<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-2096">
  <Title>GENERATING COHERENT ARGUMENTATIVE PARAGRAPHS</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
INTRODUCTION: MOTIVATION
</SectionTitle>
    <Paragraph position="0"> Ceaain types of questions require in response a statement of a conclusion and arguments to support it. In our domain, a question-answeting system offering advice to students selecting classes to plan their schedule (McKeown, 1988), should-questions, e.g., should I take AI?, fall into this class. The example shown in Figure I, extracted from a corpus of naturally occurring advising sessions that we have collected, illustrates this point.</Paragraph>
    <Paragraph position="1"> The task we consider is that of generating similar argumentative paragraphs presenting an evaluation of a course and its supporting arguments. To produce such paragraphs, a generation system must determine which arguments to include in the paragraph, how to organize them in a structured Immgraph, and how to phrase each piece of the argument. For example in Figure 1, the advisor selects the argument chain that AI provides preparation for all followup courses in the field, that the more the student is interested in AI the more he should take these followup courses and therefore, the more reason he has to take AI. This sequence of argaments forms the structure of the answer.</Paragraph>
    <Paragraph position="2"> In terms of wording, note that the conclusion that is Q Should I take AI this semester? A ~ you want to take courses like</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Natural Language Processing or
Expert S~stems or Vision
</SectionTitle>
      <Paragraph position="0"> next semester, it's very advisab~ you take AI now because that's going to ~ you a l o;. So ~ you are interested in the whole field at all, I would advise you strong~ to take AI now.</Paragraph>
      <Paragraph position="1">  supported affects the choice of expressions at many levels. We have marked in italics words that are selected in part because of the argumentative function of the proposition in which they appear. For example, saying it 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 the student to take AI.</Paragraph>
      <Paragraph position="2"> In previous work in text generation, rhetorical schemas (McKeown, 1985) and RST (rhetorical struclure theory) (Mann &amp; Thompson, 1987) have been proposed as operational techniques to produce coherent paragraphs. We have found, however, that these techniques, in their current forms, are not appropriate to address the task of generating argumentative paragraphs for two main reasons: first, RST relations are too genetic to perform argument selection and construct coherent argument chains; second, rhetorical relations in both theories do not influence directly linguistic realization and therefore cannot determine wording decisions of the type illustrated in Figure 1.</Paragraph>
      <Paragraph position="3"> We present in this paper a mechanism for planning and realizing argumentative t, aragraphs which addresses these two shortcomings. In our approach, specific argumentative relations guide both content planning and iexieal choice within the clause. Content planning is performed using two levels of argumentative relations evaluation functions and topoi (Anscombre &amp; Ducrot, 1983) to derive content from the underlying knowledge base and organize it into coherent argumentative chains. Surface realization takes advantage of the output of the paragraph stracturer to perform lexical choice at all levels of the clause.</Paragraph>
      <Paragraph position="4"> ACRES DE COLING-92, NANTES, 23-28 hol\]'r 1992 6 3 8 PRoc. OF COLING-92, NAm'ES, AUG. 23-28, 1992 In the rest of the paper, we first review previous work in paragraph planning, explaining why existing techaiques cannot be used directly in the case of argumentative paragraphs. We then present our approach, describing file content planner and the surface realization component.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
PREVIOUS WORK: SCHEMAS AND RS F
</SectionTitle>
    <Paragraph position="0"> In previous work in text generation, two methods have emerged to generate coherent paragraph-long texts: rhetorical schemas and RST (for Rhetorical Structure Theory).</Paragraph>
    <Paragraph position="1"> Schemas (McKeown, 1985) encode conventional patterns of text StlUCture. A schema is associated with a communicative goal and describes how this goal is conventionally satisfied. For example, the constituency schema is used to describe the parts of an object, and the process schema (Paris, 1987) is nsed to describe a complex process. A schema describes a sequence of rhetorical predicates where each predicate is either a primitive communicative function, which can be full'died by a single proposition, or rectursively another schema. For example the primitive predicate attributive attributes a property to an object. Each predicate is assigned a semantics in terms of a query to a knowledge base, therefore when the schema is traversed, prepositions are retrieved from the knowledge base as predicates are instantiated. The output of a schema traversal is therefore a sequence of propositions labeled by the name of the rhetorical predicate they instantiate.</Paragraph>
