Humans argue.1 This distinctive feature is at the sametime an important cognitive capacity and a powerfulsocial phenomenon. It has attracted attention and

careful analysis since the dawn of civilization, being inti-mately related to the origin of any form of social organiza-tion, from political debates to law, and of structured think-ing, from philosophy to science and arts.

As a cognitive capacity, argumentation is important forhandling conflicting beliefs, assumptions, viewpoints, opin-ions, goals, and many other kinds of mental attitudes. Whenwe are faced with a situation where we find that our infor-mation is incomplete or inconsistent, we often resort to theuse of arguments in favor and against a given position inorder to make sense of the situation. When we interact withother people we often exchange arguments in a cooperativeor competitive fashion to reach a final agreement or todefend and promote an individual position.

Articles

FALL 2017 25Copyright © 2017, Association for the Advancement of Artificial Intelligence. All rights reserved. ISSN 0738-4602

Toward Artificial Argumentation

Katie Atkinson, Pietro Baroni, Massimiliano Giacomin, Anthony Hunter, Henry Prakken, Chris Reed, Guillermo Simari,

Matthias Thimm, Serena Villata

■ The field of computational models ofargument is emerging as an importantaspect of artificial intelligence research.The reason for this is based on therecognition that if we are to developrobust intelligent systems, then it isimperative that they can handle incom-plete and inconsistent information in away that somehow emulates the wayhumans tackle such a complex task.And one of the key ways that humansdo this is to use argumentation eitherinternally, by evaluating arguments andcounterarguments‚ or externally, by forinstance entering into a discussion ordebate where arguments are exchanged.As we report in this review, recent devel-opments in the field are leading to tech-nology for artificial argumentation, inthe legal, medical, and e-governmentdomains, and interesting tools for argu-ment mining, for debating technologies,and for argumentation solvers areemerging.

Occurring continuously both in our mind and inthe social arena, argumentation pervades our intelli-gent behavior and the challenge of developing artifi-cial argumentation systems appears to be as diverseand exciting as the challenge of artificial intelligenceitself.

Indeed, this rich and important phenomenonoffers an opportunity to develop models and tools forargumentation and to conceive autonomous artificialagents that can exploit these models and tools in thecognitive tasks they are required to carry out. To thispurpose, a number of interesting lines of research arebeing investigated within artificial intelligence andseveral neighbor fields, leading to the establishmentof computational models of argument as a promisinginterdisciplinary research area. Progress in this area isexpected to contribute to significant advances in theunderstanding and modeling of various aspects ofhuman intelligence.

In this article, we review formalisms for capturingvarious aspects of argumentation, and we presentadvances in their applications, with the aim to com-municate how research is making progress towardthe goal of making artificial argumentation tech-nologies and systems a mature and widespread reali-ty. In this brief review, we are unable to discuss or citeall the relevant literature, and we suggest that theinterested reader seek more detailed coverage of thefoundations from Rahwan and Simari (2009), ofapplications from Modgil et al. (2013), and of recentdevelopments from the proceedings of the Interna-tional Conference on Computational Models ofArgument series,2 and the Argument and Computationjournal.3

Models of ArgumentComputational models of argument are being devel-oped to reflect aspects of how humans build,exchange, analyze, and use arguments in their dailylives to deal with a world where the information maybe controversial, incomplete, or inconsistent (Bench-Capon and Dunne 2007, Rahwan and Simari 2009).The diversity of the manifestations of arguments inreal life implies diversity in the relevant models tooand the impossibility to reduce the vast available lit-erature to a single reference scheme. It is possiblehowever to identify some layers that can be regardedas basic building blocks for the construction of anargumentation model. Specific modeling approachesmay differ in the selection of which layers they actu-ally use, in the way the selected layers are combined,and in the formalization adopted within each layer.

We consider the following five main layers: struc-tural, relational, dialogical, assessment, and rhetori-cal. They are described in the following and also sum-marized in Figure 1. Note that while each layer has itsown nature and distinctive traits, the boundariesbetween layers may not be so neat in some contexts,

and specific formalisms may inextricably mergetogether aspects relevant to different layers.

Structural LayerThis layer concerns the structure of the argumentsand how they are built: essentially it specifies, in agiven context, what an argument looks like, in termsof its internal structure, and which are the ingredi-ents for its construction. To exemplify, in contextswhere arguments are built from a logical knowledgebase the ingredients are the logical formulae includ-ed in the knowledge base. Then one way to buildarguments is by simply applying the logic of the lan-guage in which the knowledge base is stated to deriveconclusions. An argument here can be seen as a pair(Φ, α) where Φ is a subset of the knowledge base (a setof formulae) that logically entails α (a formula). Here,Φ is called the support, and α is the claim, of the argu-ment. Other approaches consider argument con-struction from knowledge bases as applying rules tothe formulas from the knowledge base, where therules may be defeasible. In these rule-based approach-es an argument is typically seen as a tree whose rootis the claim or conclusion, whose leaves are thepremises on which the argument is based, and whosestructure corresponds to the application of the rulesfrom the premises to the conclusion. Investigationsinto the structural layer were initiated by Pollock(1992). Prominent examples of rule-based formalismsare ASPIC+, assumption-based argumentation (ABA),and defeasible logic programming (DeLP). For a tuto-rial introduction to formalisms for structured argu-mentation, see Besnard et al. (2014).

Arguments are not built from knowledge basesonly, however. For instance, interactive systems thatacquire arguments from users may adopt theapproach of argumentation schemes (Walton, Reed,and Macagno 2008), namely stereotypical reasoningpatterns, where in addition to the premises and theclaim, a set of critical questions is considered. Criti-cal questions provide a sort of checklist of issues thatcan be raised to challenge arguments built on thebasis of a given scheme. Argumentation schemeshave also been used as a source of defeasible infer-ence rules in rule-based approaches to argument con-struction from knowledge bases. In addition, argu-mentation schemes are often considered in thecontext of argument mining (see the Argument Min-ing section) where the goal is to identify and extractthe argumentative structures embedded in a naturallanguage source, providing a machine-processablerepresentation of them.

The variety of existing argument models raises theissue of exchanging or sharing arguments among dif-ferent systems. This problem is addressed by theargument interchange format initiative (Chesñevaret al. 2006), aimed at providing an interlinguabetween various more concrete argumentation lan-guages, on the basis of a generic abstract ontology.

