Proposing Attachment Points in Argument Graphs for New Arguments Expressed in Natural Language Helmut HORACEK German Research Center for Artificial Intelligence (DFKI) Stuhlsatzenhausweg 3, D-66123 Saarbrücken, Germany email: helmut.horacek@dfki.de Abstract. Dealing with arguments in a natural debate can profit This paper is organized as follows. We first motivate the need for from formal representation techniques – in order to facilitate the supporting the incremental update of formal representations of a inspection of their role and interrelations and even reasoning natural debate. Then we outline a method that aims at finding support to determine the state of sets of arguments. An important plausible attachment points for a new argument in an argument issue in building such representations is the intended and accu- graph that represents the current state of a debate. We illustrate rate attachment of newly raised arguments in the context of the previous debate. In this paper, we propose a method for deter- this idea by a walk-through of a moderately complex example. mining likely attachment points for newly raised arguments, Finally, we discuss the state of affairs and future prospects. based on discourse concepts, such as given and new information in the natural language formulation of arguments. In the long run, the approach is likely to make the incremental building of 2 METHODOLOGY formal representations easier and it may even lead to an increase In argumentation frameworks, the proper content of an argument of the accuracy of the formal representation in some cases. which may have been made originally in NL is not accessible because its content is abstracted away in some atomic propo- 1 INTRODUCTION sition p, or even in an argument identifier. Hence, given a new argument p in a debate, the task is to refer it to the appropriate p i Building semi-formal and formal graphical representations of a in the given state of an argument graph, according to the inten- natural debate can be supported by tools, such as ARAUCARIA tion of the person who raised the argument (p → pi for a support [12] and its successors, dealt with in tutoring, e.g., for legal rea- or p → ¬pi for an attack). This may be associated with some cog- soning [1, 11] and for documenting the state of a public debate nitive load, in particular for argument graphs of increasing size. [17, 18]. One issue in using tools is the incremental update of However, because the content of the arguments is completely some intermediate state of an argument graph by a new argument. abstracted away, there is no formal support possible from the In most cases, the user of such a system is expected to pick the side of a system. appropriate attachment point for inserting the new argument, However, if arguments are available in NL form, these for- possibly supported by graphical navigation and inspection faci- mulations typically contain a number of linguistic clues, even lities, but hardly by content- and text-related concepts. This without world knowledge, on the basis of which relations bet- task may be associated with considerable burden on the side of ween arguments appear plausible. This is because arguments, the user, partially because of the size of a considerably grown though expressed in a concise and abbreviated form [2, 13] in argument graph, but also because a suitable position may not comparison to [14] are not raised in isolation, but in an ongoing always be conceptually clear. Consequently, support in seeking debate, where the person who raised the argument normally likely attachment points for the new argument may be quite wel- intends the audience to understand the underlying structure. In come, also in view of possible constraints on the new argument order to support this understanding, contributions in an argu- and its relations to previously raised ones (see [8, 9]). mentative debate are made in a coherent way as in any other NL In this paper, we propose a method for determining plausible text or conversation, by making use of cohesive measures. They attachment points for newly raised arguments, based on discour- include references, but also discourse markers, [7] distinguishes se concepts, such as given and new information in the natural between additive, adversative, causal, and temporal forms. language (NL) formulation of arguments, but also interpreting When a new argument is raised, this is typically done in a some typical argumentative roles expressed in NL. The content form where old (given) and new information is combined in the of such arguments is partially interpreted and maintained in con- NL formulation. Several linguistic theories (e.g., [10]) orient a text which yields evidence for the relations among the