Counter-Argumentation and Discourse: A Case Study Stergos Afantenos Nicholas Asher IRIT, Univ. Toulouse IRIT, CNRS, France France stergos.afantenos@irit.fr asher@irit.fr Abstract in game theoretic terms it is an equilibrium in a per- suasion game in which the addressee adopts an optimal Despite the central role that argumentation action based on the conversational history and in which plays in human communication, the compu- the arguer adopts her conversational strategy based on tational linguistics community has paid rela- the addressee’s strategy for adopting an action (Glazer tively little attention in proposing a method- and Rubinstein, 2004). Yet, despite its importance in ology for automatically identifying arguments human communication and behavior and despite the and their relations in texts. Argumentation fact that textual realizations of arguments and debates is intimately related with discourse structure, are numerous on the web, it is surprising that this area since an argument often spans more than one has received very little attention by the Computational phrase, forming thus an entity with its own Linguistics community. coherent internal structure. Moreover, argu- One domain of research in Computational Linguis- ments are linked between them either with a tics that is of particular interest for argumentation is support, an attack or a rebuttal relation. Those that of discourse. In a typical argumentation process, argumentation relations are often realized via which takes the form of a dialogue, every argument a discourse relation. Unfortunately, most of has an internal coherence meaning that it can be repre- the discourse representation theories use trees sented by a discourse graph. Moreover arguments are in order to represent discourse, a format which linked between themselves either with support, attack is incapable of representing phenomena such or rebuttal relations which are realized once again as as long distance attachments and crossed de- discourse relations linking either the whole discourse pendencies which are crucial for argumenta- subgraphs representing the arguments or parts of them. tion. A notable exception is Segmented Dis- Any attempt to automatically extract the argumentation course Representation Theory (SDRT) (Asher structure from a given text cannot afford to ignore dis- and Lascarides, 2003). In this paper we show course. Our goal in this paper is to show how argumen- how SDRT can help identify arguments and tation is intimately involved with discourse structure. their relations. We use counter-argumentation We achieve this by using counter-argumentation (fol- as our case study following Apothéloz (1989) lowing (Apothéloz, 1989; Amgoud and Prade, 2012)) and Amgoud and Prade (2012) showing how as a case study. the identification of the discourse structure can The remainder of this paper is structured as follows. greatly benefit the identification of the argu- In section 2 we present the current work in the so-called mentation structure. argumentation mining, the subfield of computational linguistics that deals with the automatic extraction of 1 Introduction the argumentation structure from texts. In section 3 we People use arguments to persuade others to adopt a tell a few words on discourse and in section 4 we show point of view or action they find beneficial to their in- how SDRT (Segmented Discourse Representatuin The- terests, or alternatively to prevent others from adopt- ory, (Asher and Lascarides, 2003)) can be applied in a ing a position or action that they find contrary to their case study focused on counter-argumentation. In sec- interests. Of course an agent may find it in her in- tion 5 we present the future work and we conclude this terest to convince an interlocutor to adopt a position paper. she herself does not believe; for instance, a seller may want to persuade a buyer that a product is worth more than she believes it is worth. Because argumentation 2 Argumentation in Computational involves an interaction between an arguer and an ad- Linguistics dressee, it involves game theoretic aspects: it is the Despite its general neglect, argumentation has been means in language for getting an agent to a position the focus of some work in Computational Linguistics. of agreement with the position one is advocating, or Teufel (1999), Teufel and Moens (2002) aim at identi- fying what they call the argumentative zones of scien- all classification or a pairwise classification. The fea- tific articles. The zones they have used include the aim tures they use are