Articles

Toward Artifcial Argumentation

Katie Atkinson, Pietro Baroni, Massimiliano Giacomin, Anthony Hunter, Henry Prakken, Chris Reed, Guillermo Simari, Matthias Thimm, Serena Villata

■ The feld of computational models of umans argue.1 This distinctive feature is at the same is emerging as an important time an important cognitive capacity and a powerful aspect of artifcial intelligence research. Hsocial phenomenon. It has attracted attention and The reason for this is based on the careful analysis since the dawn of civilization, being inti- recognition that if we are to develop mately related to the origin of any form of social organiza- robust intelligent systems, then it is tion, from political debates to law, and of structured think- imperative that they can handle incom- ing, from philosophy to science and arts. plete and inconsistent information in a way that somehow emulates the way As a cognitive capacity, argumentation is important for humans tackle such a complex task. handling conficting beliefs, assumptions, viewpoints, opin- And one of the key ways that humans ions, goals, and many other kinds of mental attitudes. When do this is to use argumentation either we are faced with a situation where we fnd that our infor- internally, by evaluating and mation is incomplete or inconsistent, we often resort to the counterarguments‚ or externally, by for use of arguments in favor and against a given position in instance entering into a discussion or order to make sense of the situation. When we interact with debate where arguments are exchanged. other people we often exchange arguments in a cooperative As we report in this review, recent devel- or competitive fashion to reach a fnal agreement or to opments in the feld are leading to tech- nology for artifcial argumentation, in defend and promote an individual position. the legal, medical, and e-government domains, and interesting tools for argu- ment mining, for debating technologies, and for argumentation solvers are emerging.

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

Occurring continuously both in our mind and in and specifc formalisms may inextricably merge the social arena, argumentation pervades our intelli- together aspects relevant to different layers. gent behavior and the challenge of developing artif- cial argumentation systems appears to be as diverse Structural Layer and exciting as the challenge of artifcial intelligence This layer concerns the structure of the arguments itself. and how they are built: essentially it specifes, in a Indeed, this rich and important phenomenon given context, what an argument looks like, in terms offers an opportunity to develop models and tools for of its internal structure, and which are the ingredi- argumentation and to conceive autonomous artifcial ents for its construction. To exemplify, in contexts agents that can exploit these models and tools in the where arguments are built from a logical knowledge cognitive tasks they are required to carry out. To this base the ingredients are the logical formulae includ- purpose, a number of interesting lines of research are ed in the knowledge base. Then one way to build being investigated within artifcial intelligence and arguments is by simply applying the logic of the lan- several neighbor felds, leading to the establishment guage in which the knowledge base is stated to derive of computational models of argument as a promising conclusions. An argument here can be seen as a pair interdisciplinary research area. Progress in this area is ( , ) where is a subset of the knowledge base (a set expected to contribute to signifcant advances in the of formulae) that logically entails (a formula). Here, understanding and modeling of various aspects of Φisα called theΦ support, and is the claim, of the argu- human intelligence. ment. Other approaches considerα argument con- In this article, we review formalisms for capturing Φstruction from knowledge basesα as applying rules to various aspects of argumentation, and we present the formulas from the knowledge base, where the advances in their applications, with the aim to com- rules may be defeasible. In these rule-based approach- municate how research is making progress toward es an argument is typically seen as a tree whose root the goal of making artifcial argumentation tech- is the claim or conclusion, whose leaves are the nologies and systems a mature and widespread reali- premises on which the argument is based, and whose ty. In this brief review, we are unable to discuss or cite structure corresponds to the application of the rules all the relevant literature, and we suggest that the from the premises to the conclusion. Investigations interested reader seek more detailed coverage of the into the structural layer were initiated by Pollock foundations from Rahwan and Simari (2009), of (1992). Prominent examples of rule-based formalisms applications from Modgil et al. (2013), and of recent are ASPIC+, assumption-based argumentation (ABA), developments from the proceedings of the Interna- and defeasible logic programming (DeLP). For a tuto- tional Conference on Computational Models of rial introduction to formalisms for structured argu- Argument series,2 and the Argument and Computation mentation, see Besnard et al. (2014). journal.3 Arguments are not built from knowledge bases only, however. For instance, interactive systems that acquire arguments from users may adopt the Models of Argument approach of argumentation schemes (Walton, Reed, Computational models of argument are being devel- and Macagno 2008), namely stereotypical reasoning oped to refect aspects of how humans build, patterns, where in addition to the premises and the exchange, analyze, and use arguments in their daily claim, a set of critical questions is considered. Criti- lives to deal with a world where the information may cal questions provide a sort of checklist of issues that be controversial, incomplete, or inconsistent (Bench- can be raised to challenge arguments built on the Capon and Dunne 2007, Rahwan and Simari 2009). basis of a given scheme. Argumentation schemes The diversity of the manifestations of arguments in have also been used as a source of defeasible infer- real life implies diversity in the relevant models too ence rules in rule-based approaches to argument con- and the impossibility to reduce the vast available lit- struction from knowledge bases. In addition, argu- erature to a single reference scheme. It is possible mentation schemes are often considered in the however to identify some layers that can be regarded context of argument mining (see the Argument Min- as basic building blocks for the construction of an ing section) where the goal is to identify and extract argumentation model. Specifc modeling approaches the argumentative structures embedded in a natural may differ in the selection of which layers they actu- language source, providing a machine-processable ally use, in the way the selected layers are combined, representation of them. and in the formalization adopted within each layer. The variety of existing argument models raises the We consider the following fve main layers: struc- issue of exchanging or sharing arguments among dif- tural, relational, dialogical, assessment, and rhetori- ferent systems. This problem is addressed by the cal. They are described in the following and also sum- argument interchange format initiative (Chesñevar marized in Figure 1. Note that while each layer has its et al. 2006), aimed at providing an interlingua own nature and distinctive traits, the boundaries between various more concrete argumentation lan- between layers may not be so neat in some contexts, guages, on the basis of a generic abstract ontology.

