Argumentation Mining in Persuasive Essays and Scientific Articles from the Discourse Structure Perspective Christian Stab† , Christian Kirschner†‡, Judith Eckle-Kohler†‡ and Iryna Gurevych†‡ † Ubiquitous Knowledge Processing Lab (UKP-TUDA) Department of Computer Science, Technische Universität Darmstadt ‡ Ubiquitous Knowledge Processing Lab (UKP-DIPF) German Institute for Educational Research www.ukp.tu-darmstadt.de Abstract (2013) applied argumentation mining in policy modeling. In this paper, we analyze and discuss ap- However, current approaches mainly focus on proaches to argumentation mining from the identification of arguments and their compo- the discourse structure perspective. We nents and largely neglect the identification of ar- chose persuasive essays and scientific ar- gumentation structures although an argument con- ticles as our example domains. By an- sists not only of a set of propositions but also ex- alyzing several example arguments and hibits a certain structure constituted by argumenta- providing an overview of previous work tive relations (Peldszus and Stede, 2013; Sergeant, on argumentation mining, we derive im- 2013). We argue in this paper that identifying ar- portant tasks that are currently not ad- gumentative relations and the argumentation struc- dressed by existing argumentation mining ture respectively is an important task for argu- systems, most importantly, the identifica- mentation mining. First, identifying argumenta- tion of argumentation structures. We dis- tive relations between argument components en- cuss the relation of this task to automated ables the identification of additional reasons for a discourse analysis and describe prelimi- given claim and thus allows the creation of valu- nary results of two annotation studies fo- able knowledge bases e.g. for establishing new cusing on the annotation of argumentation information retrieval platforms. Second, it is im- structure. Based on our findings, we derive portant to recognize which premises belong to a three challenges for encouraging future re- claim, since it is not possible to evaluate argu- search on argumentation mining. ments without knowing which premises belong to it. Third, automatically identifying the structure of 1 Introduction arguments enables novel features of applications, Argumentation mining is a recent research area such as providing feedback in computer-assisted which promises novel opportunities not only for writing (e.g., recommending reasonable usage of information retrieval, educational applications or discourse markers, suggesting rearrangements of automated assessment tools but also aims at im- argument components) or extracting argumenta- proving current legal information systems or pol- tion structures from scientific publications for au- icy modeling platforms. It focuses on automat- tomated summarization systems. ically identifying and evaluating arguments in In this paper, we analyze several examples of text documents and includes a variety of sub- argumentative discourse from the discourse struc- tasks like identifying argument components, find- ture perspective.1 We outline existing approaches ing accepted arguments and discovering argumen- on argumentation mining and discourse analysis tation structures. Researchers have already inves- and provide an overview of our current work on tigated argumentation mining in several domains. argumentation structure annotation in scientific ar- For instance, Teufel (1999) aims at identifying ticles and persuasive essays. We conclude this pa- rhetorical roles of sentences in scientific articles per with a list of challenges for encouraging future and Mochales-Palau and Moens (2011) identify 1 arguments in legal documents. Also, Feng and The examples are taken from persuasive essays which are either collected from the writing feedback section of Hirst (2011) investigated argumentation schemes http://www.essayforum.com or from the corpus in newspapers and court cases and Florou et al. compiled by Stab and Gurevych (2014) research on argumentation mining. and Stede, 2013) which indicate that the source ar- gument component is a reason or a refutation for 2 Background the target component. For instance, in both of the Philosophy and Logic proposed a vast amount of examples above, an argumentative support relation argumentation theories (e.g. Toulmin (1958), Wal- holds from the premise to the claim. The follow- ton et al. (2008), Freeman (2011)).2 The major- ing example illustrates a more complex argument ity of these theories generally agree that an ar- including one claim and three premises: gument consists of several argument components (3) “Everybody should study abroada . which can either be a premise or a claim. The sim- It’s an irreplaceable experience if you plest form of an argument includes one claim that learn standing on your own feetb since is supported by at least one premise (figure 1). you learn