<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1618">
  <Title>Towards Generating Procedural Texts: an exploration of their rhetorical and argumentative structure</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 State of the art
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.1 General typology
</SectionTitle>
      <Paragraph position="0"> Procedural texts have been studied in psycholinguistic, linguistic and didactic circles. We briefly survey various approaches here, outlining elements of interest for our objectives. null Under the heading of procedural texts, there is a quite large diversity of texts. J.M. Adam [1] notices the variability of judgments in procedural text categorization. Texts can, for example, be grouped into families according to their main objectives and style. We have, for example: a0 regulatory texts [16] that characterize expected behaviours, null a0 procedural texts [13] defined as rather liear sets of instructions, null a0 'programmatory' texts [11] which include receipes, musical scores and architectory plan, identifie how knowledge from an expert is transferred via these texts to users who are expected to follow strictly the instructions which are given.</Paragraph>
      <Paragraph position="1"> a0 instructional-prescriptive texts [21], where a quite detailed analysis of temporal and event structures is carried out, a0 injunctive texts, where [2] show the form and style used in short notices that relate e.g. fire instructions, security measures, etc., a0 advice texts [14], which include advice texts of various sorts, such as those found in large public magazines. a0 receipe texts [18], which is a domain quite well-studied, for example in language generation.</Paragraph>
      <Paragraph position="2"> Despite their goals, all these forms share common structures: specification of goals, description of lists of pre-requisites to reach the goal, and description of sequences of instructions. They also share common stylistic forms, e.g. preferences for imperative forms, and a number of typographic elements such as enumerations.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.2 Informational content
</SectionTitle>
      <Paragraph position="0"> Two works will be used as the starting point of the development of the discursive structure of procedural texts that we have elaborated (see section 3.). G.R. Bieger [7] propose a taxonomy of the contents of instructions in 9 points: inventory (objects and concepts used), description (of objects and concepts), operational (information that suggest the agent how to realize an action), spatial (spatial data about the actions), contextual, covariance (of actions, which evolve in conjunction), temporal, qualificative (manners, limits of an information), emphatic (redirects attention to another action).</Paragraph>
      <Paragraph position="1"> One of the main works in Computational Linguistics is due to [12]. She isolated 9 main structures or operations, called semantic elements from corpus analysis:  1. sequential operations: a necessary action that the agent must realize, 2. object attribute: description meant to help understand the action to realize, 3. material conditions: environment in which an action must be carried out, 4. effects: consequences of the realization of a group of operations on the world, 5. influences: explain why and how and operation must be realized, 6. co-temporal operations: expresses synchronization of operations, 7. options: optional operations, 8. preventions: describes actions to be avoided, 9. possible operations: possible operations to do in the fu null ture.</Paragraph>
      <Paragraph position="2"> She also identified 7 rhetorical relations (sequence, ccondition, elaboration, goal, result, manner, concurrence) which present those semantic elements. She positionned her work on automatic text generation : how to select the information and how to present it into a coherent text.</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
2.3 Argumentation
</SectionTitle>
      <Paragraph position="0"> Procedural texts are specific forms of discourse, satisfying constraints of economy of means, accuracy, etc. They are in general based on a specific discursive logic, made up of presuppositions, causes and consequences, goals, inductions, warnings, anaphoric networks, etc., and more psychological elements (e.g. to stimulate a user). The goal is to optimize a logical sequencing of instructions and make the user feel safe and confident with respect to the goal(s) he wants to achieve (e.g. clean an oil filter, learn how to organize a customer meeting).</Paragraph>