26 AI MAGAZINE Articles

Structural layer: How are arguments constructed?

Relational layer: What are the relationships between arguments?

Dialogical layer: How can argumentation be undertaken in dialogues?

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

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

Figure 1. Key Aspects of Argumentation.

Relational Layer strength to credibility to value-based evaluations. Arguments do not live in isolation and are linked to What relationships are signifcant and how to each other by various types of relations: the relation- identify them are highly context-dependent matters. al layer deals with identifying and formally repre- Note in particular that identifying argument rela- senting them, in view of their use in other layers or tions may be an easy mechanical procedure in set- even for descriptive and presentation purposes, since tings where arguments are formally built from a they are essential for an understanding of what is knowledge base, while in an argument mining sce- actually going on in an argumentation process. nario it is a task as challenging as the identifcation Examples of important relationships are (1) the sub- of the arguments themselves. argument and superargument relationships, indicat- Dialogical Layer ing how an argument is built incrementally on top of other arguments; (2) the attack relationship, indicat- This layer deals with the exchange of arguments ing that an argument is incompatible with another among different agents (or even between an agent argument in some sense, for example, because they and itself, in a scenario where argumentative reason- have contradictory claims, or one claim contradicts ing is conceived as a monological activity) according some premise or assumption on which the other is to formal dialogue rules. Agents may engage in the based; (3) the support relationship, intuitively mean- exchange of arguments for a variety of purposes with ing that an argument provides some backing to several dialogue types having been identifed in the another, and admitting several, even rather dissimi- literature, like inquiry, negotiation, information- lar, interpretations, depending on the actual nature seeking, deliberation, and persuasion. In all cases the of this backing; (4) a preference relationship, ranking exchange can be formalized as a dialogue game, arguments according to some criterion, and admit- which is normally made up of a set of communica- ting again a variety of instantiations ranging from tive acts called moves, and a protocol specifying

FALL 2017 27 Articles

A = Patient has A = Patient has A = Patient has 3 1 2 emphysema, which hypertension so hypertension so is a contraindication prescribe diuretics prescribe betablockers for betablockers

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 A2 because it provides a counterargument to A2.

which moves can be made at each step of the dia- feld of computational models of argument that has logue. It concerns representing and managing the attracted most research attention in the last two locutions exchanged between the agents involved, as decades, due to its generality and theoretical clean- well as specifying the contents of these locutions in ness. In particular Dung (1995) has shown the ability terms of entire arguments or components of argu- of the formalism to capture as instances several other ments. Moreover, the dialogue protocol may establish approaches, especially in the area of nonmonotonic the allowed moves on the basis of argument relation- reasoning. Dung’s approach abstracts from the origin ships. For instance, a protocol may specify that a and nature of the attack relation. A natural idea is to move is legal only if it presents an argument attack- defne this relation in terms of a more basic notion of ing an argument presented in a previous move. For confict between arguments (for example, two argu- these reasons the dialogical layer requires strict con- ments having contradictory conclusions) and a nections with the structural and relational layers. notion of relative argument strength or preference. In Moreover some dialogue protocols are defned so as the literature, there are two ways to connect these to embed an argument assessment method: in these ideas to Dung’s frameworks. The frst approach leaves cases the dialogical layer is intertwined with the Dung’s frameworks as they are but connects them assessment layer, described next. with models at the structural layer of argument to defne attack in terms of preferences or argument Assessment Layer strength while taking the structure of arguments into This layer concerns the assessment of a set of argu- account. The second approach instead extends ments and of their conclusions in order to establish Dung’s frameworks with some abstract notion of their justifcation status. The need for this layer arises argument strength or preference, while possibly also from the presence of attacks among arguments, pre- adding an abstract support relation between argu- venting them so as to be accepted altogether and call- ments. Moreover, while most approaches consider a ing for a formal method to solve the confict. This qualitative notion of acceptance, quantitative assess- problem is addressed in a principled and highly styl- ments methods are being investigated too. ized form in the context of the theory of abstract argu- Further, it must be noted that the evaluation of mentation frameworks (Dung 1995), where argu- argument acceptability is only a part, actually the ments are treated as abstract entities, deprived of any most basic one, of the assessment tasks required in an structural property and of all their relations but attack. argumentative process. In particular the fnal goal of We give an example of an argumentation framework, an agent is usually the assessment of the justifcation based on textual arguments, in fgure 2. Given its status of the statements supported by arguments, abstract nature, an argumentation framework is often which, in the end, amounts to determining what to referred to as argument graph, and this term is also believe or what to do. Since many arguments may used to refer to similar representations where addi- have the same conclusion, assessing the status of a tional relations, like support, are considered. statement involves a synthesis of the statuses of the An abstract argumentation semantics is a formal cri- arguments supporting it. As in real life, the task of terion to determine which sets of arguments, called deciding what to believe may be carried out adopting extensions, are able to survive the confict together different attitudes, ranging from extremely skeptical and can be regarded as collectively acceptable. to extremely credulous, corresponding to different Abstract is probably the sub- formal methods for statement justifcation synthesis.

