<?xml version="1.0" standalone="yes"?>
<Paper uid="W94-0308">
  <Title>Expressing Procedural Relationships in Multilingual Instructions*</Title>
  <Section position="2" start_page="0" end_page="65" type="metho">
    <SectionTitle>
Sciences Research Council Grant 319221, and the BC/DAAD
ARC Project 293.
</SectionTitle>
    <Paragraph position="0"> tAuthors are presented in alphabetical or(ler.</Paragraph>
    <Paragraph position="1"> do in this context. The complication comes when one finds, sometimes even in the same instruction mamlal, an alternate form of expression for the very same user action; such as the one shown in the following:  (2) Pull down and remove to unpack the video cable. (Apple) R ST analysts would most likely represent this with a  purpose relation, as in figure lB. Clearly, the representation formalism fails to capture the common semantics of these two examples.</Paragraph>
    <Paragraph position="2"> This problem may be only rarely encountered in a monolingual context, but is exceedingly common in a multilingual enviromnent (Delin et al., 1993). Consider the following example, taken from a trilingual, translated instruction nlanual for a step-aerobics machine:  (3) The stepping load can be altered by loosening the locking lever and changing the position of the cylinder foot. (Liftmaster) (4) Pour modifier la charge d;appui, desserrer  d6placer le pied des change the foot of the der Klemmhebel kann of the levers can durch Verschieben des ZylinderfuBes die by pvzhing of the cylinder foot the Tretbelastung ver~ndert werden.(Liftmaster) load changed be.</Paragraph>
    <Paragraph position="3"> Here we find the same information being conveyed through expressions that exhibit rather different RST  structures (see Figure 2). The English version is perhaps best characterised as a Means relation, whereas the French is a Purpose relation and the German a Circumstance. null The problem examples such as these raise for generation is twofold: (a) how do we endow our systems with the necessary flexibility to generate alternative expressions of the same basic information and (b) given this flexibility., how do we determine which of the available expressions are appropriate? The root of these problems is that, to date, research in the generation of instructions has failed to take account of the precise procedural relations that underlie the texts. In this paper, we make an excursion into the philosophical literature on Action Theory and the AI literature on plan recognition to find solidly defined procedural relations which can capture the underlying semantics of actions and their relationships in a way that will prove useful for our generation task: building multilingual instructional text generation systems. We present the results of part of an ongoing investigation of the utility of the Gene~ution and Enablement relations in contributing to our stated goal.  The philosopher Alvin Goldman identified two basic procedural relations; Generation and Enablement~ that apply between actions in a task (GoMman: 1970). These have been formalised within the Planning community and shown to be useful constructs for plan recognition (Pollack, 1986; Balkanski: 1993). In this section we review the definitions of Genelution and Enablement and describe how we have applied them in our analysis of a corpus of instructional texts in English, French and Portuguese.</Paragraph>
    <Paragraph position="4"> The Generation Relation A simple test of Generation holding between action pairs is whether it can be said that by performing one of the actions (a); the other (/3) will automatically occur (Pollack: 1986); if this is the case then it can be said that (a) gene~utes (/3). According to Goldman, Generation is a relation applying between two aet-token.~ in certain well-defined situations. For purposes of producing instructions, however, we find it necessary to view Generation as hohting between act-types: since we are dealing with actions that have not yet taken place; we are able to do this without distorting the general theory that Goldman presents (cf. Pollack. 1986). Here are some canonical examples of this relation taken from Goldman:  (6~a) John flips the switch. (cQ (b) John turns on the light. (/~) where flipping the switdl generates turning on the light and (7~a) John moves his queen to king-knight-seven. (cr) (b) John checkmates his opponent. (/3)  where moving the queen to king-knight-seven generates checkmating the opponent.</Paragraph>
    <Paragraph position="5"> Following Goldman, Generation holds in cases where: * the actions concerned are performed by the same agent (e.g., John). For reasons that will become dearer later on, we have found it necessary to qualif)' and extend this part of the definition to actions performed or perceived to be pelJolTned by the same human agent.</Paragraph>
    <Paragraph position="6"> * the relation is asymmetric: if the agent does /3 by doing a' then he does not do a' by doing/3 (e.g., the switch cannot be flipped by turning on the light). * neither action is subsequent to the other (e.g.; the light does not come on after the switch is flipped, or ~rice versa) l . A useful linguistic test for this is the invalidity of the expression '% and then i3&amp;quot;. * the actions are not co-temporal. Simply stated, if a and 3 are co-temporal, then one could validly say &amp;quot;c~&amp;quot; while also /3&amp;quot; and vice versa: for example, &amp;quot;John strummed the guitar while also wriggling his toes&amp;quot;. These actions are co-temporal, but not related by Generation. One could not validly say that John flipped on the switch while also turning on the light, or moved his queen to king-knight-seven while also checkmating his opponent.</Paragraph>