      <Paragraph position="1"> Procedural texts, from this point of view, can be analyzed not only just as sequences of mere instructions, but as efficient, one-way (i.e. no contradiction, no negotiation) argumentative discourses, designed to help a user to reach a goal, making the best decisions (see e.g. [3], [4]). This type of discourse contains a number of facets, which are all associated in a way to argumentation. Given a certain goal, it is also of much interest to compare or contrast the means used by different authors, possibly for different audiences.</Paragraph>
      <Paragraph position="2"> In most types of procedural texts, in particular social behavior, communication, etc. procedural discourse has two dimensions: an explicative component, constructed around rational and objective elements, and a seduction component whose goal is (1) to encourage the user, (2) to help him revise his opinions, (3) to enrich the goals and the purposes, by outlining certain properties or qualities or consequences of a certain action or prevention.</Paragraph>
      <Paragraph position="3"> Producing explanations is a rather synthetic activity whose goal is to use the elements introduced by knowledge explicitation mechanisms to induce generalizations, subsumptions, deductions, relations between objects or activities and the goals to reach. This is particularly visible in the lexical choices made and in the choice of some constructions, including typographic; procedural discourse is basically interactive: it communicates, teaches, justifies, explains, warns, forbids, stimulates, evaluates.</Paragraph>
      <Paragraph position="4"> Let us finally note a few NLG-oriented papers centered on the generation of arguments such as [22], which focus on generic types of arguments (ad absurdum, from cases, etc.) and [8].</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 A Discursive analysis of procedural texts
</SectionTitle>
    <Paragraph position="0"> The complexity of procedural texts leads us to proceed very gradually in the study of their structure before being able to produce even simple such texts. At an intermediate level, responding in natural language to a1 a2a5a4 a6 questions, using underspecified templates is an interesting step which can be evaluated.</Paragraph>
    <Paragraph position="1"> We collected a corpus of procedural texts from which we extracted the main structure via the design of a grammar.</Paragraph>
    <Paragraph position="2"> The corpus contains several types of procedural texts: receipes, maintenance manuals, medical notices, assembly, advice texts, etc. The grammar was enriched as the analysis progresses. Finally a second corpus was gathered to conduct a manual evaluation of the grammar. An annotation tool based on the grammar and on related marks has been developped in order to automatically evaluate on larger samples. The structures reported below essentially correspond to (1) the organization of the informational contents: how tasks are planned, according to goals and subgoals, and (2) to the argumentative strategies used (planning, progression of tasks, warnings, advices, evaluations, etc.). General principles of argumentative discourse are given e.g. in [9].</Paragraph>
    <Paragraph position="3"> In what follows, parentheses express optionality, + iteration, the comma is just a separator with no temporal connotation a priori, / is an or and the operator a1 indicates a preferred precedence. Each symbol corresponds to an XML-tag, allowing us to annotate procedural texts,.</Paragraph>
    <Paragraph position="4"> The top node is termed objective: objective a2 title, (summary), (warning)+, (prerequisites), (picture)+ a1 instruction sequences.</Paragraph>
    <Paragraph position="5"> summary a2 title+ Summary describes the global organization of the procedure, it may be useful when procedures are complex (summary can be a set of hyper-links, often pointing to titles).</Paragraph>
    <Paragraph position="6"> warning a2 text , (picture)+, (pre-requisites).</Paragraph>
    <Paragraph position="7"> pre-requisites a2 list of objects, instruction sequences.</Paragraph>
    <Paragraph position="8"> Pre-requisites describe all kinds of equipments needed to realize the action (e.g. the different constituents of a receipe) and preparatory actions. It may also include presuppositions on the user profile and abilities.</Paragraph>
    <Paragraph position="9"> picture describes a sequence of charts and/or schemas of various sorts. They often interact with instructions by e.g. making them more clear.</Paragraph>