Articles

26 AI MAGAZINE

Articles

FALL 2017 27

Relational LayerArguments do not live in isolation and are linked toeach other by various types of relations: the relation-al layer deals with identifying and formally repre-senting them, in view of their use in other layers oreven for descriptive and presentation purposes, sincethey are essential for an understanding of what isactually going on in an argumentation process.Examples of important relationships are (1) the sub-argument and superargument relationships, indicat-ing how an argument is built incrementally on top ofother arguments; (2) the attack relationship, indicat-ing that an argument is incompatible with anotherargument in some sense, for example, because theyhave contradictory claims, or one claim contradictssome premise or assumption on which the other isbased; (3) the support relationship, intuitively mean-ing that an argument provides some backing toanother, and admitting several, even rather dissimi-lar, interpretations, depending on the actual natureof this backing; (4) a preference relationship, rankingarguments according to some criterion, and admit-ting again a variety of instantiations ranging from

strength to credibility to value-based evaluations.What relationships are significant and how to

identify them are highly context-dependent matters.Note in particular that identifying argument rela-tions may be an easy mechanical procedure in set-tings where arguments are formally built from aknowledge base, while in an argument mining sce-nario it is a task as challenging as the identificationof the arguments themselves.

Dialogical LayerThis layer deals with the exchange of argumentsamong different agents (or even between an agentand itself, in a scenario where argumentative reason-ing is conceived as a monological activity) accordingto formal dialogue rules. Agents may engage in theexchange of arguments for a variety of purposes withseveral dialogue types having been identified in theliterature, like inquiry, negotiation, information-seeking, deliberation, and persuasion. In all cases theexchange can be formalized as a dialogue game,which is normally made up of a set of communica-tive acts called moves, and a protocol specifying

Figure 1. Key Aspects of Argumentation.

Structural layer: How are arguments constructed?

Relational layer: What are the relationships between arguments?

Dialogical layer: How can argumentation beundertaken in dialogues?

Assessment layer: How can a constellation of interacting argumentsbe evaluated and conclusions drawn?

Rhetorical layer: How can argumentation be tailored for an audience so that it is persuasive?

which moves can be made at each step of the dia-logue. It concerns representing and managing thelocutions exchanged between the agents involved, aswell as specifying the contents of these locutions interms of entire arguments or components of argu-ments. Moreover, the dialogue protocol may establishthe allowed moves on the basis of argument relation-ships. For instance, a protocol may specify that amove is legal only if it presents an argument attack-ing an argument presented in a previous move. Forthese reasons the dialogical layer requires strict con-nections with the structural and relational layers.Moreover some dialogue protocols are defined so asto embed an argument assessment method: in thesecases the dialogical layer is intertwined with theassessment layer, described next.

Assessment LayerThis layer concerns the assessment of a set of argu-ments and of their conclusions in order to establishtheir justification status. The need for this layer arisesfrom the presence of attacks among arguments, pre-venting them so as to be accepted altogether and call-ing for a formal method to solve the conflict. Thisproblem is addressed in a principled and highly styl-ized form in the context of the theory of abstract argu-mentation frameworks (Dung 1995), where argu-ments are treated as abstract entities, deprived of anystructural property and of all their relations but attack.We give an example of an argumentation framework,based on textual arguments, in figure 2. Given itsabstract nature, an argumentation framework is oftenreferred to as argument graph, and this term is alsoused to refer to similar representations where addi-tional relations, like support, are considered.

An abstract argumentation semantics is a formal cri-terion to determine which sets of arguments, calledextensions, are able to survive the conflict togetherand can be regarded as collectively acceptable.Abstract argumentation theory is probably the sub-

field of computational models of argument that hasattracted most research attention in the last twodecades, due to its generality and theoretical clean-ness. In particular Dung (1995) has shown the abilityof the formalism to capture as instances several otherapproaches, especially in the area of nonmonotonicreasoning. Dung’s approach abstracts from the originand nature of the attack relation. A natural idea is todefine this relation in terms of a more basic notion ofconflict between arguments (for example, two argu-ments having contradictory conclusions) and anotion of relative argument strength or preference. Inthe literature, there are two ways to connect theseideas to Dung’s frameworks. The first approach leavesDung’s frameworks as they are but connects themwith models at the structural layer of argument todefine attack in terms of preferences or argumentstrength while taking the structure of arguments intoaccount. The second approach instead extendsDung’s frameworks with some abstract notion ofargument strength or preference, while possibly alsoadding an abstract support relation between argu-ments. Moreover, while most approaches consider aqualitative notion of acceptance, quantitative assess-ments methods are being investigated too.

Further, it must be noted that the evaluation ofargument acceptability is only a part, actually themost basic one, of the assessment tasks required in anargumentative process. In particular the final goal ofan agent is usually the assessment of the justificationstatus of the statements supported by arguments,which, in the end, amounts to determining what tobelieve or what to do. Since many arguments mayhave the same conclusion, assessing the status of astatement involves a synthesis of the statuses of thearguments supporting it. As in real life, the task ofdeciding what to believe may be carried out adoptingdifferent attitudes, ranging from extremely skepticalto extremely credulous, corresponding to differentformal methods for statement justification synthesis.

Articles

28 AI MAGAZINE

Figure 2. An Example of Argumentation Framework Consisting of Three Arguments in the Medical Domain and Their Attacks.

Arguments A1 and A2 are two alternatives for treating a patient with hypertension, and A3 provides a reason against one of the options.Here, we assume that A1 and A2 attack each other because giving one treatment precludes the other, and we assume that A3 attacks A2because it provides a counterargument to A2.