    <Paragraph position="2"> While schemas label each proposition as the instantiation of a predicate, RST attempts to label the relation between propositions. RST (Mann &amp; Thompson, 1987) was t-u'st introduced as a descriptive theory aiming at enmnerating possible rhetodcal relations between discourse segments. RST relations include elaboration, anti-thesis, evidence and solutionhood. A relation connects two text spans, which can be either single propositions or recursively embedded rhetorical relations. One urgument of the relation is marked as its &amp;quot;nucleus'&amp;quot; while the others are the &amp;quot;satellites&amp;quot; and are all optional. RST was made operational as a technique for planning the structure of paragraphs in (Hovy, 1988a) and (Moore &amp; Paris, 1989). The idea is to attach a commanicative intent with each RST relation and to view the combining of relations into paragraphs as a planning process, decomposing a high-level intention into lower-level goals that eventually can be mapped to single propositions. The communicative goals associated with the leaves of the structure are then used to retrieve the content of each proposition fi'om an underlying knowledge base. By making the intentional structure of a paragraph explicit, this work follows the discourse structure theory advanced in (Grosz &amp; Sidner, 1986).</Paragraph>
    <Paragraph position="3"> Note also that, since in RST with planning, the structure of paragraphs is dynamically derived, it is possible to view schemas as the compilation of RST configurations with some information abswacted out, as pointed out in (Mann, 1987).</Paragraph>
    <Paragraph position="4"> We found that schemas and RST were not appropriate for planning and generating argumentative paragraphs because argument selection cannot be easily performed. Among the types of relations enumerated in 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 evidence links. To determine which propositions can serve as arguments and how to order them, one needs to specify precisely how arguments in the domain combine and relate to a conclusion. An RST type of approacli cannot be used alone to plan the content of an argumentative paragraph. Schemas suffer from the same limitation. null In place of a generic relation like evidence, we use specific argumentative relations called topoi (Anscombre &amp; Ducrot, 1983), e.g., the more a class is difficult, the less a student wants to take it, to perform content selection. The mechanism is detaiUed later in the paper.</Paragraph>
    <Paragraph position="5"> Rhetorical Relations and Lexical Choice While rhettwical schemas or RST have been used to determine the content of the paragraph and the ordering of the ~opositions, they have not been used to determine the surface form of the clause. We have found, however, that in argumentative paragraphs, the rhetorical function of a proposition affects its wording at many levels. Consider the following utterances, extracted from our corpus: null  (1) It requires quite a lot of prograrmning (2) It does involve some programming, but nothing outrageous.</Paragraph>
    <Paragraph position="6">  Our contention is that either (1) or (2) can be generated from the same content as input, but that the difference between the two forms is determined by the argumentative function of the clause: (1) supports the conclusion that a course should not be taken because it requires a lot of programming, which is time consuming and therefore makes the course difficult. In contrast, (2) supports the conclusion that the level of programming should not affect the decision whether to take the course. The amount of programming involved in a course can be quantified by considering how many programming assignments are required and the number of programming projects. The question is then, given this information, how to describe this information to a student: what level constitutes some programming, quite a lot of programming or a not outrageous amount of programming? Our position is that the mapping from the objective information that a course requires two programming assignments to an evaluation that it requires some programming is only partially determined by the content. It is also and over all a rhetorical decision. It is because we want to suppea a certain conclusion that we view and Act,s DE COLING-92. NAI'CfES. 23-28 AOI\]T 1992 6 3 9 PROC. OF COLING-92. NANTES, AUO. 23-28. 1992 evaluate an objective quantity as a lot or some.</Paragraph>
    <Paragraph position="7"> In addition, by looking back at examples (1) and (2), we find that this rhetorical decision also affects the choice of the main verb: the course requires programming when the evaluation of the course is negative, while it involves programming when the evaluation is positive. In (Hovy, 1988b), similar issues of lexical choice were also addressed, but different mechanisms were used to perform lexical choice and paragraph organization. null This is an instance of the general problem of expressibility discussed in (Meteer, 1990): RST and schemas in their current form do not bridge the gap between rhetorical relations and surface realization, and as a consequence, surface realization cannot take advantage of the paragraph organization to make decisions.</Paragraph>