arguments certain perspective on coherence, prominently the focus of as well as for their argumentative role - such as indicating attention, on the role of given and new information, to support whether they can be a support or an attack. We consider this reference resolution and the maintenance of discourse objects in approach a first step towards building a semantic representation a discourse history representation, as first empirically analyzed of NL arguments, focused on their argumentative role, which in [6]. Similarly, the new information in an argument provides aims at more rigour in the representation of arguments so that in its proper contribution to the debate, while the old information the long run more formal reasoning services are enabled. is intended to provide evidence about where the new argument is 22 18th Workshop on Computational Models of Natural Argument Floris Bex, Floriana Grasso, Nancy Green (eds) 16th July 2017, London, UK Function Description Procedure Propose-Attachment-Points (ArgTree,NewArg) Attachments ← empty Discourse functions forall Arg ∈ ArgTree do Given Part of an argument covered by a previous one if Author (NewArg) New New part, the complement of Given then Common ← Compare (Arg,Content(NewArg)) Access function to components of an argument else Common ← Compare (Arg,NewArg) endif MainC The main claim of an argument, without restrictions if Common MainE The main entity of an argument then endif SubE A subordinate entity of an argument endfor MainA The main assertion of an argument, without main entity return Author The witness or expert who made the argument Content The content of the argument made by someone else Procedure Compare(Arg1, Arg2) Argcommon ← empty Assessment functions if Para(New(Arg1), Arg2) then return Arg2 endif Evalu+/- The argument expresses a good or bad assessment forall ArgPart ∈ {MainC(New(Arg1)),MainE(MainC(New(Arg1))), Change↑↓ The argument expresses positive or negative change MainA(MainC(New(Arg1))), SubE(MainC(New(Arg1)))} do If Para(ArgPart,MainC(Arg2)) Linguistic "bridging functions" then endif Para One argument (portion) is a paraphrase of another one If Para(ArgPart,MainE(MainC(Arg2))) Infer One argument (portion) is inferable from another one then endif If Para(ArgPart,MainA(MainC(Arg2))) then endif Table 1. The functions to access and link substructures of arguments If Para(ArgPart,SubE(MainC(Arg2))) then endif related to in the present debate. We understand an argument as a return Argcommon endfor statement made by a participant in a debate, which may have the usual form of a support or an attack, but may also expand on the Procedure Update (Attachment,Reference,NewArg) description of an argument already made; it may elaborate the content details of another argument, or add an explicit assess- ment. As shown in [9], such statements are better combined with Given(NewArg) ← the arguments they relate to rather than forming new arguments. New(NewArg) ← Some arguments may have a special form which expresses If New(NewArg) polarity, such as evaluatives " is good/bad", and assertions then endif about changes " is in- or decreasing". Asserting evaluatives, also in combination with assertions about changes, is con- sistent with only one role of an argument. For example, if is Figure 1. The procedure for proposing attachment points for arguments claimed to be positive or desirable, and is claimed to be true or increasing, then an argument " leads to or causes " can their substructure and to link them to components of other argu- only function as a support for , but never as an attack on it. ments are listed. An argument is conceived as a proposition, While these pieces of information can readily be applied to typically a state or an event, with entities involved, possibly suggesting or restricting updates to the underlying argument with restrictions; it may be within the scope of a mental attitude, graph, actually obtaining this information is rarely easy. Refer- such as an expert opinion. Propositions are built with semantic ence to given information may be indirect, implicit or encapsul- case role fillers, abstracting away, for instance, from passive ated in some paraphrase, so that formally recognizing relations voice and function verb constructs. To start with, there are two may be difficult in some cases. Moreover, work on discourse complementing discourse functions, Given and New. The distinc- parsers, such as [5] are of limited help here, since they expect a tion between Given and New information supports the identifi- well-structured prose text in their analysis and not an cation of potential links between arguments, including evidence incrementally developing partial debate with many focus shifts; for where some piece of