<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-2048">
  <Title>THE GENERATION OF HIGH-LEVEL STRUCTURE FOR EXTENDED EXPLANATIONS</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
THE GENERATION OF HIGH-LEVEL STRUCTURE
FOR EXTENDED EXPLANATIONS
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
ABSTRACT 2
</SectionTitle>
    <Paragraph position="0"> This paper analyzes the structural features of naturally-occurring extended explanations and argues that current generation methodologies are inadequate for determining high-level structure. It presents a computational model based on the hypothesis that high-level structure - composed of a unifying framework and its associated basic blocks can be determined by bottom-up processes that attempt to satisfy speaker, listener, and compositional goals, after which top-down strategies can be used to organize the material about the selected framework.</Paragraph>
  </Section>
  <Section position="3" start_page="0" end_page="276" type="metho">
    <SectionTitle>
1 INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> In this paper, we describe the structural characteristics of extended, planned 1 explanations involving complex physical devices and present a computational model for generating such explanations. Our investigation suggests that the organizational strategies currently employed for structuring short explanations are inadequate for generating the high-level structure characteristic of that found in naturally-occurring extended explanations, which typically require several pages of text. Our computational model is based on the hypothesis that text structure is not completely recursive as others have claimed (\[GS86\], \[Rei78\], \[Po186\], \[MT88\]), but rather that the high-level structure of extended explanations is deter-.</Paragraph>
    <Paragraph position="1"> mined by processes separate from those which organize text at lower levels.</Paragraph>
    <Paragraph position="2"> Section 2.1 provides a brief overview of current models for structuring text, followed by a description of the basic block, the unit of discourse on which our model is based, in Section 2.2. Section 2.3 describes the characteristics of high-level structure of extended explanations, followed by a description of our strategy for generating this structure in Section 3; a complete description is contained in \[MCM89\].</Paragraph>
  </Section>
  <Section position="4" start_page="276" end_page="276" type="metho">
    <SectionTitle>
THE BASIC BLOCK MODEL OF
EXTENDED DISCOURSE
</SectionTitle>
    <Paragraph position="0"> ill order to generate an extended explanation, a natural language system must determine tile basic content to be conveyed; the next step is to cohesively organize this material. As anyone who has had to organize large amounts of information into a coherent text can attest, there are many possible combinations of that material, some more cohesive than others. Frequently, deciding how to orgauize a large body of material is more difficult than determining what to include. Our research is concerned with the identification of a coherent unifying framework about which an extended explanation can be organized and the criteria for selecting from among several frameworks when more than one viable alternative exists.</Paragraph>
  </Section>
  <Section position="5" start_page="276" end_page="276" type="metho">
    <SectionTitle>
2.1 OTHER APPROACHES
TEXT STRUCTURE
TO
2.1.1 COMPUTATIONAL
APPROACHES
</SectionTitle>
    <Paragraph position="0"> A number of researchers (e.g., \[GSS6\], \[MT88\], \[Rei7S\], \[PolS6\]) have argued that discom-se is con&gt; posed of hierarchically structured segments and that this structure is completely recursive in nature.</Paragraph>
    <Paragraph position="1"> Two general methodologies have been applied to the structuring of explanations: schemas (\[McK85\], \[McC85\], \[earS7\]) and rhetorical structure theory (RST) (\[HMS9\], \[HovS8\], \[MPS8\], \[MSSS\]).</Paragraph>
    <Paragraph position="2"> A schema is a discourse strategy that captures a typical pattern of discourse associated with a particular discourse purpose, (e.g., providing an analogy or evidence). Schemas can be thought of as templates composed of an ordered sequence of rhetorical predicates, which &amp;quot;characterize the predicating acts a speaker may use and delineate tile structural relation between propositions in a text.&amp;quot; ~' 2From \[McK85\] page 9 These predicates are intended to capture the structural relations that hold between clauses in a text. The predicates are used recursively, capturing the structure of text at any level.</Paragraph>
    <Paragraph position="3"> RST, developed by Mann and Thompson (\[MT88\]), was originally a tool for the analysis of text. RST claims that, except for a small number of highly-stylized forms, all coherent texts have an RST ,decomposition. RST posits a small number of relations, comparable to McKeown's rhetorical predicates, that exist between segments of text. Because each relation has associated with it well-defined intended effects and conditions necessary for it to hold, RST lends itself well to a generation methodology based on a top-down, hierarchical planning formalism (\[Sac77\]). Thus, like MeKeown's rhetorical pred-icates, RST claims to account for the structure of text at any level of the discourse hierarchy.</Paragraph>