    <Paragraph position="8"> In earlier work, we have studied the problem of generating certain connectives like but, although, because or since (Elhadad &amp; McKeown, 1990) and of generating adjectives (Elhadad, 1991). In both cases, we have found that argumentative features play an important role in the selection of appropriate wording. The importam point, is that the same argumentative features could be used to constrain both the choice of connectives between the clause and the choice of adjectives within the clause. The particular argumentative features we use are inspired from work by (Anscombre &amp; Ducrot, 1983), (Bruxelles et al, 1989) and (Broxdles &amp; Raccah, 1991).</Paragraph>
    <Paragraph position="9"> In this paper, we show how these argumentative features can be generated by a paragraph structurer, and therefore serve as a bridge between the rhetorical function of a clause and its surface realization.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
OUR APPROACH
</SectionTitle>
    <Paragraph position="0"> In order to explain how lexical choice within the clause can he affected by the rbetorical function of a proposition, we must design a text planner that annotates the propositions with information about their argumentative function. In the ADVISOR system, the following activities are performed to produce the answer to a  should-type question: 1. An expert-system determines whether tile course should be taken.</Paragraph>
    <Paragraph position="1"> 2. An evaluation system maps observations about the course from the knowledge base into evaluations that are scalar and contextdependenL null 3. The evaluation system finks these evaluations into argument chains using argumentative relations called topoi. Topoi relations are stored within the propositions as a separate feattme.</Paragraph>
    <Paragraph position="2"> 4. A paragraph stmcturer selects and organizes argumentative chains into an argumentative strategy.</Paragraph>
    <Paragraph position="3"> 5. A surface realization component maps the  argumentative strategy into a paragraph, relying on a grammar which is sensitive to the argumentative information stored in the propositions.</Paragraph>
    <Paragraph position="4"> An hnportam feature of this approach is that the mapping between information in the knowledge base and the content of the propositions is performed in two stages by two types of argumentative relations: evaluation functions and topoi. We distinguish between evaluation, which is the leap from the observation of an objective fact in the knowledge base to a context-dependent scalar evaluation, and argumentative relations, which only operate on scalar evaluations, and not on knowledge-base facts. In contrast, most other text planners simply organize propositions directly retrieved from the knowledge base.</Paragraph>
    <Paragraph position="5"> Another important feature is that we do not use generic rhetorical relations like &amp;quot;anti-thesis&amp;quot; or &amp;quot;evidence&amp;quot; but instead specific argumentative relations called topoi. Because topoi are gradual inference rules, our content planner performs a task similar to generating explanations for a rule-based expert system (McKeown &amp; Swartout, 1987). But in addition to determining content, topoi are also used to influence wording: they are added as annotations to the propositions generated by the text planner and are used by the suiface realization component to perform lexical choice.</Paragraph>
    <Paragraph position="6"> In the following sections, we detail how content planning is performed and how the grammar takes advantage of the argumentative information placed in its input to perform lexical choice.</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
CONTENT PLANNING
</SectionTitle>
    <Paragraph position="0"> Our system determines which content can be used to generate an answer in two stages using In'st evaluation functions then topoi.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Evaluation Functions
</SectionTitle>
      <Paragraph position="0"> Evaluation functions are used to map from observations of facts in the knowledge base to context~ dependent evaluations. They are domain specific and rely on the presence of a user-model. An evaluation is the rating of a knowledge-base entity on a scale. In the ADVISOR domain we have identified the relevant scales by examining a corpus of transcripts of advising sessions. We looked at all the adjectives modifying a class in these transcripts and classified them into semantic categories. The following classes were thns identified (details on this analysis are provided in (Elhadad,  Note that all of these categories are scalar and therefore define a set of dimensions along which a class can he evaluated. The task of the evaluation component is to rank a course on relevant scales. In the current implementation, r',mking is binary so a course can be in three possible states with respect to each scale: not ranked (the scale is not ,active in the current context), + (the course is high on the scale) or - (the course is low on the scale). In the current slate of the program, there is no distiaction between degrees (interesting vs. very interesting).</Paragraph>