28 AI MAGAZINE Articles

Rhetorical Layer which involves interpreting these rules. Two infuen- Normally argumentation is undertaken in some tial AI and law models of this are the HYPO system by wider context of goals for the agents involved, and Kevin Ashley and its successor CATO by Vincent so individual arguments are presented with some Aleven, which model how lawyers in common-law wider aim and according to some strategical consid- jurisdictions make use of past decisions when argu- erations. For instance, if an agent is trying to per- ing a case. Their underlying argumentation model is suade another agent to do something, then it is like- for factor- or dimension-based reasoning, where cas- ly that some rhetorical device is harnessed and this es are collections of abstract fact patterns that favor will refect the nature of the arguments used (for or oppose a conclusion, either in an all-or-nothing example, a politician may refer to investing in the fashion (factors) or to varying degrees (dimensions). future of the nation’s children as a way of persuading This work inspired subsequent formal work using the colleagues to vote for an increase in taxation). With tools of formal argumentation, resulting in formal- the roots of the study of rhetoric going back to Aris- ized versions of traditional legal argument forms totle,4 recent studies into aspects of the rhetorical lev- such as appeal to precedent and policy and the bal- el include believability and impact of arguments ancing of goals, values, and interests (Horty and from the perspective of the audience, use of threats Bench-Capon 2012). and rewards, appropriateness of advocates, and val- Finally, when the facts have been classifed, the ues of the audience. The rhetorical layer may be legal rules must be applied to them. Legal rules can absent in some contexts, for example, when neu- have exceptions or confict on other grounds. Rule- trally building arguments from a knowledge base, but based argumentation logics with preferences have can permeate all the other layers in other contexts proved useful here. since goal-oriented considerations may drive the Legal reasoning usually takes place in the context decisions of which arguments to build, taking into of a dispute between adversaries, within a prescribed account their relations with other arguments, of legal procedure. This makes the setting inherently whether, how, and when to use the arguments in a dynamic and multiparty, and raises issues of strategy dialogue, and of which assessment method (for and choice. For example, there is work on optimal example, whether a more skeptically or more credu- strategies for adversaries in debates with an adjudica- lously oriented one) to apply. tor, given their preferences over the possible out- The following sections review several prominent comes of a debate and their estimates of what the domains that exploit computational models of argu- adjudicator will likely accept. ment for the development of actual applications and, While the theoretical advances on models of legal at the same time, stimulate the relevant theoretical argument have been impressive and a number of development by providing case studies and impor- valuable prototype systems have been developed, no tant modeling challenges. systems have been deployed in everyday practice yet. One reason is the conservative attitude to technology in the legal world and its billing-by-the-hour culture, Legal Argumentation which does not stimulate innovation. Another reason The law is an obvious application domain for argu- is the fact that building realistic systems of legal argu- mentation research, since legal reasoning is essen- ment requires vast amounts of commonsense knowl- tially argumentative and to a large extent recorded in edge. However, recently things have changed. First, documents. This has led to highly stylized forms of clients of law frms increasingly demand the use of argumentation, which makes it easier to formulate modern technology. Moreover, the recent advances in and validate formal and computational models of natural language processing, machine learning, and argument than in many other domains. In this sec- data science combined with the massive digital avail- tion, we briefy discuss work and research themes in ability of legal data and information have created the this area. A more detailed survey can be found in the prospects for combining AI models of legal argument paper by Prakken and Sartor (2015). with argument mining techniques. In fact, two of the In legal cases, frst the facts have to be determined. frst argument mining projects were on legal argu- Because of the diverse nature of the evidence in most ment (Palau and Moens 2011). If this technology is cases and the need for explanation to statistical combined with AI and law’s computational models of laypeople, legal evidential reasoning is an excellent argument, then practical applications of these models test bed for combined qualitative and quantitative could be well within reach. models of defeasible reasoning. At the practical side, so-called sense-making systems have been proposed, Medical Argumentation with which crime investigators or triers of fact can structure their arguments and scenarios to make Health care is a potentially important domain for sense of a large body of evidence. developing and applying computational models of After the facts of a case have been established, they argument. It is common for health-care information must be classifed under the conditions of legal rules, to be complex, heterogeneous, incomplete, and

FALL 2017 29 Articles

inconsistent, and therefore argumentation is appeal- constitute an argument graph, and using abstract ing for those involved as it allows for important con- argumentation semantics, the winning arguments are ficts to be highlighted and analyzed and unimpor- identifed, and thereby argument-based recommen- tant conficts to be suppressed. dations for which treatments are superior can be One of the pioneers of argumentation technology, identifed. The approach has been evaluated by com- John Fox, developed a number of prototype systems parison with recommendations made in published for medical decision support such as the Capsule sys- health-care guidelines (Hunter and Williams 2012) tem (Walton et al. 1997). Capsule supports a family and it has been used to publish, in the medical liter- practitioner when she or he is about to prescribe a ature on lung cancer, a more refned systematic specifc drug for a patient. The system uses a standard review of the evidence. An ongoing study is using this database of equivalent treatments that is routinely technique in a systematic review on brain cancer for used by clinicians, and the patient records, to provide publication by Cochrane. arguments pro and con each of the alternatives. The These examples are just two of a number of appli- arguments are based on whether the patient has pre- cations of argumentation being developed for sup- viously expressed a preference for or against the alter- porting health-care professionals and patients. Fur- native, whether the patient has previously exhibited ther applications include dealing with the conficts a negative reaction to the alternative, whether there that can occur when using multiple clinical guide- is possible negative interaction with other drugs lines, supporting multidisciplinary teams of health- being taken by the patient, and the relative cost of the care professionals when dealing with diffcult clinical alternative. In a formal trial of the Capsule system, cases, and supporting medical image interpretation. 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- e-Government ance with medical guidelines. An important feature of democracies is that citizens Over recent years, there has been substantial shift can engage their governments in dialogues about in health care to evidence-based practice. This means policies. Traditionally this was done by writing letters that health-care professionals need to use the best and holding town hall debates, but over the past two available evidence to inform their decision making. decades new methods of interaction have been devel- For deciding on interventions, this normally calls for oped to exploit the benefts of current technology. evidence from randomized clinical trials. The prob- Citizens may wish to respond in several ways to pol- lem with this is that there are many such trials pub- icy proposals made by their governments. They may lished each year, and it is diffcult for clinicians to simply seek a justifcation of the proposed policy; keep abreast of this literature. To help them, there are they may wish to object to the proposed policy; or medical guidelines and systematic reviews that aggre- they may want to propose policies of their own. Such gate this evidence by providing recommendations. dialogues can be facilitated through tools to support Unfortunately, these recommendations can rapidly e-democracy, and computational models of argument become out of date, they do not take local circum- can be put to effective use in such tools. stances into account, and they normally do not con- For example, consider a local government authori- sider patients with comorbidities. To address these ty that is deciding what to build on a plot of waste- problems, an argument-based approach to aggregat- land in a community. One proposal by the local ing clinical evidence has been proposed by Hunter authority may be to permit the building of a new and Williams (2012). The framework is a formal supermarket on the grounds that this will provide approach to synthesizing knowledge from clinical tri- jobs for the local community and shopping facilities als involving multiple outcome indicators (where an for local residents. Citizens may be consulted on this outcome indicator is either positive such as the num- proposal and critique this policy as well as put for- ber of patients who survive the disease after 1 year, or ward their own proposals. For example, the local 2 years, and so on, or negative such as the proportion authority’s proposal could be critiqued by stating that of those treated who have a particular side-effect). the action of building a new supermarket will not Based on the available evidence, evidence-based argu- have the intended effect of creating jobs as there will ments are generated for claiming that one treatment be job losses from local shop owners being put out of is superior to another for a given patient. business by the supermarket. An alternative proposal Preference criteria over evidence-based arguments could be that instead of building a new supermarket, are specifed in terms of the outcome indicators, and the site should be used to build a new play center for the magnitude of those outcome indicators, in the local residents’ children. Such opinions can be evidence. Various kinds of counter-arguments attack formed into arguments by distinguishing the premis- the evidence-based arguments depending on the es (for example, there is little unemployment in the quality of evidence used (for example, evidence community and play centers promote social interac- could be attacked because a trial was not conducted tion) and conclusion (we should build a new chil- correctly). The arguments and counter-arguments dren’s play center). Argumentation-based tools can