living without depending on anyone elsec . But one who is living Claim supports Premise overseas will of course struggle with loneliness, living away from family and Figure 1: Illustration of a simple argument friendsd .” Figure 2 shows the structure of the argument in The claim3 is the central component of an ar- (3). In this example, premiseb supports the claima gument that can either be true or false. Thus, the whereas premised attacks the claima . claim is a statement that should not be accepted by the reader without additional reasons. The second supports supports a b c d component of an argument, the premise4 , under- attacks pins the plausibility of the claim. It is usually pro- vided by the proponent (writer) for convincing the Figure 2: Argumentation structure of example (3). reader of the claim. Examples (1) and (2) illustrate two simple arguments, each containing a claim (in This example illustrates three important proper- bold face) and a single premise (underlined): ties of argumentation structures: (1) “It is more convenient to learn 1. Argumentative relations can hold between about historical or art items online. non-adjacent sentence/clauses, e.g. the ar- With Internet, people do not need to gumentative attack relation from premised to travel long distances to have a real look the claima . at a painting or a sculpture, which prob- ably takes a lot of time and travel fees.” 2. Some argumentative relations are signaled by (2) “Locker checks should be made indicators, whereas others are not. For in- mandatory and done frequently be- stance, the argumentative attack relation from cause they assure security in schools, premised to the claima is indicated by the dis- make students healthy, and will make course marker ‘but’, whereas the argumenta- students obey school policies.” tive support relation from premiseb to claima is not indicated by a discourse marker. These examples illustrate that there exist argu- ment components both on the sentence level and 3. Argumentative discourse might exhibit rea- on the clause level. soning chains, e.g. the chain constituted be- Argumentative relations are usually directed re- tween argument components a, b, and c. lations between two argument components and represent the argumentation structure. There ex- 3 Argumentation Mining ist different types like support or attack (Peldszus Previous approaches on argumentation mining 2 A review of argumentation theory is beyond the scope of cover several subtasks including the separation of this paper. A survey can be found in Bentahar et al. (2010) argumentative from non-argumentative text units 3 also called conclusion (Mochales-Palau and Moens, (Moens et al., 2007; Florou et al., 2013), the 2009) 4 sometimes called support (Besnard and Hunter, 2008) or classification of argument components (with dif- reason (Anne Britt and Larson, 2003) ferent component classes) (Rooney et al., 2012; Mochales-Palau and Moens, 2009; Teufel, 1999; not tailored to a particular genre. In their exper- Feng and Hirst, 2011), and the identification iments, they identify claims, premises and non- of argumentation structures (Mochales-Palau and argumentative text units in the Araucaria corpus. Moens, 2009; Wyner et al., 2010). Feng and Hirst (2011) also use the Araucaria cor- pus for their experiments, but focus on the identi- 3.1 Separation of Argumentative from fication of argumentation schemes (Walton, 1996) Non-argumentative Text Units which are templates for arguments (e.g. argument The first step of an argumentation mining pipeline from example or argument from position to know). typically focuses on the identification of argu- Since their approach is based on features extracted mentative text units before analyzing the compo- from mutual information of claims and premises, nents or the structure of arguments. This task it requires that the argument components are re- is usually considered as a binary classification liably identified in advance. Mochales-Palau and task that labels a given text unit as argumenta- Moens (2009) report several experiments for clas- tive or non-argumentative. One of the first ap- sifying argument components. They solely focus proaches was proposed by (Moens et al., 2007). on the legal domain and in particular on legal court They focus on the identification of argumentative cases from the European Court of Human Rights text units in newspaper editorials and legal doc- (ECHR). They consider the classification of argu- uments included in the Araucaria corpus (Reed ment components as two consecutive steps. They et al., 2008). The annotation scheme utilized in utilize a maximum entropy model for identifying Araucaria is based on a domain-independent ar- argumentative text units before identifying the ar- gumentation theory proposed by Walton (1996). gumentative role (claim and premise) of the identi- A similar approach is reported by Florou et al. fied components using a