Web-based Mass Argumentation in Natural Language Adam Wyner Tom van Engers Department of Computer Science Leibniz Center for Law University of Liverpool University of Amsterdam adam@wyner.info vanengers@uva.nl ABSTRACT contributions by the public, they do not fully analyse the lin- We provide a novel framework and implementation which guistic content of the contributions nor structure the argu- integrates tools to support the acquisition of mass, distribu- mentative relationships among the statements, losing infor- tive, incremental, dynamic, argumentative knowledge in nat- mation and hindering automated reasoning. In this paper, ural language. With the Attempto Controlled English (ACE) we outline a natural language interface tool to formal argu- system, natural language statements are automatically trans- mentation systems using the Attempto Controlled English lated to a machine processable form. A discussion forum system (ACE)3 , integrated with a discussion forum to spec- allows the specification of argumentation theoretic relation- ify statement relationships, which then outputs an argument ships among statements. Statements and their relationships graph on which reasoning can be executed; the theoretical are input to a formal, implemented argumentation system, framework is detailed in [4, 5]. In the following, we briefly which calculates inferences from asserted premises. discuss an example and the implemented components. Keywords 2. EXAMPLE [5] provide a normalised example that has been adapted Natural Language Processing, Argumentation, Knowledge from the BBC’s Have Your Say online discussion of the ques- Engineering tion Should people be paid to recycle? In normalising the sentences, we have followed the lexical and syntactic conven- 1. INTRODUCTION tions of ACE. The have a sample from the 19 statements: There exist robust systems for building ontologies and in- stantiated knowledge bases using natural language input, p5: If a household pays a tax for the household’s which in part help to overcome the knowledge bottleneck garbage then the tax is unfair to the household. of translating human knowledge, expressed in natural lan- p6: Every household should pay an equal por- guage, into machine processable information. However, on- tion of the sum of the tax for the household’s tologies and knowledge bases may be debatable and inconsis- garbage. p7: No household which receives a ben- tent. Formal argumentation systems have been developed to efit which is paid by a council recycles the house- reason with inconsistent, defeasible knowledge bases [1] and hold’s garbage. p8: Every household which does [2]. Yet, such systems are abstract; translating arguments not receive a benefit which is paid by a coun- which are expressed in natural language into formal argu- cil supports a household which receives a benefit mentation frameworks is labour and knowledge intensive. which is paid by a council. A tool which supports users to build and argue about on- The statements p6, p7, and p8 are should be taken as sup- tologies and knowledge bases in natural language could find ports for the claim in p5. widespread application in important domains. Among many possible application areas, we consider public policy-making, wherein members of the public contribute statements on pol- 3. SYSTEM OUTLINE icy [3]. While current tools, e.g. argument visualisation We outline an implemented system which is theoretically DebateGraph 1 or forums Have Your Say 2 , allow web-based described in [4, 5]; the components and flow through are represented in Figure 1. We give illustrative fragments of 1 http://debategraph.org the components of the system. 2 http://www.bbc.co.uk/news/have your say/ A Discussion Forum. We have a user and a web-based threaded discussion fo- Permission to make digital or hard copies of all or part of this work for rum (which uses PhP, MySQL, and XML). As with stan- personal or classroom use is granted without fee provided that copies are dard threaded forums, the user can read statements, select not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to a statement to respond to, and enter a new statement. In republish, to post on servers or to redistribute to lists, requires prior specific addition to entering the statement itself, the user selects permission and/or a fee. the argumentation theoretic relationship between the in- International Conference on Knowledge Engineering and Knowledge Man- put statement and previous statements on the forum, where agement 2010 Lisbon, Portugal 3 Copyright 20XX ACM X-XXXXX-XX-X/XX/XX ...$10.00. http://attempto.ifi.uzh.ch sion, be taken as the core policy statements. While here we have used [2], other argumentation formalisms are possible [5]. Figure 3: Graphic Fragment of Recycling Debate 4. DISCUSSION With respect to the database of statements, relationships, assertions, and inferences, one can apply further processes such as information extraction and querying. In future work, we intend to extend the expressivity of ACE to cover deontic Figure 1: Flow of Input notions (e.g. obligation), to support interactive feedback for questions on implicit information, to give greater interac- the choices are among contradiction, premise, or conclusion. tive guidance on well-formedness of expressions, to leverage These relationships are also stored in the MySQL database. ontological information, and to enrich argumentative infor- Figure 2 shows the statements (5)-(8) in the forum. mation. Finally, we will examine the applicability of the system to a range of other domains in the social and empir- ical sciences. Acknowledgements We thank the European Commission for supporting this research in the IMPACT Project (Integrated Method for Policy making using Argument modeling and Computer as- Figure 2: Discussion List sisted Text analysis) FP7 Grant Agreement No. 247228. We thank Kiavash Bahreini for system development and graph- ics. ACE. In entering a new statement, an ACE editing window opens 5. REFERENCES which provides a predictive editor that guides the user to input well-formed statements using the ACE grammar and [1] P. M. Dung. On the acceptability of arguments and its lexicon. A controlled natural language is a formal yet expres- fundamental role in nonmonotonic reasoning, logic sive subset of a natural language with a restricted lexicon, programming and n-person games. Artificial restricted range of syntactic forms, and correlated semantic Intelligence, 77(2):321–358, 1995. interpretations. After entering the statement, it is parsed [2] T. Gordon, H. Prakken, and D. Walton. The carneades and semantically interpreted as a statement of First-order model of argument and burden of proof. Artificial Logic; the parse and semantic interpretation is stored in Intelligence, 171:875–896, 2007. an XML representation in the MySQL database for further [3] A. Macintosh. Moving toward ”intelligent” policy processing. For instance, inputting “Every household cre- development. IEEE Intelligent Systems, 24(5):79–82, ates some garbage.” automatically generates a First-order 2009. Logic representation: [4] A. Wyner, T. van Engers, and K. Bahreini. From ∀ x [[household’(x)] → ∃ y [ garbage’(y) ∧ create’(x,y) ]] policy-making statements to first-order logic. In As reported in [4], the sentences in the example are all parsed Proceedings of International Conference on Electronic and given accurate semantic interpretations in ACE. Government and the Information Systems Perspective (EGOVIS 2010), Lecture Notes in Computer Science. A Formal Argumentation System. Springer Verlag, 2010. To appear. The statements in their relationships are then used to gener- [5] A. Wyner, T. van Engers, and A. Hunter. Working on ate an argumentation graph such that, given asserted premises, the argument pipeline: Through flow issues between one can infer conclusions, which are also stored in the MySQL natural language argument, instantiated arguments, database. In Figure 3, from the premises p6, p7, and p8, we and argumentation frameworks. In Proceedings of the infer with a good likelihood (indicated by +.53), that p5 Workshop on Computational Models of Natural follows. The assertions and premises can, in a policy discus- Argument, 2010. To appear.