30 AI MAGAZINE Articles then be put to use to facilitate such debates. domain of e-government an ideal one for the study Simple tools like e-petitions5 can transform tradi- and application of tools that use computational mod- tional paper-based communication into digital com- els of argument on a large scale. munication, but recent advances have been made in the development of tools that exploit the web, such as the on-line argument mapping tools Debategraph6 Argument Mining and Debatabase,7 which enable users to model In recent years, the growth of the web, and the rap- debates by considering issues, and their pros and idly increasing amount of diverse textual data pub- cons, relevant to a debate. With such tools users can lished there, have highlighted the need for methods freely insert and modify contributions, but the argu- to identify, structure, and summarize this huge ments entered are not required to conform to any resource. Online newspapers, blogs, online debate particular semantics that would support coherence platforms, and social networks, as well as normative and argument evaluation. A comprehensive survey of and technical documents, provide a heterogeneous the state of the art in web-based argumentation tools fow of information where natural language argu- can be found in the paper by Schneider, Groza, and ments can be identifed and analyzed. The availabili- Passant (2013). ty of such data, together with the advances in natural Early collaborative decision support systems such language processing and machine learning, have sup- as Zeno (Gordon and Karacapilidis 1997) and HER- ported the rise of a new research area called argument MES (Karacapilidis and Papadias 2001) contained mining. The main goal of argument mining is the more structure by making use of the IBIS (Issue Based automated extraction of natural language arguments Information Systems) model of argument. Use of this and their relations from generic textual corpora, with model enabled a particular problem or issue to be the fnal goal to provide machine-readable structured decomposed into a number of different positions for data for computational models of argument and rea- which arguments can then be created to attack or soning engines. defend the positions until the issue is settled. Two main stages have to be considered in the typ- In recent work to consolidate different tasks rele- ical argument mining pipeline, from the unstruc- vant for e-democracy tools, on a recent European tured natural language documents toward structured project called IMPACT,8 an argumentation toolbox (possibly machine-readable) data: was created that consists of four interconnected mod- Argument Extraction: The first stage of the pipeline is to ules: an argument reconstruction tool; a structured detect arguments within the input natural language consultation tool; an argument visualisation tool; texts. The retrieved arguments will thus represent the and, a policy modeling tool. This is a decision-sup- nodes in an argument graph returned by the system. This step may be further split into two different stages port tool that makes use of structured theories of such as the extraction of arguments and the further argument representation and evaluation to enable detection of their boundaries. Many approaches have public opinion gathering on political issues from recently been applied to tackle this challenge adopting which conclusions can be drawn concerning how different methodologies like for instance support vec- government policies are presented, justifed, and tor machines, naïve Bayes classifiers, logistic regres- viewed by the users of the system. The tool uses argu- sion. mentation schemes (Walton, Reed, and Macagno Relation Extraction: The second stage of the pipeline 2008) to structure the information presented to the consists in constructing the argument graph to be users, but within the front-end interface, this struc- returned as output of the system. The goal is to identi- ture is implicit in order to facilitate ease of learning fy what are the relations holding between the argu- and use. Once opinions have been gathered, the argu- ments identified in the first stage. This is an extreme- ments generated are evaluated through the use of val- ly complex task, as it involves high-level knowledge ue-based argumentation frameworks (see the chapter representation and reasoning issues. The relations in Rahwan and Simari [2009]) to provide users with between the arguments may be of a heterogeneous nature, like attack, support, or entailment. This stage is support for which actions are justifed, according to also responsible for identifying, in structured argu- the facts and social interests promoted by the differ- mentation, the internal relations of the components ent arguments. Debates that have been modelled, of an argument, such as the connection between the using in particular the Parmenides system (see Rah- premises and the claim. Being an extremely challeng- wan and Simari [2009]), cover local issues, such as ing task, existing approaches assume simplifying whether to introduce more speed cameras on danger- hypotheses, like the fact that evidence is always asso- ous stretches of road, and wider national issues such ciated with a claim. as the UK debate about whether to ban fox hunting. To illustrate, we consider the following example This strand of work continues to be expanded within adapted from an online debate about random sobri- the Structured Consultation Tool developed as part of ety tests for drivers and their consistency with human the IMPACT project mentioned previously and the rights.10 We start with the unstructured natural lan- Carneades tools.9 guage text from which we frst aim to extract the The richness of policy debates clearly makes the arguments, and then to identify their relations:

