Ebook: Computational Models of Argument

The investigation of computational models of argument is a rich, interdisciplinary, and fascinating research field with two ultimate aims. A theoretical goal is to understand argumentation as a cognitive phenomenon by modelling it in computer programmes, while a practical goal is to support the development of computer-based systems able to engage in argumentation-related activities with human users or among themselves. These ambitious research goals involve the study of natural, artificial, and theoretical argumentation and, as such, requires openness to interactions with a variety of disciplines, such as philosophy, cognitive science, linguistics, communication studies, formal logic, game theory and mathematical graph theory.

The computational study of argumentation has two main historic origins. In 1987 John Pollock published his seminal paper Defeasible reasoning, in which he stressed the importance of reasons in the construction of arguments and gave the first systematic formal account of the evaluation of arguments given their internal structure and their relation with counterarguments. And in 1995 Phan Minh Dung’s paper On the acceptability of arguments and its fundamental role in nonmonotonic reasoning, logic programming and n-person games initiated the study of so-called abstract argumentation frameworks, which leave the nature of arguments and their relations unspecified but still allow for a rich theory of argument evaluation. This paper was in 2018 awarded the AI Journal Classic Paper Award, to recognise its role in making argumentation a mainstream research topic in artificial intelligence.

Since 2006 the biennial International Conference on Computational Models of Argument (COMMA) has provided a dedicated forum for presentation and discussion of the latest advancements in this interdisciplinary field, covering both basic research and innovative applications. The first COMMA was supported by the EU 6th Framework Programme project ASPIC and was hosted by the University of Liverpool in 2006. After the event, a steering committee promoting the continuation of the conference was established and, since then, the steady growth of interest in computational argumentation research worldwide has gone hand in hand with the development of the conference itself and of related activities by its community. Since the second edition, organized by IRIT in Toulouse in 2008, plenary invited talks by world-leading researchers and a software demonstration session became an integral part of the conference programme. The third edition, organized in 2010 by the University of Brescia in Desenzano del Garda, saw the addition of a best student paper award. The same year, the new journal Argument and Computation, closely related to the COMMA community, was started. Since the fourth edition, organized by the Vienna University of Technology in 2012, an Innovative Application Track and a section for Demonstration Abstracts were included in the proceedings. At the fifth edition, co-organized in 2014 by the Universities of Aberdeen and Dundee in Pitlochry, the main conference was preceded by the first Summer School on Argumentation: Computational and Linguistic Perspectives. The same year also saw the launch of the first International Competition on Computational Models of Argumentation (ICCMA). Since COMMA 2016, hosted by the University of Potsdam, the COMMA proceedings are Open Access. This COMMA was also the first that included additional satellite workshops in the programme. COMMA 2018 was hosted by the Institute of Philosophy and Sociology of the Polish National Academy of Sciences in Warsaw, Poland. It included an industry afternoon bringing together businesses, NGOs, academics and students interested in practical applications of argument technologies in industry.

This year COMMA is in Italy for the second time, now hosted by the University of Perugia. It is preceded by the 4th Summer School on Argumentation: Computational and Linguistic Perspectives (SSA 2020), and features a demonstrations session and three satellite workshops. The International Workshops on Systems and Algorithms for Formal Argumentation (SAFA), initiated at COMMA 2016, has its third edition, while there is a new Workshop on Argument Visualization. Finally, the well-known Workshop on Computational Models of Natural Argument, established in 2001, has its 20th edition at COMMA 2020.

Despite these continuing traditions, COMMA 2020 is different from all preceding COMMAs in one respect: because of the coronavirus pandemic that hit the world early 2020, the entire conference and its preceding summer school have to take place online. This is, of course, a huge disappointment for the local organisers and for all participants, who had been looking forward to a great conference in the beautiful city of Perugia. Nevertheless, going online secures the continuation of the COMMA conference series, allowing the presentation and discussion of the latest research results and their publication in these proceedings.