A1 = Patient hashypertension so

prescribe diuretics

A2 = Patient hashypertension so

prescribe betablockers

A3 = Patient hasemphysema, whichis a contraindication

for betablockers

Rhetorical LayerNormally argumentation is undertaken in somewider context of goals for the agents involved, andso individual arguments are presented with somewider aim and according to some strategical consid-erations. For instance, if an agent is trying to per-suade another agent to do something, then it is like-ly that some rhetorical device is harnessed and thiswill reflect the nature of the arguments used (forexample, a politician may refer to investing in thefuture of the nation’s children as a way of persuadingcolleagues to vote for an increase in taxation). Withthe roots of the study of rhetoric going back to Aris-totle,4 recent studies into aspects of the rhetorical lev-el include believability and impact of argumentsfrom the perspective of the audience, use of threatsand rewards, appropriateness of advocates, and val-ues of the audience. The rhetorical layer may beabsent in some contexts, for example, when neu-trally building arguments from a knowledge base, butcan permeate all the other layers in other contextssince goal-oriented considerations may drive thedecisions of which arguments to build, taking intoaccount their relations with other arguments, ofwhether, how, and when to use the arguments in adialogue, and of which assessment method (forexample, whether a more skeptically or more credu-lously oriented one) to apply.

The following sections review several prominentdomains that exploit computational models of argu-ment for the development of actual applications and,at the same time, stimulate the relevant theoreticaldevelopment by providing case studies and impor-tant modeling challenges.

Legal ArgumentationThe law is an obvious application domain for argu-mentation research, since legal reasoning is essen-tially argumentative and to a large extent recorded indocuments. This has led to highly stylized forms ofargumentation, which makes it easier to formulateand validate formal and computational models ofargument than in many other domains. In this sec-tion, we briefly discuss work and research themes inthis area. A more detailed survey can be found in thepaper by Prakken and Sartor (2015).

In legal cases, first the facts have to be determined.Because of the diverse nature of the evidence in mostcases and the need for explanation to statisticallaypeople, legal evidential reasoning is an excellenttest bed for combined qualitative and quantitativemodels of defeasible reasoning. At the practical side,so-called sense-making systems have been proposed,with which crime investigators or triers of fact canstructure their arguments and scenarios to makesense of a large body of evidence.

After the facts of a case have been established, theymust be classified under the conditions of legal rules,

which involves interpreting these rules. Two influen-tial AI and law models of this are the HYPO system byKevin Ashley and its successor CATO by VincentAleven, which model how lawyers in common-lawjurisdictions make use of past decisions when argu-ing a case. Their underlying argumentation model isfor factor- or dimension-based reasoning, where cas-es are collections of abstract fact patterns that favoror oppose a conclusion, either in an all-or-nothingfashion (factors) or to varying degrees (dimensions).This work inspired subsequent formal work using thetools of formal argumentation, resulting in formal-ized versions of traditional legal argument formssuch as appeal to precedent and policy and the bal-ancing of goals, values, and interests (Horty andBench-Capon 2012).

Finally, when the facts have been classified, thelegal rules must be applied to them. Legal rules canhave exceptions or conflict on other grounds. Rule-based argumentation logics with preferences haveproved useful here.

Legal reasoning usually takes place in the contextof a dispute between adversaries, within a prescribedlegal procedure. This makes the setting inherentlydynamic and multiparty, and raises issues of strategyand choice. For example, there is work on optimalstrategies for adversaries in debates with an adjudica-tor, given their preferences over the possible out-comes of a debate and their estimates of what theadjudicator will likely accept.

While the theoretical advances on models of legalargument have been impressive and a number ofvaluable prototype systems have been developed, nosystems have been deployed in everyday practice yet.One reason is the conservative attitude to technologyin the legal world and its billing-by-the-hour culture,which does not stimulate innovation. Another reasonis the fact that building realistic systems of legal argu-ment requires vast amounts of commonsense knowl-edge. However, recently things have changed. First,clients of law firms increasingly demand the use ofmodern technology. Moreover, the recent advances innatural language processing, machine learning, anddata science combined with the massive digital avail-ability of legal data and information have created theprospects for combining AI models of legal argumentwith argument mining techniques. In fact, two of thefirst argument mining projects were on legal argu-ment (Palau and Moens 2011). If this technology iscombined with AI and law’s computational models ofargument, then practical applications of these modelscould be well within reach.

Medical ArgumentationHealth care is a potentially important domain fordeveloping and applying computational models ofargument. It is common for health-care informationto be complex, heterogeneous, incomplete, and

Articles

FALL 2017 29

inconsistent, and therefore argumentation is appeal-ing for those involved as it allows for important con-flicts to be highlighted and analyzed and unimpor-tant conflicts to be suppressed.

One of the pioneers of argumentation technology,John Fox, developed a number of prototype systemsfor medical decision support such as the Capsule sys-tem (Walton et al. 1997). Capsule supports a familypractitioner when she or he is about to prescribe aspecific drug for a patient. The system uses a standarddatabase of equivalent treatments that is routinelyused by clinicians, and the patient records, to providearguments pro and con each of the alternatives. Thearguments are based on whether the patient has pre-viously expressed a preference for or against the alter-native, whether the patient has previously exhibiteda negative reaction to the alternative, whether thereis possible negative interaction with other drugsbeing taken by the patient, and the relative cost of thealternative. In a formal trial of the Capsule system,with 42 clinicians using 36 simulated cases, the sys-tem was shown to help clinicians improve the quali-ty of their prescribing and to improve their compli-ance with medical guidelines.

Over recent years, there has been substantial shiftin health care to evidence-based practice. This meansthat health-care professionals need to use the bestavailable evidence to inform their decision making.For deciding on interventions, this normally calls forevidence from randomized clinical trials. The prob-lem with this is that there are many such trials pub-lished each year, and it is difficult for clinicians tokeep abreast of this literature. To help them, there aremedical guidelines and systematic reviews that aggre-gate this evidence by providing recommendations.Unfortunately, these recommendations can rapidlybecome out of date, they do not take local circum-stances into account, and they normally do not con-sider patients with comorbidities. To address theseproblems, an argument-based approach to aggregat-ing clinical evidence has been proposed by Hunterand Williams (2012). The framework is a formalapproach to synthesizing knowledge from clinical tri-als involving multiple outcome indicators (where anoutcome indicator is either positive such as the num-ber of patients who survive the disease after 1 year, or2 years, and so on, or negative such as the proportionof those treated who have a particular side-effect).Based on the available evidence, evidence-based argu-ments are generated for claiming that one treatmentis superior to another for a given patient.

