Accepting Freeman's criterion means that an argument- ation that has “modal linkage” has convergent argument structure and an argument that has “relevance linkage” has linked argument structure. That means that dialectically complex argument structures like multiple or cumulative coor- dinatively compound argumentation will be logically conver- gent. A Cumulative Coordinatively compound argumentation structure cannot have a logical structure of relevance linkage, since “doubting sufficiency” presupposes that the antagonist consider that the doubted argument was at least relevant for the standpoint in question. From a logical point of view, they share no mediate concept, therefore they are only modally “linked” and therewith independently relevant for the standpoint, hence they are convergent (if assuming Freeman’s criterion). Single argumentation in the pragma-dialectical system will, from a logical point of view, always have linked argument structure since all propositions are necessary in order to be a reason at all. Even though an argumentation structure with either contrasting types of argumentschemes or contrasting instances of the same type of argument scheme is a necessary condition for being a multiple argumentation structure, it is not a sufficient condition. Each argument must also be sufficient on dialectical grounds before you can claim that they have a multiple argu- ment structure. Whether the protagonist thinks that each by itself is a sufficient defense for the standpoint in the critical discussion is an assessment of the context and various argumentative clues (Snoeck Henkemans 1997).
Evaluating an argument begins by identifying the reasoning pattern it is based on. These com- mon reasoning patterns are conceptualised within the field of argumentation theory as ‘argumentschemes’ (Section 2). While corpus-linguistic ap- proaches have gained traction in the study of argu- mentation – partly motivated by the rise of ‘argu- ment mining’ (Stede and Schneider, 2018) – these have generally focused on aspects of argumenta- tive discourse other than argumentschemes (such as the use of rhetorical figures of speech (Har- ris and Di Marco, 2017)). The empirical study of argumentschemes would greatly benefit from quantitative data in the form of annotated text cor- pora. Existing corpora annotated with argumentschemes, however, tend to be based on restricted typologies, be of limited size, or suffer from poor validation (Section 3).
The annotators involved in the project were nine graduate students with no specific background in Linguistics or Argumentation. Three different an- notators have been assigned to the annotation of each essay. The task consisted in annotating the “support” relations between premise-claim, claim- major claim, and premise-premise with one of the middle level argumentation schemes given in Fig- ure 2 or NoArgument. For the identification of the middle level argumentschemes, annotators were provided with an heuristic procedure and asked to look for linguistic clues as a further confirma- tion for their choices. We included the label of NoArgument to account for potential cases where premises/claims in support of claims/major claims do not actually instantiate any inferential path and cannot, hence, be considered proper arguments. For example, in the following pair of clauses: “[This, therefore, makes museums a place to en- tertain in people leisure time] PREMISE . [People
How to Disagree About ArgumentSchemes Argumentation also takes place outside contexts of real dialogue. When a court gives an argument for a ruling, for ex- ample, it leaves the parties little room for questions, objections, and counter-arguments. Typically, the parties only have a chance to make their cases prior to the ruling. They may be able to appeal, but eventually a judicial decision will settle the case. The idea that legal argumentation can be depicted as a dialogue is not altogether false, but it can be very misleading. Advocates exchange arguments in relatively dynamic fashion only in proto- typical adversarial systems, but even there they would have lim- ited opportunity to counter the court’s reasoning. Legal argu- mentation does not always occur between equal parties with ample space for dialogue. The law is characterized by formality, hierarchy, authority, and a strong institutional incentive toward settlement. 12 In making reference to Sumner, for example, Jus- tice Wilson advanced a terse argument from authority that could certainly be improved. Additional premises could be added to the argument, critical questions might be tackled—but these are all possible improvements. A scheme that portrays how the court really argued in that case would have to be a simpler scheme. A reader of the court’s judgment might also use critical questions to guide her own, private assessment of the court’s reasoning; but, again, these critical questions should not figure in a depic- tion of the argument as given by the court.
Specifically, we propose guidelines for the annotation of argumentschemes for both SUPPORT and ATTACK rela- tions using the Argumentum Model of Topics framework (Rigotti and Morasso, 2010) (Section 2). These guide- lines, based on semantic principles, are scalable to other text genres as well as languages. In addition, we present a new annotation tool for argumentschemes with a user- friendly interface to support the annotation process (Sec- tion 3). The reliability of the guidelines is tested through a pilot annotation project on top of 40 microtexts, ob- taining moderate agreement (Section 4). Finally, we re- port on initial experiments on the mapping of rhetorical discourse relations and types of argumentschemes (Sec- tion 5). The annotation guidelines, the corpus and the an- notation tool are available at http : //angcl.ling.uni − potsdam.de/resources/argmicro.html.