The COMMA 2020 programme reflects the interdisciplinary nature of the field, and its contributions range from theoretical to practical (although most are theoretical). Theoretical contributions include new formal models, the study of formal or computational properties of models, designs for implemented systems and experimental research. Practical papers include applications to medicine, law, crime investigation, chatbots and online product reviews. The conference respects its historic origins by providing both abstract and structured accounts of argumentation. Some papers propose formal argument schemes for specific forms of argument. Many papers focus on the evaluation of arguments or their conclusions given a body of arguments, with a continuation of a recent trend to study gradual (e.g. probabilistic) notions of evaluation. Other papers focus on the dialogical processes by which argumentation proceeds, sometimes from a game-theoretical point of view. The focus on argument mining, which first appeared at COMMA 2016, is continued while an emerging trend this year is the use of argumentation for explainable AI.

The three invited talks also reflect the diverse nature of the field. Professor Catarina Dutilh Novaes from the Free University Amsterdam discusses the role of adversariality in argumentation from a social-epistemology perspective. Professor John Horty of the University of Maryland gives a logical analysis of defeasible reasoning about open-textured predicates in natural language and legal theory. Finally, Professor Chris Reed of the University of Dundee covers a broad spectrum from philosophical foundations via algorithmic research to technological applications.

Finally, we acknowledge the work of all those who have contributed in making the conference and its satellite events a success. We would like to thank IOS Press for publishing these proceedings and continuing to make them Open Access. As local and international sponsors of the conference, we would like to thank in random order Fondazione Cassa di Risparmio di Perugia, Gruppo Nazionale per il Calcolo Scientifico (GNCS-INdAM), the Artificial Intelligence Journal (funding scheme for promoting AI research), the Department of Mathematics and Computer Science (University of Perugia), the University of Perugia, Confcommercio Umbria, Associazione Nazionale Imprese ICT (Assintel), and iter innovazione terziario. We also thank the Italian Association for Artificial Intelligence (AIxIA), which supported the best student paper award. We thank the invited speakers Catarina Dutilh Novaes, John Horty and Chris Reed for their insightful and inspiring talks. We acknowledge steady support and encouragement by the COMMA Steering Committee, and are very grateful to the Programme Committee and additional reviewers, whose invaluable expertise and efforts have led to the selection of 28 full papers and 13 short papers out of a record total of 78 submissions, and 13 demonstration abstracts. The submission and reviewing process has been managed through the Easychair conference system, which we acknowledge for supporting COMMA since the first edition. Our thanks also to the COMMA 2020 workshop organisers Jodi Schneider, Matthias Thimm, Fabian Sperrle and their co-organisers, and to the summer school programme chair Massimiliano Giacomin. Last but not least, we thank all the authors and participants for contributing to the success of the conference with their hard work and commitment.

Henry Prakken (Programme chair)

Stefano Bistarelli (Conference chair)

Francesco Santini (Conference co-chair and demo chair)

Carlo Taticchi (Publicity chair)

Utrecht/Perugia, July 2020

Since at least the 1980s, the role of adversariality in argumentation has been extensively discussed. Some authors criticize adversarial conceptions and practices of argumentation and instead defend more cooperative approaches, both on moral and on epistemic grounds. Others retort that argumentation is inherently adversarial, and that the problem lies not with adversariality per se but with overly aggressive manifestations therof. In this paper, I defend the view that specific instances of argumentation are (and should be) adversarial or cooperative proportionally to pre-existing conflict. What determines whether an argumentative situation should be primarily adversarial or primarily cooperative are contextual features and background conditions, in particular the extent to which the parties involved have prior conflicting or convergent interests and goals. I articulate a notion of adversariality in terms of the relevant parties pursuing conflicting interests, and argue that, while cooperative argumentation is to be encouraged whenever possible, conflict as such is an inevitable aspect of human sociality and thus cannot be completely eliminated.

I will discuss some of the problems presented by open textured predicates for the semantics of natural language, as well as in legal theory. I will then (i) sketch an account of constraint in common law, (ii) suggest that this account can be adapted to help us understand open textured predicates as well, (iii) talk a bit about the reasoning involved in reaching decisions that satisfy this account of constraint, and (iv) show how this reasoning can be modeled in a simple defeasible logic.