information has first been introduced. however, discourse parsers may help in interpreting rhetorical Both Given and New are structured propositions corresponding relations between adjacent propositions. to the entire argument, with annotations indicating the active In order to support the recognition of cases as described parts, in a complementary way. In Given, the portions of an above, we define substructures of NL arguments in an abstract argument whose content is covered by some previous argument form, to be used by an interpretation procedure. In Table 1, the are marked, that is, this part serves as a reference to express functions needed to access components of arguments, to check linking between arguments. New is then what is not covered by 18th Workshop on Computational Models of Natural Argument 23 Floris Bex, Floriana Grasso, Nancy Green (eds) 16th July 2017, London, UK Given. Then some functions are defined which provide access to an argument's substructure that are potential candidates for iden- 1. Every householder should pay tax for the garbage which the tifying relations between arguments via their components. Com- householder throws away. ponents of an argument include its main claim (MainC), disre- 2. No householder should pay tax for the garbage which the garding restrictions which may be expressed in subordinate householder throws away. clauses. MainC can be decomposed into its main entity (MainE), 3. Paying tax for garbage increases recycling. typically an agent, and the assertion ascribed to it (MainA), that 4. Recycling more is good. is (MainC without MainE) and other subordinate entities (SubE). 5. Paying tax for garbage is unfair. If an argument is embedded in a propositional attitude, that is, 6. Every householder should be charged equally. there is one external person referred to to which the argument 7. Every householder who takes benefits does not recycle. content is attributed, the functions Author and Content are used 8. Every householder who does not take benefits pays for every to pull out the reference to that person and to the embedded argu- householder who does take benefits. ment. Moreover, arguments may take the specific form of " 9. Professor Resicke says that recycling reduces the need for is positive/negative" or " increases/decreases". This may new garbage dumps. have consequences on the role of arguments related to such argu- 10. A reduction of the need for new garbage dumps is good. ments. Functions Evalu+, Evalu-, Change↑ and Change ↓ check 11. Professor Resicke is not objective. for these forms. Finally, there are two functions which are in- 12. Professor Resicke owns a recycling company. tended to bridge variations in wording and reference to some 13. A person who owns a recycling company earns money from piece of information: Para is a two-place function which yields recycling. true if two assertions or entities used as parameters are para- 14. Supermarkets create garbage. phrases of one another, that is, they are semantically equivalent; 15. Supermarkets should pay tax. this may be verified, for instance, by systems checking for logi- 16. Supermarkets pass the taxes for the garbage to the consumer. cal entailment, such as [3], in both directions, or by paraphrase checking systems. Infer is intended to cover more general cases, but since concrete criteria which inferences are adequate to be Figure 2. The sequence of arguments in Wyner's running example embedded in discourse references is a widely open question, we do not investigate this case here further, 3 A RUNNING EXAMPLE The procedure searching for proposed attachment points is given in pseudo-code in Figure 1. It consists of the main proce- In this section, we introduce the running example used by Wyner dure Propose-Attachment-Points, with a subprocedure Compare. and his co-authors [17, 18] to show how state-of-the-art natural This yields an ordered list of hypotheses, from which the person language processing methods can be applied to build abstracted who produced the argument can choose. Once an attachment representations to be used by argumentation frameworks [4] point is confirmed, the book keeping procedure Update is called. under some simplifications – the restriction to controled Eng- As an initialization, the point of debate, which constitutes the lish, and user cooperation to specify the role and scope of newly argument tree at the beginning, is completely marked as New. introduced arguments, from the perspective of how adequately the Then the procedure Propose-Attachment-Ponits is invoked for assertions to be ultimately incorporated into an argumentation each newly raised argument NewArg. NewArg is compared with framework are categorized and attached to the incrementally con- the