Issues related to housing stock retrofitting

Get Complete Project Material File(s) Now! »

Argumentation for decision support


In the introduction, we hypothesized that the argumentative approach, and more precisely, abstract argumentation framework (AAF), could be a relevant decision support for participatory approach. In order to be able to justify this assertion, we must first explain what is the Abstract Argumentation Framework (AAF). It requires to introduce the notions of argument and argumentation, before analyzing how researchers in mathematics and computer science have grasped these notions to compose a new field of research: the computational argumentation. We will study this field, tracing a state of the art of computational argumentation tools in order to define whether the literature proposes methods or tools that meet the needs of our decision problem. Without wanting to divulge too much about this analysis, it will lead to the lack of tools adapted to real-time multi-actors decision support (i.e. participatory framework). This will lead us to suggest that abstract argumentation could be an interesting solution to meet this need, but also we will expose the limits that we will have to overcome with it. This will allow us to conclude regarding the existing works, but also to pose the challenges that the method we will propose in the following chapters, will have to meet in order to meet our need for decision support.

Introduction of argumentation

What is argumentation

Argumentation in philosophy

The word « argument » comes from the Latin arguere (to make bright, enlighten, make known, prove, etc.) (Harper, 2020). In usual language, as defined by Cambridge Dictionary, an argument is1: (1) a disagreement, or the process of disagreeing; or (2) a reason or reasons why you support or oppose an idea or suggestion, or the process of explaining these reasons. By extension, argumentation is a set of arguments used to explain something or to persuade people.
The study of argumentation was already an important topic in antiquity (Breton and Gauthier, 2000). The argumentation theory has known different developments through philosophy and sciences History, and notably in the middle of the XXth century with the works of Toulmin Toulmin (2003) and Perelman and Olbrechts-Tyteca Perelman and Olbrechts-Tyteca (2000) studying the natural arguments in a more descriptive way (Tremblay, 2012). It led to different approaches, which can be combined, to study argumentation (Tremblay, 2012): pragmatic rhetorical, pragmatic, dialectical and logical. The pragmatic approach aims at studying argumentation in a broader context of communication; in this approach, the explicit and implicit expression of arguments is considered. The rhetorical approach studies the impact of argumentation on its audience (e.g. effectiveness in persuasion). The dialectical approach analyzes the context of discursive exchanges between stakeholders and are interested in the structure of the exchanges and its internal rules. Finally, the logical approach analyzes the form and validity of arguments, as well as the supporting or attacking relationships between them.
In the present work, we are interested in argumentation in a decision-making context, from the expression of arguments in natural language to the determination of the validity of arguments. Indeed, decision-making can be seen as an activity aiming at building valid arguments through which the actors involved in a decision-making process, will seek to establish that a subset of possible options, possibly containing only one option, is the best. In a decision support context, several approaches are relevant: the pragmatic approach, which allows us to account for the way an argument is expressed in natural language, the dialectical approach which is interested in the structure of exchanges, and the logical approach which seeks to determine the validity of arguments; the latter, as we shall see, is the one most explored in mathematics/computing science, because it is the easiest to implement in computer logic. According to this perspective, we can define an argument as a set of statements consisting of a conclusion and at least one premise, linked together by a logical connection (Breton and Gauthier, 2000). An argument will support or discredit an option or another argument.