Computational models of argument have vast potential to transform human reasoning and decision-making wherever it occurs taking theories rooted in philosophy, developing algorithms in data science, natural language processing and AI, and engineering solutions that could end up on a phone in everyone’s pocket. Fulfilling that potential, however, is enormously challenging. Sometimes, what’s required is overhauling our most fundamental theories to accommodate real world phenomena: arguments in the real world, for example, most typically occur in multi-party contexts, so new theories have had to be developed to account for and handle dialogical, dialectical and interactional aspects of argumentation, whilst still supporting formally well-understood phenomena such as abstraction and acceptability, audiences and values, lexical semantics and argument structure.

At other times, though, what’s required is forging ahead with a pragmatic compromise at the theoretical level that sacrifices a complete computational account of all facets of argumentation, but which nonetheless helps tackle some specific problem. Applications for supporting argumentation in domains as diverse as law, science and intelligence analysis have adopted this tack, delivering prototypes that demonstrate the potential of argument technology in different sectors.

At yet other times the problem is more a practical one: how on Earth do we assemble datasets of argumentation large enough for training supervised machine learning algorithms (let alone large enough for sheer statistical learning)? Or how can we develop, ab initio, linguistic annotation methods that can keep up with live debate? Right across its broad range of competence, the field of computational models of argument has had to pull itself up by its bootstraps, developing its own working methods, requirements, data standards, software tooling, research challenges and vocabulary.

Then again, sometimes what’s required is hard academic slog to drive forward performance: the new field of argument mining is an excellent example where progress is being made in leaps and bounds, even as the challenges are being broadened from domain specific to domain independent, monolingual to multilingual, monological to dialogical. It is the determined inspiration of those working in argument mining that is responsible for results starting to come through that represent acceptable performance on realistic tasks.

But perhaps the greatest challenge, though, is what in commercial terms is known as route to market. How do we get the fruits of our labours into the hands of the hundreds of millions of people who could benefit from it? Whether contributing to the quality of national and international debate, helping the general public identify fake news, improving counterterrorism threat analysis, or enhancing democratic processes or whether nudging arguments in a pub to be a bit more accurate, helping separating couples reach more acceptable agreements, or offering an elderly parent some advice on the latest Covid rumour: wherever argument plays a role, argument technology has the potential to improve matters. Neither developing new philosophical theory nor building new phone apps (nor anything in between) is enough on its own, but with a clearer game plan for the community as a whole there is an opportunity for us to start to fulfil the potential we have collectively for making a significant difference in the world.

Chatbots are versatile tools that have the potential of being used for computational persuasion where the chatbot acts as the persuader and the human agent as the persuadee. To allow the user to type his or her arguments, as opposed to selecting them from a menu, the chatbot needs a sufficiently large knowledge base of arguments and counterarguments. And in order to make the user change their current stance on a subject, the chatbot needs a method to select persuasive counterarguments. To address this, we present a chatbot that is equipped with an argument graph and the ability to identify the concerns of the user argument in order to select appropriate counterarguments. We evaluate the bot in a study with participants and show how using our method can make the chatbot more persuasive.

In this paper we present an argumentation-based approach to representing and reasoning about a domain of law that has previously been addressed through a machine learning approach. The domain concerns cases that all fall within the remit of a specific Article within the European Court of Human Rights. We perform a comparison between the approaches, based on two criteria: ability of the model to accurately replicate the decision that was made in the real life legal cases within the particular domain, and the quality of the explanation provided by the models. Our initial results show that the system based on the argumentation approach improves on the machine learning results in terms of accuracy, and can explain its outcomes in terms of the issue on which the case turned, and the factors that were crucial in arriving at the conclusion.

In the last years, several empirical approaches have been proposed to tackle argument mining tasks, e.g., argument classification, relation prediction, argument synthesis. These approaches rely more and more on language models (e.g., BERT) to boost their performance. However, these language models require a lot of training data, and size is often a drawback of the available argument mining data sets. The goal of this paper is to assess the robustness of these language models for the argument classification task. More precisely, the aim of the current work is twofold: first, we generate adversarial examples addressing linguistic perturbations in the original sentences, and second, we improve the robustness of argument classification models using adversarial training. Two empirical evaluations are addressed relying on standard datasets for AM tasks, whilst the generated adversarial examples are qualitatively evaluated through a user study. Results prove the robustness of BERT for the argument classification task, yet highlighting that it is not invulnerable to simple linguistic perturbations in the input data.