genuine part (New) of all previously raised arguments in structed argument graph (see Figure 2 for the list of assertions, ArgTree, including the point of debate. The restriction to the and Figure 3 for the argument graph built out of them). In the New part of previous arguments to be compared is motivated by argument graph, node labels refer to argument numbers in Figure the preference to arguments where content has been introduced 2, full arrows represent support links, dashed arrows represent (is New) over those where it merely refers to (Given). A distinc- attack links. tion is made as to whether the NewArg expressed a propositional When a user raises a new argument, he also specifies the attitude (Author) or not, to select the proper content for the com- argument to which the new one is related and the category of that parison. The procedure Compare carries out the comparison, for relation. Since humans generally tend to be sloppy in their for- the whole argument and for its components MainC, MainE, mulations, express pieces of information in limited degrees of MainA, and its SubEs, checking whether any of these parts are explicitness, especially in inference-rich discourse, and may find semantically equivalent in some combination. Successful com- it hard to precisely identify semantic relations in a given con- parisons are collected and at the end sorted by the degree of com- text, we can expect a number of problems associated with user monality. There may be one, several or no candidates. Once the specifications of this kind. intended attachment point is picked by the person who raised the Later, a transformation method has been proposed [8, 9] argument (which may be one of the attachment points proposed which leads to a variation of this argument graph (see Figure 4), or another one), the procedure Update inserts NewArg in the argu- that attempts to avoid ontological discrepancies and dupli- ment tree and assigns its components the states of Given or New. cations, to uncover implicit information and to choose relations In addition, Update proposes the conflation of two arguments if between assertions that are as conceptually accurate as possible. the new argument expresses only an assessment. No attempt is In two cases, two nodes are combined into a single one (3 and 4, made yet to check consistency of argumentative roles. as well as 5 and 6), In addition, some changes in the arguments 24 18th Workshop on Computational Models of Natural Argument Floris Bex, Floriana Grasso, Nancy Green (eds) 16th July 2017, London, UK 2 1 2 1 5/6 15 5 15 3/4 4 3 10 9 14 16a 10 14 16 8 7 9a 10 11 7 6 8 11 12 13 12 13 Figure 3: The original argument graph by Wyner [17, 18] Figure 4: The revised argument graph according to [8, 9] are made, thereby introducing new arguments, such as 9a, which argument (3) is not validated, since Paying tax for garbage is represents the argumentation scheme relying on expert opinion marked as Given in this argument. and changing arguments, such as 16a, which can be paraphrased 6. Every householder should be charged equally. by "customers pay for the garbage" (a logical consequence of The best match for this argument is between its main entity argument 16), Finally, the structure of the graph may be changed and the main entity of the point of debate. The contrast in some parts, to obtain semantically more accurate relations, between unfair and should be charged equally as a reason for a such as direct rather than indirect attacks. In the following, we link would require too much inference capabilities. refer to each version of the argument graph and discuss relations, 7. Every householder who takes benefits does not recycle. discrepancies and support for obtaining ontologically more For this argument, there are two possible attachments: (1) accurate representations where appropriate Every householder with the main entity of the debate, and (2) the main assertion does not recycle with the New part of ar- gument (3), as a weakly related paraphrase, at least. Recogni- 4 WALKING THROUGH THE EXAMPLE zing some sense of the apparently intended rule, "it is unfair, because some householders do not recycle" is out of reach. In this section, we illustrate the envisioned effects of our 8 . Every householder who does not take benefits pays for every method, exemplified by Wyner's running example. We sketch householder who does take benefits. the incremental building of a new argument graph, geared by the This argument has a stronger relation to the previous one (7) proposals for attachment points at every newly raised argument. via Every householder who does (not) take benefits than to The first two assertions, the points of the debate, we treat as a the main point of debate, which