Argumentation in debate

Since Antiquity, it has been considered that argumentation requires a situation of controversy in which the different stakeholders express arguments in order to defend their opinion or attack the adverse opinion (Emediato and Damasceno-Morais, 2019). Argumentation is based on representations (topics) that somehow connect the participants, either through sharing (doxic connivance) or through opposition (dissonance) (Emediato and Damasceno-Morais, 2019). Stakeholders can provide arguments or be the aim of arguments. But if argumentation is built around controversy, it is also built on agreements and consensus, allowing participants to be members of a community of discourse (Perelman and Olbrechts-Tyteca, 2000).
So, argumentation takes place in an interactional framework and in a space of interdiscursivity (Emediato and Damasceno-Morais, 2019). Interdiscursivity refers to how discourses circulate, or alternatively drift, and how they are dialectically exported (decontextualized) and imported (recontextualized) between various sites and occasions of enunciation, and notably during a debate (Tracy et al., 2015). Argumentation implies an intersubjective investment in which the arguing person position herself/himself in a social dialogue, regarding consensus (belonging to a community) and disensus (reaction to the discourse of others). Plantin Plantin (2009) presents the argumentation as an activity taking place in a space organized by a tension between enunciative and interactional work; the speaker built a continuous planned intervention in which she/he chains together good reasons and exposes a « coherent world »; the consistency of this coherent world is not insured, because the coherence is only internal to the stakeholders vision. The argumentation is then the opposition between the « coherent worlds » defend by the stakeholders, or a sharing of antagonistic visions on the world which is more or less coherent.
Through debates and controversies, one can thus not only analyze the debated elements that meet with consensus or dissensus, which is useful in a decision-making framework, but also better understand the positioning of the different stakeholders, their points of view on the subject of the debate and on the world, and their preferences. Debates and the analysis of the induced argumentation is therefore a very rich tool in terms of decision support. It is for this reason, and because arguments carry an internal logic, that the field of artificial intelligence has become interested in argumentation.

Argumentation in computer science

In the 1990s, the study of argumentation theory was extended to computer science and artificial intelligence (Bench-Capon and Dunne, 2007). Then, Argumentation has become a major subject of artificial intelligence in the last two decades (Diller et al., 2018). Computational argumentation has largely benefit from the computer development and a very large number of tools have emerged in the last twenty years. This led to the development of numerous approaches and models based on
arguments. These models and approaches have different objectives (decision-aiding, analyze speech…) and different interpretations and uses of the notion of argument according to their domain of application. It composes a rich corpus of knowledge spreading from different disciplines (computer science, linguistic, etc.). The common point of all these tools is to propose a formalization of the arguments and argumentation. Some are only interested in modelling and representation of the argument (e.g. argument mapping), others allow to make different inferences with (e.g. defining acceptable arguments, shared arguments…) and to draw conclusions from an argumentative debate, or to automatically extract arguments from natural language (argument mining). In this section, we will analyze computational argumentation, regarding firstly the models of argument, then, the argument visualization and finally the logic reasoning.

Models of arguments

Computational argument used different models of argument, depending on the aim of the tool. Literature provides several theoretical models of arguments; the most used are Toulmin argument model Toulmin (2003) and IBIS(Ebadi et al., 2009) (Figure 1.1.1). Toulmin’s argument model provides a set of categories allowing to represent the logical structure of arguments organized in a graphical form (Figure 1.1.1a). Toulmin’s model is composed of six concepts: Datum, Claim, Warrant, Rebuttal, Backing and Qualifier. A datum is a fact or an observation about the situation under discussion. It is the basis of a claim, which corresponds to controversial, observation, prediction, or characterization. Datum and Claim are linked by a Warrant which corresponds to an inference rule. The tuple Data, Claim, W arrant corresponds to the inferential core of the argument. The Qualifier defines the stance toward a claim or registers the degree of certainty. Rebuttal indicates the conditions under which the claim can be taken as true. Backing of an argument is some knowledge structure which serves to justify the warrant. The IBIS (Issue-based information system) model consists of three elements (Figure 1.1.1b): Issue, Position and Argument. Issues are questions which are at the center of the debate, e.g. “What should we do about X?” where X is the issue that is of interest to a group. Positions are potential answers to the issue. Arguments can be Pros (arguments for) or Cons (arguments against) a position. IBIS elements are usually represented as nodes, and associations between elements are represented as directed edges (arrows).
These models are really useful to analyze arguments in a debate or as a theoretic support to build an argument. But the use of such argument models (Toulmin or IBIS) in a context of a debate in participatory approaches is quite complex. Indeed, the different stakeholders have limited knowledge of theoretical argumentation and it has been shown that conflict resolution cannot be solved with complex argument structures due to the human argumentation process (Sillince and Saeedi, 1999). Moreover, since participatory approach requires real time debate, to argue and identify the type of argument at the same time would drastically increase the complexity for non-experts. Therefore, regarding our decision problem, we suggest to use a simpler model. Some works propose for instance to use only a dual consideration of arguments: an argument supports or defeats an idea (pros and cons principle) (Amgoud et al., 2008). But these works does not consider the nature of the argument itself (e.g. is the argument is the object of the debate or a statement about this object). It could lead to some misunderstanding of the notion of argument. As a conclusion, we state that the literature lacks of an intermediary model of argument, sufficiently simple to be used in real time by any stakeholders, but sufficiently complete to avoid misunderstanding about its meaning.

