.

Definition 6

A labelling for a T-AF is a function  assigning to each proposition a degree of acceptance. Let denote the set of all labellings.

Properties of labellings and actors can then be expressed as fuzzy formulas over T-AFs. Entities can be grouped based on the similarity of their beliefs, e.g. in order to reason about group dynamics.

Definition 7 (Similarity)

Two actors  in a T-AF are considered similar if they have voiced similar belief.

Definition 8 (Attack & Support)

A proposition  is attacked under a given labelling  if there is an attack (complex or atomic) with accepted givens, rejected exceptions, and claim :

The definition of support is similar but with .

In Dung-style AAFs (1), the semantics is mostly defined in terms of a notion of defense, which needs to be suitable adapted in the presence of supports and exceptions:

Definition 9 (Defense)

A proposition  is defended under a labelling  if every attack against is defeated. That is, either a given of the attack is attacked or an exception is supported by accepted propositions:

In AAFs, the notion of attack induces a notion of conflict-freeness. For T-AFs, we generalize this concept to a notion of consistency effectively incorporating both attack and support.

Definition 10 (Consistency)

A labelling  for a T-AF is consistent if the assigned labels are in accordance with the argumentation actions of the system.

(To see that this does indeed generalize the standard notion of conflict-freeness, introduce a unary attack scheme and equip it with mutual exclusion between the given and the claim as the schema interpretation.)

Remark 2

Note that can be seen as the evaluation of the formula under the labelling  but we prefer to write it as defined because we later need to talk about different labellings in the same formula.

The extensions in the style of Dung [19] become fuzzy predicates defined in terms of our definition of defense and consistency — that is, a given labelling will not be absolutely stable, preferred etc., but rather have these properties to a certain degree, in accordance with our general consideration of weights. We emphasize that this applies in particular also to preferred and grounded labellings, which are now ‘maximal’ or ‘least’ not in an absolute but in a fuzzy sense.

Definition 11 (Admissible Labelling)

A labelling  is admissible if it is consistent and each accepted proposition is defended in :

Definition 12 (Stable Labelling)

A labelling  is stable if it is consistent and each rejected proposition is attacked under :

Definition 13 (Preferred Labelling)

A labelling is preferred if it is a maximal admissible labelling:

Definition 14 (Complete Labelling)

A labelling is complete if it is admissible and accepts each defended proposition:

Definition 15 (Grounded Labelling)

A labelling is grounded if it is a least complete labelling of the T-AF:

Apart from translating the definitions of Dung AAFs [19] we can define exemplary properties of the actors in the dialogue specific to our extended setting.

Definition 16 (Agreement)

A labelling  agrees with an actor  if ’s beliefs are retained in :

Having defined agreement opens possibilities to express interesting properties like rationality of an actor.

Definition 17 (Rationality)

An actor  is rational if her belief can be extended to as a consistent position:

Definition 18 (Justified Trust)

An actor  has justified trust if the belief of actors trusts can be consistent in the same labelling:

“Aktoren denen x glaubt” ist jetzt etwas unschön finde ich.

An actor  having justified trust not only expresses that the actors trusts have rational beliefs but also that these beliefs are consistent with eachother.

Definition 19 (Trust Compliance)

A labelling  is trust compliant for an actor  if belief of actors trusted by  is reflected in :

Trust compliance expresses that trusts her trusted actors for validity in the sense of Demolombe [18] and reflects that in her beliefs (the labelling ). Combining all trusted actors via disjunction captures a similar notion to parallel path composition as defined by Parsons et al. [32].

4 Aspic+ Translation

Given that the argument map in T-AIF is very similar to AIF and hence can be easily converted, we obtain ASPIC+ semantics using an existing translation [6]. That translation can be improved with the additional information available in our extended formalism to obtain per-actor semantics from the translation.

For constructing the argumentation theory we modify the translation [6] by forming a separate knowledge base for each actor. An ordering on the formulas in the knowledge base can be derived from the trust network via the accumulated trust and commitment of each author. This semantically results in different systems for each participant that reflect her own beliefs in the propositions as well as her trust in the other entities participating in the discussion.

5 Conclusion and future research

We have proposed the Trichotomic , aimed at representing argumentation resulting from dialogue between multiple actors, capturing aspects from all three areas of the aristotelian trichotomy. Our format is inspired by the AIF and especially its extensions AIF+ and S-AIF. The ethos aspect is captured by relating entities via trust and allowing their ethical properties to be involved in arguments. Relating entities to their illocutions incorporates the strength of their commitment as an aspect from the domain of pathos. Finally, the argument map represents the logos aspect of argumentation, capturing the logical connections between the propositions. Given that the AIF and the existing extending formalisms are specified using the Web Ontology Language (OWL), we plan to formalize our format in an ontology as well.

