Advances in Argument Mining

Chris Reed Katarzyna Budzynska Centre for Argument Technology Institute for Philosophy & Sociology University of Dundee, UK Polish Academy of Sciences [email protected] Warsaw, Poland [email protected]

1 Description composing an argument into its constituent parts, determining how those parts are connected and This course aims to introduce students to an excit- structured – and how they connect with other ar- ing and dynamic area that has witnessed remark- guments and argument parts – and finally, evalu- able growth over the past 36 months. Argument ating the quality of those connections. Thorough mining builds on opinion mining, sentiment anal- overviews are provided in, e.g., (Stede and Schnei- ysis and related to tasks to automatically extract der, 2018; Lippi and Torroni, 2016) with a wide not just what people think, but why they hold the range of more detailed themes covered elsewhere, opinions they do. From being largely beyond the such as premise-conclusion recovery (Stab and state of the art barely five years ago, there are now Gurevych, 2017), types of argumentation pattern many hundreds of papers on the topic, millions (Walton et al., 2008), relationships between se- of dollars of commercial and research investment, mantic and argumentative structures (Becker et al., and the 6th ACL workshop on the topic will be in 2017), how ethos of speakers interacts with argu- Florence in 2019. The tutors have delivered tuto- ment structure (Duthie and Budzynska, 2018) and rials on argument mining at ACL 2016, at IJCAI automated assessment of argument persuasiveness 2016 and at ESSLLI 2017; for ACL 2019, we have (Carlile et al., 2018). developed a tutorial that provides a synthesis of the major advances in the area over the past three Growth. From just a handful of papers in years. print around 2010, argument mining has grown rapidly with Google Scholar now reporting around Argument and debate form cornerstones of 2,000 articles mentioning the topic in their title. civilised society and of intellectual life. Processes ACL, EMNLP and NAACL have included over of argumentation run our governments, structure 50 articles on argument mining in the past three scientific endeavour and frame religious belief. years alone, in addition to the 92 articles pub- Recognising and understanding argument are cen- lished at the ACL workshop series on Argument tral to decision-making and professional activity Mining (co-founded by Reed). Argument min- in all walks of life, which is why we place them at ing has been building momentum both within the the centre of academic pedagogy and practice; it’s ACL community and further afield, with both spe- why such a premium is placed upon these skills; cialist conferences (such as Computational Mod- and it’s why rationality is one of the very defining els of Argument, COMMA, co-chaired in 2018 by notions of what it is to be human. Budzynska) and generalist conferences (such as As our understanding of how arguments are as- IJCAI) devoting increasing time to papers, work- sembled, are interpreted and have impact has im- shops and graduate-level training on argument proved, so it has become possible to frame compu- mining. The creation and publication of datasets tational questions about how it might be possible has been an important contributor to the vitality for machines to model and replicate the processes of the field, with papers at LREC and in LRE in- involved in identifying, reconstructing, interpret- creasing the breadth of this foundational aspect to ing and evaluating reasoning expressed in natural the field (Abbott et al., 2016). By the same token, language arguments. This, then, is argument min- new SEMEVAL tasks have also started to set the ing: identifying that an argument is present, de- goalposts and shape robust comparative evalua-

39 Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics: Tutorial Abstracts, pages 39–42 Florence, Italy, July 28 - August 2, 2019. c 2019 Association for Computational Linguistics tions (https://competitions.codalab. dialogical argumentation. To consolidate under- org/competitions/17327). standing of the material in each of these 20 minute blocks, the first part concludes with a 30 minute Challenge. Argument mining is a particularly practical session in which students will get an op- challenging task and is an exciting domain in portunity to apply the theory to an example drawn which to work because it is increasingly clear that from a real-life setting. deep learning and distributional techniques alone In the second part, we will cover techniques for are not delivering the same kind of successes that argument mining from (i) straightforward applica- have been enjoyed in some other areas of NLP. tion of machine learning techniques; through (ii) Many labs working with algorithms for argument use of BiLSTMs in particular for exploiting se- mining are finding that hybrid approaches that in- quence structure latent in argument presentation; tegrate rule-based and statistical methods are more to (iii) the development of hybrid approaches to likely to deliver the strongest performance. As a argument mining. We will again encourage deep result, recent argument mining has been pushing understanding from the students through a short the boundaries of approaches to NLP in general. practical implementation exercise making use of R. Argument mining in the press. The past year The outline syllabus runs thus: has seen a rapidly accelerating public profile for argument technologies in general, with Reed com- Part A: Foundations missioned to produce articles that have appeared • A1 (20 mins). Theory of argument structure in Newsweek (arg.tech/newsweek) and on – linked, convergent, serial, divergent, rebut, the BBC (arg.tech/bbcnews), and media undercut – indicators – enthymemes – logos, events such as IBM’s Project Debater launch ethos, pathos. (e.g., www.wired.com/story/now-the- computer-can-argue-with-you). The • A2 (20 mins). Semantic types – argumenta- BBC too has commissioned technology that in- tion schemes – ADU segmentation – datasets, cludes the first live deployment of argument min- corpora and shared tasks. ing (www.bbc.co.uk/taster/pilots/ evidence-toolkit-moral-maze) in sup- • A3 (20 mins). Argument in dialogue – Infer- porting identification of fake news. The tutorial ence Anchoring Theory – reported speech – will make use of these high-profile applications complex and implicit speech acts. of argument mining to contextualise and motivate • A4 (30 mins). Practical session: Analysing the topics covered. natural argument.