Support Vector Machine. (2013). In their experiments, they classify text segments crawled with a focused crawler as either 3.3 Identification of Argumentation containing an argument or not. They focus on the Structures identification of arguments in the policy model- ing domain for facilitating decision making. For Currently, there are only few approaches aiming that purpose, they utilize several discourse mark- at the identification of argumentation structures. ers and features extracted from the tense and mood For instance, the approach proposed by Mochales- of verbs. Palau and Moens (2011) relies on a manually Although the separation of argumentative from created context-free grammar (CFG) and on the non-argumentative text units is an important step presence of discourse markers for identifying a in argumentation mining, it merely enables the de- tree-like structure between argument components. tection of text units relevant for argumentation and However, the approach relies on the presence of does not reveal the argumentative role of argument discourse markers and exploits manually created components. rules. Therefore, it does not accommodate ill- formatted arguments (Wyner et al., 2010) and is 3.2 Classification of Argument Components not capable of identifying implicit argumentation The classification of argument components aims structures which are common in argumentative at identifying the argumentative role (e.g. claims discourse. Indeed, Marcu and Echihabi (2002) and premises) of argument components. found that only 26% of the evidence relations in One of the first approaches to identify argument the RST Discourse Treebank (Carlson et al., 2001) components is Argumentative Zoning proposed by include discourse markers. (Teufel, 1999). Each sentence is classified as one Another approach was presented by Cabrio and of seven rhetorical roles including e.g. claim, re- Villata (2012). They identify relations between ar- sult or purpose using structural, lexical and syn- guments of an online debate platform for identify- tactic features. The underlying assumption of this ing accepted arguments and to support the interac- work is that argument components extracted from tions in online debates. In contrast to the work of a scientific article provide a good summary of its Mochales-Palau and Moens (2011), this approach content. Rooney et al. (2012) also focus on the aims at identifying relations between arguments identification of argument components but in con- (macro-level) and not between argument compo- trast to the work of Teufel (1999) their scheme is nents (micro-level). 4 Argumentation and Discourse Analysis uses a different set of discourse relations than the PDTB (Prasad et al., 2008). Discourse analysis aims at identifying discourse It is still an open question how the proposed dis- relations that hold between adjacent text units with course relations relate to argumentative relations. text units being sentences, clauses or nominaliza- Although, there are preliminary findings that indi- tions (Webber et al., 2012). Since text units might cate that there are certain similarities (Cabrio et be argument components and discourse relations al., 2013), approaches like RST and PDTB aim are often closely related to argumentative rela- at identifying general discourse structures and are tions, previous work in automated discourse anal- not tailored to argumentative discourse. ysis is highly relevant for argumentation mining. The difference of the relations is best illustrated 4.1 Discourse Relations and Argumentative by the work of Biran and Rambow (2011), which Relations is to the best of our knowledge the only approach that focuses on the identification of distinct argu- Most previous work in automated discourse anal- mentative relations. The authors argue that exist- ysis is based on corpora annotated with general ing definitions of discourse relations are only us- discourse relations, most notably the Penn Dis- able as a building block for argumentation mining course Treebank (PDTB) (Prasad et al., 2008) and that there are no distinct argumentative rela- and the Rhetorical Structure Theory (RST) Dis- tions included in existing approaches. Therefore, course Treebank (Carlson et al., 2003). Whereas they combine 12 relations from the RST Discourse RST represents the discourse structure as a tree, Treebank (Carlson et al., 2001) to a single argu- the PDTB allows more general graph structure. mentative support relation for identifying justifi- For the annotation of discourse relations in the cations in online discussions. PDTB, two different types of discourse relations were distinguished: implicit and explicit relations. 4.2 Discourse Markers and Indicators of Whereas explicit discourse relations are indicated Argumentative Relations by discourse markers, implicit discourse relations are not indicated by discourse markers and the There is a large body of previous research in lin- identification of those relations requires more so- guistics on the role of discourse markers, sig- phisticated methods. nalling discourse relations (e.g.