FALL 2017 31 Articles

Random breath tests to public vehicle drivers can hard- ing tools used to analyze, aggregate, synthesize, struc- ly be called an invasion of privacy or an investigation ture, summarize, and reason about arguments in without due cause, because public safety is at stake. texts, with more sophisticated natural language pro- Random tests are routinely carried out by many train cessing (NLP) methods. However, and considering the and bus companies and are being introduced on air- complexity of the task, to do so it is still necessary to lines as well. The same applies for other drivers, who are a major liability to the safety and lives of other reach a deeper level of understanding of the inner drivers. workings of natural language, and evolve new meth- ods expanding the ones currently found in natural Randomly testing employees cannot be considered an language processing. invasion of privacy. People who have to take random Moreover, to tackle these challenging tasks, high- breath tests to drive trucks or fly planes as part of their quality annotated corpora are needed for use as a jobs are taking the test as part of their job. They are being paid and must do what their employer wants training set for any kind of aforementioned identif- them to do in order to keep their job. Searching ran- cation. These corpora are mainly composed by three dom people outside of the context of employment different elements: an annotated data set that repre- with no suspicion of a crime is very different as it sents the gold standard whose annotation has been erodes civil liberties and sets a dangerous precedent. checked and validated by expert annotators and is used to train the system for the required task (that is, The frst goal of the argument mining pipeline con- arguments or relations extraction), a set of guidelines sists in extracting the arguments from this text. In the to explain in a detailed way how the data has been previous example, we highlight the four arguments annotated, and fnally, the unlabelled raw corpus that that can be identifed: can be used to test the system after the training Random breath tests to public vehicle drivers can hard- phase. The reliability of a corpus is ensured by the cal- ly be called an invasion of privacy or an investigation culation of the interannotator agreement that meas- without due cause, because public safety is at stake [A ]. ures the degree of agreement in performing the anno- 1 12 Random tests are routinely carried out by many train tation task among the involved annotators. Current and bus companies and are being introduced on air- prototypes of argument mining systems require to be lines as well. The same applies for other drivers, who are trained against the data the task is addressed to, and a major liability to the safety and lives of other drivers the construction of such annotated corpora remains among the most time-consuming activities in this [A2]. Randomly testing employees cannot be considered an pipeline. For an exhaustive state of the art review on argu- invasion of privacy [A3]. People who have to take ran- dom breath tests to drive trucks or fy planes as part ment mining techniques and applications, we refer of their jobs are taking the test as part of their job. the reader to the paper by Lippi and Torroni (2016). They are being paid and must do what their employ- er wants them to do in order to keep their job. Search- Debating Technologies ing random people outside of the context of employment with no suspicion of a crime is very different as it erodes There is a long tradition of -aided debate

civil liberties and sets a dangerous precedent [A4]. systems with roots in e-democracy, decision support, Given these four arguments (that is, A1, A2, A3, and and so on. These systems have two things in com- A4), the relations among them have to be identifed. mon: frst, they implement idiosyncratic and new Let us consider for the explanatory purpose of this dialogical structures, or games, with little reuse or example that the two relations we aim at identifying incremental development. The second is that little or are the attack and the support relations only. In this no contribution to the debate itself is offered by the case, we have that, taking into account the temporal machine. The system role has been one of support dimension of the debate to decide the direction of the and facilitation only. With a variety of techniques for

relations, argument A3 supports argument A1, and automatically mining argument structures from both 11 argument A4 attacks argument A2. monological and dialogical resources, an exciting It is worth noticing that the identifcation of the new possibility is opened up for not just supporting arguments and their relations is much more subtle and enhancing new human-human arguments but and ambiguous than what emerges from this also developing new human-machine arguments: explanatory example, and may often be a matter of this is the space of debate technology, a specifc sub- interpretation that current state-of-the-art argument feld of argument technology in general. mining systems cannot tackle yet. For instance, argu- Several systems have demonstrated stand-alone

ment A1 can be considered as a subargument of argu- applications of debate technology focusing on ment A2 as “The same applies …” refers to A1, and domains of use such as pedagogy (Pinkwart and argument A3 can be interpreted as a kind of persua- McLaren 2012), in which both responsibility for the sive statement meant to strengthen argument A4. structuring of a debate and its automatic furthering To address this kind of issue and build more capa- are taken on by the machine. The key bottleneck in ble applications, it is necessary to enhance the exist- such systems, however, is the availability of data.