Preference criteria over evidence-based argumentsare specified in terms of the outcome indicators, andthe magnitude of those outcome indicators, in theevidence. Various kinds of counter-arguments attackthe evidence-based arguments depending on thequality of evidence used (for example, evidencecould be attacked because a trial was not conductedcorrectly). The arguments and counter-arguments

constitute an argument graph, and using abstractargumentation semantics, the winning arguments areidentified, and thereby argument-based recommen-dations for which treatments are superior can beidentified. The approach has been evaluated by com-parison with recommendations made in publishedhealth-care guidelines (Hunter and Williams 2012)and it has been used to publish, in the medical liter-ature on lung cancer, a more refined systematicreview of the evidence. An ongoing study is using thistechnique in a systematic review on brain cancer forpublication by Cochrane.

These examples are just two of a number of appli-cations of argumentation being developed for sup-porting health-care professionals and patients. Fur-ther applications include dealing with the conflictsthat can occur when using multiple clinical guide-lines, supporting multidisciplinary teams of health-care professionals when dealing with difficult clinicalcases, and supporting medical image interpretation.

e-GovernmentAn important feature of democracies is that citizenscan engage their governments in dialogues aboutpolicies. Traditionally this was done by writing lettersand holding town hall debates, but over the past twodecades new methods of interaction have been devel-oped to exploit the benefits of current technology.Citizens may wish to respond in several ways to pol-icy proposals made by their governments. They maysimply seek a justification of the proposed policy;they may wish to object to the proposed policy; orthey may want to propose policies of their own. Suchdialogues can be facilitated through tools to supporte-democracy, and computational models of argumentcan be put to effective use in such tools.

For example, consider a local government authori-ty that is deciding what to build on a plot of waste-land in a community. One proposal by the localauthority may be to permit the building of a newsupermarket on the grounds that this will providejobs for the local community and shopping facilitiesfor local residents. Citizens may be consulted on thisproposal and critique this policy as well as put for-ward their own proposals. For example, the localauthority’s proposal could be critiqued by stating thatthe action of building a new supermarket will nothave the intended effect of creating jobs as there willbe job losses from local shop owners being put out ofbusiness by the supermarket. An alternative proposalcould be that instead of building a new supermarket,the site should be used to build a new play center forlocal residents’ children. Such opinions can beformed into arguments by distinguishing the premis-es (for example, there is little unemployment in thecommunity and play centers promote social interac-tion) and conclusion (we should build a new chil-dren’s play center). Argumentation-based tools can

Articles

30 AI MAGAZINE

then be put to use to facilitate such debates.Simple tools like e-petitions5 can transform tradi-

tional paper-based communication into digital com-munication, but recent advances have been made inthe development of tools that exploit the web, suchas the on-line argument mapping tools Debategraph6

and Debatabase,7 which enable users to modeldebates by considering issues, and their pros andcons, relevant to a debate. With such tools users canfreely insert and modify contributions, but the argu-ments entered are not required to conform to anyparticular semantics that would support coherenceand argument evaluation. A comprehensive survey ofthe state of the art in web-based argumentation toolscan be found in the paper by Schneider, Groza, andPassant (2013).

Early collaborative decision support systems suchas Zeno (Gordon and Karacapilidis 1997) and HER-MES (Karacapilidis and Papadias 2001) containedmore structure by making use of the IBIS (Issue BasedInformation Systems) model of argument. Use of thismodel enabled a particular problem or issue to bedecomposed into a number of different positions forwhich arguments can then be created to attack ordefend the positions until the issue is settled.

In recent work to consolidate different tasks rele-vant for e-democracy tools, on a recent Europeanproject called IMPACT,8 an argumentation toolboxwas created that consists of four interconnected mod-ules: an argument reconstruction tool; a structuredconsultation tool; an argument visualisation tool;and, a policy modeling tool. This is a decision-sup-port tool that makes use of structured theories ofargument representation and evaluation to enablepublic opinion gathering on political issues fromwhich conclusions can be drawn concerning howgovernment policies are presented, justified, andviewed by the users of the system. The tool uses argu-mentation schemes (Walton, Reed, and Macagno2008) to structure the information presented to theusers, but within the front-end interface, this struc-ture is implicit in order to facilitate ease of learningand use. Once opinions have been gathered, the argu-ments generated are evaluated through the use of val-ue-based argumentation frameworks (see the chapterin Rahwan and Simari [2009]) to provide users withsupport for which actions are justified, according tothe facts and social interests promoted by the differ-ent arguments. Debates that have been modelled,using in particular the Parmenides system (see Rah-wan and Simari [2009]), cover local issues, such aswhether to introduce more speed cameras on danger-ous stretches of road, and wider national issues suchas the UK debate about whether to ban fox hunting.This strand of work continues to be expanded withinthe Structured Consultation Tool developed as part ofthe IMPACT project mentioned previously and theCarneades tools.9

The richness of policy debates clearly makes the

domain of e-government an ideal one for the studyand application of tools that use computational mod-els of argument on a large scale.

Argument MiningIn recent years, the growth of the web, and the rap-idly increasing amount of diverse textual data pub-lished there, have highlighted the need for methodsto identify, structure, and summarize this hugeresource. Online newspapers, blogs, online debateplatforms, and social networks, as well as normativeand technical documents, provide a heterogeneousflow of information where natural language argu-ments can be identified and analyzed. The availabili-ty of such data, together with the advances in naturallanguage processing and machine learning, have sup-ported the rise of a new research area called argumentmining. The main goal of argument mining is theautomated extraction of natural language argumentsand their relations from generic textual corpora, withthe final goal to provide machine-readable structureddata for computational models of argument and rea-soning engines.

Two main stages have to be considered in the typ-ical argument mining pipeline, from the unstruc-tured natural language documents toward structured(possibly machine-readable) data:

Argument Extraction: The first stage of the pipeline is todetect arguments within the input natural languagetexts. The retrieved arguments will thus represent thenodes in an argument graph returned by the system.This step may be further split into two different stagessuch as the extraction of arguments and the furtherdetection of their boundaries. Many approaches haverecently been applied to tackle this challenge adoptingdifferent methodologies like for instance support vec-tor machines, naïve Bayes classifiers, logistic regres-sion.