    <Paragraph position="7"> * if the agent had not done a, then he would not have done/3; * there is a (set of) condition(s) &amp;quot;i (say, George making sure that the light bulb was screwed in) such that According to Goldman, this holds even if the light does not come on for a few seconds at'ter the switch was flipped.  - the conjunction of a and 7 entails/3; - if 3&amp;quot; had not obtained: the performance of a' would not have led to/3.</Paragraph>
    <Paragraph position="8"> A GeneTution relation between action pairs can receive a variety of linguistic realisations. Consider the following semantically equivalent instructions:  (8) Checkmate your opponent by moving your queen to king-knight seven.</Paragraph>
    <Paragraph position="9"> (9) Move your queen to king-knight seven to check, rate your opponezkt.</Paragraph>
    <Paragraph position="10"> (10) I~ you want to checkmate your opponent, move your queen to king-knight seven.</Paragraph>
    <Paragraph position="11">  Two points are worthy of note here. The first is that although causation plays a key r61e in GeneTution, /3 is not in a strict sense caused by a'; rather,/3 is the action of (to use Goldman:s term) 'bringing about' the state that is the direct result of performing cr under specific conditions (7)- Referring back to example 6: John's flipping of the switch causes the event of the light coming on; John's 'bringing about' of this event can be captured by the expression 'turning on the light:.</Paragraph>
    <Paragraph position="12"> The second point to be noted is that in the domain of instructions for mechanical and electronic objects there is also a hidden agent the machine itself, as an agent acting on behalf of the user. This leads to the possibility of equivalence of expressions such as ':I deleted the file&amp;quot; and ':My coml)uterdeleted the file&amp;quot;. 'We consider these to be equally valid expressions of the generated part of a Generation relation. For example, we also count instructions such as the following (for using a cooking utensil) (11) Imprimez un mouvement de rotation &amp;quot;~ Impart a movement of ~vtation to la garniture: elle se d~.nlonte.(Darty) the plate, it itself removes.</Paragraph>
    <Paragraph position="13"> as instances of Generation. Expressions of generated actions with non-human agents are relatively uncommon in English instructions but (as will be discussed later) this is not true of all languages.</Paragraph>
    <Paragraph position="14"> The Enablement Relation Instructions also inchtde expressions such as the folh)wing: null  (12) You must restart your Macintosh to use your new software.(Mierosoft) (13) Before using your data cartridge: allow it to acch:mate to your operating environment for eight hours or fo7&amp;quot; the amount of time it has been exposed to dissimilar conditions whichever is less.(3M) (14) 2 Double-click tl~e Installer icon with the mouse button.</Paragraph>
    <Paragraph position="15"> The installer displays an introductory screen.</Paragraph>
    <Paragraph position="16"> 3 Click the OK button.(Microsoft) (15) In order to operate remote control: you m,u, st first  have noted or&amp;quot; memorised your code.(BeU) These are clearly not instances of Generation, since the execution of any member of the action pair does not 'automatically ~ result in the execution of the other. In each case: some intermediating action(s) is required. Following Goldman (1970), this type of relation between action pairs is most commonly referred to as Enablement (Pollack, 1986; Balkanski: 1993; Di Eugenio, 1992). Goldman does not provide a formal definition of this relation, but Pollack gives a useful general definition: &amp;quot; ... when a enables/3, then the agent needs to do something more than o to guarantee that/3 will be done.&amp;quot; (Pollack, 1986: page 40). A more detailed definition is provided by Balkansld: ': ... Action Ai enables action A3 if and only if there is a set of conditions: C, such that one of the conditions in C: Ci: holds as a result of the performance of Ai: and either there is a third action A~ such that A2 conditionally generates A3 under C; or C is the executability condition of A3.&amp;quot; (Balkansld: 1993, page 41).</Paragraph>
    <Paragraph position="17"> Enablement applies to all four examples above because in them: one action brings about conditions that are necessary: but not necessarily sufficient for the subsequent performance of the other. In eexample 12: the reader must restart the Macintosh in order to be in a position to be able to use the new software. There are in fact other actions necessary for the use of the software, but they are not mentioned. Acclimating the cartridge in example 13 is not sufficient to make it in use. Similarl)~ in example 14, clicking on the Installer icon only brings about a set of conditions (in this case, the appropriate menu), for the user to click the OK button. In example  7th International Generation Workshop * Kennebunkport, Maine * June 21-24, 1994 15: the reader nmst know the code in order to be able to operate the remote control: although knowing the code is not enough.</Paragraph>