2 Type Break As argument mining has been covered at an ACL tutorial previously, in 2016. This tutorial is classi- Part B: Applications fied as ’Introductory’ and the syllabus is designed • B1 (20 mins). Simple machine learning. IOB to minimise preprequisites. We aim, however, to schema for segmentation – classifiers for arg- focus heavily on results from the past three years nonarg – classifiers for premise-conclusion. during which time significant progress has been made. • B2 (20 mins). Advanced machine learning. Word embeddings for argumentation 3 Outline schemes – BiLSTM models for argumenta- The tutorial is structured in two parts, each of tion sequence patterns. which mixes lecturing with practical work. In the • B3 (20 mins). Hybrid approaches. Argu- first part, we will cover theory of argument struc- ment structure parsing – illocutionary structure from (i) the basics in argumentation theory; ture parsing – dialogical priors. through (ii) recent results in computational models of argument; to (iii) the latest (as yet largely • B4 (30 mins). Practical session: Argument unpublished) techniques that allow modelling of mining in R.

40 All materials will be made available at a a tutorial on Argument Mining at IJCAI 2016, dedicated tutorial website as they were for our and with Reed an extensive week-long course ACL2016, IJCAI 2016 and ESSLLI 2017 tutori- at the 29th European Summer School in Logic, als. Language, and Information (ESSLLI2017). This website will be located at Chris Reed (Centre for Argument Technology, http://arg.tech/acl2019tut. University of Dundee, [email protected], www.arg.tech). Chris is Full Professor of 4 Prerequisites Computer Science and Philosophy at the Univer- The tutorial is intended to be accessible to most sity of Dundee, where he heads the Centre for Ar- ACL attendees, so has straightforward prerequi- gument Technology. Chris has been working at sites: the overlap between argumentation theory and artificial intelligence for over twenty years, has won • Basic familiarity with supervised machine over £6m of funding from government and com- learning techniques and the way they are em- mercial sources and has over 200 peer-reviewed ployed and assessed papers in the area (including papers in ACL, COL- ING, IJCAI, ECAI and AAAI) and five books. • Experience of using R will be an advantage, He has also been instrumental in the development but is not required, for practical session B4. of the Argument Interchange Format, an international standard for computational work in the area; Attendees are expected to have or to be working he is spear-heading the major engineering effort towards a PhD in computational linguistics or a behind the Argument Web; and he was a found- closely cognate area, but no previous experience ing editor of the Journal of Argument & Compu- of academic investigation of argument is expected. tation. He was co-organiser of COMMA 2014, of 5 Tutors the first ACL workshop on Argumentation Mining in 2014, was the chair of the third workshop on Katarzyna Budzynska (Computational Argument Mining with ACL in 2016, and has re- Ethos Lab, Polish Academy of Sci- cently won funding for a £1m project on the topic ences, [email protected], in collaboration with IBM. With Gurevych, Stein www.computationalethos.org). and Slonim, he delivered a tutorial on Argument Katarzyna is an associate professor in phi- Mining at ACL 2016 which was extremely well losophy at the National Polish Academy of attended, and followed that with a course at ESS- Sciences and an associate professor in com- LLI 2017 with Budzynska. puting at the University of Dundee (UK). Her work focuses on communication structures of Acknowledgments argumentation, dialogue and ethos. She has Our work in developing this tutorial has been sup- published two books and over 80 peer-reviewed ported in part in the UK by EPSRC under grant papers including articles in journals such as EP/N014781/1. Artificial Intelligence, Association for Computing Machinery (ACM TOIT) and Synthese. Katarzyna founded a national movement, the Polish School References of Argumentation, and sits in the steering commit- tees of a new initiative, the European Conference Rob Abbott, Brian Ecker, Pranav Anand, and Mari- lyn A. Walker. 2016. Internet Argument Corpus 2.0: on Argumentation (ecargument.org), and An SQL schema for dialogic social media and the the ArgDiaP Association for Argumentation, corpora to go with it. In Language Resources and Dialogue and Persuasion (argdiap.pl). Most Evaluation Conference (LREC), Portoroz,ˇ Slovenia. recently, she established her research group – Maria Becker, Alexis Palmer, and Anette Frank. 2017. the Computational Ethos Lab which develops Semantic clause types and modality as features for innovative technologies to process the use of argument analysis. Journal of Argument and Com- ethos in natural language in order to predict the putation, 8(2):95–112. results of presidential elections, detect trolls and Winston Carlile, Nishant Gurrapadi, Zixuan Ke, and cyber-bullies in social media, and uncover poten- Vincent Ng. 2018. Give me more feedback: Anno- tial terrorist threats. With Villata, she delivered tating argument persuasiveness and related attributes

41 in student essays. In Proceedings of the 56th Annual Meeting of the Association for Computational Lin- guistics, pages 621–631. Association for Computa- tional Linguistics. Rory Duthie and Katarzyna Budzynska. 2018. A deep modular rnn approach for ethos mining. In Proceed- ings of the Twenty-Seventh International Joint Con- ference on Artiﬁcial Intelligence, IJCAI-18, pages 4041–4047. Marco Lippi and Paolo Torroni. 2016. Argumentation mining: State of the art and emerging trends. ACM Trans. Internet Technology, 16(2):10:1–10:25. Christian Stab and Iryna Gurevych. 2017. Parsing argumentation structures in persuasive essays. Com- putational Linguistics, 43(3):619–659. Manfred Stede and Jodi Schneider. 2018. Argumen- tation Mining. Synthesis Lectures on Human Lan- guage Technologies. Morgan Claypool. Douglas Walton, Chris Reed, and Fabrizio Macagno. 2008. Argumentation Schemes. Cambridge Univer- sity Press.