Relation Extraction: The second stage of the pipelineconsists in constructing the argument graph to bereturned as output of the system. The goal is to identi-fy what are the relations holding between the argu-ments identified in the first stage. This is an extreme-ly complex task, as it involves high-level knowledgerepresentation and reasoning issues. The relationsbetween the arguments may be of a heterogeneousnature, like attack, support, or entailment. This stage isalso responsible for identifying, in structured argu-mentation, the internal relations of the componentsof an argument, such as the connection between thepremises and the claim. Being an extremely challeng-ing task, existing approaches assume simplifyinghypotheses, like the fact that evidence is always asso-ciated with a claim.

To illustrate, we consider the following exampleadapted from an online debate about random sobri-ety tests for drivers and their consistency with humanrights.10 We start with the unstructured natural lan-guage text from which we first aim to extract thearguments, and then to identify their relations:

Articles

FALL 2017 31

Random breath tests to public vehicle drivers can hard-ly be called an invasion of privacy or an investigationwithout due cause, because public safety is at stake.Random tests are routinely carried out by many trainand bus companies and are being introduced on air-lines as well. The same applies for other drivers, whoare a major liability to the safety and lives of otherdrivers.

Randomly testing employees cannot be considered aninvasion of privacy. People who have to take randombreath tests to drive trucks or fly planes as part of theirjobs are taking the test as part of their job. They arebeing paid and must do what their employer wantsthem to do in order to keep their job. Searching ran-dom people outside of the context of employmentwith no suspicion of a crime is very different as iterodes civil liberties and sets a dangerous precedent.

The first goal of the argument mining pipeline con-sists in extracting the arguments from this text. In theprevious example, we highlight the four argumentsthat can be identified:

Random breath tests to public vehicle drivers can hard-ly be called an invasion of privacy or an investigationwithout due cause, because public safety is at stake [A1].Random tests are routinely carried out by many trainand bus companies and are being introduced on air-lines as well. The same applies for other drivers, who area major liability to the safety and lives of other drivers[A2].

Randomly testing employees cannot be considered aninvasion of privacy [A3]. People who have to take ran-dom breath tests to drive trucks or fly planes as partof their jobs are taking the test as part of their job.They are being paid and must do what their employ-er wants them to do in order to keep their job. Search-ing random people outside of the context of employmentwith no suspicion of a crime is very different as it erodescivil liberties and sets a dangerous precedent [A4].

Given these four arguments (that is, A1, A2, A3, andA4), the relations among them have to be identified.Let us consider for the explanatory purpose of thisexample that the two relations we aim at identifyingare the attack and the support relations only. In thiscase, we have that, taking into account the temporaldimension of the debate to decide the direction of therelations, argument A3 supports argument A1, andargument A4 attacks argument A2.

11

It is worth noticing that the identification of thearguments and their relations is much more subtleand ambiguous than what emerges from thisexplanatory example, and may often be a matter ofinterpretation that current state-of-the-art argumentmining systems cannot tackle yet. For instance, argu-ment A1 can be considered as a subargument of argu-ment A2 as “The same applies …” refers to A1, andargument A3 can be interpreted as a kind of persua-sive statement meant to strengthen argument A4.

To address this kind of issue and build more capa-ble applications, it is necessary to enhance the exist-

ing tools used to analyze, aggregate, synthesize, struc-ture, summarize, and reason about arguments intexts, with more sophisticated natural language pro-cessing (NLP) methods. However, and considering thecomplexity of the task, to do so it is still necessary toreach a deeper level of understanding of the innerworkings of natural language, and evolve new meth-ods expanding the ones currently found in naturallanguage processing.

Moreover, to tackle these challenging tasks, high-quality annotated corpora are needed for use as atraining set for any kind of aforementioned identifi-cation. These corpora are mainly composed by threedifferent elements: an annotated data set that repre-sents the gold standard whose annotation has beenchecked and validated by expert annotators and isused to train the system for the required task (that is,arguments or relations extraction), a set of guidelinesto explain in a detailed way how the data has beenannotated, and finally, the unlabelled raw corpus thatcan be used to test the system after the trainingphase. The reliability of a corpus is ensured by the cal-culation of the interannotator agreement that meas-ures the degree of agreement in performing the anno-tation task among the involved annotators.12 Currentprototypes of argument mining systems require to betrained against the data the task is addressed to, andthe construction of such annotated corpora remainsamong the most time-consuming activities in thispipeline.

For an exhaustive state of the art review on argu-ment mining techniques and applications, we referthe reader to the paper by Lippi and Torroni (2016).

Debating TechnologiesThere is a long tradition of computer-aided debatesystems with roots in e-democracy, decision support,and so on. These systems have two things in com-mon: first, they implement idiosyncratic and newdialogical structures, or games, with little reuse orincremental development. The second is that little orno contribution to the debate itself is offered by themachine. The system role has been one of supportand facilitation only. With a variety of techniques forautomatically mining argument structures from bothmonological and dialogical resources, an excitingnew possibility is opened up for not just supportingand enhancing new human-human arguments butalso developing new human-machine arguments:this is the space of debate technology, a specific sub-field of argument technology in general.

Several systems have demonstrated stand-aloneapplications of debate technology focusing ondomains of use such as pedagogy (Pinkwart andMcLaren 2012), in which both responsibility for thestructuring of a debate and its automatic furtheringare taken on by the machine. The key bottleneck insuch systems, however, is the availability of data.

Articles

32 AI MAGAZINE

Many of the resources available in the largest open-ly accessible datasets provided by the Argument Web(searchable at aifdb.org; see Bex et al. [2013]), alsoindicate argument provenance and authorship. Byassociating arguments with arguers, a ready-builtmechanism becomes available for populating agentknowledge bases. This forms the foundation of theArvina system, shown in figure 2, which makes use ofa general-purpose platform for executing dialoguegames or protocols. There are two approaches to suchgeneralized dialogue execution that allow systems todeliver mixed-initiative argumentation wherebyhumans and software agents can play one of a num-ber of debating dialogue games on a level playingfield (one in which, indeed, infrastructure may haveno way of distinguishing human from software play-ers). The first, lightweight, approach extends existingprogramming languages with a small set of commu-nicative coordination constructs. The advantage ofthis approach is that systems for debating technolo-gy can be rapidly prototyped with few new conceptsrequired. The problem is that for engineering practi-cal systems, it leaves a very large amount of dialogi-cal componentry to be defined by researchers anddevelopers. The solution to this problem is offered inrich dialogue execution, by which all the usual com-ponents of dialogue games (participants, commit-ments, turn taking, backtracking, and others) arebaked in to a rich domain-specific language, which,though less elegant than a lightweight approach,gives the developer a much more extensive languagefor engineering. This is the approach taken in theDGDL language used by Arvina. Though such anapproach lacks the general-purpose flexibility thatmight be hoped for from some future system capableof full NLP understanding, it provides a flexible, intu-itive, and naturalistic mechanism for navigating com-plex information spaces (such as climate change,abortion, civil liberties, and so on).