    <Paragraph position="4"> While these methods have proven to be effective for organizing short pieces of text, we maintain that they are inadequate for generating the characteristic structure of extended explanations at the level of the primary segments, which occupy the first level of the discourse hierarchy. We contend that the characteristics exhibited by the primary segments of extended explanations, to be described in the next section, cannot be captured by recursive processes. Rather, we maintain that high-level structure must be gener-ated 'by a separate, bottom-up process, after which recursive organizational strategies can be applied at lower levels.</Paragraph>
  </Section>
  <Section position="6" start_page="276" end_page="276" type="metho">
    <SectionTitle>
2.1.2 RHETORIC
</SectionTitle>
    <Paragraph position="0"> Rhetoric, the formal study of the art of good writing, provides general strategies for organizing text at a high level that are absent from the computational models. Analysis - &amp;quot;the method of explanation whereby a subject is divided into its separate component parts ''3 - is possibly the most instrumental of these strategies. There are no hard-andfast rules for determining what constitutes an appropriate analysis of a subject. As \[WA60\] observes, a subject may be classified in as many ways as it has characteristics/parts/stages/etc. However, there are three criteria which experts (\[WA60\], \[Are75\], \[Tho57\], \[Dan67\], \[KP66\]) mutually consider essential for a satisfactory organizational strategy:</Paragraph>
  </Section>
  <Section position="7" start_page="276" end_page="277" type="metho">
    <SectionTitle>
1. The
</SectionTitle>
    <Paragraph position="0"> tent (e.g. scheme should be logical; a single, consiscriterion should be used for tile analysis time, steps in a process).</Paragraph>
    <Paragraph position="1"> 2. The scheme should exhaust all of the possibilities; everything to be conveyed should be encompassed by the scheme.</Paragraph>
    <Paragraph position="2"> 3. The resultant categories should be mutually exclusive; nothing should belong to more than one. 3\[Are75\] page 107 While the type of explanation with which this paper is concerned exhibits a high-level organization reflective of these criteria, the criteria by themselves do not provide the specificity necessary for computational generation. These guidelines include no suggestions for dealing with situations in which no logical, all-inclusive framework Call bc identified, nor do they offer suggestions for selecting among several organizational schemes which meet the prescribed criteria equally well. Furthermore, the guidelines are not sufficient in-and-of themselves to account for all of the observed phenomena discussed in the following sections.</Paragraph>
    <Section position="1" start_page="276" end_page="277" type="sub_section">
      <SectionTitle>
2.2 BASIC BLOCKS
</SectionTitle>
      <Paragraph position="0"> Our model is based on a discourse unit which we have termed a basic block. A basic block consists of  two elements: 1. an organizational focus, such as a person or location, and 2. a set of concepts related to that focus.  The focus is what makes a cohesive unit of the material in the block; it is the thread cominon to all of this material, whether directly or indirectly. A basic block will be realized as a primary segment of text which occupies the first level of the discourse hierarchy. In a coherent discourse, the loci on which the basic blocks are based are themselves related, each representing a different aspect of some unifying framework. These points are demonstrated by the testimony fi'om which the basic block in Figure 1 was extracted 4.</Paragraph>
      <Paragraph position="1"> This block references a particular time frame: zero to thirty seconds of the accident at Three Mile Island 5. The remaining blocks of that testimony are similarly constructed around time frames, e.g., one to six minutes, six minutes to one hour, etc. Observed frameworks demonstrate a gamut of types: properties of the concepts (location, time), planning strategies in which events are involved (medical diagnosis), and characteristics that are not only inherent in the material but also due in part to the speaker's perception of them (significant factors). There appears to be no limit to what can constitute an acceptable framework, only that it is derived from the material itself and not from an independent device solely concerned with text structure. What may be a potential framework for one set of material may be totally inadequate for another. Note that these features are reflective of the guidelines suggested by analysis. In addition to forming a cohesive unit, basic block structure is explicitly distinguished in the following two ways. First, it is often explicitly marked. In  Now, what happened at Three Mile Island was that a feedwater transient was experienced, and by that I mean, simply, we lost feedwater to the plant momentarily. null Now, wittl loss of feedwater to the steam generator, the plant will experience a momentary pressurization above its normal pressure. This pressurization is released by a relief valve located at the top of the pressurizer. It simply opens and lets out a little bit of stean't to take care of the excess pressure that is interruption- null Then at 15 seconds into the event-keeping in mind that the valves opened maybe 5 seconds into the event-at 15 seconds the pressure started coming down because the valve had opened and cut off the pressure.</Paragraph>