Argument(ation) Mining (AM) is the research area which aims at extracting argument components and predicting argumentative relations (i.e., support and attack) from text. In particular, numerous approaches have been proposed in the literature to predict the relations holding between arguments, and application-specific annotated resources were built for this purpose. Despite the fact that these resources were created to experiment on the same task, the definition of a single relation prediction method to be successfully applied to a significant portion of these datasets is an open research problem in AM. This means that none of the methods proposed in the literature can be easily ported from one resource to another. In this paper, we address this problem by proposing a set of dataset independent strong neural baselines which obtain homogeneous results on all the datasets proposed in the literature for the argumentative relation prediction task in AM. Thus, our baselines can be employed by the AM community to compare more effectively how well a method performs on the argumentative relation prediction task.

This paper provides an initial exploration on the relationships between PageRank and gradual argumentation semantics. After showing that PageRank, directly interpreted as an argumentation semantics for support frameworks, fails to satisfy some generally desirable properties, we propose a novel approach to reconstruct PageRank as gradual semantics of a suitably defined bipolar argumentation framework, while satisfying these desirable properties. The theoretical advantages of the approach are complemented by an illustration of its potential application to support the generation of better explanations of PageRank scores for end users.

Attack-Support Argumentation Framework (ASAF) is an extension of the Bipolar Argumentation Framework that allows for attacks and supports not only between arguments but also targeting attacks and supports at any level. In this paper we propose an incremental approach for computing the skeptical preferred acceptance in dynamic ASAFs. Specifically, we investigate how the skeptical acceptance of a goal element (an argument, an attack, or a support) evolves when a given ASAF is updated by adding or retracting an argument, an attack, or a support, and propose an incremental algorithm for solving this problem. Our approach relies on identifying a portion of the given ASAF which is sufficient to determine the status of the goal w.r.t. the updated ASAF. We experimentally evaluate our approach showing that it outperforms the computation from scratch on average.

The paper investigates gradual semantics that are able to deal with similarity between arguments. Following the approach that defines semantics with evaluation methods, i.e., a couple of aggregation functions, the paper argues for the need of a novel function, called adjustment function. The latter is responsible for taking into account similarity when it is available. It aims at reducing the strengths of attackers according to the possible similarities between them. The reason is that similarity is seen as redundancy that should be avoided, otherwise a semantics may return inaccurate evaluations of arguments. The paper proposes a novel adjustment function that is based on the well-known weighted h-Categorizer, and investigates its formal properties.

We examine different methods of handling argument consistency and minimality in logical argumentation frameworks, showing that both properties may (and sometimes even should) be omitted from the definition of arguments. In process, we consider the adequacy of attack rules to the underlying logics.

Abstract dialectical frameworks (ADFs) are one of the most powerful generalization of classical Dung-style AFs. In this paper we show how to use ADFs if we want to deal with acceptance conditions changing over time. We therefore introduce so-called timed abstract dialectical frameworks (tADFs) which are essentially ADFs equipped with time states. Beside a precise formal definition of tADFs and an illustrating example we prove that KleeneÂŘs three-valued logic K3 facilitate the evaluation of acceptance functions if we do not allow multiple occurrences of atoms.

The paper provides an initial study on how ranking semantics in argumentation have to be handled when leaving the purely abstract setting. We employ claim-augmented frameworks where each argument is associated to a claim it stands for. We propose liftings from argument- to claim-level in two veins: for desired properties and for actual rankings. Our main contribution is to investigate whether the satisfaction of properties by argument-based ranking semantics carries over to the lifted, claim-based, variants of the corresponding properties and semantics.

There are two intuitive principles governing belief formation and argument evaluation that can potentially clash. After arguing that adopting them unrestrictedly leads to an infinite regress, we propose a formal framework in which qualified versions of both principles can be subscribed without falling into such a regress. The proposal integrates tools from two different traditions: structured argumentation and awareness epistemic logic. We show that our formalism satisfies certain rationality postulates and argue that the rest of them can be seen as too ideal when modelling resource-bounded agents.

The concept of strong admissibility plays an important role in some of the dialectical proof procedures that have been stated for grounded semantics. As the grounded extension is the (unique) biggest strongly admissible set, to show that an argument is in the grounded extension it suffices to show that it is in a strongly admissible set. We are interested in identifying a strongly admissible set that minimizes the number of steps needed in the associated dialectical proof procedure. In the current work, we look at the computational complexity of doing so.