These complex information spaces are often suffi-ciently intricate and detailed that users —both thosewho have engaged in the debate and also those whoare reviewing it post hoc — need additional mecha-nisms to make sense of the otherwise potentiallyoverwhelming deluge of data. This has led to two dis-tinct approaches. The first involves a range of debateanalytics that aggregate, calculate, and interrogatethe argument structures created in a debate allowinginsight into, for example, strong and weak argu-ments, stimulating and boring participants, centraland peripheral issues, and so on. The secondapproach focuses on augmented debate, adding richstreams of additional information concerning thedebate presented visually and often tied to a videorecording. One of the most sophisticated and rich sys-tems of augmented debate with analytics is the EDVproject, which was trialled with the 2014 UK electiondebates. In both cases, situated, linguistic and dialog-ical metrics (such as dominance and relationships

between speakers) are used in combination with met-rics based on structured argumentation (which yieldinsights into the inferential structures created by par-ticipants) and those based on abstract argumentation(which can contribute to assessing debatewide fea-tures such as which arguments are winning).

As debate technology starts to mature, we are thusseeing not only increasingly sophisticated systems forsupporting and contributing to debates, but alsocomplementary systems for making sense of the largedatasets that result.

Argumentation SolversThe scenarios outlined previously require the use ofeffective systems for solving various problems relatedto computational argumentation. Recall theapproach of abstract argumentation discussed in theassessment layer subsection (see also figure 2). Thisapproach models argumentation scenarios as graphsand a central question is how to determine the exten-sions of such a graph, that is, sets of arguments thatcan collectively be accepted, given the attack relationrepresented by the arcs. The literature offers variousways to formally define this concept of acceptability,but the computational problem of extracting anysuch a set from a given graph is usually hard to solveand can exhibit complexity beyond P and NP. Forexample, deciding, for a given graph and a givenargument, whether the argument is contained in allextensions under the so-called preferred semantics(Dung 1995) is ΠP

2-complete, which is regarded ashighly infeasible. Roughly speaking, while (deter-ministic) algorithms to NP-hard problems usuallyrequire exponential worst-case runtime, algorithmsfor ΠP

2-complete problems may even exhibt superex-ponential worst-case run time. However, solutions tothese kind of problems are required by applicationsutilizing argumentative decision procedures, andrecently, the community started to address thesechallenges by developing specific argumentationsolvers for both abstract and structured argumenta-tion settings.

Solvers for abstract argumentation are usually gen-eral-purpose tools similar to SAT-solvers (for a reviewof SAT-solvers, see Gomes et al. [2008]) and solve thecomputational problem of determining acceptablearguments from a given graph. Solvers for this settingusually fall into one of two categories, the reduction-based approach and the direct approach (Charwat etal. 2015). In the reduction-based approach, the prob-lem at hand is translated into another formalism(such as SAT) and specialized solvers for that formal-ism are used to solve the original problem. In thedirect approach, the peculiarities of abstract argu-mentation frameworks are exploited to directly solvethe problem at hand without the use of another for-malism.

Systems addressing the structured argumentation

Articles

FALL 2017 33

setting (see the Structural Layer previously men-tioned) are additionally concerned with problemsrelated to argument construction and defeat discov-ery. In many application scenarios, knowledge is rep-resented as facts and rules or, more generally, as for-mulas in some logic. In order to apply argumentationtechnology, arguments have to be constructed bycombining formulas and conflicts between differentarguments have to be detected.13, 14, 15, 16 Many sys-tems for structured argumentation generate argu-mentation graphs such as the one shown in figure 2as output and use abstract argumentation solvers forthe actual determination of acceptable arguments.However, as actual application contexts may requirethe user to specify facts and rules rather than theinduced arguments, systems for structured argumen-tation are a key element for the adoption of argu-mentation technology in actual applications.

In order to evaluate the state of the art of argu-mentation solvers, the International Competition onComputational Models of Argumentation (ICCMA)17

was initiated in 2014 and organized its first contest in2015 (Thimm et al. 2016). For the first contest, thefocus was on problems related to abstract argumenta-tion and solvers were evaluated based on their run-time performance for computing extensions or decid-ing on acceptance of arguments with respect tocomplete, preferred, stable, and grounded semanticsof abstract argumentation, (compare with Dunn[1995]). There were 18 participating systems and the

best performing ones achieved significant improve-ments with respect to the state of the art. Based onthese encouraging results, a second contest will beheld in 2017.

ConclusionsDeveloping artificial tools that capture the humanability to argue is an ambitious research goal, and itmay ultimately prove to be as difficult as developingAI in general.

As described in this article, current researchaddresses a range of applications like law, medicine, e-government, debating, where argument-basedapproaches have shown to be beneficial for intelli-gent activities like sense making and decision mak-ing. These areas witness an increasing integration ofproactive support and automated reasoning capabili-ties, complementing the functionalities offered byother useful but more passive tools like argumentvisualization systems.

Even more sophisticated roles for artificial argu-mentation tools are sought in the medium term andare the subject of recent research initiatives. Forinstance, there is a growing interest in developingcomputational persuasion systems able to assist peo-ple in making better choices in their daily activities.Consider scenarios such as a doctor persuading apatient to drink less alcohol, a road safety expert per-

Articles

34 AI MAGAZINE

Figure 3. Arvina.

suading drivers to not text while driving, or an onlinesafety expert persuading users of social media sites tonot reveal too much personal information online.These all involve the persuader finding the right argu-ments to use with respect to the persuadee’s knowl-edge, priorities, and biases. Using artificial argumen-tation to build automated persuaders provides severalinteresting research challenges the community isstarting to tackle. For example, the Framework forComputational Persuasion project18 is developing acomputational model of argument for behaviorchange in health care. This kind of application callsfor the development of rhetorical and dialogical lay-ers in figure 1.