    <Paragraph position="10"> Instruction sequences is structured as follows: instruction sequences a2 instseq a1 discursive connector a1 instruction sequences / instseq.</Paragraph>
    <Paragraph position="11"> instseq is then of one of four main types below: instseq a2 (goal), imperative linear sequence / (goal), optional sequence / (goal), alternative sequence / (goal), imperative co-temporal sequence.</Paragraph>
    <Paragraph position="12"> Goal may contain, besides the target itself motivations, manners, references etc.</Paragraph>
    <Paragraph position="13"> Each type of instruction sequence is defined as follows: imperative linear sequence a2 instruction a1 (temporal mark), imperative linear sequence/ instruction. (e.g. inspect carefully if the filter is clean and then open the valve) optional sequence a2 conditional expression, imperative linear sequence. (e.g. if you prefer a stronger flavor, add curry powder and cream.) alternative sequence a2 (conditional expression), (argument) imperative linear sequence, (alternativeopposition mark) a1 instseq / (conditional expression, instseq)+. (e.g. if you can locate the COM1 port, then ... otherwise, or if you wish to be more cautious or cannot locate it, dismount ....).</Paragraph>
    <Paragraph position="14"> imperative co-temporal sequence a2 imperative linear sequence a1 co-temporal mark, a1 imperative co-temporal sequence / instruction.</Paragraph>
    <Paragraph position="15"> A co-temporal sequence relates instructions which must be realized at the same time.</Paragraph>
    <Paragraph position="16"> Finally, Instruction is the lower level and has the following structure, with recursion on objective: instruction a2 (iterative expression), action, (reference)+, (goal)+, (manner)+, (motivation), (limit), (picture)+, (warning) / objective.</Paragraph>
    <Paragraph position="17"> Instructions can be complex since they may contain their own goals, warnings and pictures. If an instruction is complex it is analyzed as an objective.</Paragraph>
    <Paragraph position="18"> As an illustration, the annotation of an alternative sequence which is analysed as &amp;quot;embedded conditions&amp;quot; is given, page 4.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 Rhetorical structures
</SectionTitle>
      <Paragraph position="0"> Rhetorical structures play several roles in our approach. They first give a semantics to the discursive structure syntax given above. They also contribute to enhancing the production of well-designed responses [12]. They are also useful, as shall be seen below, to allow for the integration of procedural texts dealing with similar objectives or goals, but this is an extremely difficult task. Finally, they are used to answer questions with a higher accuracy by clearly identifying e.g. instruments (for the instrumental how), risks (via the warnings) and equipment needed (via the prerequisites).</Paragraph>
      <Paragraph position="1"> The RST [15] is a descriptive theory that specifies 23 possible relations showing how two portions of a text are linked.</Paragraph>
      <Paragraph position="2"> Previous work on procedural texts [12], [20], [19], used limited RST relations and suggested additional relations that fit procedural texts, which we use for our own analysis (limit, alternative, concurrence). We identified 16 rhetorical relations among which we introduced 6 new relations from our corpora analysis:  compulsory and the other depends on the subject will or on the situation itself (steam the fish for 10 minutes and put it 5 minutes in the oven if you want it to turn brown).</Paragraph>
      <Paragraph position="3"> a0 Reference: holds between an action and a segment which provides the localisation (in the text or in related texts via hyperlink) of the detailed procedure, (remove the reductor (see page 18)); This relation occurs also between the summary (which contains the subgoals of the global objective) and the related instruction sequences.</Paragraph>
      <Paragraph position="4"> a0 Prevention: is usually a relation between an action and its warnings. Satellites include expressions such as: be careful not to ..., and 'don't' expressions (cut the wood planks, don't draw any line!).</Paragraph>
      <Paragraph position="5"> a0 Prerequisites: occur between an action or an objective and a list of of entities and instruments or a set of ac- null tions without which the action or the objective cannot be realized (changing a car wheel : to change a wheel is not difficult, with the proviso of having in one's car the good tools : wheel brace, jack, clean rag, torch (if dark), warning triangle).</Paragraph>