Our main contribution in this paper is to provide a formalism for representing and reasoning about more than just the logos aspect of argumentation. This enables a very natural treatment, e.g., of Ad Hominem arguments, which affect the acceptance of all arguments made by the attacked actor. Having the dialogue game [27, 36] in the representation can interact nicely with ethical aspects of argumentation by e.g. penalizing trustworthiness when illogical moves are made or undefendable stances are taken. The representation is also a good basis for deriving a per-participant semantics incorporating voiced beliefs and trust relations of actors to approximate what they are likely to believe.

We have proposed T-AFs as an exemplary simplified version of T-AIF, aimed at defining properties profiling the participating actors. T-AFs do not include the dialogue part of T-AIF, and there are meta-arguments — that is, arguments talking about the discussion itself, which in fact occur rather frequently — that become expressible when the dialogue is included in the structure. An example of this kind are arguments against an actor’s credibility on the premise that she has retracted previous statements she had high commitment to when counterarguments were given.

Since the tooling around the AIF ecosystem only supports some features of T-AIF, we plan to provide tool support for creating, storing, and visualizing argument structures in our format. The envisioned tool support will also include reasoning capabilities beyond the described ASPIC+ translation for working on a given T-AIF graph directly.

As maybe apparent from our choice of examples, we plan to evaluate our formalism on a dataset of argumentative discourse on Twitter, specifically concerning Brexit (pre-referendum). Ultimately, we intend to construct T-AIF representations automatically from written discussions using a structured argumentation mining algorithm.


  • [1] L. Amgoud and J. Ben-Naim. Evaluation of arguments in weighted bipolar graphs. pp. 25–35.
  • [2] L. Amgoud and J. Ben-Naim. Weighted bipolar argumentation graphs: Axioms and semantics. In J. Lang, ed.,

    International Joint Conference on Artificial Intelligence, IJCAI 2018

    , pp. 5194–5198.
  • [3] L. Amgoud and P. Besnard. Bridging the gap between abstract argumentation systems and logic. In L. Godo and A. Pugliese, eds., Scalable Uncertainty Management, SUM 2009, vol. 5785 of LNCS, pp. 12–27. Springer.
  • [4] Aristotle. On Rhetoric: A Theory of Civic Discourse. OUP USA.
  • [5] P. Besnard and A. Hunter. A logic-based theory of deductive arguments. 128:203–235.
  • [6] F. Bex, H. Prakken, and C. Reed. A formal analysis of the aif in terms of the aspic framework. In Computational Models of Argument, COMMA 2010, pp. 99–110.
  • [7] G. Boella, D. M. Gabbay, L. W. N. van der Torre, and S. Villata. Support in abstract argumentation. In Computational Models of Argument, COMMA 2010, pp. 111–122.
  • [8] F. Boltužić and J. Šnajder. Fill the gap! analyzing implicit premises between claims from online debates. In Argument Mining, ArgMining@ACL 2016.
  • [9] I. Boudhar, F. Nouioua, and V. Risch. Handling preferences in argumentation frameworks with necessities. In International Conference on Agents and Artificial Intelligence, ICAART 2012, pp. 340–345.
  • [10] K. Budzynska, M. Janier, J. Kang, C. Reed, P. Saint-Dizier, M. Stede, and O. Yaskorska. Towards argument mining from dialogue. In Computational Models of Argument, COMMA 2014, pp. 185–196.
  • [11] K. Budzynska, M. Janier, C. Reed, and P. Saint-Dizier. Theoretical foundations for illocutionary structure parsing. 7(1):91–108.
  • [12] K. Budzynska and C. Reed. Speech acts of argumentation: Inference anchors and peripheral cues in dialogue. In Proceedings of 11th International Conference on Computational Models of Natural Argument (CMNA 2011).
  • [13] M. Caminada and L. Amgoud. On the evaluation of argumentation formalisms. 171(5-6):286–310.
  • [14] C. Castelfranchi and R. Falcone. Principles of trust for MAS: Cognitive anatomy, social importance, and quantification. In Multi Agent Systems, ICMAS 1998, pp. 72–79. IEEE Comp. Soc.
  • [15] C. Cayrol and M. Lagasquie-Schiex. On the acceptability of arguments in bipolar argumentation frameworks. In L. Godo, ed., Symbolic and Quantitative Approaches to Reasoning with Uncertainty, ECSQARU 2005, vol. 3571 of Lecture Notes in Computer Science, pp. 378–389. Springer.
  • [16] C. Chesñevar, J. McGinnis, S. Modgil, I. Rahwan, C. Reed, G. Simari, M. South, G. Vreeswijk, and S. Willmott. Towards an argument interchange format. 21(04):293.
  • [17] A. Cohen, S. Parsons, E. I. Sklar, and P. McBurney. A characterization of types of support between structured arguments and their relationship with support in abstract argumentation. 94:76–104.
  • [18] R. Demolombe. Reasoning about trust: A formal logical framework. pp. 291–303.
  • [19] P. M. Dung.

