Scientific argumentation detection as limited-domain intention recognition Simone Teufel Computer Laboratory University of Cambridge 15 JJ Thomson Avenue, Cambridge, UK Simone.Teufel@cl.cam.ac.uk Abstract 2008; Walton et al., 2008; Green, 2014). We are here interested in a definition close to discourse structure, We describe the task of intention-based text and concentrate in particular on the recognition of understanding for scientific argumentation. prototypical argumentation steps in scientific expo- The model of scientific argumentation pre- sition. We posit that these argumentation steps can sented here is based on the recognition of be defined at an abstract level so that world knowl- 28 concrete rhetorical moves in text. These moves can in turn be associated with higher- edge is not required for their recognition. level intentions. The intentions we aim to There is a clear connection between our goal and model operate in the limited domain of sci- intention recognition. Fully understanding every as- entific argumentation and justification; it is pect of an author’s argumentation requires the recog- the limitation of the domain which makes our nition of all of their intentions, which in turn means intentions predictable and enumerable, unlike that we would have to model, generalise over, and general intentions. do inference with general world knowledge. This We explain how rhetorical moves relate to is of course an AI-hard task fraught with many the- higher-level intentions. We also discuss oretical and practical problems; consider the sym- work in progress towards a corpus annotated with limited-domain intentions, and speculate bolic AI work on this and closely related problems about the design of an automatic recognition (e.g., Schank and Abelson, 1977; Pollack, 1986, system, for which many components already 1990; Norvig, 1989; Cohen et al., 1990 and Car- exist today. berry, 1990). We will propose instead to reframe argumenta- tion detection as a limited-domain intention recog- 1 Introduction nition task. The basic building blocks of our model Automatically recognising the structure of an argu- of an argument are instances of higher-level inten- ment is an attractive and challenging task, which has tions which the authors are likely to have had when received interest for a long time from the AI as well they were writing their paper. The representation we as the natural language processing community, and suggest for intentions does not contain any proposi- recently from both communities together in a joint tional content based on arbitrary world knowledge. effort. Because arguments are global text structur- Instead, our intentions are represented as generalised ing devices, argument recognition has the potential propositions such as “Our solution is better than the to advance text understanding and the many real-life competition’s”. Such speech acts realise parts of the tasks that could profit from it. author’s intention of persuading the reader that the There are various definitions of what an argu- work described in the paper is novel and significant. ment is (Toulmin, 1958; Cohen, 1984; Dung, 1995; When during processing we encounter the sentence Brüninghaus and Ashley, 2005; Besnard and Hunter, To our knowledge, our system is the first one aimed at building semantic lexicons from raw text tain sentiment, in the form of “good” vs. “bad” situ- without using any additional semantic knowledge. ations, as well as successful vs. failed problem solv- (9706013, S-171) ing acts. our representation only registers the author’s inten- As far as the representation of time in the events tion of staking a novelty claim for their new work. and states described in rhetorical moves is con- The proposition is generalised in that the proposi- cerned, another simplification is possible: it suffices tional content of the novelty, i.e., the fact that the to model three points in time, the time before the authors built the first lexicon from raw text without authors’ research activity begins (t0 ), and the times any additional semantic knowledge, is not encoded. during (t1 ) and after (t2 ) their research activity. Of This detail is not important at the level of abstraction course, the real actions by the authors that gave rise we have in mind. to the research in the paper are spread in time in far The simplification of argument recognition into more complex ways, but a scientific paper is a so- a limited-domain intention recognition problem is cial construct (Bazerman, 1985). The telling of “the possible because of the high degree of convention- story” follows the convention that all research acts alisation of scientific argumentation. Following associated with the paper happen simultaneously, Swales (1990), we call explicit statements such as and that they transform an earlier state of the world the above novelty claim “rhetorical moves”. Rhetor- into a new (better) one. ical