Formalizing argumentation

In a debate between different stakeholders, each argument is not only related to the subject matter, but also to some of the other proposed arguments forming an argument structure. This process of argumentation is informal and does not follow logical or mathematical reasoning (van Bruggen et al., 2003). This lead generally to ill-structured problems (Reitman, 1965, Simon, 1973) which: 1) have an ambiguous and incomplete problem specification; 2) lack clear-cut criteria to evaluate whether a solution has been reached, implying that there are no stopping rules; 3) make use of several potential information sources that may be used to represent problem spaces although it is unclear which ones should be used and how they should be integrated; and 4) have neither a complete enumeration of applicable operators nor a predetermined path from initial state to goal state. This issue lead to the development of different tools and methods aiming at representing the argumentative structure; these tools belong to the Computer-Supported Argumentation Visualization (CSAV) domain. It has a double objective: ensure that the participant to the debate is able to understand the structure of the argumentation and ensure that this structure is consistent according to logical reasoning, which is interpretable by computers. In this perspective different argument modeling and visualisation software propose to formalize, model and display the argument structure (Kirschner et al., 2003, van Bruggen et al., 2003, Yun et al., 2016). All these tools allow to provide an argument mapping which is a visual representation of the structure of an argument. The argument map varies according to the approach and model used, but they can include the conclusions, the premises, the objections, the counterarguments, the rebuttals (Davies, 2012)… In terms of application, the argument mapping tools (and CSAV) can be used to analyze an argumentative debate, to teach argumentation (e.g. in a legal course), to help structuring arguments… This capacity of representation of the argumentation is really interesting in a debate, in order to help the facilitator to structure the debate. Furthermore, it can be completed with a tool able to perform inference on the argumentative structure.

Argument inferred information

The computational argumentation systems can provide more information than just a formalization of the arguments. Different approaches, methods and tools were provided in order to infer knowledge from the formalized argument or set of arguments. These inferences can consist in providing information on the state of arguments (e.g. accepted or rejected) and/or to define a « winner » (Karacapilidis et al., 2009). As we previously exposed with the argument definition (i.e. a set of statements consisting of a conclusion and at least one premise, linked together by a logical connection), argument lays on a logical structure. The same result can be given regarding the argumentation which is composed of a set of connected arguments. They are two levels of validity regarding argument: internal (is argument valid in itself) and external (is argument valid in relation with other arguments). So, two types of approaches exist in computational argumentation (Besnard et al., 2020): structured argumentation and abstract argumentation. Structured argumentation are about building arguments and identifying their relationships, whereas abstract argumentation consider that a set of arguments in interaction is given without considering the way they were built or their meaning. Structural argumentation question the internal validity of an argument: an argument is structurally valid if considering that all its premises are true, the conclusion is necessarily true. At the opposite, an argument is non-valid if even the premises are true, the conclusion is not necessarily true. A sound argument is a valid argument with true premises, whereas an unsound argument has at least one false premise. Abstract argumentation focuses on computing the validity of an argument, in relation to other arguments. In abstract argumentation, the internal consistency of the argument is not questioned; an argument in itself is assumed to be valid. However, an argument which is attacked by a valid argument is considered as non valid. The validity of the arguments is then defined from the set of arguments depending on their relation (attack and support). These two approaches produce a high number of works proposing to use structural argumentation logic (e.g. (Franqueira and Horsman, 2020, Hahn, 2020, Yu et al., 2015)) or abstract argumentation logic (e.g. (Amgoud et al., 2008, Dung, 1995, Flouris and Bikakis, 2019)) to assess the validity of an argument. Some works (but many less) are positioned at the interface of these two approaches as (Prakken, 2010). An exhaustive state of the art regarding works using one of these two approaches, or a hybrid approach, is not within our reach. However, it is worth noting that many works in artificial intelligence have attempted to propose, based on internal or external logic, to evaluate the validity of arguments. We will be able to largely rely on this work to reach our goal.