divided into general ones concerning of the paper, general scientific background, description all schemes (features reflecting textual surface form) or of the authors’ previous work, comparison with other specific ones for each scheme (mostly cue phrases and works, etc. They are using a naive bayes model trying patterns). to classify each sentence into one of the predefined cat- Cabrio and Villata (2012a; 2012b) take a different egories using mostly surface features (position, length, stance. Their goal is to use Dung’s (1995) abstract ar- etc) and whether the sentence contains title words or gumentation framework in order to detect a set of ac- words scoring high in terms of tf.idf . cepted arguments from online debates. They extract Palau and Moens (2009) have recently attempted ar- arguments from Debatopedia2 using textual entailment gumentation mining, or the identification of arguments techniques. More precisely, if a sentence T entails in a text. They assume that an argument consists of another sentence H then they consider that there is a a series of premises and a conclusion. Premises and support relation between the two sentences (and thus conclusions are represented by propositions in the text. points of views) otherwise there is an attack relation. Of course, not all propositions in a given text are part They use the open source software package EDITS3 in of an argument. In order to tackle the problem of ar- order to perform textual entailment. In order then to gumentation mining the authors break it into a series identify the set of arguments that would be acceptable of subtasks. Initially they are interested in perform- by a an external observer the authors use Dung’s (1995) ing a binary classification of each proposition into ei- abstract argumentation framework. In essence an argu- ther a proposition participating in an argument or not. ment belongs to the aforementioned set if all the argu- Propositions that are positively classified are then sent ments attacking it are rejected. An argument is rejected to a second classifier which determines whether it is a if at least one accepted argument attacks it. premise or a conclusion. For both classification tasks they use a maximum entropy model and the Araucaria 3 Discourse corpus1 as well as a corpus extracted from the euro- The little prior work on argumentation has ignored dis- pean court of human rights. The features they use for course structure, and we think this is a mistake. A com- the first classifier include surface features ({1, 2, 3}- plete discourse structure of a dialogue will determine grams, punctuation, sentence and word length), POS how each interlocutor’s contribution relates to other information (adverbs, verbs and modal auxiliaries) and contributions, both her own and that of other dialogue syntactic parsing. The second classifier uses again sur- participants. This structure already by itself is crucial face features, POS tags for the subject and main verb, to determining the structure of an argument—which at- simple rhetorical and argumentative patterns as well as tacks are directed towards which other contributions. the results of the first classifier (although no structured Moreover, an argument is not just a sequence of attacks prediction is attempted which would probably be more but a much more complex structure. For one thing, ar- appropriate, given that the two classifiers are not inde- guments contain support moves as well; a good persua- pendent). Of course, once one has identified the propo- sion strategy is to explain why one’s claims are true, sitions that are premises and conclusions, one does not but another is to provide background that will enable yet have the full arguments. In order to get them, the the addressee to understand one’s reasons, and yet an- authors create a simple CFG grammar which tries to other is to provide more details about the claims them- get the tree structure of arguments. The authors do not selves. All of these ”strategies” involve in fact rhetori- attempt to detect the relations (e.g. support, attack, re- cal moves that are different and that may be appropri- buttal) that connect the arguments between each other. ate in different situations. A discourse structure makes The Araucaria corpus is used by Feng and Hirst plain these different types of moves through the use of (2011) as well but their goal is not performing ar- different discourse relations. gumentation mining. Instead they focus on the task In effect, discourse structure has the promise to give of classifying arguments