moves are well-documented in various disci- These simplifications allow us to define the plines: they occur frequently, and they can be enu- 28 rhetorical moves in Figure 11 . We also give some merated and classified, as applied linguists have examples of rhetorical moves from the chemistry, done in some detail for several disciplines (e.g., My- computational linguistics and agriculture literature, ers, 1992; Hyland, 1998; Salager-Meyer, 1992). which were sourced from our annotated corpora. Swales also coined the expression “research The overall argumentation structure we propose space” – a cognitive construct consisting of scien- concerns the author’s argument that their research tific problems, methods and research acts that au- was worthy of publication, and all of its subargu- thors use when they locate their research with re- ments – which, at its heart, is always the same ar- spect to historical approaches and current trends. gument. Argument recognition then corresponds to When we faced the decision of which types a guess as to which strategy the author pursued in of semantic participants to encode in our repre- making this argument. This process will have to sentation of rhetorical moves, we tried to achieve be driven by a bottom-up recognition of rhetorical as much generalisation as possible, in line with moves, as these are the only explicitly expressed the Knowledge Claim Discourse Model (KCDM, parts of the argument. This will trigger a simple Teufel, 2010). In fact, the core semantic participants form of inference as to which higher-level intention in rhetorical moves can be reduced to just two sets – might have been present during the writing of the US (the paper’s authors) and THEM (everybody else paper. who has ever published). In previous work, we have used a robust classi- When it comes to the states and events expressed fication model called Argumentative Zoning (AZ; in rhetorical moves, we maximally generalise again Teufel, 2000, 2010; Teufel et al. 2009, O’Seaghdha and end up with four classes of predicates, where and Teufel 2014), that turns some aspects of the the classes are defined based on the number of par- more general argumentation recognition model of ticipants in the logical act expressed in the move. the KCDM into a simple sentence classification task. We differentiate statements about the authors’ own In AZ, rhetorical moves with a similar function were work (US); statements about others’ previous work bundled together into 7 (in later versions 15 or 6) (THEM); statements about the connection between flat classes or zones, and each sentence was classi- the authors’ work with previous work (US and fied into one of these on the basis of surface features, THEM); and finally statements about the research space and the authors’ position in it. Another rele- 1 An earlier version of the list of moves appears in Teufel vant observation is that rhetorical moves often con- (1998). I. Properties of research space including sequence information. This way of phras- R-1 Problem addressed is a problem ing the problem allows for tractable recognition and R-2 New goal/problem is new evaluation. AZ classification has been shown to lead R-3 New goal/problem is hard to stable and reliable annotation on several scientific R-4 New goal/problem is important/interesting disciplines, and it is also demonstrably useful for a R-5 Solution to new problem is desirable set of applications such as the detection of new ideas R-6 No solution to new problem exists in a large scientific area, summarisation, search, and II. Properties of new solution (US) writing assistance. R-7 New solution solves problem R-8 New solution avoids problems Nevertheless, AZ is only a flat approximation of R-9 New solution necessary to achieve goal a larger argumentation model of scientific justifica- R-10 New solution is advantageous tion. The work presented here is a departure from R-11 New solution has limitations AZ in that it aims to model the stages of scientific R-12 Future work follows from new solution argumentation in a more informative, finer-grained III. Properties of existing solution (THEM) way. H-1 Existing solution is flawed H-2 Existing solution does not solve problem 2 The role of citations in the argument H-3 Existing solution introduces new problem H-4 Existing solution solves problem The reader may have noticed that the rhetorical H-5 Existing solution is advantageous moves in parts III and IV of Fig. 1, which are con- IV. Relationships between existing cerned with statements about THEM (i.e., other pub- and new solutions (US and THEM) lished authors), are closely connected to citation H-6 New solution is better than existing solution function2 . In fact, we have in the past attempted H-7 New solution avoids problems (when existing the recognition of some of the H-moves as an iso- does not) H-8 New goal/problem/solution is different from lated task, in the form of citation