READ  The routes for CO2 storage and utilization

Argument mining

Argumentation mining aims at automatically extracting arguments from unstructured textual documents. It becomes an important domain of research since mid-2010s, due to its potential in processing information in innovative ways (Lippi and Torroni, 2016). In a decision-making context, the interest of such approach is very important. It could allow, during a debate, to automatically model the arguments expressed by the participants, using predefined structure (e.g. Toulmin argument model, abstract argumentation…). Argument mining process is composed of two stages: Arguments’ extraction and relations’ prediction. The first stage consists in the identification of arguments within the natural language text. This involves the detection of the argument components (e.g. claim, premises) and the identification of their textual boundaries. The second stage consists in predicting what are the relations holding between the arguments identified in the first stage. This second stage is particularly challenging. It requires a high-level knowledge representation and reasoning issues.
Many approaches shave recently been proposed to address these two stages as Support Vector Machines (SVM) (Niculae et al., 2017, Stab and Gurevych, 2017), Naïve Bayes classifiers (Duthie et al., 2016), or other methods and techniques inherited from deep learning (Lytos et al., 2019) … But the remaining work to make them really operational in a context such as a real time debate between citizens is still enormous. We are only in the early stages and no work can accomplish such a task today. This is all the more true since most of the work is positioned on a text in English and to our knowledge, there is no argument mining work attacking the French language. Obviously, language is not neutral in the use of these approaches and the multilinguality issue is still to solve (Cabrio and Villata, 2018). The use of argument mining in our context is then a perspective but not yet possible. In our work, the arguments must be manually structured according to a predefined argument structure.

Conclusion on Argumentation

In this section, we explored the notions of argument and argumentation, from concepts used and analyzed by philosophers and linguists to their used in Artificial Intelligence. Obviously, these notions are particularly rich and this section has just scratched the surface of their meaning, but it has laid the groundwork for how these concepts can be used in this work. Before getting to grips with them, we need to study the way in which researchers have used the argumentation through the development of tools dedicated to different objectives, and in particular to decision support. Thus, the next section proposes a state of the art of argumentation tools for decision support.

A state of the art on argumentation tools for decision support


In this part, we investigate tools for argumentation, specifically dedicated to decision-making. In order to establish a perimeter for this study, we consider only tools which are presented or discussed in scientific publications. The research is then based on an analysis of the scientific literature: paper presenting a specific tool or comparing/analyzing a set of tools. The research methodology was based on the following steps: (1) Identification of the target: computational tools for argumentation-based decision support, (2) Scientific database search, using the keywords ”Tool and Argumentation and Computer and Decision-making” on Scopus (33 documents) and ISI Web of Science (22 documents). (3) Selection of tools: we extracted all the tool described or cited in these articles, (4) Tool Analysis. To ensure the relevance of analyzed paper, the following conditions to limit the set of papers were used: (i) we consider only papers from journals and conferences, (ii) we consider only papers written in English, (iii) we limit the perimeter to papers presenting implemented tools and (iv) we consider autonomous tools (e.g. we exclude extension calculation engines for abstract argumentation frameworks as Bistarelli et al. (2018)). Based on these considerations, 25 articles were collected. This set of papers presents 37 tools that will be analyzed in this section. Table 1.1.1 provides the list of tools and the corresponding references. We distinguished 7 main purposes for the tools:
• CSAV (Computer-Supported Argument Visualization): enables to formalize and graphically represent the different arguments
• CSCA (Computer-Supported Collaborative Argumentation): allows a group of stakeholders to formalize arguments
• CDM (Collaborative Decision-Making): allows a group of stakeholders to make decisions based on arguments
• CP (Collaborative Planning): allows a group of stakeholders to collaboratively plan a project
• ABDM (Argument-Based Decision-Making): supports decision-making, basing the decision on arguments
• ABD (Argument-Based Design): supports design process for a product, basing it on arguments
• AR (Argumentation Reasoning): enables to automatically perform reasoning on arguments