      <Paragraph position="2"> The valve should have reclosed when it got back down to about 2,250 pounds; it did not reclose. The pressure proceeded to come on down. At about 30 seconds into the event, this water here started disappearing, of course, because now you are continuing to remove a very large amount of heat here, which is then coming off as secondary side steam generation, and this water will proceed to disappear if you do not start replacing it.</Paragraph>
      <Paragraph position="3"> And the auxiliary feedwater which normally comes on to make sure this does not go dry came on at about 30 seconds into the event. And at least the pumps were running. So this picture here is just the first 30 ,~econds and-or thereabouts.</Paragraph>
      <Paragraph position="4"> And this figure here is indicative of the situation from  reference to its focus: &amp;quot;So this picture here is just the first 30 seconds and-or thereabouts.&amp;quot; The subsequent block ii~ this testimony is also well marked, this time by its initial sentence (which is the last sentence of the figure): &amp;quot;And this figure here is indicative of the situation from 1 to 6 minutes.&amp;quot; The basic block structure is further distinguished by certain forms of repetition, whose use is closely tied to the basic block structure. While a brief allusion to these forms is made below, a detailed discussion is contained in \[MCM89\].</Paragraph>
    </Section>
    <Section position="2" start_page="277" end_page="277" type="sub_section">
      <SectionTitle>
2.3 CHARACTERISTICS
</SectionTitle>
      <Paragraph position="0"/>
    </Section>
  </Section>
  <Section position="8" start_page="277" end_page="278" type="metho">
    <SectionTitle>
OF COHERENT, HIGH-LEVEL
STRUCTURE
</SectionTitle>
    <Paragraph position="0"> Given tile existence of a high-level structure as evidenced by explicit markings and repetition, we must consider how such a framework is chosen. If the only consideration were the identification of a set of related foci which can partition the material to be conveyed, then any set of related concepts about which the material can be cohesively organized would suffice. Consider the motivation behind the block of Figure 1. On the surface, the events in this block all occurred within the first thirty seconds of the accident. However, it is doubtful whether the driving force behind the construction of this block was to communicate the time frame in which these events occurred; rather, what is of importance is their relative sequence in the total series of events, their cause-effect relations, and their impact on the resulting accident. One could argue that the individual events in this block represent a cause-effect chain, and hence their mutual grouping; but this chain is continued in the next block of the testimony. Thus, this argument alone cannot account for this segmentation.</Paragraph>
    <Paragraph position="1"> Apparently, other factors beyond the ability to cohesively juxtapose clauses contribute to the segmentation and the high-level framework about which it is constructed.</Paragraph>
    <Paragraph position="2"> In all of the dialogues examined, the blocks are of approximately the same size. Apparently, speakers choose organizations that tend to be well-balanced.</Paragraph>
    <Paragraph position="3"> However, balance does not seem to be the only criterion taken into consideration; if it were, we could expect to see perfectly balanced organizations in which each of the concepts to be conveyed is realized by its own basic block in addition to those well-bahmced organizations found in our text analysis whose blocks are comprised of many concepts. The size of the component basic blocks seems to be a further consideration in determining what constitutes an acceptable framework. No lengthy discourse organized about a single primary segment has been observed, nor have any in which primary segments are composed of single concepts. The &amp;quot;ideal&amp;quot; size of a basic block seems to be three or four paragraphs in length for an explanation of three pages.</Paragraph>
    <Paragraph position="4"> We contend that the high-level structure of extended explanations reflects the characteristics of an &amp;quot;ideal&amp;quot; framework in which: The basic block loci about which the material is organized reflect various aspects of the unifying framework.</Paragraph>
    <Paragraph position="5"> * Basic blocks are the same size.</Paragraph>
    <Paragraph position="6"> * The length of each basic block approximates the ideal size.</Paragraph>
    <Paragraph position="7"> Our basic block model of discourse posits that the attainment of each of these features, termed compositional goals, is instrumental in the selection of a high-level framework.</Paragraph>
    <Paragraph position="8"> However, not all observed explanations conform to the basic block model as presented thus far. For example, in explanations that are concluded by a final  summary, one would expect tile summary to emphasize the major points stressed by the speaker in the body of the text. In fact, there is a strong correlation between the segments comprising such a summary and the foci about which the basic blocks are constructed (see \[MCM89\]). Ilowever, some explanations have been examined in which material that appears in a summary has not been included in the text. Another apparent exception has been observed in explanations that exhibit a well-defined high-level structure that accounts for all of the basic blocks except for the final one, which is at best loosely related to the others.</Paragraph>