into argumentation schemes a much more detailed picture of the nature and struc- (Walton et al., 2008). Araucaria arguments contain en- ture of argumentation. At the moment, we don’t know thymemes annotated by human subjects which Feng exactly what that picture is. But by pursuing the anal- and Hirst (2011) remove. Moreover, each argument ysis of dialogues in terms of argument structure and is annotated with various argumentation schemes but discourse structure we can find out. the authors keep only the ones that are annotated with Walton’s schemes. They keep only the 5 more frequent 4 Counter-Argumentation: A Case schemes. In total they have 393 arguments which they classify into one of five schemes. Concerning the clas- Study sification method, they use the C4.5 algorithm imple- To illustrate our point in the previous section, we il- mented in Weka in order to perform either a one-vs- lustrate how constructed examples of different sorts of 1 2 http://araucaria.computing.dundee.ac. http://debatepedia.idebate.org uk/ 3 http://edits.fbk.eu arguments given by Apothéloz (1989) look from a dis- — [No, she worked hard.]3 [Her eyes have bags course structure point of view. Apothéloz (1989) iden- underneath them.]4 tified four different modes of arguing against a given 1 argument. In this work an argument is simply a pair C(x) : R(y) where R represents the function of reason Expl. and x its content and C the function of conclusion and 2 3 x its content. x and y can be either propositions, con- Correction Expl.∗ clusions or enthymemes. Given the above, Apothéloz (1989) distinguishes between four different modes of 4 arguing against a given argument C(x): (3) — [Clara works hard]1 [because she is 1. disputing the plausibility or the truth of a reason, ambitious.]2 that is the propositions used in y — [It is not out of ambition that Clara works hard.]3 [She is not ambitious.]4 2. disputing the completeness of the reason 1 Ack. 3. disputing the relevance of the reason with respect Expl. to the conclusion, and 2 3 Correction 4. disputing the argumentative orientation of the rea- Elab. son by showing the reason presented is rather in 4 favor of the conclusion’s opposite. Nonetheless, Apothéloz (1989) completely ignores the In all three examples, the second speaker does not internal structure that the arguments have. In the fol- challenge her interlocutor concerning her conclusion lowing we analyse the different modes of counter- (EDU 1 in all three cases). In fact, in the example (3) argumentation that Apothéloz (1989) provides, giving the second speaker explicitly acknowledges the content examples found in (Amgoud and Prade, 2012). Our of the conclusion (Acknowledgment(1, 3)). Instead goal is to show how discourse analysis can help the the second speaker’s disagreement is always with the field of computational linguistics not only detect re- truth value of the reason behind the conclusion. This lations between arguments but also analyse the inter- takes the form of a Correction relation between the nal structure of an argument. In the following, we are first speaker’s EDU representing the reason (EDU 2 in using the Segmented Representation Discourse Theory all cases) and the second speaker’s counter-argument (SDRT) (Asher and Lascarides, 2003). For the sake (EDU 3 for examples (2) and (3) and CDU π1 for ex- of representation, discourse is represented as a hyper- ample (1)). For the last two examples the speaker pro- graph with discourse relations being the edges of the vides additional reason for her beliefs either by means graph and Elementary Discourse Units (EDUs) being of an Elaboration relation or an Explanation∗ rela- nodes containing only one element, while Complex tion. This last relation signals an explanation of why b Discourse Units (CDUs) are nodes containing more said that Clara worked hard. It is an explanation of a than one simple elements (Asher et al., 2011). speech act and provides epistemic grounds for the con- tent of the assertion. Note that in all the above exam- Disputing the plausibility of a reason ples the Correction discourse relation amounts to an When one disputes the plausibility of a reason essen- attack relation. tially it amounts to proving that the reason is false. Disputing the completeness of a reason Apothéloz (1989) provides three different ways of showing that; we illustrate them with the following ex- In the second mode of counter-argumentation that amples. Apothéloz (1989) has identified, the second speaker does not attack the truthfulness of the reason but rather (1) — [Clara