Generating Coherent Argumentative Paragraphs

Q Should I take AI this semester? We address the problem of generating a coherent A If you want to take courses like paragraph presenting arguments for a conclusion in a Natural Language Processing or text generation system. Existing text planning techExpert Systems or Vision niques are not appropriate for this task for two main next semester, reasons: they do not explain how arguments can be it’s very advisable you take AI linked together in a linear presentation order and they do because not explain how the rhetorical function of a proposition that’s going to help you a lot So if you are interested affects its wording. in the whole field at all, I would advise you strongly We present a mechanism to generate argumentative to take AI now. paragraphs where argumentative relations constrain not only the rhetorical structure of the paragraph, but also Figure 1: An argumentative paragraph the surface form of each proposition. In our approach, a text planner relies on a set of specific argumentative relations to extract information from the knowledge reason he has to take AI. This sequence of arguments base, to map it to scalar and context dependent evaluaforms the structure of the answer. tions and to organize it into chains of arguments. The same information used for planning is also used by the In terms of wording, note that the conclusion that is surface realization component to perform lexical choice supported affects the choice of expressions at many at all the levels of the clause (connectives, main verb, levels. We have marked in italics words that are selected adverbial adjuncts, adjectives and determiners). The in part because of the argumentative function of the mechanism is implemented in the ADVISOR II system proposition in which they appear. For example, saying it using FUF, an extended functional unification formalism. is very advisable as opposed to it is OK, deciding to add strongly and selecting a lot instead of somewhat are all decisions motivated by the advisor’s goal of convincing INTRODUCTION: MOTIVATION the student to take AI. Certain types of questions require in response a statement of a conclusion and arguments to support it. In our In previous work in text generation, rhetorical domain, a question-answering system offering advice to schemas (McKeown, 1985) and RST (rhetorical strucstudents selecting classes to plan their schedule ture theory) (Mann & Thompson, 1987) have been (McKeown, 1988), should-questions, e.g., should I take proposed as operational techniques to produce coherent AI?, fall into this class. The example shown in Figure 1, paragraphs. We have found, however, that these techextracted from a corpus of naturally occurring advising niques, in their current forms, are not appropriate to adsessions that we have collected, illustrates this point. dress the task of generating argumentative paragraphs for two main reasons: first, RST relations are too generic The task we consider is that of generating similar to perform argument selection and construct coherent arargumentative paragraphs presenting an evaluation of a gument chains; second, rhetorical relations in both course and its supporting arguments. To produce such theories do not influence directly linguistic realization paragraphs, a generation system must determine which and therefore cannot determine wording decisions of the arguments to include in the paragraph, how to organize type illustrated in Figure 1. them in a structured paragraph, and how to phrase each piece of the argument. For example in Figure 1, the We present in this paper a mechanism for planning advisor selects the argument chain that AI provides and realizing argumentative paragraphs which addresses preparation for all followup courses in the field, that the these two shortcomings. In our approach, specific armore the student is interested in AI the more he should gumentative relations guide both content planning and take these followup courses and therefore, the more lexical choice within the clause. Content planning is 1Reprinted from Proceedings of COLING’92, August 1992, Nantes, France. performed using two levels of argumentative relations knowledge base. By making the intentional structure of evaluation functions and topoi (Anscombre & Ducrot, a paragraph explicit, this work follows the discourse 1983) to derive content from the underlying knowledge structure theory advanced in (Grosz & Sidner, 1986). base and organize it into coherent argumentative chains. Note also that, since in RST with planning, the structure Surface realization takes advantage of the output of the of paragraphs is dynamically derived, it is possible to paragraph structurer to perform lexical choice at all view schemas as the compilation of RST configurations levels of the clause. with some information abstracted out, as pointed out in (Mann, 1987). In the rest of the paper, we first review previous work in paragraph planning, explaining why existing We found that schemas and RST were not aptechniques cannot be used directly in the case of arpropriate for planning and generating argumentative gumentative paragraphs. We then present our approach, paragraphs because argument selection cannot be easily describing the content planner and the surface realization performed. Among the types of relations enumerated in component. RST, only two would apply to the analysis of argumentative paragraphs: evidence and thesis-antithesis. If these relations were to be composed into a paragraph structure, they would yield a chain of undistinguished PREVIOUS WORK: SCHEMAS AND RST evidence links. To determine which propositions can In previous work in text generation, two methods serve as arguments and how to order them, one needs to have emerged to generate coherent paragraph-long texts: specify precisely how arguments in the domain combine rhetorical schemas and RST (for Rhetorical Structure and relate to a conclusion. An RST type of approach Theory). cannot be used alone to plan the content of an argumentative paragraph. Schemas suffer from the same limitaSchemas (McKeown, 1985) encode conventional pattion. terns of text structure. A schema is associated with a communicative goal and describes how this goal is conIn place of a generic relation like evidence, we use ventionally satisfied. For example, the constituency specific argumentative relations called topoi (Anscombre schema is