The 37 tools propose, each, a support for decision-making, more or less directly, based on the modeling of arguments. However, few are really relevant regarding our application context, i.e. real-time debate support in a participatory perspective. Collaborative planning (CP) and Argument-Based Design (ABD) are specifically dedicated to a precise decision task (respectively planning and design) which does not correspond to our objective which considers participatory decision-making in a more general way. Some other tools, such as EcoBioCAP is based on a database that is not generic and includes some specificities related to the application domain (biodegradable and bio-sourced packaging for EcoBioCAP). Its use in another context would thus require adaptation work for each new domain of application. Furthermore, it is not currently available or freely accessible online. This last point is a major limitation of the use of several of these tools. Indeed, some tools were developed into a project and were not maintain and/or made available after the end of their respective project. Moreover, our application context requires a tool capable of formalizing arguments with an interface usable in real time debate. Some of the cited tools do not provide such an interface and require to enter command lines (e.g. SI-COBRA, ABEL, DeLP) or text files (e.g. Argumap, SAKE). These tools are not relevant in our context.
We can divide the remaining tools into two categories: argument representation tools and argument reasoning tools. The first category contains tools which aim only to formalize arguments and argument debate. Most of CSAV, CSCA and CDM belong to this category. For instance, Argvis, CoPe_it! and SPeCS do not provide any inference engine. These tools are useful to structure and understand argument structure, but not to generate new knowledge from inference. We assume that being able to perform automated reasoning on the argument structure is very useful in a real-time debate context. Indeed, this can make it possible to highlight, in real time, the points of conflict, the preferred solutions, the key arguments, etc. Such information can enrich a debate.
Regarding applications, two questions arise: Are these tools actually used (i.e. on real case studies) and if yes, with which purpose and in which context. A first observation is that among the 37 tools that are referenced, only few have been the subject of an application (or have at least one application referenced in the articles). For instance, literature provides no application for ConArg. This demonstrates that many of these tools are positioned at a theoretical level without being intended to be used for real applications. Few domains have been addressed by these tools, e.g. housing market trend (Pendo), cryptography (ABEL)… None of these tools have been used for real-time debate case-study.


None of the tools presented here are able to meet the requirements needed (real time argument input, inference engine, output usable in real-time debate, etc.). It is therefore necessary to develop an argument computational tool dedicated to real-time debates. In this perspective, we make the assumption that abstract argumentation framework provides enough inference rules – when a semantic is given to arguments (i.e. hybrid approach between structural and abstract argumentation) – to find out either a solution in an debate or pinpoint a conflict.

Argumentation frameworks

 Abstract argumentation framework

The AAF or Dung’s Argumentation Framework (DAF), developed by Dung (1995), is based on the principle that an argument is considered acceptable as long as no other argument attacks it or if attacked, it is defended by other arguments. Dung defines an AAF as an oriented graph where nodes are arguments and edges are attacks between arguments (Figure 1.1.2). The term ”argument” should not be misunderstood with the common interpretation such as ”a set of statements composed of a conclusion and at least one premise, linked together by a logical link” (Breton and Gauthier, 2000). In Dung’s framework, an argument is an abstract notion which has no semantic meaning; it can represent a data, a proposal or anything else. The notion of attack has also no direct semantic meaning; it is also an abstract concept. Based on these principles, Dung provides different definitions.


Definition 1. An argumentation framework is a pair AF = A, R with:
• A, a set of arguments,
• R, the binary relation on A such as for an argument ai attacking aj , (ai, aj) ∈ R.
Definition 2. A set of arguments S ⊆ A is said conflict-free iff ∀ai, aj ∈ S, the arguments do not attack each other: (ai, aj) ∈/ R.
Definition 3. An argument ai ∈ A is acceptable with respect to a set S ⊆ A iff ∀aj ∈ A, if aj attacks ai then aj is attacked by S. Being attacked by a set S means being attacked by an argument in S. Definition 4. A set of arguments S ⊆ A is admissible if and only if S is conflict-free and each argument in S is acceptable with respect to S. The empty set is always admissible.