‘because’, ‘there- Take as an example the argumentation structure fore’, ‘since’, etc.) in discourse analysis. Most discussed in section 2. previous investigations of discourse markers are based on the PDTB (Prasad et al., 2008) and on the “Everybody should study abroada . It’s RST Discourse Treebank (Carlson et al., 2003). an irreplaceable experience if you learn However, a critically discussed question in this standing on your own feetb since you context is the definition of discourse markers. Are learn living without depending on any- discourse markers in the sense of indicators mark- one elsec . But one who is living over- ing discourse relations just words like ‘because’, seas will of course struggle with lone- ‘therefore’, ‘since’? Taboada (2006) investigates liness, living away from family and the role of discourse markers in corpora annotated friendsd .” with discourse relations according to the RST. In her discussion of related work on discourse mark- Whereas the argument components b and c, as ers in linguistics, she concludes that there are well as c and d are related through the discourse many lexical and linguistic devices signalling dis- marker ‘since’ (signalling an explicit CAUSE rela- course relations beyond discourse markers, such tion) and ‘but’ (signalling an explicit CONTRAST as the mood (e.g. indicative or conjunctive) or the relation), the discourse relation JUSTIFY between modality (e.g. possibility, necessity) of a sentence. a and b is an implicit relation. In particular, for argumentative discourse, the Existing approaches of discourse analysis pro- role of indicators, such as discourse markers, is not posed different sets of discourse relations, and well-understood yet, which is due to the lack of there is currently no consensus in the literature corpora annotated with argumentation structures. about the ‘right’ set of discourse relations. For Recently, Tseronis (2011) summarized interme- instance, the RST (Mann and Thompson, 1988) diate results of a corpus-based analysis of argu- mentative moves, aiming at the identification of there is no prior work on identifying argumenta- linguistic surface cues that act as argumentative tion structures on a fine-grained level in scientific markers. According to Tseronis (2011), any sin- full-texts yet (see section 3.3). gle or complex lexical expression can act as an Due to the lack of evaluation datasets, we are argumentative marker, and it can either mark an performing an annotation study with four annota- argumentative relation (i.e., connecting two argu- tors, two domain experts and two annotators who ments or argument components) or signal a certain developed the annotation guidelines. Our dataset argumentative role, such as a claim or a premise. consists of about 20 scientific full-texts from the Moreover, he observed that also sequential pat- educational domain. For the annotation study, terns of argumentative markers indicate particular we developed our own Web-based annotation tool argumentative moves, for instance, first stating the (see figure 3 for a screenshot). The annotation common ground (e.g., using the marker it is un- tool allows to label argument components directly derstandable ...) and then presenting an attack to in the text with different colors and to add differ- this common ground (e.g., using a marker such as ent relations (like support or attack) between ar- nevertheless). gument components. The resulting argumentation structure is visualized as a graph (see figure 3). 5 Argumentation Structure Annotation Next, we plan to develop weakly supervised Our research in argumentation mining is mo- machine learning methods to automatically anno- tivated by the (1) information access and (2) tate scientific publications with argument compo- computer-assisted writing perspective. Currently, nents and the relations between them. The first we are conducting two annotation studies, focused step will be to distinguish non-argumentative parts on analyzing argumentation structures in scientific (for example descriptions of the document struc- articles and persuasive essays. In the following ture) from argumentative parts (see section 3.1). subsections we provide an overview of the (pre- The second step will be to identify support and at- liminary) results. tack relations between the argument components. In particular, we will explore lexical features, such 5.1 Argumentation Structures in Scientific as discourse markers (for example ‘hence’, ‘so’, Articles ‘for that reason’, ‘but’, ‘however’, see section 4), and semantic features, such as text similarity or One of the main goals of any scientific publica- textual entailment. tion is to present new research results to an expert audience. In order to emphasize the novelty and 5.2 