32 AI MAGAZINE Articles

Many of the resources available in the largest open- between speakers) are used in combination with met- ly accessible datasets provided by the Argument Web rics based on structured argumentation (which yield (searchable at aifdb.org; see Bex et al. [2013]), also insights into the inferential structures created by par- indicate argument provenance and authorship. By ticipants) and those based on abstract argumentation associating arguments with arguers, a ready-built (which can contribute to assessing debatewide fea- mechanism becomes available for populating agent tures such as which arguments are winning). knowledge bases. This forms the foundation of the As debate technology starts to mature, we are thus Arvina system, shown in fgure 2, which makes use of seeing not only increasingly sophisticated systems for a general-purpose platform for executing dialogue supporting and contributing to debates, but also games or protocols. There are two approaches to such complementary systems for making sense of the large generalized dialogue execution that allow systems to datasets that result. deliver mixed-initiative argumentation whereby humans and software agents can play one of a num- ber of debating dialogue games on a level playing Argumentation Solvers feld (one in which, indeed, infrastructure may have The scenarios outlined previously require the use of no way of distinguishing human from software play- effective systems for solving various problems related ers). The frst, lightweight, approach extends existing to computational argumentation. Recall the programming languages with a small set of commu- approach of abstract argumentation discussed in the nicative coordination constructs. The advantage of assessment layer subsection (see also fgure 2). This this approach is that systems for debating technolo- approach models argumentation scenarios as graphs gy can be rapidly prototyped with few new concepts and a central question is how to determine the exten- required. The problem is that for engineering practi- sions of such a graph, that is, sets of arguments that cal systems, it leaves a very large amount of dialogi- can collectively be accepted, given the attack relation cal componentry to be defned by researchers and represented by the arcs. The literature offers various developers. The solution to this problem is offered in ways to formally defne this concept of acceptability, rich dialogue execution, by which all the usual com- but the computational problem of extracting any ponents of dialogue games (participants, commit- such a set from a given graph is usually hard to solve ments, turn taking, backtracking, and others) are and can exhibit complexity beyond P and NP. For baked in to a rich domain-specifc language, which, example, deciding, for a given graph and a given though less elegant than a lightweight approach, argument, whether the argument is contained in all gives the developer a much more extensive language extensions under the so-called preferred semantics P for engineering. This is the approach taken in the (Dung 1995) is 2-complete, which is regarded as DGDL language used by Arvina. Though such an highly infeasible. Roughly speaking, while (deter- approach lacks the general-purpose fexibility that ministic) algorithmsΠ to NP-hard problems usually might be hoped for from some future system capable require exponential worst-case runtime, algorithms P of full NLP understanding, it provides a fexible, intu- for 2-complete problems may even exhibt superex- itive, and naturalistic mechanism for navigating com- ponential worst-case run time. However, solutions to plex information spaces (such as climate change, theseΠ kind of problems are required by applications abortion, civil liberties, and so on). utilizing argumentative decision procedures, and These complex information spaces are often suff- recently, the community started to address these ciently intricate and detailed that users —both those challenges by developing specifc argumentation who have engaged in the debate and also those who solvers for both abstract and structured argumenta- are reviewing it post hoc — need additional mecha- tion settings. nisms to make sense of the otherwise potentially Solvers for abstract argumentation are usually gen- overwhelming deluge of data. This has led to two dis- eral-purpose tools similar to SAT-solvers (for a review tinct approaches. The frst involves a range of debate of SAT-solvers, see Gomes et al. [2008]) and solve the analytics that aggregate, calculate, and interrogate computational problem of determining acceptable the argument structures created in a debate allowing arguments from a given graph. Solvers for this setting insight into, for example, strong and weak argu- usually fall into one of two categories, the reduction- ments, stimulating and boring participants, central based approach and the direct approach (Charwat et and peripheral issues, and so on. The second al. 2015). In the reduction-based approach, the prob- approach focuses on augmented debate, adding rich lem at hand is translated into another formalism streams of additional information concerning the (such as SAT) and specialized solvers for that formal- debate presented visually and often tied to a video ism are used to solve the original problem. In the recording. One of the most sophisticated and rich sys- direct approach, the peculiarities of abstract argu- tems of augmented debate with analytics is the EDV mentation frameworks are exploited to directly solve project, which was trialled with the 2014 UK election the problem at hand without the use of another for- debates. In both cases, situated, linguistic and dialog- malism. ical metrics (such as dominance and relationships Systems addressing the structured argumentation

FALL 2017 33 Articles

Figure 3. Arvina.

setting (see the Structural Layer previously men- best performing ones achieved signifcant improve- tioned) are additionally concerned with problems ments with respect to the state of the art. Based on related to argument construction and defeat discov- these encouraging results, a second contest will be ery. In many application scenarios, knowledge is rep- held in 2017. resented as facts and rules or, more generally, as for- mulas in some logic. In order to apply argumentation technology, arguments have to be constructed by Conclusions combining formulas and conficts between different Developing artifcial tools that capture the human 13, 14, 15, 16 arguments have to be detected. Many sys- ability to argue is an ambitious research goal, and it tems for structured argumentation generate argu- may ultimately prove to be as diffcult as developing mentation graphs such as the one shown in fgure 2 AI in general. as output and use abstract argumentation solvers for As described in this article, current research the actual determination of acceptable arguments. addresses a range of applications like law, medicine, e- However, as actual application contexts may require government, debating, where argument-based the user to specify facts and rules rather than the approaches have shown to be benefcial for intelli- induced arguments, systems for structured argumen- gent activities like sense making and decision mak- tation are a key element for the adoption of argu- mentation technology in actual applications. ing. These areas witness an increasing integration of In order to evaluate the state of the art of argu- proactive support and automated reasoning capabili- mentation solvers, the International Competition on ties, complementing the functionalities offered by Computational Models of Argumentation (ICCMA)17 other useful but more passive tools like argument was initiated in 2014 and organized its frst contest in visualization systems. 2015 (Thimm et al. 2016). For the frst contest, the Even more sophisticated roles for artifcial argu- focus was on problems related to abstract argumenta- mentation tools are sought in the medium term and tion and solvers were evaluated based on their run- are the subject of recent research initiatives. For time performance for computing extensions or decid- instance, there is a growing interest in developing ing on acceptance of arguments with respect to computational persuasion systems able to assist peo- complete, preferred, stable, and grounded semantics ple in making better choices in their daily activities. of abstract argumentation, (compare with Dunn Consider scenarios such as a doctor persuading a [1995]). There were 18 participating systems and the patient to drink less alcohol, a road safety expert per-