function classifica- existing tion (CFC; Teufel et al., 2006); others (Garzone and H-9 New goal/problem is harder than existing Mercer, 2000; Cohen et al., 2006) have used other goal/problem schemes for similar citation classification tasks. H-10 New result is different from existing result Where, how often, and how authors cite previous H-11 New claim is different from/clashes with exist- work is an important aspect of their overall scientific ing claim argument. For instance, the authors might choose H-12 Agreement/support between existing and new claim one of the possible articles types (review, research H-13 Existing solution provides basis for new solu- paper, pioneer work etc) to support a particular point tion in their overall argument. The choice of a particu- H-14 Existing solution provides part of new solution lar pioneer paper might signal their intellectual her- H-15 Existing solution (adapted) provides part of itage. They might tell us who their rivals are, and new solution who uses similar methods for a different goal (i.e., H-16 Existing solution is similar to new solution not rivals), whose infrastructure they borrow, and Recently, R-4 the use of imines as starting materials whose work supports theirs and vice versa. These in the synthesis of nitrogen-containing compounds has questions will crucially influence where in the text attracted a lot of interest from synthetic chemists.(1) (physically and logically in terms of the argumenta- (b200198e) tion) a given citation will occur. H-4 This account makes reasonably good empirical As a result of all this, it is often possible to de- predictions, though H-2 it does fail for the following termine some citations as being particularly central examples: . . . (9503014, S-75) to the authors’ paper. This information, if it could H-12 Greater survival of tillers under irrigated con- ditions agrees with other reports in barley [4,28] and be automatically determined from text in a reliable wheat [10,13,26]. (A027) 2 These 16 moves also follow a different naming scheme, where the move name starts with the letter “H” – historically, Figure 1: Rhetorical moves; some examples such moves were called “hinge” moves, as opposed to the “R” (“rhetorical”) moves in parts I and II of Fig 1. way, would vastly improve bibliographic search. It research goals, one may simply show that no other also has the potential to improve bibliometric assess- work is similar enough to one’s goal: new goals (cre- ments of a piece of work’s impact, e.g. in the sense ated at t1 ) cannot be compared to existing state-of- of Borgman and Furner (2002), White (2004), and the-art, which is frozen in time at t0 . (Novelty is Boyak and Klavans (2010). a rare example of a high-level intention which can be left to the reader to infer, or alternatively stated 3 Higher-level intentions explicitly as move R-2 or R-6.) Note that each citation that has an H-type rhetor- There are some rhetorical moves that at first glance ical move associated with it automatically strength- seem to make litte sense. Stating H-5, praise of other ens the claim that the authors are knowledgeable in people’s work, might comparatively weaken the au- the field (one of the important subgoals of HLG-4, thor’s own knowledge claim. Similarly, stating H-9, soundness). Under our model, citations without any the fact that the author’s research goal is harder than associated H-move are not contributing to this goal, other people’s goal, might prompt the criticism that as a knowledgeable author must be able to state the the authors have simply chosen their goal badly – relationship of the current work to earlier work. (A had they chosen an easier goal, the solution might simple statement of similarity with somebody else’s have been easier, or achieved better results. work should barely count, but has been given a However, rhetorical moves must be interpreted as “weak” move, H-16, because we encountered it so part of the larger picture of the overall scientific ar- frequently in our corpus studies.) gument. Scientific writing can be seen as one big From Fig. 2 we can now see why stating H-5 can game where an author’s overall goal is to success- be a good strategic move even though it praises other fully manoeuvre their paper past the peer review, so people’s work – it supports HLG-4 (soundness of that it can be published. methodology) via the sub-argument that by includ- According to the conventions of peer review, there ing praise-worthy existing work, the authors make is a small set of criteria for acceptance – the authors sure they use the best methods currently available. need to show that the problem they address is justi- Similarly, the statement that one’s goal is harder fied (High-Level-Goal 1 or HLG-1 for short), that than somebody else’s motivates that the authors’ their knowledge claim is significant (HLG-2) and chosen problem is justified (HLG-1) and significant novel (HLG-3), and