Identifying Argumentation Structures for importance of the research findings, scientists usu- Computer-Assisted Writing ally build up an argumentation structure that pro- vides numerous arguments in favor of their results. The goal of computer-assisted writing is to pro- The goal of this annotation study is to automati- vide feedback about written language in order cally identify those argumentation structures on a to improve text quality and writing skills of au- fine-grained level in scientific publications in the thors respectively. Common approaches are for educational domain and thereby to improve infor- instance focused on providing feedback about mation access. A potential use case could be an spelling and grammar, whereas more sophisti- automated summarization system creating a sum- cated approaches also provide feedback about dis- mary of important arguments presented in a scien- course structures (Burstein et al., 2003), readabil- tific article. ity (Pitler and Nenkova, 2008), style (Burstein and Up to now only coarse-grained approaches like Wolska, 2003) or aim at facilitating second lan- Argumentative Zoning (Teufel et al., 2009; Li- guage writing (Chen et al., 2012; Huang et al., akata et al., 2012; Yepes et al., 2013) have been 2012). developed for argumentation mining in scientific Argumentative Writing Support is a particu- publications. These approaches classify argument lar type of computer-assisted writing that aims at components according to their argumentative con- providing feedback about argumentation and thus tribution to the document (see section 3.2) but they postulates methods for reliably identifying argu- do not consider any relations between the argu- ments. Besides the recognition of argument com- ment components. To the best of our knowledge, ponents, the identification of the argumentation Figure 3: Screenshot of the annotation tool for argumentation structure annotation in scientific full- texts: The left side includes the text of a scientific article and the argument components marked with different colors and labels (a1-a7). The graph visualization on the right side illustrates the argumentation structure. Each node represents an argument component connected with several relations (‘support’, ‘attack’, ‘sequence’). structure is crucial for argumentative writing sup- topic and the author’s stance is crucial for anno- port, since it would open novel possibilities for tating arguments. According to these findings, providing formative feedback about argumenta- we defined a top-down annotation process start- tion. On the one hand, an analysis of the argu- ing with the major claim and drilling-down to the mentation structure would enable the recommen- claims and the premises so that the annotators are dation of more meaningful arrangements of argu- aware of the author’s stance and the topic before ment components and a reasonable usage of dis- annotating other components. Using this strategy, course markers. Both have been shown to increase we achieved an inter-rater agreement of αU = argument comprehension and recall, and thus the 0.725 for argument components and α = 0.81 quality of the text (Anne Britt and Larson, 2003). for argumentative relations indicating that the pro- On the other hand, by identifying which premises posed scheme and annotation process successfully belong to a claim, it would be possible to advice guides annotators to substantial agreement. For the author to add additional support in her/his ar- more details about this annotation study, we re- gumentation to improve the persuasiveness. fer the interested reader to (Stab and Gurevych, Following this vision, we conducted an anno- 2014), which includes a detailed description of the tation study with three annotators to model ar- annotation scheme, an analysis of inter-annotator gument components and the argumentation struc- agreements on different granularities and an er- ture in persuasive essays at the clause-level. The ror analysis. The corpus as well as the annotation corpus includes 90 persuasive essays which we guidelines are freely available to encourage future selected from essayforum.com. Our annotation research.6 scheme includes three argument components (ma- jor claim, claim and premise) and two argumen- tative relations (support and attack). For defining 5 We used Krippendorff’s αU (Krippendorff, 2004) for the annotation guidelines and the annotation pro- measuring the agreement since there are no predefined mar- cess we conducted a preliminary study on a cor- bles in our study and annotators had also to identify the boundaries of argument components. pus of 14 short text snippets with five non-trained 6 http://www.ukp.tu-darmstadt.de/data/ annotators and found that information about the argumentation-mining 6 Challenges (6) “Random locker checks should be made obligatory.a Locker checks help Existing approaches of argumentation mining students stay both