In the longer term, there are exciting possibilitiesfor developing artificial agents able to use argumen-tation as a general pattern of interaction with otheragents, exactly like humans are able to argue withother humans to achieve collectively useful behav-iors. Consider a situation where heterogeneous robotsneed to work together to survey a situation such as alarge building on fire. Exactly like in a team of fire-fighters, each robot will have direct perception ofsome local situation and will need to exchange infor-mation and coordinate actions with other robots in adynamic environment where, altogether, informa-tion will always be incomplete and inconsistent and,consequently, goals and action plans might need tobe revised at any moment. Different capabilities ofthe team members will have to be taken into accounttoo. These features call for high-level arguing capa-bilities, applicable in a variety of contexts among het-erogeneous agents whose unique common propertymight be the capability to argue itself. In this senseartificial argumentation promises, in the long term,to provide a sort of universal social glue for linkingtogether, in a plug and play and cooperative manner,robots and any other kind of intelligent agents.

AcknowledgementsThis work has been partially supported by EPSRCgrants EP/N008294/1 and EP/N014871/1, by EUH2020 research and innovation programme underthe Marie Sklodowska-Curie grant agreement No.690974 for the project MIREL: MIning and REason-ing with Legal texts, and by funds provided by theInstitute for Computer Science and Engineering, Uni-versidad Nacional del Sur, Argentina.

Notes

1. “Humans argue” is a truism. Either you already believe itor you would need to argue against it.2. comma.csc.liv.ac.uk/.3. www.iospress.nl/journal/argument-computation/. 4. plato.stanford.edu/entries/logic-informal/.5. petition.parliament.uk/.6. debategraph.org/.7. idebate.org/debatabase.

8. www.policy-impact.eu/.9. carneades.github.io/.10. debatepedia.idebate.org/en/index.php/Debate: Randomalcohol breath tests for drivers.11. Argument A2 has to be read as [Random breath tests to]other drivers [can hardly be called an invasion of privacy oran investigation without due cause as they are] a major lia-bility to the safety and lives of other drivers.12. The number of involved annotators should be > 1 inorder to allow for the calculation of this measure and, as aconsequence, produce a reliable resource.13. lidia.cs.uns.edu.ar/delp client/.14. toast.arg-tech.org.15. tweetyproject.org.16. robertcraven.org/proarg.17. argumentationcompetition.org.18. www.computationalpersuasion.com.

ReferencesBench-Capon, T., and Dunne, P. 2007. Argumentation inArtificial Intelligence. Artificial Intelligence 171(10–15): 619–641.Besnard, P.; Javier García, A.; Hunter, A.; Modgil, S.; Prakken,H.; Simari, G.; and Toni, F. 2014. Introduction to StructuredArgumentation. Argument and Computation 5(1): 1–4. doi.org/10.1080/ 19462166.2013.869764Bex, F.; Lawrence, J.; Snaith, M.; and Reed, C. 2013. Imple-menting the Argument Web. Communications of the ACM56(10): 951–989. doi.org/10.1145/2500891Charwat, G.; Dvorak, W.; Gaggl, S.; Wallner, J.; and Woltran,S. 2015. Methods for Solving Reasoning Problems inAbstract Argumentation — A Survey. Artificial Intelligence220: 28–63. doi.org/10.1145/2500891Chesñevar, C.; McGinnis, J.; Modgil, S.; Rahwan, I.; Reed, C.;Simari, G.; South, M.; Vreeswijk, G.; and Willmott, S. 2006.Towards an Argument Interchange Format. Knowledge Engi-neering Review 21(4): 293–316. doi.org/10.1017/S0269888906001044Dung, P. 1995. On the Acceptability of Arguments and ItsFundamental Role in Nonmonotonic Reasoning, Logic Pro-gramming and n-Person Games. Artificial Intelligence 77(2):321–358. doi.org/10.1016/0004-3702(94)00041-XGomes, C.; Kautz, H.; Sabharwal, A.; and Selman, B. 2008.Satisfiability Solvers. In Handbook of Knowledge Representa-tion, ed. F. Van Harmelen, V. Lifschitz, and B. Porter, 89–134.Amsterdam, The Netherlands: Elsevier.Gordon, T., and Karacapilidis, N. 1997. The Zeno Argumen-tation Framework. In Proceedings of the International Confer-ence on Artificial Intelligence and Law (ICAIL ’97), 10–18. NewYork: Association for Computing Machinery. doi.org/10.1145/261618.261622Horty, J., and Bench-Capon, T. 2012. A Factor-Based Defini-tion of Precedential Constraint. Artificial Intelligence and Law20: 181–214. doi.org/10.1007/s10506-012-9125-Hunter, A., and Williams, M. 2012. Aggregating EvidenceAbout the Positive and Negative Effects of Treatments. Arti-ficial Intelligence in Medicine 56(3): 173–190. doi.org/10.1016/j.artmed.2012.09.004Karacapilidis, N., and Papadias, D. 2001. Computer Sup-ported Argumentation and Collaborative Decision Making:The HERMES System. Information Systems 26(4): 259–277.

Articles

FALL 2017 35

tee of the COMMA conferences on Compu-tational Models of Argument. He is on theeditorial board of several journals, includingArtificial Intelligence (currently as an associ-ate editor).

Chris Reed is a professor of computer sci-ence and philosophy at the University ofDundee in Scotland, where he heads theCenter for Argument Technology. Reed hasbeen working at the overlap between argu-mentation theory and artificial intelligencefor more than 20 years, has obtained 6 mil-lion in funding, and has over 150 peer-reviewed papers in the area. He collaborateswith a wide range of partners such as IBMand the BBC and is also active in publicengagement and commercialization ofresearch, having served as executive director(CTO, CSO, and CEO) of three startup com-panies.

Guillermo R. Simari is a full professor inlogic for computer science and artificialintelligence at the Universidad Nacional delSur (UNS), Baha Blanca, Argentina. He stud-ied mathematics at the same university andreceived a master of science in computer sci-ence and a Ph.D. in computer science fromWashington University in Saint Louis, USA.The focus of his research is on the formalfoundations and effective implementationof defeasible reasoning systems forautonomous agents. He chairs the ArtificialIntelligence Research and DevelopmentLaboratory (LIDIA).