34 AI MAGAZINE Articles suading drivers to not text while driving, or an online 8. www.policy-impact.eu/. safety expert persuading users of social media sites to 9. carneades.github.io/. not reveal too much personal information online. 10. debatepedia.idebate.org/en/index.php/Debate: Random These all involve the persuader fnding the right argu- alcohol breath tests for drivers. ments to use with respect to the persuadee’s knowl- 11. Argument A2 has to be read as [Random breath tests to] edge, priorities, and biases. Using artifcial argumen- other drivers [can hardly be called an invasion of privacy or tation to build automated persuaders provides several an investigation without due cause as they are] a major lia- interesting research challenges the community is bility to the safety and lives of other drivers. starting to tackle. For example, the Framework for 12. The number of involved annotators should be > 1 in Computational Persuasion project18 is developing a order to allow for the calculation of this measure and, as a computational model of argument for behavior consequence, produce a reliable resource. change in health care. This kind of application calls 13. lidia.cs.uns.edu.ar/delp client/. for the development of rhetorical and dialogical lay- 14. toast.arg-tech.org. ers in fgure 1. 15. tweetyproject.org. In the longer term, there are exciting possibilities 16. robertcraven.org/proarg. for developing artifcial agents able to use argumen- 17. argumentationcompetition.org. tation as a general pattern of interaction with other 18. www.computationalpersuasion.com. agents, exactly like humans are able to argue with other humans to achieve collectively useful behav- References iors. Consider a situation where heterogeneous robots Bench-Capon, T., and Dunne, P. 2007. Argumentation in need to work together to survey a situation such as a Artifcial Intelligence. Artifcial Intelligence 171(10–15): 619– large building on fre. Exactly like in a team of fre- 641. fghters, each robot will have direct perception of Besnard, P.; Javier García, A.; Hunter, A.; Modgil, S.; Prakken, some local situation and will need to exchange infor- H.; Simari, G.; and Toni, F. 2014. Introduction to Structured mation and coordinate actions with other robots in a Argumentation. Argument and Computation 5(1): 1–4. doi. dynamic environment where, altogether, informa- org/10.1080/ 19462166.2013.869764 tion will always be incomplete and inconsistent and, Bex, F.; Lawrence, J.; Snaith, M.; and Reed, C. 2013. Imple- consequently, goals and action plans might need to menting the Argument Web. Communications of the ACM be revised at any moment. Different capabilities of 56(10): 951–989. doi.org/10.1145/2500891 the team members will have to be taken into account Charwat, G.; Dvorak, W.; Gaggl, S.; Wallner, J.; and Woltran, too. These features call for high-level arguing capa- S. 2015. Methods for Solving Reasoning Problems in bilities, applicable in a variety of contexts among het- Abstract Argumentation — A Survey. Artifcial Intelligence erogeneous agents whose unique common property 220: 28–63. doi.org/10.1145/2500891 might be the capability to argue itself. In this sense Chesñevar, C.; McGinnis, J.; Modgil, S.; Rahwan, I.; Reed, C.; artifcial argumentation promises, in the long term, Simari, G.; South, M.; Vreeswijk, G.; and Willmott, S. 2006. to provide a sort of universal social glue for linking Towards an Argument Interchange Format. Knowledge Engi- neering Review 21(4): 293–316. doi.org/10.1017/ together, in a plug and play and cooperative manner, S0269888906001044 robots and any other kind of intelligent agents. Dung, P. 1995. On the Acceptability of Arguments and Its Acknowledgements Fundamental Role in Nonmonotonic Reasoning, Logic Pro- gramming and n-Person Games. Artifcial Intelligence 77(2): This work has been partially supported by EPSRC 321–358. doi.org/10.1016/0004-3702(94)00041-X grants EP/N008294/1 and EP/N014871/1, by EU Gomes, C.; Kautz, H.; Sabharwal, A.; and Selman, B. 2008. H2020 research and innovation programme under Satisfability Solvers. In Handbook of Knowledge Representa- the Marie Sklodowska-Curie grant agreement No. tion, ed. F. Van Harmelen, V. Lifschitz, and B. Porter, 89–134. 690974 for the project MIREL: MIning and REason- Amsterdam, The Netherlands: Elsevier. ing with Legal texts, and by funds provided by the Gordon, T., and Karacapilidis, N. 1997. The Zeno Argumen- Institute for and Engineering, Uni- tation Framework. In Proceedings of the International Confer- versidad Nacional del Sur, Argentina. ence on Artifcial Intelligence and Law (ICAIL ’97), 10–18. New York: Association for Computing Machinery. doi.org/10. Notes 1145/261618.261622 Horty, J., and Bench-Capon, T. 2012. A Factor-Based Defni- 1. “Humans argue” is a truism. Either you already believe it tion of Precedential Constraint. Artifcial Intelligence and Law or you would need to argue against it. 20: 181–214. doi.org/10.1007/s10506-012-9125- 2. comma.csc.liv.ac.uk/. Hunter, A., and Williams, M. 2012. Aggregating Evidence 3. www.iospress.nl/journal/argument-computation/. About the Positive and Negative Effects of Treatments. Arti- fcial Intelligence in Medicine 56(3): 173–190. doi.org/10.1016/ 4. plato.stanford.edu/entries/logic-informal/. j.artmed.2012.09.004 5. petition.parliament.uk/. Karacapilidis, N., and Papadias, D. 2001. Computer Sup- 6. debategraph.org/. ported Argumentation and Collaborative Decision Making: 7. idebate.org/debatabase. The HERMES System. Information Systems 26(4): 259–277.