physically and men- mainly focus on the identification of argument tally healthy.b It discourages students components (section 3). Based on the examples from bringing firearms and especially analyzed in section 2 and on the experience gained drugs.c ” in our annotation studies (section 5), we identified the following challenges for future research in ar- In this argumentation structure, a can be clas- gumentation mining that have not been addressed sified as a claim. However, without being aware adequately by previous work. of the argument component a, b becomes a claim Segmentation: Most of the existing approaches which is supported by premise c. The same situa- are based on the sentence-level. However, for an- tion can be found in example (3) in section 2. If we alyzing arguments, a more fine-grained segmenta- look at the argument components b and c in isola- tion is needed (Sergeant, 2013). Apart from the tion, we can classify b as claim. However, looking sentence level, in real world data argument com- at the whole example, the argument component a ponents exist on the clause level or can spread over is the claim, supported by the premise b. The same several sentences. For instance, example (4) il- holds for the argument components c and a which lustrates that a single sentence can contain multi- would be connected by a support relation if they ple argument components (claim in bold face and are considered in isolation. Both examples illus- premise underlined) (see also example (2) in sec- trate that the context is crucial for classifying ar- tion 2). In example (5) the premise consists of two gument components as claims or premises and for sentences, because both sentences are needed to identifying the argumentation structure. Although, represent and support the “different opinions” in Stab and Gurevych (2014) proposed an annotation the claim. process that facilitates these decisions in manual annotation studies of persuasive essays, it is still (4) “Eating apples is healthy which has an open issue how to model the context in order to to do with substrates which prevent can- improve the performance of automatic argumenta- cer and other diseases.” tion mining methods. (5) “There are different opinions about Ambiguity of Argumentation Structures: coffee. Some people say they need it to The most important challenge for identifying argu- stay awake. Other people think it’s un- mentation structures is ambiguity, since there are healthy.” often several possible interpretations of argumen- tation structures which makes it hard or even im- It is an open question if existing segmentation possible to identify one correct interpretation. In approaches can be used for reliably identifying the previous examples, we have already seen that the boundaries of argument components. In example classification of argument components depends on (4) we find two times the word “which”. This the context and the considered argument compo- makes it hard for a segmenter to split the sentence nents respectively. However, even if we consider correctly in only two parts. On the other hand, all components of an argument, there might be the combination of sentences (example (5)) also several reasonable interpretations of its structure. requires more elaborated techniques that are able For instance, the structure of example (6) can be to identify sentences that are related and only form interpreted in three different ways (figure 4). In the in combination the support of a particular claim. first interpretation, the argument component c sup- Context Dependence: The context is crucial ports argument component b and argument com- for identifying arguments, their components and ponent b supports argument component a, whereas argumentation structures. As illustrated by Stab in the second interpretation argument components and Gurevych (2014), it is even a hard task for hu- b and c both support argument component a. The man annotators to distinguish claims and premises third interpretation contains all possible argumen- without being aware of the context. For instance, tative relations from the first and second interpre- the following three argument components consti- tation combined, and thus represents a graph struc- tute a reasoning chain in which c is a premise for ture (in contrast to a tree structure). b and b a premise for a: The ambiguity of argumentation structures rep- a dressing the problem of ambiguous argumentation support a a structures. In particular, the ambiguity of argu- mentation structure poses an important issue for b support support support support support b c b support c future work. c References Figure 4: Several interpretations of the argumen- M. Anne Britt and Aaron A. Larson. 2003. Construct- tation structure of example (6). ing representations of arguments. Journal of Mem- ory and Language, 48(4):794–810. resents a major challenge for argument anno- Jamal Bentahar, Bernard Moulin, and Micheline Bélanger. 