Immediately following a <content> element, a partial semantic interpretation of that content may be given in an <entities-props> element. This element may contain <entity> tags for entities that have been introduced into the discourse in the preceding <content> element. The ID attribute of an <entity> uniquely identifies it in the discourse. Since an entity may have been introduced earlier, the annotator must determine if mentions are coreferential. In Figure 2 the first <entities-props> element shown describes the preceding <content> as introducing two discourse entities, assigned the IDs group1 and pheno1 by the annotator. The paraphrase attribute of <entity> and other elements is used to help the human reader. An <entities-props> element also may contain propositions, marked <prop>. A proposition consists of a relation name used in the definition of argumentschemes in our catalogue and the entity ID of its arguments, e.g., have_pheno(group1, pheno1). A set of six semantic relations is used: have_geno, have_pheno, have_protein, difference, similar, and cause. Propositions may be negated. Although entities and relations were manually extracted, this is a stop-gap approach until NLP tools can assist in semantic extraction.
is going to rain tonight.” Then, according to this view, the oppo- nent is committed to the following connection premise as a pre- sumption: “If Erwin is a meteorologist and Erwin says that it’s going to rain tonight, and that it’s going to rain tonight is a me- teorological statement, then it is going to rain tonight.” Given the status of this connection premise as a presumption, the op- ponent is allowed to challenge it, but she must then, upon the proponent’s request, provide a validation for this challenge, which is an argument for the thesis that challenging, and thereby retracting commitment to, this presumption is, within the con- text at hand, a legitimate move. For example, she might allege that Dutch meteorologists have a bad track record, which defeats the use of authority arguments in this special context, Erwin be- ing Dutch. In addition, after having challenged this presumptive connection premise, the opponent must, upon the proponent’s request, provide an explanation of her challenge of the connec- tion premise. For example, she might state that she reckons with the possibility that Erwin is making a mistake, possibly as a re- sult of lack of sleep, or of a neurological disorder, or she might, again, put the quality of Dutch meteorology into question.
The scheme can be applied when there is knowledge of plausible hazards, e.g. identified by hazard analysis. To illustrate the scheme in action, we use the portable insulin pump system presented in (Sommerville 2007, p.196). The system is concerned with reading the blood sugar (glucose) sensor, computing the insulin requirements and controlling the micro pump which causes the insulin to be delivered automatically. Hazard analysis of the system reveals the following eight hazards: (1) insulin overdose, (2) insulin underdose, (3) power failure due to exhausted battery, (4) electrical interference with other medical equipment, (5) parts of machine break off in body, (6) infection caused by introduction of machine, (7) poor sensor and actuator contact, and (8) Allergic reaction to materials or insulin. To demonstrate that the insulin pump is acceptable safe to operate, a plausible strategy is to apply the scheme of argument from hazard avoidance and the argument instance will be as follows.
cation of Argumentation Schemes” by Fabrizio Macagno. Manfred Kraus concentrates on answering the question of how certain arguments can be deductive, yet defeasible through his alternative solution of viewing “arguments by analogy within a greater range of argument types that derive from comparisons and similarities” (p. 172). Argumentative schemes as prototypi- cal combinations of two separate and distinct abstractions are addressed by Fabrizio Macagno through his problematization of the shortcomings found in existing classifications. Instead, he proposes a new model prioritizing the pragmatic purpose of ar- gument in dialogue. Though the shortest of these sections, the transition from evaluation of argument to argumentschemes, helps draw out important nuanced differences in the ways con- struction matters for argument effectiveness.
Compared to other approaches to argumentschemes (e.g. Pragma-Dialectical approach), the AMT Model, taking into account the level of material premises, makes it possible to explain the applicability of the inferential connection at work (maxim) to the actual context of the argumentation. The AMT Model, guaranteeing the context-boundedness of arguments, is a useful tool to analyse the argumentschemes under question: the degree of defeasibility of the standpoint depends on the implementation of the maxim in the material component. Moreover, the partition of the material component into two levels mirrors the semantic components at the basis of apparire’s semantic structure which play a major role in the building up of inferences: the endoxa corresponds to the expression of the perceiver’s knowledge about the relevance of the data in drawing the inference (see B Dox section 3) and the
Note that once we have a hybrid argumentation-based logical framework that includes new informal argumentschemes the ques- tion arises of how strong are these new schemes? A partial answer to this, given implicitly by the results of the challenge, is simply that the same priority preference as in the case for the reformulation of formal logic in terms of argumentation applies, namely that these argumentschemes are stronger than the “hypotheses scheme”, i.e. the scheme where we support a conclusion without any premises on which to rest this claim. In addition, any argument scheme that di- rectly or explicitly provides or supports factual information is stronger than arguments which do not do so. Again this new prefer- ence is cognitively valid as humans are observed to have greater trust to factual information rather than information derived through some chain of construction. It is important to note that there is no relative preference between formal and informal schemes. In this sense, within such hybrid reasoning frameworks, the formal argumentschemes lose their strictness, they become defeasible like the infor- mal ones!