used to describe the parts of an object, and the & Ducrot, 1983), e.g., the more a class is difficult, the process schema (Paris, 1987) is used to describe a comless a student wants to take it, to perform content selecplex process. A schema describes a sequence of tion. The mechanism is detailled later in the paper. rhetorical predicates where each predicate is either a primitive communicative function, which can be fulfilled by a single proposition, or recursively another schema. Rhetorical Relations and Lexical Choice For example the primitive predicate attributive attributes While rhetorical schemas or RST have been used to a property to an object. Each predicate is assigned a determine the content of the paragraph and the ordering semantics in terms of a query to a knowledge base, of the propositions, they have not been used to determine therefore when the schema is traversed, propositions are the surface form of the clause. We have found, however, retrieved from the knowledge base as predicates are inthat in argumentative paragraphs, the rhetorical function stantiated. The output of a schema traversal is therefore of a proposition affects its wording at many levels. Cona sequence of propositions labeled by the name of the sider the following utterances, extracted from our corrhetorical predicate they instantiate. pus: While schemas label each proposition as the instan(1) It requires quite a lot of programming tiation of a predicate, RST attempts to label the relation between propositions. RST (Mann & Thompson, (2) It does involve some programming, but nothing 1987) was first introduced as a descriptive theory aiming outrageous. at enumerating possible rhetorical relations between discourse segments. RST relations include elaboration, Our contention is that either (1) or (2) can be anti-thesis, evidence and solutionhood. A relation congenerated from the same content as input, but that the nects two text spans, which can be either single proposidifference between the two forms is determined by the tions or recursively embedded rhetorical relations. One argumentative function of the clause: (1) supports the argument of the relation is marked as its ‘‘nucleus’’ conclusion that a course should not be taken because it while the others are the ‘‘satellites’’ and are all optional. requires a lot of programming, which is time consuming and therefore makes the course difficult. In contrast, (2) RST was made operational as a technique for plansupports the conclusion that the level of programming ning the structure of paragraphs in (Hovy, 1988a) and should not affect the decision whether to take the course. (Moore & Paris, 1989). The idea is to attach a communicative intent with each RST relation and to view the The amount of programming involved in a course combining of relations into paragraphs as a planning can be quantified by considering how many programprocess, decomposing a high-level intention into lowerming assignments are required and the number of prolevel goals that eventually can be mapped to single gramming projects. The question is then, given this inpropositions. The communicative goals associated with formation, how to describe this information to a student: the leaves of the structure are then used to retrieve the what level constitutes some programming, quite a lot of content of each proposition from an underlying programming or a not outrageous amount of tative relations called topoi. Topoi relations programming? are stored within the propositions as a separate feature. Our position is that the mapping from the objective 4. A paragraph structurer selects and orinformation that a course requires two programming asganizes argumentative chains into an arsignments to an evaluation that it requires some gumentative strategy. programming is only partially determined by the content. It is also and over all a rhetorical decision. It is because 5. A surface realization component maps the we want to support a certain conclusion that we view and argumentative strategy into a paragraph, evaluate an objective quantity as a lot or some. relying on a grammar which is sensitive to the argumentative information stored in the In addition, by looking back at examples (1) and (2), propositions. we find that this rhetorical decision also affects the choice of the main verb: the course requires programAn important feature of this approach is that the ming when the evaluation of the course is negative, mapping between information in the knowledge base and while it involves programming when the evaluation is the content of the propositions is performed in two positive. In (Hovy, 1988b), similar issues of lexical stages by two types of argumentative relations: evaluachoice were also addressed, but different mechanisms tion functions and topoi. We distinguish between were used to perform lexical choice and paragraph orevaluation, which is the leap from the observation of an ganization. objective fact in the knowledge base to a contextdependent scalar evaluation, and argumentative relaThis is an instance of the general problem of exprestions, which only operate on scalar evaluations, and not sibility discussed in (Meteer, 1990): RST and schemas in on knowledge-base facts. In contrast, most other text their current form do not bridge the gap between rhetoriplanners simply organize propositions directly retrieved cal relations and surface realization, and as a confrom the knowledge base. sequence, surface realization cannot take advantage of the paragraph organization to make decisions. Another important feature is that we do not use generic rhetorical relations like ‘‘anti-thesis’’ or In earlier work, we have studied the problem of ‘‘evidence’’ but instead specific argumentative relations generating certain connectives like but, although, called topoi. Because topoi are gradual inference rules, because or since (Elhadad & McKeown, 1990) and of our content planner performs a task similar to generating generating adjectives (Elhadad, 1991). In both cases, we explanations for a rule-based expert system (McKeown have found that argumentative features play an important & Swartout, 1987). But in addition to determining conrole in the selection of appropriate wording. The importent, topoi are also used to influence wording: they are tant point, is that the same argumentative features could added as annotations to the propositions generated by the be used to constrain