2010. A taxonomy of argumentation tation studies and consequently the creation of models used for knowledge representation. Artifi- reliable gold standards for argumentation min- cial Intelligence Review, 33(3):211–259. ing. In all annotation studies we know, exactly one annotation is considered to be correct which Philippe Besnard and Anthony Hunter. 2008. El- ements of argumentation, volume 47. MIT press means that other possibly correct interpretations Cambridge. are considered as incorrect and therefore down- grade the results for the inter annotator agree- Or Biran and Owen Rambow. 2011. Identifying jus- tifications in written dialogs by classifying text as ment and the performance of automatic classi- argumentative. International Journal of Semantic fiers. Consequently, it might be interesting to Computing, 05(04):363–381. explore different evaluation methods. For in- stance, evaluation schemes used in automatic text Jill Burstein and Magdalena Wolska. 2003. Toward Evaluation of Writing Style: Finding Overly Repet- summarization could be considered as an alterna- itive Word Use in Student Essays. In Proceedings tive. In text summarization, inter annotator agree- of the Tenth Conference on European Chapter of the ment for human-generated summaries is particu- Association for Computational Linguistics - Volume larly low, and hence, each human-generated sum- 1, EACL ’03, pages 35–42, Budapest, Hungary. mary is considered valid for evaluating an auto- Jill Burstein, Daniel Marcu, and Kevin Knight. 2003. matic summarization system (Nenkova and McK- Finding the WRITE Stuff: Automatic Identification eown, 2012). of Discourse Structure in Student Essays. IEEE In- telligent Systems, 18(1):32–39. 7 Conclusion Elena Cabrio and Serena Villata. 2012. Natural lan- In this paper, we showed that existing approaches guage arguments: A combined approach. In Pro- to argumentation mining mainly focus on the iden- ceedings of the 20th European Conference on Artifi- cial Intelligence, ECAI ’12, pages 205–210, Mont- tification of argument components and largely ne- pellier, France. glect the identification of argumentation struc- tures, although this task is crucial for many Elena Cabrio, Sara Tonelli, and Serena Villata. 2013. From discourse analysis to argumentation schemes promising applications, e.g., for building novel ar- and back: Relations and differences. In João Leite, gument related knowledge bases. By examining TranCao Son, Paolo Torroni, Leon Torre, and Ste- several examples, we derived characteristic prop- fan Woltran, editors, Computational Logic in Multi- erties of argumentation structures. We discussed Agent Systems, volume 8143 of Lecture Notes in the relation of discourse analysis and argumen- Computer Science, pages 1–17. Springer Berlin Hei- delberg. tation structure and showed that previous works in discourse analysis are not capable of identify- Lynn Carlson, Daniel Marcu, and Mary Ellen ing argumentation structures, because discourse Okurowski. 2001. Building a discourse-tagged cor- pus in the framework of rhetorical structure theory. relations do not cover all argumentative relations In Proceedings of the Second SIGdial Workshop on and are limited to relations between adjacent text Discourse and Dialogue - Volume 16, SIGDIAL ’01, units. Based on our observations, we derived three pages 1–10, Aalborg, Denmark. challenges for encouraging future research, i.e., Lynn Carlson, Daniel Marcu, and Mary Ellen (i) identifying the boundaries of argument compo- Okurowski, 2003. Building a discourse-tagged cor- nents, (ii) modeling the context of argument com- pus in the framework of rhetorical structure theory, ponents and argumentative relations, and (iii) ad- chapter 5, pages 85–112. Springer. Mei-Hua Chen, Shih-Ting Huang, Hung-Ting Hsieh, Marie-Francine Moens, Erik Boiy, Raquel Mochales Ting-Hui Kao, and Jason S. Chang. 2012. FLOW: Palau, and Chris Reed. 2007. Automatic detection A First-language-oriented Writing Assistant Sys- of arguments in legal texts. In Proceedings of the tem. In Proceedings of the ACL 2012 System 11th International Conference on Artificial Intelli- Demonstrations, ACL ’12, pages 157–162, Jeju Is- gence and Law, ICAIL ’07, pages 225–230, New land, Korea. York, NY, USA. ACM. Vanessa Wei Feng and Graeme Hirst. 2011. Classi- Ani Nenkova and Kathleen McKeown. 2012. A survey fying arguments by scheme. In Proceedings of the of text summarization techniques. In Mining Text 49th Annual Meeting of the Association for Com- Data, pages 43–76. putational Linguistics: Human Language Technolo- gies - Volume 1, HLT ’11, pages 987–996, Strouds- Andreas Peldszus and Manfred Stede. 2013. From burg, PA, USA. Association for Computational Lin- Argument Diagrams to Argumentation Mining in guistics. Texts: A Survey. International Journal of Cogni- tive Informatics and Natural Intelligence (IJCINI), Eirini Florou, Stasinos Konstantopoulos, Antonis 7(1):1–31. Koukourikos, and Pythagoras Karampiperis. 