      <Paragraph position="1"> Rmlking is accomplished by using simple rules which determine under which conditiolts objective facts stored iu the knowledge b,ase ean become convincing evidence for an evaluation. Figure 2 shows three evaluation rules used in the current system.</Paragraph>
      <Paragraph position="2">  U (user.programming -) checks if in the current slate of the user model the system has evidence that the user dislikes programming.</Paragraph>
      <Paragraph position="3"> K(class.programming-hw &gt; 0) isaquerytothe knowledge base to determine whether the class has some programming assigmenls. An assertiou of the form E ( c las s. programaning + ) is a. positive evaluation of the course Oil tile programming scale. If none of the rules shown in Figure 2 are activated, the programming scale will remain non-activated.</Paragraph>
      <Paragraph position="4"> If the first rule is activated, n proposition attributing a number of programming assignments to the class is added to the paragraph being planned. In addition, this content is annotated by an evaluation on the programming scale. Tile output of the evaluation system is therefore a set of propositions annotated by evaluations along each of the activated scales.</Paragraph>
      <Paragraph position="5">  Once the course has been evaluated on the activated scales, the evaluation system considers relations between the scales. We use the notion of topoi as defined in (Anscombre &amp; Ducrot, 1983) to describe such relations. Topoi are gradual inference rules of the form &amp;quot;the more/less X is P, the more/less Y is Q.&amp;quot; Figure 3 shows sample topoi used in the ADVISOR system.</Paragraph>
      <Paragraph position="6">  Topoi play the role of dletoricM relations ill RST by explaining the relation between two propositions ill it paragraph. But they are different ill that they mc very specific relations ,as opposed to generic relations like &amp;quot;anti-thesis&amp;quot; or &amp;quot;evidence&amp;quot;. They can therefore he used to determine the content of the answer and the order in which arguments should be presented.</Paragraph>
      <Paragraph position="7"> But the most in~pomant feature of topoi for our puro poses is that they can he related to lexical choice ill a natural way. In (Bruxelles et al, 1989) ,'rod (Bruxelles &amp; Raccah, 1991) it is suggested that lexical items can be defined ht p,'wt by their argumentative potential. For example, it is part of the definition of the verb &amp;quot;to reo quire&amp;quot; as used in our domain, that its subject is evaluated oil the scale of difficulty. This argumentative connotation explains tile contrast between (3) and (4), ill a context where both are addressed to a student who enjoys progr, unming:  (3) ? At least Al requires prograrmning, so it's easy. (4) At least AI involves programming, so it's easy.</Paragraph>
      <Paragraph position="8">  The same scales are u~d both in topoi taxi ill our lexieal description. They therefore serve `as a bridge between the rhetodeal structure of tile p~wagraph and lexiced choice.</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="0" end_page="0" type="metho">
    <SectionTitle>
A GRAMMAR SENSITIVE TO
ARGUMENTATIVE CONSTRAINTS
</SectionTitle>
    <Paragraph position="0"> The output of the evaluation system is a list of chains of acceptable argamentative derivations supporting tile conclusion that a course should be taken or Blot. Each proposition in the chain is annotated by a featore AO lbr Argumentative Orientation which indicates how it relates to the surrounding propositions. Figure 4 shows a sample proposition using the notation of fmmtional descriptions (FDs) used in fimcfioual unification grammars. null This input represents the proposition that AI covers (among others) a set of topics in the area of theory (namely, logic), and the AO feature indicates that this proposition is used as an argument for the conclusion that AI is a difficult course, by virtue of the topos  part of the topos is shown in abbreviation). Details on the role of each fields in the AO feature and on the representation of quantification are provided in (EIhadad, 1992).</Paragraph>
    <Paragraph position="1"> Because of this AO specification, the grammar will choose appropriately realization (5) instead of (6):  (5) AI requires a lot of programming (6) AI involves some programming.</Paragraph>
    <Paragraph position="2">  The realization component is implemented in FUF, all extended functional unification grammar formalism which we have implemented (Elhadad, 1990, Elhadad, 1992). In the grammar we use, lexical choice and syntactic realization ate interleaved. For example, the choice of the verb is handled by the altemation shown in Figure 5. In this Figme, the notation alt indicates a disjunction between alternatives; ralt indicates a random alternation, and is used to indicate that the grammar does not account for the difference between the alternatives; the curly braces notation in pairs of the form ((go) value ) indicates that the go feature is not embedded in the lexical verb constituent unified with the grammar but rather is a top level feature within the clause.</Paragraph>