will fail her exams.]1 [She did not its completeness. Here are some examples. work hard]2 — [Clara?!]3 [She worked non-stop.]4 (4) — [Clara will fail her exams.]1 [She did not work hard]2 1 — [Clara will not fail her exams.]3 [She is very Expl. smart.]4 2 π1 Correction 1 3 Correction Expl. Expl. 3 4 2 4 Continuation In this example, the second speaker neither affirms nei- (2) — [Clara will fail her exams.]1 [She did not ther denies the reason, i.e. the fact that Clara didn’t work hard]2 work hard. Instead, she is ignoring it (manifested by the fact that no discourse relation exists between EDUs π1 2 and 3 or 4). Instead she corrects the conclusion of the first speaker by providing more evidence which lead to the contrary. Again, the Correction discourse relation Correction 1 2 connects two arguments and serves as an attack argu- Expl. Ack. mentative relation. 3 4 (5) — [Paul is in his office ]1 [because his car is in Contrast the carpark.]2 — [But the car is in the carpark]3 [because it Here the second speaker acknowledges the reason of has a mechanical problem and is undriveable.]4 the first person, as seen by the discourse relation be- tween EDUs 2 and 3, but then shows that there is a con- trast between this and her conclusion, disagreeing thus π1 π2 with the whole argument. It is important to note once Correction again that in this example, as the preceding ones, the discourse analysis enables us to clearly pinpoint which 1 2 3 4 elements of the first argument are accepted and which Expl. Expl. are attacked by the second speaker. In this case both arguments (as before) are thor- (8) — [Clara will fail her exams.]1 [She did not oughly supported by an Explanation discourse re- work hard]2 lation. Moreover the second speaker even explic- — [She will not fail her exams]3 [because she itly agrees with the reason given by the first one did not work hard,]4 [but rather because of the (Acknowledgment(2, 3)) but she disagrees with the stress.]5 whole argument (note the Correction relation be- tween the two CDUs) since she judges that the reason Correction is not enough and provides more evidence (EDU 4) to π1 3 back her disagreement up. ¬Expl. Expl. (6) — [This object is red]1 [since it looks red.]2 1 2 4 5 — [But the object is illuminated by a red Explanation Contrast light.]3 Counterevidence This is a very interesting example. As the discourse π1 3 analysis shows the undirected cycle that is produced between EDUs 3, 4 and 5 enables the second speaker to Contrast explain why she disagrees with the whole of the initial 1 2 Expl. statement. Now, this example is quite more complicated to ana- Disputing the argumentative orientation of a lyze. There is a contrast between the object’s looking reason red, which generates the expectation that it is red, and In the final mode of counter-argumentation that the fact that the object is illuminated by a red light, Apothéloz (1989) has proposed the second speaker which would tend to put that expectation in doubt. But does not dispute neither the reason nor the conclusion. putting the expectation into doubt also puts into doubt Instead she argues that the reason corroborates towards the causal relation supposed by the first speaker be- the opposite of the conclusion. This can be illustrated tween 1 and 2. with the following example. Disputing the relevance of a reason (9) — [Running a marathon is exhausting.]1 [The In the third mode of counter-argumentation that whole body undergoes too much stress.]2 Apothéloz (1989) has identified concerns the second — [That’s precisely what makes it nice!]3 speaker does not attack the truthfulness or the com- Acknowledgment π1 3 pleteness of a reason but instead its relevance. Be- Correction low are some examples of this mode of counter- argumentation. 1 2 Expl. (7) — [Clara will fail her exams.]1 [She did not work hard]2 5 Discussion and Future Work — [Indeed, she did not work hard,]3 [but not working hard is not a reason to necessarily fail In the previous section we have showed via the use one’s exams.]4 of a case study how the use of a discourse represen- tation theory can help us represent in fine detail the phenomena that take place during argumentation—in 4 this particular case, counter argumentation during a di- alogue. In order to represent discourse we have chosen 1 to use the Segmented Discourse Representation Theory (SDRT) of Asher and Lascarides (2003). This choice was made after careful consideration of the phenomena 3 present during argumentation as well as the expressive power of other discourse representation theories. 