Semantics of acceptation

Dung called extension a set of arguments sharing common properties. Each extension is built according to different rules called semantics.
Definition 5. A set S ⊆ A is a preferred extension of an argumentation framework if and only if S is admissible and maximal (for set inclusion).
Definition 6. A set S ⊆ A is a stable extension if and only if S is admissible and ∀ai ∈ S, ∀aj ∈ S, (ai, aj) ∈ R.
Lemma 1. A stable extension is therefore an admissible set that attacks all the arguments that do not belong to it. It means that a stable extension is a preferred extension but the reciprocal is false. Definition 7. A set S ⊆ A is a complete extension iff S is admissible and each acceptable argument with respect to S belongs to S.
Lemma 2. A preferred extension is complete.
Definition 8. A set S ⊆ A is a grounded extension if it is the smallest (for set inclusion) complete extension. A grounded extension is unique.
Based on these definitions, Dung provided two inference rules in order to define the accepted arguments: credulous and skeptical. In a credulous inference, an argument is accepted if it belongs to at least one preferred extension. In a skeptical one, an argument is accepted if it belongs to the grounded extension. To sum up, the acceptability of arguments according to Dung’s AAF as its belonging to preferred (credulous) or grounded (skeptical) extension. Compute these extensions allows, thus, to assess the validity of the arguments.

Table of contents :

1 Argumentation 
1.1 Argumentation for decision support
1.1.1 Introduction
1.1.2 Introduction of argumentation What is argumentation Argumentation in philosophy Argumentation in debate Argumentation in computer science Models of arguments Formalizing argumentation Argument inferred information Argument mining Conclusion on Argumentation
1.1.3 A state of the art on argumentation tools for decision support Methodology Analysis Conclusion
1.1.4 Argumentation frameworks Abstract argumentation framework Definitions Semantics of acceptation Existing tools for extensions computation Derived argumentation frameworks Value-base argumentation framework Bipolar argumentation framework Probabilistic Argumentation Framework Which Argumentation framework for our decision problem? Using an abstract argumentation framework in collective decisions
1.1.5 Conclusion
1.2 An argumentation interface for participative approach
1.2.1 Introduction
1.2.2 Designing an argumentation model for non-experts : Argued Discussion Model
1.2.3 Analyzing debate in real-time with AIPA General principles Translation rules Conclusion translation StatementAgainst translation StatementFor translation Support translation Optimistic translation Providing indications regarding the graph status
1.2.4 Implementation
1.2.5 Applications Application 1: real-time debate Application 2: participatory approach
1.2.6 Conclusion
1.3 Coupling MCDA methods with AIPA: ArguedMCDA 52
1.3.1 Introduction
1.3.2 MultiCriteria decision analysis Why multicriteria? The different approaches of MCDA Synthetic criterion Synthetic outranking Interactive process Which approach or method should be used?
1.3.3 Methodology Why combining MCDA with argumentation ArguedMCDA Process
1.3.4 Application to a simple illustrative case-study Discussion on alternatives Discussion on criteria Discussion on criteria importance Running MCDA method Discussion on MCDA results Conclusion on the example
1.3.5 Conclusion
2 Building retrofitting 
2.1 Context of application
2.1.1 Introduction
2.1.2 Issues related to housing stock retrofitting Societal issue Operational issue
2.1.3 Decision problem characteristics Stakeholders Social landlord Tenant Work team Decision process Alternatives Criteria and Objectives Complexity of the decision problem
2.1.4 State of the art regarding the social housing stock retrofitting
2.1.5 The REHA-PARCS proposition A three-steps process Reference buildings identification Optimization Decision support
2.1.6 Conclusion
2.2 Application to a case study 84
2.2.1 Introduction
2.2.2 Applying AMHORE to a case study Presentation of the case study Preamble Example of the construction of an AIPA graph for criterion selection, step by step Results Step 1. Criteria selection Step 2. Preferences parameters Step 3. Running the MCDA Step 4. Discussion the MCDA results
2.2.3 Discussion Discussion related to the AMHORE process Interest of the ArguedMCDA approach Leading the discussions Added value of WebAIPA interface during the debate Methodological discussion MCDA method Dung AAF Explicative argument Sensitivity and validation
2.2.4 Conclusion
Summary of the chosen approach
Methodological perspectives
Applicative perspecives
Past projects
Current projects
Conclusion [Français]
Résumé de l’approche choisie
Perspectives méthodologiques
Perspectives d’application
Projets passés
Projets en cours


Related Posts