that the research methodology (HLG-2), and additionally strengthens HLG-4 (via they use is sound (HLG-4). If valid evidence for the the claim that the authors know their field well). fulfilment of these criteria is presented, the peer re- This illustrates that a rhetorical move can support view cannot justifiably reject the paper. more than one high-level intention. Fig. 2 spells out how the overall argument for validity is put together from high- and medium- 4 Knowledge representation of moves and level intentions and rhetorical moves3 . Rhetorical moves in Fig. 2 appear in shaded boxes (H- and R- intentions type moves in different shades of grey). Above the What has been said so far raises the question of rhetorical moves, we see a simple representation of which knowledge representation is most suited for the intentions posited in the model. For simplic- modelling intentions and rhetorical moves. Design- ity and readability, Fig. 3 repeats the same network ing a propositional logic that expresses the full se- without rhetorical moves. The arrows in both figures mantics of rhetorical moves and of higher-level in- express the “supports” relationship in argumentation tentions is a task that goes far beyond the current theory. For instance, in order to argue for the novelty paper; it requires a thorough design of the semantics of one’s work, a state-of-the-art comparison may or of objects and events/states in this limited domain, may not be necessary – this depends on whether one as well as an appropriate type of inference. Nev- describes the research goal as new or not. For new ertheless, we will sketch some of the principles of 3 An earlier version of this diagram appears as Fig.3.1.7 in what might be usefully encoded. Teufel (2000, p.105). The THEM entities would need to be grounded to Figure 2: Argumentation network (including rhetorical moves). Valid contribution to science HLG-1 Justification HLG-3 Novelty HLG-4 Soundness P well-motivated HLG-2 Significance Solution New Good usage Knowledgable R-1 R-5 Clash with literature P Known No solution to P exists P New Support We use it Existing is good R-4 H-10 H-11 P big enough Comparison to SoA R-6 R-2 H-12 H-13 H-14 H-4 H-5 Others tried to solve P R-11 R-12 Solution works They are bad We are better Suff. different H-15 H-17 Suff. # citations R-3 R-10 R-7 R-8 R-9 H-3 H-1 H-2 H-7 H-6 H-16 H-18 H-8 H-9 Valid contribution to science HLG-1 Justification HLG-3 Novelty HLG-2 Significance HLG-4 Soundness P well-motivated Solution New Clash with literature Support Knowledgable Good usage No solution to P exists P Known P New Suff. # citations Existing is good We use it P big enough Comparison to SoA Suff. different Others tried to solve P They are bad We are better Solution works Figure 3: Argumentation network (excluding rhetorical moves). citations, possibly also to more general entities such R-5 solution(s) ∧ solve(s, p, t1) ∧ good(a, t2) ∧ as “many linguists in the 1970s”. Entities would aspect(a, s) ∧ problem(p) ∧ address(US, p) need to be tracked throughout the paper, for in- R-12 problem(p1) ∧ cause(s, p1, t1 ) ∧ solution(s) ∧ solve(s, p) ∧ problem(p) ∧ address(US, p stance by performing co-reference. We would also H-1 solution(s1) ∧ own(THEM, s1 ) ∧ bad(a, t0 ) need to represent problems, solutions and goals as ∧ aspect(a, s) ∧ solve(s1 , p) ∧ problem(p) ∧ atomic types, i.e., the fact that they are considered address(US, p problems, solutions and goals, rather than their con- H-7 solution(s1) ∧ own(THEM, s1 ) ∧ solution(s) tent. (The system should keep pointers to the textual ∧ own(US, s) ∧ 6 solve(s1 , p, t0 ) (∧ strings that express this content, so that down-stream solves(s, p, t1 ) processing or human users can gain access to this in- H-15 own(THEM, s1 ) ∧ solution (s1 ) ∧ solution formation.) (s2 ) ∧ change(US, s1 , s2 , t1 ) ∧ use(US, s2 , t1 ) The exact representation of a proposition is open Figure 4: Sketch of knowledge representation for se- to speculation at this point, but moves would likely lected rhetorical moves be decomposed into atomic clauses. Events and properties in the limited domain (such as changing a solution into another one, or the fact that one solu- As an example of what the representation might tion is better than another) would be associated with look like, Fig. 4 expresses five moves in a simple a time; for instance all actions that logically happen prepositional logic. Here, ownership of solutions during the research act presented in the paper would (by US or THEM) is expressed directly, as are sim- be associated with t1 . ple relationships between solutions, problems, re- Inference could be performed by a theorem sults and claims. Consider move H-15, for instance prover, which could inhibit or further activate the – adapting somebody else’s solution means taking potentially possible “supports” relationships given it, changing it into something else, and then using in Fig. 1, by taking