      <Paragraph position="6"> a0 Concurrence: occurs between two rival co-temporal actions (to choose the best computer, run the program A on Mac, at the same time run the program B on PC. If Mac detects the component before the PC, then use Mac, otherwise use PC).</Paragraph>
      <Paragraph position="7"> a0 Development: allows for the identification of the procedure and sub-procedures in a text. It usually links the titles or the goals to the instruction sequences.</Paragraph>
      <Paragraph position="8"> The chart below summarizes, for the rhetorical relations we use, the elements in our grammar which are involved.</Paragraph>
      <Paragraph position="9">  Argumentation is found in the expression of procedural text objectives, in the expression of disjunction, alternatives, warnings, and within instructions.</Paragraph>
      <Paragraph position="10"> Let us review here the four major forms of arguments we frequently found in corpora. We outline here the main conceptual and syntactic structures that characterize each of these forms. Verb classes referred to are in general those specified in WordNet [10]: a0 'Objective or goal' arguments: are the most usual ones. They usually introduce a set of instructions or more locally an instruction. Their target is the &amp;quot;goal&amp;quot; symbol of the grammar. They basically introduce causality between a goal and the set of instructions that realize it. The abstract schemas are the following: (1) purpose connector-infinitive verb, (2) causal connectordeverbal and (3) titles.</Paragraph>
      <Paragraph position="11"> - purpose connectors : pour, afin de, etc. (to, in order to) (e.g. to remove the bearings, for lubrification of the universal joint shafts).</Paragraph>
      <Paragraph position="12"> - titles : infinitive verbs or deverbals (e.g. engine dismount). null a0 Prevention arguments: embedded either in a 'positive' or a 'negative' formulation. Their role is basically to explain and to justify. Negative formulation is easy to identify: there are prototypical expressions that introduce the arguments. Negative formulation follows the abstract schemas : (1) negative causal connector-infinitive risk verbs; (2) causal connector-modal +VP(negative polarity, infinitive); (3) negative causal mark-risk verb class VP; (4) causal connector-VP(with negation); (5) causal connector-prevention verb.</Paragraph>
      <Paragraph position="13"> - negative causal connectors: sous peine de, sinon, car sinon, sans quoi, etc. (otherwise, under the risk of) (e.g. sous peine d'attaquer la teinte du bois).</Paragraph>
      <Paragraph position="14"> - risk verb class: risquer, causer, nuire, commettre etc. (e.g. pour ne pas commettre d'erreur).</Paragraph>
      <Paragraph position="15"> - prevention verbs: 'eviter, pr'evenir, etc. (e.g. afin d''eviter que la carte se d'echausse lorsqu'on la visse au ch^assis, gloss: in order to prevent the card from skipping off its rack).</Paragraph>
      <Paragraph position="16"> - causal connector and negation: de facon `a ne pas, pour ne pas, pour que ... ne ...pas etc. (in order not to) (e.g. pour ne pas le rendre brillant, gloss: in order not to make it too bright).</Paragraph>
      <Paragraph position="17"> - modal VP: pouvoir, pouvoir-^etre (e.g. car il peut ^etre us'e pr'ematur'ement par la d'efaillance d'un autre, gloss: because it may be prematurely worn due to the failure of another component).</Paragraph>
      <Paragraph position="18"> Positive formulation marks are the same as for the first category of arguments described above. We have the following abstract schemas: (1) purpose mark-infinitive verb; (2) causal subordination mark-subordinate proposition, (3) causal mark-proposition: - purpose marks: afin de, pour (so as to, for).</Paragraph>
      <Paragraph position="19"> - causal marks: car, c'est pourquoi etc. (e.g. car ceux-ci sont les plus d'elicats).</Paragraph>
      <Paragraph position="20"> - causal subordination marks: afin que, pour que, etc.</Paragraph>
      <Paragraph position="21"> (so that, for).</Paragraph>
      <Paragraph position="22"> - the verbs encountered are usually of conservative type : conserver, maintenir, etc.</Paragraph>
      <Paragraph position="23"> To discriminate arguments using purpose marks from those of the first class, we can use a reformulation criterion. Positive prevention arguments can be reformulated to a negative form using negative causal connectors or verbal inferences (e.g. afin que la semence adh`ere bien au sol a2 car sinon la semence n'adh`erera pas au sol (gloss: in order for the crop to adhere to the ground / otherwise the crop will not adhere to the ground)).</Paragraph>