In ASPIC-style structured argumentation an argument can rebut another argument by attacking its conclusion. Two ways of formalizing rebuttal have been proposed: In restricted rebuttal, the attacked conclusion must have been arrived at with a defeasible rule, whereas in unrestricted rebuttal, it may have been arrived at with a strict rule, as long as at least one of the antecedents of this strict rule was already defeasible. One systematic way of choosing between various possible definitions of a framework for structured argumentation is to study what rationality postulates are satisfied by which definition, for example whether the closure postulate holds, i.e. whether the accepted conclusions are closed under strict rules. While having some benefits, the proposal to use unrestricted rebuttal faces the problem that the closure postulate only holds for the grounded semantics but fails when other argumentation semantics are applied, whereas with restricted rebuttal the closure postulate always holds. In this paper we propose that ASPIC-style argumentation can benefit from keeping track not only of the attack relation between arguments, but also the relation of deductive joint support that holds between a set of arguments and an argument that was constructed from that set using a strict rule. By taking this deductive joint support relation into account while determining the extensions, the closure postulate holds with unrestricted rebuttal under all admissibility-based semantics. We define the semantics of deductive joint support through the flattening method.

An important approach to abstract argumentation is the labeling-based approach, in which one makes use of labelings that assign to each argument one of three labels: in, out or und. In this paper, we address the question, which of the twenty-seven functions from the set of labels to the set of labels can be represented by an argumentation framework. We prove that in preferred, complete and grounded semantics, eleven label functions can be represented in this way while sixteen label functions cannot be represented by any argumentation framework. We show how this analysis of label functions can be applied to prove an impossibility result: Argumentation frameworks extended with a certain kind of weak attack relation cannot be flattened to the standard Dung argumentation frameworks.

Baumann, Brewka and Ulbricht recently introduced weak admissibility as an alternative to Dung’s notion of admissibility, and they use it to define weakly preferred, weakly complete and weakly grounded semantics of argumentation frameworks. In this paper we analyze their new semantics with respect to the principles discussed in the literature on abstract argumentation. Moreover, we introduce two variants of their new semantics, which we call qualified and semi-qualified semantics, and we check which principles they satisfy as well. Since the existing principles do not distinguish our new semantics from the ones of Baumann et al., we also introduce some new principles to distinguish them. Besides selecting a semantics for an application, or for algorithmic design, our new principle-based analysis can also be used for the further search for weak admissibility semantics.

In this work we revisit computational aspects of strongly admissible semantics in Dung’s abstract argumentation frameworks. First, we complement the existing complexity analysis by focusing on the problem of computing strongly admissible sets of minimum size that contain a given argument and providing NP-hardness as well as hardness of approximation results. Based on these results, we then investigate two approaches to compute (minimum-sized) strongly admissible sets based on Answer Set Programming (ASP) and Integer Linear Programming (ILP), and provide an experimental comparison of their performance.

Generalizing the attack structure in argumentation frameworks (AFs) has been studied in different ways. Most prominently, the binary attack relation of Dung frameworks has been extended to the notion of collective attacks. The resulting formalism is often termed SETAFs. Another approach is provided via abstract dialectical frameworks (ADFs), where acceptance conditions specify the relation between arguments; restricting these conditions naturally allows for so-called support-free ADFs. The aim of the paper is to shed light on the relation between these two different approaches. To this end, we investigate and compare the expressiveness of SETAFs and support-free ADFs under the lens of 3-valued semantics. Our results show that it is only the presence of unsatisfiable acceptance conditions in support-free ADFs that discriminate the two approaches.

In this paper, we provide a detailed analysis of PageRank to determine the relevance of arguments along with content- and knowledge-based methods from the field of natural language processing. We do not only show how the cross-linking of arguments is only slightly involved in the recognition of relevance, we rather show how basic common knowledge and reader-involving methods outperform the purely structure-related PageRank. The methods we propose are based on the latest research and correlate strongly with human awareness regarding the relevance of arguments. Altogether, we show that PageRank does not fully capture the relevance of arguments and must be extended by a contextual level in order to take concepts of natural language into account at the web level, as they are unavoidably involved in argumentation.