2013. Argument extraction for supporting public policy Emily Pitler and Ani Nenkova. 2008. Revisiting formulation. In Proceedings of the 7th Workshop on readability: A unified framework for predicting text Language Technology for Cultural Heritage, Social quality. In Proceedings of the Conference on Em- Sciences, and Humanities, pages 49–54, Sofia, Bul- pirical Methods in Natural Language Processing, garia, August. Association for Computational Lin- EMNLP ’08, pages 186–195. guistics. Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Milt- James B. Freeman. 2011. Argument Structure: Repre- sakaki, Livio Robaldo, Aravind Joshi, and Bonnie sentation and Theory, volume 18 of Argumentation Webber. 2008. The Penn Discourse TreeBank 2.0. Library. Springer. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08), Chung-chi Huang, Ping-che Yang, Mei-hua Chen, pages 28–30, Marrakech, Morocco. Hung-ting Hsieh, Ting-hui Kao, and Jason S. Chang. 2012. TransAhead: A Writing Assistant for CAT Chris Reed, Raquel Mochales-Palau, Glenn Rowe, and and CALL. In Proceedings of the Demonstrations at Marie-Francine Moens. 2008. Language resources the 13th Conference of the European Chapter of the for studying argument. In Proceedings of the Sixth Association for Computational Linguistics, EACL International Conference on Language Resources ’12, pages 16–19, Avignon, France. and Evaluation, LREC ’08, pages 2613–2618, Mar- rakech, Morocco. Klaus Krippendorff. 2004. Measuring the Reliability of Qualitative Text Analysis Data. Quality & Quan- Niall Rooney, Hui Wang, and Fiona Browne. 2012. tity, 38(6):787–800. Applying kernel methods to argumentation min- Maria Liakata, Shyamasree Saha, Simon Dobnik, ing. In Proceedings of the Twenty-Fifth Interna- Colin Batchelor, and Dietrich Rebholz-Schuhmann. tional Florida Artificial Intelligence Research So- 2012. Automatic recognition of conceptualization ciety Conference, FLAIRS ’12, pages 272–275, zones in scientific articles and two life science ap- Marco Island, FL, USA. plications. Bioinformatics, 28(7):991–1000. Alan Sergeant. 2013. Automatic argumentation ex- William C. Mann and Sandra A. Thompson. 1988. traction. In Proceedings of the 10th European Se- Rhetorical structure theory: Toward a functional the- mantic Web Conference, ESWC ’13, pages 656–660, ory of text organization. Text, 8(3):243–281. Montpellier, France. Daniel Marcu and Abdessamad Echihabi. 2002. An Christian Stab and Iryna Gurevych. 2014. Annotat- unsupervised approach to recognizing discourse re- ing argument components and relations in persua- lations. In Proceedings of the 40th Annual Meeting sive essays. In Proceedings of the 25th International on Association for Computational Linguistics, ACL Conference on Computational Linguistics (COLING ’02, pages 368–375. 2014), page (to appear), Dublin, Ireland, August. Raquel Mochales-Palau and Marie-Francine Moens. Maite Taboada. 2006. Discourse Markers as Sig- 2009. Argumentation mining: The detection, classi- nals (or Not) of Rhetorical Relationsteu. Journal of fication and structure of arguments in text. In Pro- Pragmatics, 38(4):567–592. ceedings of the 12th International Conference on Ar- tificial Intelligence and Law, ICAIL ’09, pages 98– Simone Teufel, Advaith Siddharthan, and Colin Batch- 107, New York, NY, USA. ACM. elor. 2009. Towards discipline-independent ar- gumentative zoning: evidence from chemistry and Raquel Mochales-Palau and Marie-Francine Moens. computational linguistics. In Proceedings of the 2011. Argumentation mining. Artificial Intelligence 2009 Conference on Empirical Methods in Natural and Law, 19(1):1–22. Language Processing, pages 1493–1502. Simone Teufel. 1999. Argumentative Zoning: Infor- mation Extraction from Scientific Text. Ph.D. thesis, University of Edinburgh. Stephen E. Toulmin. 1958. The uses of Argument. Cambridge University Press. Assimakis Tseronis. 2011. From connectives to argu- mentative markers: A quest for markers of argumen- tative moves and of related aspects of argumentative discourse. Argumentation, 25(4):427–447. Douglas Walton, Chris Reed, and Fabrizio Macagno. 2008. Argumentation Schemes. Cambridge Univer- sity Press. Douglas N Walton. 1996. Argumentation schemes for presumptive reasoning. Routledge. Bonnie Webber, Mark Egg, and Valia Kordoni. 2012. Discourse structure and language technology. Natu- ral Language Engineering, 18:437–490, 10. Adam Wyner, Raquel Mochales Palau, Marie-Francine Moens, and David Milward. 2010. Approaches to text mining arguments from legal cases. In Semantic Processing of Legal Texts, volume 6036 of Lecture Notes in Computer Science, pages 60–79. Springer. Antonio Jimeno Yepes, James G. Mork, and Alan R. Aronson. 2013. Using the argumentative structure of scientific literature to improve information access. In Proceedings of the 2013 Workshop on Biomedical Natural Language Processing, pages 102–110.