both the choice of connectives text planner and are used by the surface realization combetween the clause and the choice of adjectives within ponent to perform lexical choice. the clause. The particular argumentative features we use are inspired from work by (Anscombre & Ducrot, 1983), In the following sections, we detail how content (Bruxelles et al, 1989) and (Bruxelles & Raccah, 1991). planning is performed and how the grammar takes adIn this paper, we show how these argumentative features vantage of the argumentative information placed in its can be generated by a paragraph structurer, and therefore input to perform lexical choice. serve as a bridge between the rhetorical function of a clause and its surface realization. CONTENT PLANNING Our system determines which content can be used to OUR APPROACH generate an answer in two stages using first evaluation In order to explain how lexical choice within the functions then topoi. clause can be affected by the rhetorical function of a proposition, we must design a text planner that annotates the propositions with information about their argumenEvaluation Functions tative function. In the ADVISOR system, the following Evaluation functions are used to map from obseractivities are performed to produce the answer to a vations of facts in the knowledge base to contextshould-type question: dependent evaluations. They are domain specific and 1. An expert-system determines whether the rely on the presence of a user-model. An evaluation is course should be taken. the rating of a knowledge-base entity on a scale. In the ADVISOR domain we have identified the relevant scales 2. An evaluation system maps observations by examining a corpus of transcripts of advising sesabout the course from the knowledge base sions. We looked at all the adjectives modifying a class into evaluations that are scalar and contextin these transcripts and classified them into semantic dependent. categories. The following classes were thus identified 3. The evaluation system links these evalua(details on this analysis are provided in (Elhadad, tions into argument chains using argumen1991)): Argumentative Relations: Topoi Once the course has been evaluated on the activated • Goodness scales, the evaluation system considers relations between • Importance the scales. We use the notion of topoi as defined in (Anscombre & Ducrot, 1983) to describe such relations. • Level Topoi are gradual inference rules of the form ‘‘the more/less X is P, the more/less Y is Q.’’ Figure 3 shows • Difficulty sample topoi used in the ADVISOR system. • Workload • Domain: programming and mathematical workload + / difficulty + workload + / time-required + Note that all of these categories are scalar and therefore define a set of dimensions along which a class can difficulty + / workload + be evaluated. The task of the evaluation component is to difficulty + / time-required + difficulty + / take rank a course on relevant scales. In the current implementation, ranking is binary so a course can be in programming + / time-required + three possible states with respect to each scale: not ranked (the scale is not active in the current context), + interest + / take + (the course is high on the scale) or (the course is low on importance + / take + the scale). In the current state of the program, there is no distinction between degrees (interesting vs. very Figure 3: Sample topoi used in ADVISOR interesting). Ranking is accomplished by using simple rules Topoi play the role of rhetorical relations in RST by which determine under which conditions objective facts explaining the relation between two propositions in a stored in the knowledge base can become convincing paragraph. But they are different in that they are very evidence for an evaluation. Figure 2 shows three evaluaspecific relations as opposed to generic relations like tion rules used in the current system. ‘‘anti-thesis’’ or ‘‘evidence’’. They can therefore be used to determine the content of the answer and the order in which arguments should be presented. If U(user.programming.-) & K(class.programming-hw > 0) then E(class.programming +) But the most important feature of topoi for our purposes is that they can be related to lexical choice in a If U(user.programming.*) & natural way. In (Bruxelles et al, 1989) and (Bruxelles & K(class.programming-hw = 0) Raccah, 1991) it is suggested that lexical items can be then E(class.programming -) defined in part by their argumentative potential. For If U(user.programming.+) & example, it is part of the definition of the verb ‘‘to reK(class.programming-hw > 3) quire’’ as used in our domain, that its subject is then E(class.programming +) evaluated on the scale of difficulty. This argumentative connotation explains the contrast between (3) and (4), in a context where both are addressed to a student who Figure 2: Sample evaluation rules enjoys programming: (3) ? At least AI requires programming, so it’s easy. U(user.programming -) checks if in the current state of the user model the system has evidence that (4) At least AI involves programming, so it’s easy. the user dislikes programming. K(class.programming-hw > 0) is a query to the The same scales are used both in topoi and in our knowledge base to determine whether the class has some lexical description. They therefore serve as a bridge beprogramming assigments. An assertion of the form tween the rhetorical structure of the paragraph and lexE(class.programming +) is a positive evaluation ical choice. of the course on the programming scale. If none of the rules shown in Figure 2 are activated, the programming scale will remain non-activated. A GRAMMAR SENSITIVE TO If the first rule is activated, a proposition attributing a ARGUMENTATIVE CONSTRAINTS number of programming assignments to the class is The output of the evaluation system is a list of chains added to the paragraph being planned. In addition, this of acceptable argumentative derivations supporting the content is annotated by an evaluation on the conclusion that a course should be taken or not. Each programming scale. The output of the evaluation sysproposition in the chain is annotated by a feature AO for tem is therefore a set of propositions annotated by Argumentative Orientation which indicates how it relates evaluations along each of the activated scales. to the surrounding propositions. Figure 4 shows a sample proposition using the notation of functional