    <Paragraph position="18"> Note that in some cases of Enablern.ent (e.g.: in 12 and 13)~ the intermediate action can in fact be performed in  the absence of the enabling condition; the result will simply be that the enabled action will not occur. In others however (e.g.: examples 14 and 15), the intermediate action is impossible to perform when the enabling condition does not hold.</Paragraph>
    <Paragraph position="19"> An important property of Enablement which distinguishes it from Generation is that the enabled and enabling actions can be carried out by different human agents. Goldman alludes to this in his discussion of example 6: where he introduces a necessary condition for the Generation of John turning on the light: George's screwing the bulb into the socket.</Paragraph>
    <Paragraph position="20"> As in the case of some of the Generation examples we saw before, actions in Enablement relations can be expressed as an activity carried out by a human agent or by. say: a machine or in terms of the associated state of the agent. For example, example 15 could equally have been expressed as the following:  (16) In order teJ operate remote control: you must know your code.</Paragraph>
    <Paragraph position="21"> (17) For the remote control to operate, you must first  have noted or memorised your code.</Paragraph>
    <Paragraph position="22"> A Formalisation From our discussion so far, it is clear that the relations of Gene,ution and Enablement should be identifiable aspects of plans for perfornfing procedures. We employ a plan formalism that is a simple extension of STRIPS-styled operators developed by Fikes (1971) and expanded in the NOAH system (Sacerdoti: 1977). In our representation, plans comprise: goal(s): an action (or set of actions) which motivate(s) the use of the plan; constraints: states which nmst hoht before a plan can be employed. Constraints cannot be achieved through planning; side-effects: states which arise as unplanned effects of carrying out a plan; a body: an action or action complex which executes the plan; if these are not l/rimitive: they can themselves be achieved through another plan; preconditions: an action or action complex which, when carried out, leads to conditions necessary for the successful execution of the plan; i.e: the body will be executable but its execution will not generate the goal (even if the constraints hold) unless the precondition is realisable. Preconditions can be planned for} -&amp;quot;This type of preconditions relates to Pollack's (1986) g~neration-enablin.q ~:ondition.</Paragraph>
    <Paragraph position="23"> Plans must minimally have a body and goal. Actions: in turn, comprise: constraints: these have the same properties as the constraints on plans; side-effects: these also have the same properties as the side-effects on plans; effects: states which arise from the bringing-about of the action; preconditions: like the preconditions of plans; these are either an action or an action complex and can be planned for. Unlike the preconditions of plans, however, their effects must hold before the action can take place (i.e., the action cannot result if the precondition does not hold). 3 Actions must minimally have an effect. Finally: states can have evidences. These are phenomena which signify that the state holds.</Paragraph>
    <Paragraph position="24"> As in all STRIPS-based planning formalisms, plans in our representation can have associated sub-plans. In our case: sub-plans arise through the body of a plan: the preconditions of a plan: or through the preconditions of an action.</Paragraph>
    <Paragraph position="25"> Having set out this scheme, we can now proceed to identify instances of Generation and Enablement within our representation for instructions: a generates ~ iff a' is the body of a plan e whose goal is B.</Paragraph>
    <Paragraph position="26"> a enables/) if a is a precondition of a plan e and is the goal of plan e: or if/3 is the body of e and a' is a precondition of/3.</Paragraph>
    <Paragraph position="27"> Figure 3 shows graphically the relationship between the notion of a plan and the Generation and Enablement relations.</Paragraph>
    <Section position="1" start_page="63" end_page="65" type="sub_section">
      <SectionTitle>
Generation and Enablement in
Instructions
</SectionTitle>
      <Paragraph position="0"> \'Ve have shown through examples that Generation and Enablement can be expressed in a variety of ways. In instruction generation, it is clearly important to signal the correct relationship between actions, since failure to do so could lead to situations where the user diligently carries out the action without achieving the desired effect.</Paragraph>
      <Paragraph position="1"> For example, if the following Enablement relation (18) To clean the case: disconnect the power plug.(Microsoft) is misperceived as Generation (as it might be were it to be expressed as &amp;quot;Clean the case by disconnecting the power plug&amp;quot;): the user would believe that by disconnecting the plug she would be guaranteed to achieve her goal of cleaning the case; this would only result in frustration, bewilderment and a still dirty case. We are thus :~This Lotion is closely related to Balkanski's (1993) deftnition of ezecutabiIity condition.</Paragraph>