the plausibility of a particular the changed solution. Some moves, such as R-6 inference into account, in the light of the textual ev- and R-9, look like they might require quantification, idence encountered. which exceeds the expressivity of simple predicate Axioms could directly encode some of the rules logic. of the scientific publication game, such that the ex- Several aspects of the moves’ semantics are not istence of a problem is a bad state, that of a solution explicitly expressed in text; they could even be mod- is a good state, but that a solution needing something elled as presuppositions. For instance, R-7 states else is a bad state again. Temporal inference could that a rival’s solution does not solve one’s problem, require axioms such as things that persist at a cer- which presupposes that the author’s solution does, tain time also persist in later times, unless they are otherwise it would not be a relevant statement. R-7 changed. thereby implicitly invokes a comparison between the author’s approach and the rivals’, which is won by forms inference as to which higher-level intention is the authors. Crucially, whether or not the authors’ supported by currently activated rhetorical moves. successful problem-solving is explicitly mentioned The output of such an analysis would be a par- in the text or not is optional. Another example is the tially activated network expressing the overall ar- need to know whether a problem mentioned in a cer- gument likely to be followed in the paper, where tain rhetorical move is actually the problem that the each node in the network is annotated with a authors address in the current paper. This is often more or less instantiated knowledge representation. decisive, because the knowledge claim of the paper The activated network can be considered as an is connected exclusively to this particular problem. automatically-derived explanation for the place in In some part of the paper, the authors give us the in- the research space where the authors situate them- formation which problem it is that they address, but selves. they will typically not repeat this elsewhere. Newly-derived, intermediate levels of informa- It is the discourse model’s job to accumulate the tion should be additionally available from such an information about the identity of important prob- analysis, as a side-effect of this hybrid style of lems in its knowledge representation. This can be recognition. For instance, coreference resolution is done either via coreference or via some other mech- an important aspect of analysis and contributes to the anism that infers that the discourse is still concerned superficial features. It could also feed into a mech- with the same problem. This may seem a very hard anism that determines which of the cited previous task, but at least it is not doomed in principle: in approaches is central to the argumentation in the pa- earlier work we managed to train non-experts in per- per, which of these the authors present as their main forming similar inferences and judgements during rivals or collaborators, and which aspects of existing AZ annotation, using no world knowledge, only dis- work they criticise or praise. course cues. It is quite obvious that a solution to this task would be immediately useful for a host of appli- 5 Design of a recogniser cations in search, summarisation and the teaching of scientific writing. As the system would be able How could all this be recognised in unlimited text? to associate textual statements with the correspond- The recognition of rhetorical moves would drive ing likely intentions it recognised, it could produce a recognition with this model; as the only visible parts justification for its overall analysis of the argument. of the argument, rhetorical moves correspond to the Operating as a text critiquer, such a system could bottom-up element. In contrast, high-level inten- point out badly-expressed instances of well-known tions form the top-down, a priori expectations. They argumentation patterns, e.g. missing or weak evi- can only ever be inferred, because the authors typi- dence for particular high-level intentions. cally leave them implicit, so their recognition will Appealing though such applications are, the main never be made with absolute certainty. point of the analysis laid out here is the development A hybrid statistical-symbolic recogniser of scien- of a theory of text understanding of naturally occur- tific argumentation could instantiate the network in ring arguments in scientific text. Given the state of Fig. 2 on the fly for each new incoming paper, and current NLP technology, some of the intermediate keep a knowledge base of propositions derived dur- levels of recognition necessary for this seem to us to ing recognition. Whenever one of the moves is de- be within reach in the near future. tected, the activation of its associated box is trig- gered. Statistically trained recognisers