      <Paragraph position="24"> a0 Performing arguments: These arguments are less imperative than the others, they are rather advices, evaluations. The corresponding abstract schemas are: (1) causal connector-performing NP; (2) causal connector-performing verb; (3) causal connector-modalperforming verb; (4) performing proposition. - performing verbs: e.g. permettre, am'eliorer, etc.</Paragraph>
      <Paragraph position="25"> (allow, improve).</Paragraph>
      <Paragraph position="26"> - performing NP: e.g. Pour une meilleure finition; pour des raisons de performances.</Paragraph>
      <Paragraph position="27"> - performing proposition: e.g. Have small bills. It's easier to tip and to pay your fare that way.</Paragraph>
      <Paragraph position="28"> a0 Threatening arguments: These arguments have a strong impact on the user's intention to realize the instruction provided, the instruction is made compulsory by using this kind of argument. This is the injuctive form. It follows the following schema: (1) otherwise connector-consequence proposition; (2) otherwise negative expression-consequence proposition - otherwise connectors: sinon.</Paragraph>
      <Paragraph position="29"> - otherwise negative expression: si ... ne ...pas... (e.g. si vous ne le faites pas, nous le p'erimerons automatiquement apr`es trois semaines en ligne, if you do not do it, we will revoke it immediately).</Paragraph>
      <Paragraph position="30"> Besides these four main types of arguments, we found some forms of stimulation-evaluation (what you only have to do now...), and evaluation.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5 Injunctive forms
</SectionTitle>
    <Paragraph position="0"> Let us now say a few words about interesting syntactic and morphological characteristics. First, we found no sign of author positioning: there is no use of personal pronoun like 'I' or 'We'. However, the author's enunciation is made visible in French by the use of imperative and infinitive verbal forms.</Paragraph>
    <Paragraph position="1"> The most important form is certainly the injunctive discourse.</Paragraph>
    <Paragraph position="2"> It characterizes certain modalities of discourse: orders, preventions, warnings, avoidances, advices. These all have a strong volitive and deontic dimension.</Paragraph>
    <Paragraph position="3"> Injunctive discourse shows how the author of an procedural text imposes his point of view to the user. The goal is that the user knows how to execute it in a way as explicit and less ambiguous as possible. The user is assumed to have the required competences to realize it.</Paragraph>
    <Paragraph position="4"> Procedural texts are an example of a logic of action. Injunction is particularly frequent in cooking receipes, security notices, etc. Its strength is measured via the illocutionary force of the statement. In general we observed that infinitive or imperative modes are used in French. Some examples of injunction forms are given below, from which we could construct dedicated NLG templates: a0 infinitive: Mettre la poudre dans le verre (put the powder in the glass).</Paragraph>
    <Paragraph position="5"> a0 imperative: Enlevez la bague sup'erieure du bol d'articulation `a l'aide d'un burin (gloss: remove upper bushing from socket using a chisel).</Paragraph>
    <Paragraph position="6"> a0 modal verbs: Vous devez enduire la face int'erieure du pivot de p^ate d''etanch'eit'e SILICOMET (gloss: you must coat internal face of pivot with SILICOMET sealing compound).</Paragraph>
    <Paragraph position="7"> a0 preference expresssion : &amp;quot;il est conseill'e de ...&amp;quot;, &amp;quot;nous vous recommandons de ...&amp;quot;, &amp;quot;il est pr'ef'erable de ...&amp;quot; (it is advised to, we recommend that).</Paragraph>
    <Paragraph position="8"> a0 negative infinitive form: Ne pas utiliser de facon prolong'ee sans avis m'edical do not use on the long term without medical advice.</Paragraph>
    <Paragraph position="9"> In everyday life, we encounter many injunctions posted in public areas. In French, these injunctions follow in general these regular structural schemas: a0 deverbal-infinitive (e.g. d'efense d'afficher (gloss: stick no bills).</Paragraph>
    <Paragraph position="10"> a0 courtesy formula-negative infinitive (e.g. pri`ere de ne pas fumer (gloss: no smoking (please)).</Paragraph>
  </Section>
class="xml-element"></Paper>