    On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games.

  • [20] P. M. Dung and P. M. Thang. Representing the semantics of abstract dialectical frameworks based on arguments and attacks. pp. 1–19.
  • [21] A. Herzig, E. Lorini, J. F. Hübner, and L. Vercouter. A logic of trust and reputation. 18(1):214–244.
  • [22] J. Lawrence, M. Snaith, B. Konat, K. Budzynska, and C. Reed. Debating technology for dialogical argument: Sensemaking, engagement, and analytics. 17(3):1–23.
  • [23] T. Lukasiewicz and U. Straccia. Managing uncertainty and vagueness in description logics for the semantic web. 6:291–308.
  • [24] P. McBurney and S. Parsons. Agent ludens: Games for agent dialogues. In Game-Theoretic and Decision-Theoretic Agents (GTDT 2001): Proceedings of the 2001 AAAI Spring Symposium, pp. 70–77.
  • [25] P. McBurney and S. Parsons. Games that agents play: A formal framework for dialogues between autonomous agents. 11(3):315–334.
  • [26] S. Modgil. Revisiting abstract argumentation frameworks. In E. Black, S. Modgil, and N. Oren, eds., Theory and Applications of Formal Argumentation, TAFA 2013, vol. 8306 of Lecture Notes in Computer Science, pp. 1–15. Springer.
  • [27] S. Modgil and J. McGinnis. Towards characterising argumentation based dialogue in the argument interchange format. pp. 80–93.
  • [28] S. Modgil and H. Prakken. The aspic+ framework for structured argumentation: a tutorial. 5(1):31–62.
  • [29] T. Mossakowski and F. Neuhaus. Modular Semantics and Characteristics for Bipolar Weighted Argumentation Graphs.
  • [30] G. Ogunniye, A. Toniolo, and N. Oren. Meta-argumentation frameworks for multi-party dialogues. pp. 585–593.
  • [31] S. Parsons, K. Atkinson, Z. Li, P. McBurney, E. Sklar, M. Singh, K. Haigh, K. Levitt, and J. Rowe. Argument schemes for reasoning about trust. 5(2-3):160–190.
  • [32] S. Parsons, Y. Tang, E. Sklar, P. McBurney, and K. Cai. Argumentation-based reasoning in agents with varying degrees of trust. In Autonomous Agents and Multiagent Systems, AAMAS 2011, pp. 879–886.
  • [33] C. Perelman and L. Olbrechts-Tyteca. The New Rhetoric: A Treatise on Argumentation. University of Notre Dame Press.
  • [34] H. Prakken. An abstract framework for argumentation with structured arguments. 1(2):93–124.
  • [35] C. Reed, S. Wells, K. Budzynska, and J. Devereux. Building arguments with argumentation: the role of illocutionary force in computational models of argument. In Computational Models of Argument, COMMA 2010, pp. 415–426.
  • [36] C. Reed, S. Wells, J. Devereux, and G. Rowe. Aif+: Dialogue in the argument interchange format. In Computational Models of Argument, COMMA 2008, pp. 311–323.
  • [37] L. Rodenhäuser. A matter of trust.
  • [38] M. Snaith and C. Reed. Toast: Online aspic+ implementation. In Computational Models of Argument, COMMA 2012, pp. 509–510.
  • [39] S. Toulmin. The Uses of Argument. Cambridge University Press.
  • [40] G. Vreeswijk. Abstract argumentation systems. 90(1-2):225–279.
  • [41] J. Wagemans. Constructing a periodic table of arguments.
  • [42] D. Walton. Argumentation Schemes for Presumptive Reasoning. Routledge.
  • [43] D. Walton. Logical Dialogue-Games and Fallacies. University Press of America Inc.
  • [44] D. Walton and E. Krabbe. Commitment in Dialogue: Basic Concepts of Interpersonal Reasoning. State University of New York Press.
  • [45] D. Walton, C. Reed, and F. Macagno. Argumentation Schemes. Cambridge University Press.
  • [46] S. Wells and C. A. Reed. A domain specific language for describing diverse systems of dialogue. 10(4):309–329.

Appendix 0.A Appendix










Upper Ontology














Forms Ontology
Figure 2: The AIF+ Ontology. The boxes are concepts and the arrows represent the inclusion relation. L-nodes, TA-nodes, and Transition Schemes [36] as well as YA-nodes and Illocutionary Schemes [35] were added to AIF.