based on su- 6 Conclusions perficial features and evidence from tens of thou- sands of analysed papers provide a confidence value This paper promotes robust text understanding of for the recognition of each move, which is translated scientific articles in a deeper manner than is cur- into the strength of activation.The symbolic part of rently practiced, as this would lead to more infor- the recogniser keeps track of the logic representation mative, symbolic representations of argument struc- accumulated up to that point in processing, and per- turing. Mature technologies exist for determining specific scientific entities such as gene names (cf. in the process of corroborating the argumentation- the review by Campos et al., 2014) and specific level observations by corpus annotation of rhetori- events such as protein–gene interactions (e.g., Reb- cal moves. This initially takes the form of adding holz et al., 2005). In contrast to our work, such ap- information to already existing AZ- and CFC-level proaches are domain-specific and only recognise a annotation, with the aim of constructing a full-scale small part of the entities or relationships modelled rhetorical move annotation. Higher-level goals will here. A different line of research associates text then be annotated as a second step. pieces with the research phase or information struc- Practical work also concerns building the recog- ture a given statement belongs to, where information nisers of rhetorical moves. Several such recognis- structure is defined in terms of methods, results, con- ers already exist and will be refined in future work. clusions etc, as in the work of Liakata et al. (2010), It will be interesting to study exactly when infer- Guo et al (2013) and Hirohata et al. (2008). A re- ence about higher-level intentions becomes neces- lated task, hedge detection in science, has been es- sary, and which kinds of constraints can be derived tablished and competitively evaluated (see Farkas et from the argumentation network and the knowledge al. (2010) for an overview of the respective CoNLL representation so as to usefully guide the inference shared task). While these two approaches (informa- mechanism. tion structuring and hedge recognition) are domain- independent like ours, the analysis presented here aims at a deeper, more informative representation of References relationships between general entities in the research Charles Bazerman. 1985. Physicists reading physics, space. schema-laden purposes and purpose-laden schema. At the other end of the spectrum, we are aware Written Communication, 2(1):3–23. of at least one deeper analysis of argument struc- Philippe Besnard and Anthony Hunter. 2008. Elements ture in science than ours, which is manual and of argumentation. MIT Press. Christine L. Borgman and Jonathan Furner. 2002. Schol- takes world-knowledge into account, namely Green arly communication and bibliometrics. In Annual re- (2014); our approach differs from hers in that we opt view of information science and technology: Vol. 36, to model argumentation in a domain- and discipline- pages 3–72. Information Today, Medford, NJ. independent manner, which is automatic but neces- Kevin W. Boyack and Richard Klavans. 2010. Co- sarily at a far shallower level. citation analysis, bibliographic coupling, and direct Our claims in this paper include that a logical sci- citation: Which citation approach represents the re- entific argument structure exists and can be inter- search front most accurately? Journal of the Ameri- preted by a human reader, even in light of ambiguity can Society for Information Science and Technology, 61(12):2389–2404. and although only some steps of the argumentation Stefanie Brüninghaus and Kevin D. Ashley. 2005. Gen- are explicitly stated. We have also claimed that this erating legal arguments and predictions from case type of analysis holds for all disciplines in principle, texts. In Proceedings of the 10th international confer- but certainly for all empirical sciences. We further ence on Artificial intelligence and law, pages 65–74. claim that a substantial part of the argumentation in a David Campos, Srgio Matos, and Jos Lus Oliveir. 2014. well-written paper is recognisable to a reader even if Current methodologies for biomedical named en- they do not have any domain knowledge. These are tity recognition. In Mourad Elloumi and Albert Y. rather strong claims: It is not even clear whether hu- Zomaya, editors, Biological Knowledge Discovery mans can recognise the explicit argumentation parts, Handbook: Preprocessing, Mining and Postprocess- ing of Biological Data. Wiley. let alone the inferred ones. We therefore need to sub- Sandra Carberry. 1990. Plan Recognition in Natural stantiate the claims with annotation experiments. Language Dialogue. MIT Press, Cambridge, MA. In our work to date, we have made empirical Philip R. Cohen, Jerry Morgan, and Martha E. Pollack, observations about argumentation structure in syn- editors. 