Argument mining is an emerging research area which introduces new challenges in natural language processing and generation. Argument mining re- search applies to written texts, e.g. (Mochales Palau et ali.., 2009), (Kirschner et ali., 2015), for example for opinion analysis, e.g. (Villalba et al., 2012), me- diation analysis (Janier et al. 2015) or transcribed ar- gumentative dialog analysis, e.g. (Budzynska et ali., 2014), (Swanson et ali., 2015). The analysis of the NLP techniques relevant for argument mining from annotated structures is analyzed in e.g. (Peldszus et al. 2016). Annotated corpora are now available, e.g. the AIFDB dialog corpora or (Walker et al., 2012). These corpora are very useful to understand how argumentation is realized in texts, e.g. to iden- tify argumentative discourse units (ADUs), linguis- tic cues (Nguyen et al., 2015), and argumentation strategies, in a concrete way, possibly in association with abstract argumentation schemes, as shown in e.g. (Feng et al., 2011). In natural language genera- tion, argument generation started as early as (Zuck- erman et ali. 2000). Finally, reasoning aspects re- lated to argumentation analysis are developed in e.g. (Fiedler et al., 2007) and (Winterstein, 2012) from a formal semantics perspective. Abstracting over ar- guments allows to construct summaries and to in- duce customer preferences or value systems.
To the best of our knowledge there is no ap- proach that decomposes propositions into fine- grained components and uses them to determine argument structure. Our Decompositional Argu- ment Mining (DAM) identifies argument struc- ture by exploiting similarity (between C and A) and relations between the polarities of OC and OA. Our first hypothesis is then the AR between EDUs is governed by the relations between their functional components. For instance, the support relation between (2) and (9) from Table 1 is a function of the similarity between C of (9) “cook- ing; potato; burger” and A of (2) “food” and the agreement between the polarities of their respec- tive opinion expressions (i.e. the opinions “have an opportunity; interesting” and “better” are both positive). Similarly the support relation between (6) and (7) is the function of the similarity be- tween A of (6) “job” and C of (7) “job” and the agreement between the polarities of their respec- tive opinion expressions (i.e. “are losing” and “are fleeing” are both negative). The attack rela- tion between (10) and (11) is the function of the similarity between C of (10) “advertising” and A of (11) “advertising” and the contradiction be- tween the polarities of the opinion on A of (10) “should be prohibited” and the opinion on C of (11) “needs”.
Comparing the two learned argument word lists gives us interesting insights. The lists have 142 common words with 9 discourse connectives (e.g. ‘therefore’, ‘despite’), 72 content words (e.g. ‘re- sult’, ‘support’), and 61 stop words. 30 of the com- mon argument words appear in top 100 features of AltAD, but only 5 are content words: ‘conclusion’, ‘topic’, ‘analyze’, ‘show’, and ‘reason’. This shows that while the two argument word lists have a fair amount of common words, the transferable part is mostly limited to function words, e.g. discourse connectives, stop words. In contrast, 270 of the 285 unique words to AltAD are not selected for top 100 features, and most of those are popular terms in aca- demic writings, e.g. ‘research’, ‘hypothesis’, ‘vari- able’. Moreover AD’s top 100 features have 20 ar- gument words unique to the model, and 19 of those are content words, e.g. ‘believe’, ‘agree’, ‘discuss’, ‘view’. These non-transferable parts suggest that ar- gument words should be learned from appropriate seeds and development sets for best performance. 5 Related Work
animals, or that human animals can think, or that you are the thinking thing sitting in your chair, in a way that does not imply that those claims are true. (That is the point of the linguistic hypotheses I mentioned earlier.) What to do? Well, I invite you to compare the thinking-animal argument with the transplant argument. Which is more likely: That there are no animals? That no animal could ever think? That you are one of at least two intelligent beings sitting in your chair? Or that you would not, after all, go along with your transplanted cerebrum?
For the second assignment (Essay 2), all stu- dents were provided with the same three journal articles relating to uses of AI in education. They were asked to summarize, in one hundred and fifty to two hundred and fifty words, key issues relating to the use of AI in teaching and learn- ing as stated in the articles. Then, they had to write a three to five hundred word argument es- say addressing the question: Should artificial in- telligence be used in teaching and learning? Both essays were assessed using a rubric designed by the three collaborators, based on existing rubrics: SRI’s Source-Based Argument Scoring Attributes (AWC) (Gallagher et al., 2015) and Ferretti’s well known argument rubric (Ferretti et al., 2000). The four dimensions and their weights (in parenthe- ses) are shown below. Each dimension or sub- dimension was rated on a 6-point scale ([0 to 5]; see Appendix A which gives the rubric for Essay 1.)
First, we propose an improved similarity mea- sure between argument positions of two predicates that take semantically similar arguments. For ex- ample, someone possibly arrested can also surren- der him/herself, that is, objects of “arrest” and sub- jects of “surrender (oneself)” are occupied by se- mantically similar nouns. Gerber and Chai (2010) proposed analysis of English nominal predicates with this similarity to take discourse context into account. However, the similarity measure they used has drawbacks: it requires a co-reference resolver and a large number of documents. We improve their similarity measure alleviating these drawbacks by using argument position. We detail previous work on capturing discourse context in Section 2, and our proposal in Section 3.1.