Matthias Thimm is a senior lecturer at theUniversity of Koblenz-Landau, Germany. Hereceived his Ph.D. from the University ofDortmund in 2011 on the topic of proba-bilistic reasoning with inconsistencies andhis habilitation degree from the Universityof Koblenz-Landau in 2016. His researchinterests include reasoning with uncertainand inconsistent information, in particularusing approaches from computational mod-els of argumentation and inconsistencymeasurement.

Serena Villata is a researcher at CNRS. Herwork focuses on argumentation theory andspans from formal models of argumentationto natural language–based and emotion-based ones. She is the author of more than60 scholarly publications including 11 jour-nal articles and 4 book chapters. She wasinvited to coauthor a chapter on processingargumentation in natural language texts inthe Handbook of Formal Argumentation. Shecoorganized the seminar Frontiers and Con-nections Between Argumentation Theoryand Natural Language Processing in July2014 (Bertinoro, Italy), one of the firstattempts to put together researchers fromcomputational linguistics and from argu-mentation theory.

doi.org/10.1016/S0306-4379(01)00020-5Lippi, M., and Torroni, P. 2016. Argumenta-tion Mining: State of the Art and EmergingTrends. ACM Transactions on Internet Tech-nology. 10:1–10:25.Modgil, S.; Toni, F.; Bex, F.; Bratko, I.;Chesñevar, C.; Dvorák, W.; Falappa, M.; Fan,X.; Gaggl, S.; García, A.; González, M.; Gor-don, T.; Leite, J.; Mozina, M.; Reed, C.;Simari, G.; Szeider, S.; Torroni, P.; andWoltran, S. 2013. The Added Value of Argu-mentation. In Agreement Technologies, ed. S.Ossowski, 357–403. Berlin: Springer.doi.org/10.1007/978-94-007-5583-3_21Palau, R. M., and Moens, M. 2011. Argu-mentation Mining. Artificial Intelligence inLaw 19(1): 1–22. doi.org/10.1007/s10506-010-9104-xPinkwart, N., and McLaren, B. M., eds. 2012.Educational Technologies for Teaching Argu-mentation Skills. Sharjah, UAE: Bentham Sci-ence Publishers.Pollock, J. 1992. How to Reason Defeasibly.Artificial Intelligence 57(1): 1 – 42. doi.org/10.1016/0004-3702(92)90103-5Prakken, H., and Sartor, G. 2015. Law andLogic: A Review from an ArgumentationPerspective. Artificial Intelligence 227: 214–245. doi.org/10.1016/j.artint.2015.06.005Rahwan, I., and Simari, G., eds. 2009. Argu-mentation in Artificial Intelligence. Berlin:Springer.Schneider, J.; Groza, T.; and Passant, A.2013. A Review of Argumentation for theSocial Semantic Web. Semantic Web 4(2):159–218.Thimm, M.; Villata, S.; Cerutti, F.; Oren, N.;Strass, H.; and Vallati, M. 2016. SummaryReport of the First International Competi-tion on Computational Models of Argu-mentation. AI Magazine 37(1): 102–104.doi.org/10.1609/aimag.v37i1.2640Walton, D.; Reed, C.; and Macagno, F. 2008.Argumentation Schemes. Cambridge: Cam-bridge University Press. doi.org/10.1017/CBO9780511802034Walton, R.; Gierl, C.; Yudkin, P.; Mistry, H.;Vessey, M.; and Fox, J. 1997. Evaluation ofComputer Support for Prescribing CAPSULEUsing Simulated Cases. British Medical Jour-nal 315(7111): 791–795. doi.org/10.1136/bmj.315.7111.791

Katie Atkinson is a professor of computerscience and head of the Department of Com-puter Science at the University of Liverpool,UK. Her research concerns computationalmodels of argument, with a particular focuson persuasive argumentation in practical rea-soning and how this can be applied indomains such as law, e-democracy, and agentsystems. Atkinson’s work covers both theo-retical and applied aspects; recent collabora-

tive projects have concerned the develop-ment of intelligent tools for a law company,and realization of tools to support e-democ-racy and legal knowledge-based systems.

Pietro Baroni is a full professor of computerscience and engineering at the Departmentof Information Engineering of the Universi-ty of Brescia, Italy. He received the “Laurea”degree in mechanical engineering from theUniversity of Brescia and a masters in infor-mation technology, from CEFRIEL, Milan.He is author of more than 120 scientificpapers in the area of AIe and knowledge-based systems, with a main focus on theoryand applications of computational argu-mentation. He was a founding member ofthe steering committee of the Computation-al Models of Argument (COMMA) confer-ence series, served as program chair of COM-MA 2016, and is currently coeditor-in-chiefof the Argument & Computation journal.

Massimiliano Giacomin is an associate pro-fessor in computer science and engineeringat the University of Brescia (Italy). Hereceived an M.S. in electronics engineeringfrom the University of Padova (Italy) and aPh.D. degree in information engineeringfrom the University of Brescia. His researchinterests are in the field of AI, includingknowledge representation and automatedreasoning (in particular, argumentation the-ory, fuzzy constraints, temporal reasoning),multiagent systems, and knowledge-basedsystems. He is an author of more than 100papers in these research areas. He is a mem-ber of the editorial board of Argument andComputation.

Anthony Hunter is a professor of artificialintelligence, and head of the Intelligent Sys-tems Research Group, in the Department ofComputer Science, University College Lon-don (UCL). His research interests are onaspects of inconsistency, argumentation,and knowledge merging.

Henry Prakken is a lecturer in artificialintelligence at the Computer ScienceDepartment of Utrecht University and pro-fessor in legal informatics and legal argu-mentation at the Law faculty of the Univer-sity of Groningen. He has master’s degreesin law (1985) and philosophy (1988) fromthe University of Groningen. In 1993 heobtained is Ph.D. degree (cum laude) at theFree University Amsterdam with a thesistitled Logical Tools for Modelling LegalArgument. His main research interests con-cern computational models of argumenta-tion and their application in multiagent sys-tems, legal reasoning, and other areas.Prakken is a past president of the Interna-tional Association for AI and Law, of theJURIX Foundation for Legal Knowledge-Based Systems, and of the steering commit-

Articles

36 AI MAGAZINE