1990. Intentions in Communication. MIT thetic chemistry, computer science, computational Press, Cambridge, MA. linguistics, and agriculture, but many of these are A.M. Cohen, W.R. Hersh, K. Peterson, and Po-Yin Yen. confined to the level of AZ or CFC. We are now 2006. Reducing workload in systematic review prepa- ration using automated citation classification. Jour- Conference on Computational Linguistics (COLING nal of the American Medical Informatics Association, 2014), Dublin, Ireland. 13(2):206–219. Martha E. Pollack. 1986. A model of plan inference Robin Cohen. 1984. A computational theory of the that distinguishes between the beliefs of actors and ob- function of clue words in argument understanding. In servers. In Proceedings of the 24th Annual Meeting of Proceedings of the 10th International Conference on the Association for Computational Linguistics (ACL- Computational Linguistics(COLING-84), pages 251– 86), pages 207–214, New York, US. 255. Martha E. Pollack. 1990. Plans as complex mental at- Phan Minh Dung. 1995. On the acceptability of argu- titudes. In P.R. Cohen, J. Morgan, and M.E. Pollack, ments and its fundamental role in nonmonotonic rea- editors, Intentions in Communication, pages 77–103. soning, logic programming, and n-person games. Ar- MIT Press, Cambridge, MA. tificial Intelligence, 77:321–357. Dietrich Rebholz-Schuhmann, H Kirsch, and F Couto. Richrd Farkas, Veronika Vincze, Gyrgy Mra, Jnos Csirik, 2005. Facts from textis text mining ready to deliver? and Gyrgy Szarvas. 2010. The conll-2010 shared PLoS Biol, 3(2). doi:10.1371/journal.pbio.0030065. task: learning to detect hedges and their scope in nat- Francoise Salager-Meyer. 1992. A text-type and move ural language text. In Proceedings of CoNLL ’10: analysis study of verb tense and modality distributions Shared Task Proceedings of the Fourteenth Confer- in medical English abstracts. English for Specific Pur- ence on Computational Natural Language Learning — poses, 11:93–113. Shared Task. Roger C. Schank and Robert P. Abelson. 1977. Scripts, Mark Garzone and Robert E. Mercer. 2000. Towards Goals, Plans and Understanding. Lawrence Erlbaum, an automated citation classifier. In Proceedings of Hillsdale, NJ. the 13th Biennial Conference of the CSCI/SCEIO (AI- John Swales, 1990. Genre Analysis: English in Aca- 2000), pages 337–346. demic and Research Settings. Chapter 7: Research ar- Nancy L. Green. 2014. Towards creation of a corpus ticles in English, pages 110–176. Cambridge Univer- for argumentation mining the biomedical genetics re- sity Press, Cambridge, UK. search literature. In Proc. of the First Workshop on Simone Teufel, Advaith Siddharthan, and Colin Batche- Argumentation Mining, ACL 2014. lor. 2009. Towards discipline-independent argumen- Yufan Guo, Roi Reichart, and Anna Korhonen. 2013. tative zoning: Evidence from chemistry and computa- Improved information structure analysis of scientific tional linguistics. In Proceedings of EMNLP-09, Sin- documents through discourse and lexical constraints. gapore. In Proceedings of NAACL-2013, Atlanta, US. Simone Teufel. 1998. Meta-discourse markers and Kenji Hirohata, Naoaki Okazaki, Sophia Ananiadou, and problem-structuring in scientific articles. In Proceed- Mitsuru Ishizuka. 2008. Identifying sections in sci- ings of the ACL-98 Workshop on Discourse Struc- entific abstracts using conditional random fields. In ture and Discourse Markers, pages 43–49, Montreal, Proceedings of the Third International Joint Confer- Canada. ence on Natural Language Processing (IJCNLP 2008), Simone Teufel. 2000. Argumentative Zoning: Infor- pages 381–388, Hyderabad, India. ACL Anthology mation Extraction from Scientific Text. Ph.D. thesis, Ref. I08-1050. School of Cognitive Science, University of Edinburgh, Ken Hyland. 1998. Persuasion and context: The prag- Edinburgh, UK. matics of academic metadiscourse. Journal of Prag- Simone Teufel. 2010. The Structure of Scientific Arti- matics, 30(4):437–455. cles: Applications to Citation Indexing and Summa- Maria Liakata, Simone Teufel, Advaith Siddharthan, and rization. CSLI Publications. Colin Batchelor. 2010. Corpora for conceptualisation Stephen Toulmin. 1958. The Uses of Argument. Cam- and zoning of scientific papers. In In: Proceedings of bridge University Press. LREC-10, Valetta, Malta. Douglas Walton, Chris Reed, and Fabrizio Macagno. Greg Myers. 1992. In this paper we report...—speech 2008. Argumentation Schemes. Cambridge University acts and scientific facts. Journal of Pragmatics, Press. 17(4):295–313. Howard D. White. 2004. Citation analysis and discourse Peter Norvig. 1989. Marker passing as a weak method analysis revisited. Applied Linguistics, 25(1):89–116. for text inferencing. Cognitive Science, 13(4):569– 620. Diarmuid O’Seaghdha and Simone Teufel. 2014. Un- supervised learning of rhetorical structure with un- topic models. In Proceedings of the 25th International