2 Take for example the Rhetorical Structure Theory (RST, Mann and Thompson (1988)), which is the most where edges with arrows denote support relations and widely cited and used discourse representation theory edges with circles denote undercuts. The RST graph currently. In RST, as in SDRT, the basic units are the for the above dialogue is the following: same, namely EDUs.4 In RST adjacent EDUs can E VIDENCE be linked together with rhetorical relations in order to form what in RST’s jargon are called spans. Spans can be linked with rhetorical relations either with other ✠ E VIDENCE A NTITHESIS adjacent EDUs or adjacent spans. We keep on em- ✠ ❘ phasizing the word “adjacent” since this constitutes in 1 2 3 4 our opinion (but see also (Peldszus and Stede, 2013)) As we can see, the structural properties of those two a limitation of RST since it does not allow this the- graphs are completely different and the use of RST for ory to have long distance dependencies, a crucial phe- argumentative analysis does not seem to be a promis- nomenon in argumentation. SDRT does not have this ing path to follow. On the other hand, SDRT neatly limitation. Consider example (7). In this simple ex- follows the argumentation graph (we have used the box ample the Correction relation—which, incidentally, is representation of SDRT here) making it thus more ap- the backbone of the second speaker’s attack—holds be- propriate for use in argumentative analysis. tween non-adjacent EDUs. Even if the first speaker’s argument was much longer, or if the second speaker elaborated on the fact that Clara did not work hard (and 1 Correction Correction thus we had many EDUs intervening between π1 and 4) Expl. 3 4 it wouldn’t influence the fact that the complex segment 2 π1 would be attached to EDU 4. Such long distance attachments are impossible with SDRT which requires that each EDU or span is attached to an adjacent EDU or span. At this point we would like to say a few words on The second problem that RST has as far as the rep- the computational extraction of discourse structures. resentation of argumentative structures is concerned, is Most of the published work currently is using the RST that it cannot correctly represent rebuttals. This is prob- framework. This is due to two facts. Firstly there are lem that is also reported by Peldszus and Stede (2013) more annotated data available for RST and secondly so we are using their example, slightly modified in or- the problem is computationally less demanding since der to illustrate this point. Consider the following dia- decisions are always made locally (attachments can be logue: either left or right of a given span) which renders this framework more simple and thus more attractive to re- (10) — [We should tear the building down.]1 [It is searchers. Of course, this implies that all long distance full of asbestos.]2 attachements are completely lost, an aspect which is — [It is possible to clean it up.]3 crucial, as we have seen, for argumentation. — [But that would be forbiddingly expensi- Muller et al. (2012) have recently attempted extrac- ve!]4 tion of SDRT structures using data from the ANNODIS corpus (Afantenos et al., 2012), annotated with SDRT structures, with state of the art results. The authors at- The argumentation graph that results from this dia- tack the problem of predicting SDRT discourse struc- logue, according to the scheme proposed in (Peldszus tures by making some simplifications to the objects that and Stede, 2013) is the following: they need to predict, namely they eliminate CDUs by 4 There is a big difference as far as EDUs are concerned making the assumption that, semantically speaking, at- between the two theories. In SDRT EDUs can be embedded tachment to a CDU amounts to attaching to its head— the one within the other whilst RST does not allow it. that is the uppermost and leftmost EDU. They have thus structures reminiscent of dependency graphs in syntac- tic analysis. The authors perform structured prediction on the de- pendency graphs they produced which can be broken down into two steps. Initially they learn local prob- units and their semantics. In Contstraints in Dis- ability distributions for attaching and labeling EDUs, course (CID 2011), Agay-Roches Rouges, France. based on naive bayes and logistic regression models. Elena Cabrio and Serena Villata. 