    <Paragraph position="9"> We postulate that a speaker attempts to identify a framework that is capable of coherently expressing all of the material he wishes to convey while satisfying the compositional goals equally well. However, a speaker is rarely blessed with such an ideal situation. The problem appears to be one of finding a satisfactory, rather than an optimal, unifying framework. We hypothesize that the satisfaction of some goals will be sacriffced ill favor of others so that a framework that provides the best overall solution can be achieved. This relaxation of goal constraints explains the above anomalies and the variability of size and balance observed in basic block structure.</Paragraph>
  </Section>
  <Section position="9" start_page="278" end_page="278" type="metho">
    <SectionTitle>
3 A MODEL FOR GENERATION
</SectionTitle>
    <Paragraph position="0"> We have developed a computational ntodel of discourse generation that captures the structural characteristics observed in naturally-occm:ring explanations. Our model is based on the hypothesis that the high-level structure of a discourse can be determined by bottom-up processes that attempt to satisfy speaker, listener, and compositional goals.</Paragraph>
    <Paragraph position="1"> Once this organization has been established, top-down processes are used to organize the information into basic blocks and to supplenlent that information based on the choice of framework. Only after the basic block structure has been established will detailed organization within a block and realization into text proceed. The remainder of this section describes our strategy for identifying an organizational framework and the resulting basic blocks given an initial set of concepts to be conveyed. We conclude by establishing our model within the context of a complete text generating system.</Paragraph>
    <Section position="1" start_page="278" end_page="278" type="sub_section">
      <SectionTitle>
3.1 IDENTIFICATION OF CANDI-
</SectionTitle>
      <Paragraph position="0"/>
    </Section>
  </Section>
  <Section position="10" start_page="278" end_page="279" type="metho">
    <SectionTitle>
DATE FRAMEWORKS
</SectionTitle>
    <Paragraph position="0"> We suggest that a speaker, when organizing nil extended explanation, will be faced with one of the following general situations: (r)IIe already has a well-defined organizational structure in which the material to be conveyed has already been organized.</Paragraph>
    <Paragraph position="1"> * lie has a set of goals about which he wants to structure the discourse. In this situation, the goals serve as basic block loci about which he must now attempt to partition his material.</Paragraph>
    <Paragraph position="2"> * The speaker is starting &amp;quot;from scratch&amp;quot;; he needs to find a unifying framework about which he can structure what he wants to say. It is this situation which our strategy intends to capture.</Paragraph>
    <Paragraph position="3"> Our strategy assumes that the generation process starts with some initial state of affairs and a communicative goal to be achieved. Based on this, a rough set of information to be conveyed will be iden-. tiffed. The selection of this material is not our concern; rather, we are concerned with determining a reasonable high-level structure for this information. We assume that the material to be conveyed has been categorized into at least two levels of importance: that which must be included at all costs (tile kernel), and that which would be nice to include, but due to time, style, or coherence may be left out. The process of identifying candidate frameworks begins with the kernel.</Paragraph>
    <Paragraph position="4"> The implication of the discussion in Section 2 is that a unifying framework will be some feature to which all of the kernel concepts can be related. In our model, we a.ssmne that the domain knowledge is maintained as a hierarchical network. Such a representation suggests that a node to which all of these concepts converge could serve as the unifying framework. Tile basic block loci, which represent various aspects of this framework, would then correspond to children of this node.</Paragraph>
    <Paragraph position="5"> Given a set of propositions to be conveyed, one is randonfly selected from the kernel and a trace is performed upwards from each of its arguments through the hierarchy. The traversal of the hierarchy is pertbrmed using generalization links, e.g., ISA, IS-PART, and SUBG()AL, incrementing counters associated with each node that is traversed. Additionally, CAUSE/EFFECT links can be followed when events are encountered. Property links, such as color and size, are used only if they were included as part of the kernel. Once the traversal has been completed for each element of the kernel, the nodes are ranked according to their counters; this will give some idea of tile number of concepts that converge on each.</Paragraph>
    <Paragraph position="6"> At this point, the general characteristics of basic blocks can be used to eliminate some obvious noncandidate frameworks. Since no extended explanations have been observed to be composed of a single, massive basic block, a node that has the same value as its parent can be eliminated from consideration as a candidate framework. By the same token, since no organization has been observed in which a basic block is constructed for each concept, any node that has a number of children on the same order as tlle number of concepts in the kernel carl also be eliminated from consideration. Furthermore, nodes  at which very few concepts converge can be disregarded since an acceptable framework must account for most (ideally all) of the kernel.</Paragraph>