      <Paragraph position="2">  Our approach to this problem is to look to the data: we are analysing a corpus of commercially-produced instructions for manual tasks to extract the mappings between Generation and En.ablement and linguistic form.</Paragraph>
      <Paragraph position="3"> The results of our study inform our design of a system for generating draft instructions in multiple languages front a common; language-independent: representation.</Paragraph>
      <Paragraph position="4"> ~'Ve describe some of our results in this section, based on our analyses of English: French and Portuguese instructions. null The data we report are from original texts written in the particular language being addressed; they are not the product of translation. Our method of analysis is to identify instances of Gene~utiou and Enablement from the corpus and to register its grammatical form. We report here only on data for which we all agree on the chosen procedural relationship.</Paragraph>
      <Paragraph position="5"> In analysing the data, our main concern has been to  determine: 1. What are the commonly used mappings front Generation and Enablement onto language? 2. Which forms are ambiguous between Generation and Enablement and how can we exploit this information for generation? 3. What are the factors that influence the 'appropriate null ness: of a selected form? The English Results The English corpus is made up of 451 clauses (approximately 3500 words) taken from 9 instruction manuals intended for non-expert readers. A detailed accounting of the grammatical forms used to express Generation relations is shown in table 1. Similar statistics for Enablement relations are shown in table 2.</Paragraph>
      <Paragraph position="6"> The general results of these mappings are shown graphically in figure 4, and as can be seen: there are grammatical forms which tend to be exclusively used for either Genelution or Enablement (imperative + by, sequence, and before conditions) and other forms that are common to both relations (to infinitives and :for phrases).</Paragraph>
      <Paragraph position="7"> The most common forms shown in tables 1 and '2 appear to be clear patterns of use which can be exploited in text generation, rather than just random selections or artifacts of this particular corpus.</Paragraph>
      <Paragraph position="8"> Expressing Generation in English In our corpus: GeneTution relations tend to be expressed with either the imperative + by: the imperative + to infinitive, or one of a set of :for phrases. These forms, which together constitute 80% of the Gene~zttion relations identiffed, are exemplified here in examples from manuals for a Macintosh colour monitor and for a Bell telephone answering machine: (19) If your computer system does cause interference to radio or television reception: try to correct the interference by using one or more of the following men-</Paragraph>
    </Section>
    <Section position="2" start_page="65" end_page="65" type="sub_section">
      <SectionTitle>
for Enablement Relations
</SectionTitle>
      <Paragraph position="0"> sures:(Apple) (20) To interrupt playback: rotate the VOLUME knob towards the rear, past the 'click:.(Bell) (21) See your authorized Apple dealer for information on screen save programs.(Apple) In example 19, the action of using &amp;quot;one or more of the following measures&amp;quot; generates the action of correcting the interference. They are performed by the same human agent, asymmetric, one does not occur before the other, and there exist the specified conditions -f; the same pattern is repeated in examples 20 and 21.</Paragraph>
      <Paragraph position="1"> The first question is what to make of the gaps? Why are Gene,ution relationships so infrequently expressed using sequences of clauses, and never expressed using before clauses? This clearly follows from the definition of Generation that neither action can be subsequent to the other, a relationship that is strongly implied by sequences. There are cases where sentence sequences are used to express Generation, but these are infrequent and occur in very specific syntactic situations (Vander Linden, 1993). In general; grammatical forms which typically function as expressions of sequence should be avoided when expressing Generation relationships because they tend to mislead the reader by suggesting the performance of two distinct actions.</Paragraph>
      <Paragraph position="2"> Concerning the question of identifying the contexts in which the imperative + by, the to infinitive, and the for phrase arise, it has been observed that syntactic constraints and lexical availability appear to motivate the use of the imperative + by form and the for phrasal forms (Vander Linden, 1993).</Paragraph>
      <Paragraph position="3"> Expressing Enablement in English Enablement relations tend, in 89% of the cases, toward before conditions: one of a set of sequence forms; imperative + to infinitive and other for phrasal types. Examples of  these forms are given here: (22) Quit all applications before beginning the following procedure.(Microsoft) (23) Switch offat the mains, disconnect the power adaptor and remove the cassette. Wipe the tape heads and pinch roller gently with a cotton-wool bud dipped in tape head cleaning fluid or methylated spirits.(Bell) (24) Return to seat to place calls.(Airfbue) (25) PLEASE SAVE THESE INSTRUCTIONS FOR</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>