is merely via Every house- union since one is the negation of the other. The subsequently holder. Hence, a supporting relation between arguments (7) raised arguments are dealt with as follows: and (8) is suggested as in Figure 4 rather than these argu- 3. Paying tax for garbage increases recycling. ments being sister nodes as in Figure 3. In this argument, its main entity (Paying tax for garbage) is 9 . Professor Resicke says that recycling reduces the need for considered a paraphrase of the claim's main assertion (Should new garbage dumps. pay tax for the garbage, disregarding modality). Increases In the content of this embedded argument, the main entity, recycling then becomes the New part of this argument. recycling, gives rise to two potential attachment points - 4. Recycling more is good. the main entity of argument (3) and, to a less direct extent, The main entity of this argument (Recycling more) is assess- does not recycle in argument (7). ed as a paraphrase of the New part of the previous argument 10. A reduction of the need for new garbage dumps is good. (increases recycling), the only match. Since the argument In relation to the previous argument, this argument is structu- conforms to the Evalu+ pattern, its embedding in the previ- rally almost identical to argument (4) in connection with ous argument is proposed to yield a support of the positive argument (3): it matches the New part of argument (9), only variant of the point of dabate. This conflation of arguments adding an evaluation. Similarly as with the previous pair of corresponds to the version of the argument graph in Figure 4. arguments, their conflation in the representation is propos- 5. Paying tax for garbage is unfair. ed, which is even more compact than the version in Figure 4. The main entity of this argument (Paying tax for garbage) is 11. Professor Resicke is not objective. a paraphrase of the main assertion of the point of debate, Through reference to a previously introduced person, the pro- yielding is unfair as the New part. Note that a reference to posed attachment point prominenty stands out here. 18th Workshop on Computational Models of Natural Argument 25 Floris Bex, Floriana Grasso, Nancy Green (eds) 16th July 2017, London, UK 12. Professor Resicke owns a recycling company. ence. In addition, precisely elaborating the access functions to The main entity in this argument, Professor Resicke can components of an argument, such as main entity, are to be done. refer to its previous references in arguments (9) and (11); the The success of the method will largely depend on how the stronger connection between is not objective and owns a predicates Para in Infer can be fleshed out so that a reasonable recycling company is not recognizible. share of references can be established with acceptable effort. 13. A person who owns a recycling company earns money from In the future, we intend to address these issues of formali- recycling. zation, as well as to investigate the application to larger corpora For this argument, the main entity matches the New part of of argumentative texts. A useful extension of the method is a the previous argument. It is therefore proposed to expand on more fine-grained elaboration of preferences on attachment it (Figure 4), rather than to become its sister node (Figure 3) points, geared by the focus of attention, similar to models of 14. Supermarkets create garbage. ordinary discourse. Moreover, distinctions among modalities For this assertion, only the embedded entity garbage matches may yield evidence for preferred structures among arguments. with components of other arguments; but only in its role of Finally, knowledge about argumentation structures [15, 16], an embedded concept of the point of debate, this is consider- such as argumentation schemes may be exploited. ed successful. In argument (3), garbage belongs to the Given part so that this argument is not a suitable attachment point. 15. Supermarkets should pay tax. REFERENCES This argument can be attached to the previous one since they [1] K. Ashley, N. Pinkwart, C. Lynch, and V. Aleven. Learning by Dia- share the main entity, Supermarkets. In addition, the point of gramming Supreme Court Oral Arguments. In Proceedings of the 11th International Conference on Artificial Intelligence and Law, 271- debate also qualifies as an attachment point, because it shares 275, New York (NY), ACM Press, (2007). the main assertion should pay tax with the new argument. [2] R. Cohen. Analyzing the Structure of Argumentative Discourse. The preference among these two, including the relation Computational Linguistics 13(1-2), 11-24, (1987). [3] I. Dagan, O. Glickman and B. Magnini. The PASCAL Recognosing between arguments (14) and (15) is hard