2012a. Combin- They effectively thus create a complete graph where ing textual entailment and argumentation theory for each node represents an EDU and each arc a probabil- supporting online debates interactions. In Proceed- ity of attachment. The authors then move on to the de- ings of the 50th Annual Meeting of the Association coding phase where the goal is to extract the graph that for Computational Linguistics (Volume 2: Short Pa- approaches the reference object. They use two decod- pers), pages 208–212, Jeju Island, Korea, July. As- ing approaches based on A∗ and Maximum Spanning sociation for Computational Linguistics. Tree (MST) algorithms. Elena Cabrio and Serena Villata. 2012b. Natural Language Arguments: A Combined Approach. In Closing this paper we would like to state that one of 20th European Conference on Artificial Intelligence the main reasons that extraction of argumentative struc- (ECAI 2012), Montpellier, France. tures has not been more widely explored by the com- putational linguistics community is due to the fact that Phan Minh Dung. 1995. On the acceptability of ar- few annotated corpora exist. We believe that a project guments and its fundamental role in nonmonotonic with the goal of jointly annotating argumentative and reasoning, logic programming and n-person games. discourse structures is crucial for the advancement of Artificial Intelligence, 77:321–357. this field, as well as other fields such as automatic sum- Vanessa Wei Feng and Graeme Hirst. 2011. Classi- marization (Afantenos et al., 2008), question answer- fying arguments by scheme. In Proceedings of the ing, etc. 49th Annual Meeting of the Association for Com- putational Linguistics: Human Language Technolo- gies, pages 987–996, Portland, Oregon, USA, June. References Association for Computational Linguistics. Stergos D. Afantenos, Vangelis Karkaletsis, Panagiotis Jacob Glazer and Ariel Rubinstein. 2004. On opti- Stamatopoulos, and Constantin Halatsis. 2008. Us- mal rules of persuasion. Econometrica, 72(6):1715– ing Synchronic and Diachronic Relations for Sum- 1736, November. marizing Multiple Documents Describing Evolving Events. Journal of Intelligent Information Systems, William C. Mann and Sandra A. Thompson. 1988. 30(3):183–226, June. Rhetorical Structure Theory: Towards a Functional Theory of Text Organization. Text, 8(3):243–281. Stergos Afantenos, Nicholas Asher, Farah Bena- mara, Myriam Bras, Cecile Fabre, Mai Ho-Dac, Philippe Muller, Stergos Afantenos, Pascal Denis, and Anne Le Draoulec, Philippe Muller, Marie-Paul Nicholas Asher. 2012. Constrained decoding for Pery-Woodley, Laurent Prevot, Josette Rebeyrolles, text-level discourse parsing. In Proceedings of COL- Ludovic Tanguy, Marianne Vergez-Couret, and ING 2012, pages 1883–1900, Mumbai, India, De- Laure Vieu. 2012. An empirical resource for discov- cember. The COLING 2012 Organizing Committee. ering cognitive principles of discourse organisation: Raquel Mochales Palau and Marie-Francine Moens. the annodis corpus. In Nicoletta Calzolari (Con- 2009. Argumentation mining: The detection, clas- ference Chair), Khalid Choukri, Thierry Declerck, sification and structure of arguments in text. In Pro- Mehmet Uğur Doğan, Bente Maegaard, Joseph Mar- ceedings of the 12th International Conference on Ar- iani, Jan Odijk, and Stelios Piperidis, editors, Pro- tificial Intelligence and Law, ICAIL ’09, pages 98– ceedings of the Eight International Conference on 107, New York, NY, USA. ACM. Language Resources and Evaluation (LREC’12), Is- tanbul, Turkey, may. European Language Resources Andreas Peldszus and Manfred Stede. 2013. From ar- Association (ELRA). gument diagrams to argumentation mining in texts: A survey. International Journal of Cognitive Infor- Leila Amgoud and Henri Prade. 2012. Can ai models matics and Natural Intelligence (IJCINI), 7(1):1–31. capture natural language argumentation? Interna- tional Journal of Cognitive Informatics and Natural Simone Teufel and Marc Moens. 2002. Summariz- Intelligence, 6(3):19–32, July. ing Scientific Articles: Experiments with Relevance and Rhetorical Status. Computational Linguistics, Denis Apothéloz. 1989. Esquisse d’un catalogue des 28(4):409–445, December. formes de la contre-argumentation. Travaux du Cen- tre de Recherches Sémiologiques, 57:69–86. Simone Teufel. 1999. Argumentative Zoning: Infor- mation Extraction from Scientific Articles. PhD The- Nicholas Asher and Alex Lascarides. 2003. Logics of sis, University of Edinburgh. Conversation. Studies in Natural Language Process- ing. Cambridge University Press, Cambridge, UK. Douglas Walton, Chris Reed, and Fabrizio Macagno. 2008. Argumentation Schemes. Cambridge Univer- Nicholas Asher, Antoine Venant, Philippe Muller, and sity Press. Stergos D. Afantenos. 2011. Complex discourse