    <Paragraph position="3"> The fragment shown in Figure 5 specifies how the grammar can map from an input concept e-contain to a verb expressing this relation. The grammar for this relation contains two branches: in the first branch, the verbs &amp;quot;require&amp;quot; and &amp;quot;demand&amp;quot; are described as being argumentatively marked on the scale of difficulty. They can therefore be selected to project an evaluation on their subject. Note that the choice between &amp;quot;require&amp;quot; and &amp;quot;demand&amp;quot; is arbitrary, as indicated by the ralt constrnct - it is not explained by the grammar. The second branch describes the verbs &amp;quot;contain&amp;quot; and &amp;quot;involve&amp;quot; as neutral verbs, that 60 not add any connotation.</Paragraph>
    <Paragraph position="4"> When there is an argumentative connotation, the grammar specifies which participant in the clause is affected by the argumentative evaluation (for both verbs in the example, the subject of the verb is the entity that carries the evaluation). Similar lexical descriptions for adjectives are described in (Elhadad, 1991).</Paragraph>
    <Paragraph position="5"> The part of the grammar generating the syntactic structure of the clause is inspired both by systemic grammars ( (Halliday, 1985) and especially (Fawcett, 1987)) ACRES DE COLING-92, NANTES, 23-28 AOtT 1992 6 4 2 PRec. OF COLING-92, NANTES, AUG. 2.3-28, 1992 \[or the semantic features of the input and by HPSG (Pollard &amp; Sag, 1987) for the overall flow of control. It has been extended to account for the flow of argumentative information from lexical items to constituents and to the clause. For example, inserting an adjective argumentatively marked as the describer of a noun group creates an argumentative orientation feature at the level of the noun group which is then percolated to the clause in which the noun group is a participant.</Paragraph>
    <Paragraph position="6"> Finally, the clause grammar has been extended with a clause complex constituent which determines connective selection and clause combining (an extension of (McKeown &amp; Elhadad, 1991)). A clause complex is represented as an FD with features directive and subordinate (a notion similar to the RST distinction between nucleus and satellite). As discussed in (Ethadad &amp; McKeown, 1990), there are many different conneclives expressing argumentative relations. For example all of the following connectives can be used to express an evidence relation: because, since, therefore, so, as a consequence, then. The choice offered to the analyst is then: (i) to ignore the differences between such close counectives; (ii) to define a single rhetorical relation for each connective or (iii) to determine the choice of connective on other factors than the rhetorical relation alone. We adopt this later approach, and conclude that the output of the pat'agmph structurer must not determine the connectives, as is generally done by schema or RST bused planners. Instead we take advantage of how our text planner labels each proposition with information about its rhetorical function to determine which connective is most ,appropriate in combination with the other pragmatic factors discussed in (Elliadad &amp; McKeown, 1990). In this paper, we have also explained how the argumentative features needed to select connectives are broduced by the content planner.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
User Profile:
</SectionTitle>
      <Paragraph position="0"> because it requires a lot of work and it is pretty mathematical, but it is an interesting course, because it covca's many nip topics. and it offers lots of programming hws.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Implementation
</SectionTitle>
      <Paragraph position="0"> The content planner is fully implemented. In the surface realization component, the clause grammar is fully implemented with account fi)r argumentative leatares in adjectives, verbs and adverbial adjuncts. A 'large portion of the grammar covers the determiner sequence and how the choice of determiners like &amp;quot;many&amp;quot;, &amp;quot;most&amp;quot;, &amp;quot;few&amp;quot; etc. has an influence ou tile argunleatative orientation of the close. The grammar for connectives is separately implemented but not yet merged with the rest of the grammar.</Paragraph>
      <Paragraph position="1"> Tile grammar is quite large: the current version includes 580 disjunctions; it covers simple and complex clauses, interrogation, negation, a complex tense system, relative clauses, control and raising, coordination with some forms of ellipsis. We have extended FUF by adding sophisticated conlrol devices (Elhadad &amp; Robin, 1992), making it possible to ha*vile sucl, -large grammars. In particular, we are able to deal with tile non-kw~al constraints across constituent boundaries i,nposed by argumentative relations in an efficient way.</Paragraph>
      <Paragraph position="2"> Figure 6 shows the type of paragraphs obtained when all the pieces of the surface realization component are put together.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>