FALL 2017 35 Articles doi.org/10.1016/S0306-4379(01)00020-5 tive projects have concerned the develop- tee of the COMMA conferences on Compu- Lippi, M., and Torroni, P. 2016. Argumenta- ment of intelligent tools for a law company, tational Models of Argument. He is on the tion Mining: State of the Art and Emerging and realization of tools to support e-democ- editorial board of several journals, including Trends. ACM Transactions on Internet Tech- racy and legal knowledge-based systems. Artifcial Intelligence (currently as an associ- ate editor). nology. 10:1–10:25. Pietro Baroni is a full professor of computer Modgil, S.; Toni, F.; Bex, F.; Bratko, I.; science and engineering at the Department Chris Reed is a professor of computer sci- Chesñevar, C.; Dvorák, W.; Falappa, M.; Fan, of Information Engineering of the Universi- ence and philosophy at the University of X.; Gaggl, S.; García, A.; González, M.; Gor- ty of Brescia, Italy. He received the “Laurea” Dundee in Scotland, where he heads the don, T.; Leite, J.; Mozina, M.; Reed, C.; degree in mechanical engineering from the Center for Argument Technology. Reed has Simari, G.; Szeider, S.; Torroni, P.; and University of Brescia and a masters in infor- been working at the overlap between argu- Woltran, S. 2013. The Added Value of Argu- mation technology, from CEFRIEL, Milan. mentation theory and artifcial intelligence mentation. In Agreement Technologies, ed. S. He is author of more than 120 scientifc for more than 20 years, has obtained 6 mil- Ossowski, 357–403. Berlin: Springer. papers in the area of AIe and knowledge- lion in funding, and has over 150 peer- doi.org/10.1007/978-94-007-5583-3_21 based systems, with a main focus on theory reviewed papers in the area. He collaborates Palau, R. M., and Moens, M. 2011. Argu- and applications of computational argu- with a wide range of partners such as IBM mentation Mining. Artifcial Intelligence in mentation. He was a founding member of and the BBC and is also active in public Law 19(1): 1–22. doi.org/10.1007/s10506- the steering committee of the Computation- engagement and commercialization of 010-9104-x al Models of Argument (COMMA) confer- research, having served as executive director Pinkwart, N., and McLaren, B. M., eds. 2012. ence series, served as program chair of COM- (CTO, CSO, and CEO) of three startup com- Educational Technologies for Teaching Argu- MA 2016, and is currently coeditor-in-chief panies. of the Argument & Computation journal. mentation Skills. Sharjah, UAE: Bentham Sci- Guillermo R. Simari is a full professor in ence Publishers. Massimiliano Giacomin is an associate pro- logic for computer science and artifcial Pollock, J. 1992. How to Reason Defeasibly. fessor in computer science and engineering intelligence at the Universidad Nacional del Artifcial Intelligence 57(1): 1 – 42. doi.org/10. at the University of Brescia (Italy). He Sur (UNS), Baha Blanca, Argentina. He stud- 1016/0004-3702(92)90103-5 received an M.S. in electronics engineering ied mathematics at the same university and Prakken, H., and Sartor, G. 2015. Law and from the University of Padova (Italy) and a received a master of science in computer sci- Logic: A Review from an Argumentation Ph.D. degree in information engineering ence and a Ph.D. in computer science from Perspective. Artifcial Intelligence 227: 214– from the University of Brescia. His research Washington University in Saint Louis, USA. 245. doi.org/10.1016/j.artint.2015.06.005 interests are in the feld of AI, including The focus of his research is on the formal knowledge representation and automated foundations and effective implementation Rahwan, I., and Simari, G., eds. 2009. Argu- reasoning (in particular, argumentation the- of defeasible reasoning systems for mentation in Artifcial Intelligence. Berlin: ory, fuzzy constraints, temporal reasoning), autonomous agents. He chairs the Artifcial Springer. multiagent systems, and knowledge-based Intelligence Research and Development Schneider, J.; Groza, T.; and Passant, A. systems. He is an author of more than 100 Laboratory (LIDIA). 2013. A Review of Argumentation for the papers in these research areas. He is a mem- Matthias Thimm is a senior lecturer at the Social Semantic Web. Semantic Web 4(2): ber of the editorial board of Argument and University of Koblenz-Landau, Germany. He 159–218. Computation. Thimm, M.; Villata, S.; Cerutti, F.; Oren, N.; received his Ph.D. from the University of Strass, H.; and Vallati, M. 2016. Summary Anthony Hunter is a professor of artifcial Dortmund in 2011 on the topic of proba- Report of the First International Competi- intelligence, and head of the Intelligent Sys- bilistic reasoning with inconsistencies and tion on Computational Models of Argu- tems Research Group, in the Department of his habilitation degree from the University mentation. AI Magazine 37(1): 102–104. Computer Science, University College Lon- of Koblenz-Landau in 2016. His research doi.org/10.1609/aimag.v37i1.2640 don (UCL). His research interests are on interests include reasoning with uncertain aspects of inconsistency, argumentation, and inconsistent information, in particular Walton, D.; Reed, C.; and Macagno, F. 2008. and knowledge merging. using approaches from computational mod- Argumentation Schemes. Cambridge: Cam- els of argumentation and inconsistency bridge University Press. doi.org/10.1017/ Henry Prakken is a lecturer in artifcial measurement. CBO9780511802034 intelligence at the Computer Science Walton, R.; Gierl, C.; Yudkin, P.; Mistry, H.; Department of Utrecht University and pro- Serena Villata is a researcher at CNRS. Her Vessey, M.; and Fox, J. 1997. Evaluation of fessor in and legal argu- work focuses on argumentation theory and Computer Support for Prescribing CAPSULE mentation at the Law faculty of the Univer- spans from formal models of argumentation Using Simulated Cases. British Medical Jour- sity of Groningen. He has master’s degrees to natural language–based and emotion- nal 315(7111): 791–795. doi.org/10.1136/ in law (1985) and philosophy (1988) from based ones. She is the author of more than bmj.315.7111.791 the University of Groningen. In 1993 he 60 scholarly publications including 11 jour- obtained is Ph.D. degree (cum laude) at the nal articles and 4 book chapters. She was Katie Atkinson is a professor of computer Free University Amsterdam with a thesis invited to coauthor a chapter on processing science and head of the Department of Com- titled Logical Tools for Modelling Legal argumentation in natural language texts in puter Science at the University of Liverpool, Argument. His main research interests con- the Handbook of Formal Argumentation. She UK. Her research concerns computational cern computational models of argumenta- coorganized the seminar Frontiers and Con- models of argument, with a particular focus tion and their application in multiagent sys- nections Between Argumentation Theory on persuasive argumentation in practical rea- tems, legal reasoning, and other areas. and Natural Language Processing in July soning and how this can be applied in Prakken is a past president of the Interna- 2014 (Bertinoro, Italy), one of the frst domains such as law, e-democracy, and agent tional Association for AI and Law, of the attempts to put together researchers from systems. Atkinson’s work covers both theo- JURIX Foundation for Legal Knowledge- computational linguistics and from argu- retical and applied aspects; recent collabora- Based Systems, and of the steering commit- mentation theory.

36 AI MAGAZINE