    <Paragraph position="7"> The remaining nodes represent the potential candidates for the unifying framework. In all likelihood, none of these candidates will account for all of the kernel concepts. However, it may be possible to connect such concepts to a candidate by finding a link between these concepts and ones which are already associated with a candidate, possibly via a property they have in common. Additionally, it may be possible to include such material in a final summary or in a &amp;quot;catchall&amp;quot; block. That speakers do introduce such material in this way is evidenced by our transcript analysis. We claim that the inclusion of such material in a final summary is warranted if the speaker initially intended to include a final summary and if the amount of material is of the same order as that comprising the individual segments of the summary; for the creation of a catchall block to be viable, there must be sufficient material to construct an adequate basic block and the material must form a cohesive unit. 6</Paragraph>
    <Section position="1" start_page="279" end_page="279" type="sub_section">
      <SectionTitle>
3.2 SELECTION OF THE UNIFY-
</SectionTitle>
      <Paragraph position="0"/>
    </Section>
  </Section>
  <Section position="11" start_page="279" end_page="279" type="metho">
    <SectionTitle>
ING FRAMEWORK
</SectionTitle>
    <Paragraph position="0"> Once potential fl'ameworks have been identified, each must be evaluated according to how well it meets the criteria described in Section 2.3: * ttow thoroughly does a candidate account for the selected material? ,, How uniformly does a candidate distribute the concepts among the resulting basic blocks? ,, IIow closely do the generated blocks conform to the ideal size? Additionally, a candidate may be evaluated ~ to how well it meets the needs of a user model.</Paragraph>
    <Paragraph position="1"> The actual blocks are constructed around a candidate's immediate descendents, not the candidate itself. So, for example, if a candidate were the node workstation.s, its children - actual workstations at which things occurred - will become the basic block fool. The ideal case is one in which each child of the candidate accounts for approximately the same number of concepts. The balance each candidate achieves can be determined by comparing the counters of its children. At the same time, the candidate will be rated according to how closely its blocks conform to the ideal size. Lastly, we can rate the candidates on how well they meet the demands, if any, imposed by a user model. For example, if we know the user is familiar with the location in which the events occurred, basic blocks based on location should be given higher ratings; organizing an explanation around a framework with which the listener is familiar will fadlitate his assimilation of the information.</Paragraph>
    <Section position="1" start_page="279" end_page="279" type="sub_section">
      <SectionTitle>
3.3 GENERATION OF TEXT
</SectionTitle>
      <Paragraph position="0"> At this point, the high-level structure of the explanation has been determined. In the remainder of this paper, we will attempt to place our strategy in the context of a complete system by giving a brief overview of the remaining phases of our model.</Paragraph>
      <Paragraph position="1"> Once a fl'amework has been selected, the concepts to be conveyed can be partitioned about each of the resulting foci. At this point, there may be some concepts for which the framework does not account. The system must determine whether to include them as a separate, trailing basic block, as part of a final summary, or to eliminate them altogether (refer to Section 3.1). In addition, depending on several factors (e.g., verbosity constraints, choice of framework, etc.), those concepts of moderate import may be integrated into the framework by performing a traversal of the networks or by finding links to concepts already incorporated by the framework (as described in Sections 3.1 and 3.2). These steps complete the process of partitioning the material about the unifying framework.</Paragraph>
      <Paragraph position="2"> The next step is to determine the order of presentation for the blocks. The order in which the basic blocks are presented may be inlposed by the materim itself or by some external factor. If no overriding considerations exist, the material may present obvious &amp;quot;natural&amp;quot; choices, such as sequential ordering by time or by cause-effect. On the other hand, there may be a predetermined preference for a particular order, e.g., by their importance as determined by some external process.</Paragraph>
      <Paragraph position="3"> Once the blocks have been ordered, text generation can proceed. We intend to use an established, low-level text structuring strategy (e.g., RST or schemas) for this phase. We anticipate that additional information may be included in order to satisfy low-level discourse goals (e.g., supplying information requisite for the user's comprehension of the material to be presented); thus, we do not claim that the entire contents of a text have been determined before structuring is performed. Rather than organizing and realizing text for the blocks en mass, we have adopted a strategy in which these steps are performed for each block individually. Text for one block is presented to the user before proceeding to the organization and generation of text for the next block. This is to preclude wasted effort which may result from reorganization necessitated by a user's questions.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>