to assess. None of Textual Entailment Challenge. In J. Quiñonero-Candela, I. Dagan, B. the approaches addressing the running example proposed Magnini and F. d'Alché-Buc (Eds.), Machine Leaning Challenges. criteria for adjoining argument (15) in between argument (14) Lecture Notes in Computer Science, Vol. 3944, pp. 177-190, (2006). [4] P. M. Dung. On the Acceptability of Arguments and its Fundamental and the point of debate, as the human did. Role in Nonmonotonic Reasoning, Logic Programming and N-person 16. Supermarkets pass the taxes for the garbage to the consumer. Games'. Artificial Intelligence 77(2), 321–358, (1995). [5] V. W. Feng and G. Hirst. Text-based Discourse Parsing with Rich This argument has a stronger relation to the previous argu- Linguistic Features. Proceedings of the 50th Annual Meeting of the ment via its main entity Supermarkets, than to the point of Association for Computational Linguistics (ACL'12), pp 60-68, (2012). debate by its embedded entity the taxes for the garbage. Since [6] B. Grosz and C. Sidner. Attention, Intention and the Structure of Discourse. Computational Linguistics 12(3), 175-204, (1986). argument (15) is previously established in a place so that [7] M. A. K. Halliday and R. Hasan. Cohesion in English. Longman, argument (14) depends on it, Supermarkets is only accessible London and New York, (1976). for reference in argument (15), where it is New. [8] H. Horacek, Towards Bridging Between Natural Language and Logic-Based Representations of Natural Arguments. In Proc. of the CMNA 12, Workshop at 20th European Conference on Artificial Intelligence, pp. 21-25, (2012). 6 CONCLUSION AND DISCUSSION [9] H. Horacek, Gathering and Using Linguistic Evidence for Deter- mining the Scope and Role of Natural Language Arguments. Joint In this paper, we have described a method that aims at identify- CMNA 14 and 1st International Workshop on Methodologies for Research on Legal Argumentation, Workshop at JURIX, (2014). ing and proposing attachment points for newly raised NL argu- [10] K. Lamprecht. Information Structure and Sentence Form. Cambridge, ments in a given state of an argument graph. The method makes Cambridge University Press, (1994). use of some linguistic concepts to select components in the NL [11] N. Pinkwart, C. Lynch, K. Ashley and V. Aleven, Re-evaluating LARGO in the Classroom: Are Diagrams Better Than Text for formulation of arguments and to interpret their argumentative Teaching Argumentation Skills? Lecture Notes in Computer Science, role, prominently Given and New information. As it has been Volume 5091, Intelligent Tutoring Systems, pp. 90-100, (2008). [12] G. Roew, F. Macagno, C. Reed, D. Walton. Araucaria as a Tool for demonstrated by a small example, the choices made are mostly Diagramming Arguments in Teaching and Studying Philosophy, reasonable, but not always as (possibly) intended by the human. Teaching Philosophy 29(2), (2008). In addition, some relations between arguments have been found [13] J. Sadock. Modus Brevis: The Truncated Argument. In Papers from the 13th Regional Meeting, Chicago Linguistic Society, 545-554, (1977). to be superior to those chosen by the human in some cases. We [14] S. Toulmin, The Use of Argument, (1958). believe that our method is not only useful to support navigation [15] D. Walton. Argumentation Schemes for Presumptive Reasoning. Erl- for the incremental building of argument graphs; it can also baum, Mahwah, N.J. (1996). [16] D. Walton, C. Reed, and F. Macagno. Argumentation Schemes. Cam- support the categorization and placement of newly raised argu- bridge Univ. Press, (2008). ments, and, in the long run, enable logical reasoning services. [17| A. Wyner, T. van Engers and A. Hunter. Working on the Argument Pipeline: Through Flow Issues between Natural Language Argument, Admittedly, the degree of elaboration is still on some kind of Instantiated Arguments, and Argumentation Frameworks, In Proc. of anecdotal level so far. The interpretation of some of the predi- the ECAI 2010 Workshop on Computational Models of Natural cates used in the formalization need to be fleshed out more Argument, Lisbon, Portugal, August (2010). [18] A. Wyner, T. van Engers, and K. Bahreini. From Policy-Making Sta- precisely. In particular, this comprises the Para predicate, that tements to First-order Logic'. In K. Normann Andersen, E. Frances- determines whether two assertions are paraphrases of one coni, A. Gronlund, and T. M. van Engers, editors, EGOVIS, volume another or somehow close to it to justify an argumentative refer- 6267 of Lecture Notes in Computer Science, 47–61, Springer, (2010). 26 18th Workshop on Computational Models of Natural Argument Floris Bex, Floriana Grasso, Nancy Green (eds) 16th July 2017, London, UK