<?xml version="1.0" standalone="yes"?>
<Paper uid="A92-1009">
  <Title>Automatic Generation of On-Line Documentation in the IDAS Project*</Title>
  <Section position="3" start_page="0" end_page="65" type="intro">
    <SectionTitle>
2 Architecture
</SectionTitle>
    <Paragraph position="0"> A simplified version of IDAS's architecture is shown in  ATE software is intercepted by the Listener, 1 which detects mentions of ATE components and extracts information about the user's task (e.g., what test program he is running). Mentioned components are added to the discourse in-focus list, and are also made mousable in  the output window; if the user clicks on one, he invokes IDAS and the Listener creates an initial query about that component, i.e., an initial point in question space (Section 2.1). The question space point is given to IDAS's NL generation system, which generates a response using three modules: content determination, which picks relevant information out of the knowledge base to communicate to the user; text planning, which converts this information into an expression in SPL, the ISI Sentence Planning Language \[Kasper, 1989\]; and surface realization, which produces a surface form, i.e., an annotated text string.</Paragraph>
    <Paragraph position="1"> The annotations consist of text-formatting commands (e.g., Begin-New-Line) and hypertext specifications.</Paragraph>
    <Paragraph position="2"> The annotated text string is given to the Hypertert Inlcrface system, which presents it to the user in a hypertext window; this window also includes buttons for hyperschema follow-up questions (Section 4.1). If the user clicks on a mouse-sensitive word or a button, the point in question space that corresponds to this query is passed to the NL generation system, and the process iterates.</Paragraph>
    <Section position="1" start_page="64" end_page="65" type="sub_section">
      <SectionTitle>
2.1 Question Space
</SectionTitle>
      <Paragraph position="0"> Question space is the set of queries that can be given to IDAS's NL generation system; IDAS's hypertext system can be viewed as a tool that enables a user to move around question space until he finds a point that gives him the information he is looking for. A point in question-space is a tuple with five components:  ally physical ATE components, but can in some cases be actions or other knowledge-base entities. * Task: The user's task, e.g., Operations or ReplacePart. null * User-Expertise: The user's expertise level, e.g., Novice or Skilled.</Paragraph>
      <Paragraph position="1"> * Discourse-in.focus: The set of in-focus objects for  referring expression generation \[Grosz and Sidner, 1986\].</Paragraph>
      <Paragraph position="2"> For example, the question space point (What-is-it, DC-Power-Supply-23, Operations, Skilled, {VXI-Chassis-36, DC-Power-Supply-23}) represents the query &amp;quot;What is the DC Power Supply&amp;quot; when asked by a user of Skilled expertise who is engaged in an Operations task with the discourse context containing the objects VXI.Chassis-36 and DC-Power-Supply-23.</Paragraph>
      <Paragraph position="3"> The NL Generation component would in this case produce the response ~How-do-l-perform-the-task is interpreted as How-do-Iuse-it for Operations tasks, How-do-I-replace-it for Replace-Part tasks, etc.</Paragraph>
      <Paragraph position="4">  &amp;quot;The DC power supply is a black Elgar AT8000 DC power supply.&amp;quot; Variations in the above tuple would be processed as follows: * Component: If a different component had been specified, IDAS would have generated another response that communicated colour, manufacturer, and model-number information, as specified by the content-determination rule for What-is-it questions asked during Operations tasks (Section 3.2). For example, if the component had been Printer-12, the generated text would have been &amp;quot;The printer is a white Epson LQ-1010 printer.&amp;quot; * Basic Question: A different response pattern (i.e., content-determination rule) would have been used for a different basic question. For example, if the basic question had been What-is-its-purpose, the response would have been &amp;quot;The DC power supply provides DC power for the UUT.&amp;quot; * Task: A different response pattern would also have been used if a different task had been specified. For example, for the What-is-it question, if the user's task had been Replace-Part instead of Operations, colour would have been omitted but a part number would have been included, e.g., &amp;quot;Tile DC power supply is an Elgar AT8000 DC power supply with part number OPT-EP2.&amp;quot; * User-Expertise: Tile What-is-its-purpose response would have been phrased differently if the user's expertise level had been Novice instead of Skilled: 'unit under test' would have been used instead of 'UUT', 'power' instead of'DC power', and 'the black power supply' instead of 'the DC power supply', giving: null &amp;quot;The black power supply provides power for the unit under test.&amp;quot;  * Discourse-in-focus: The discourse-in-focus list does not affect the above responses, but it would affect the response to Where-is-it. The response to Where-is-it under the original discourse-in-focus list would have been:  &amp;quot;The DC power supply is below the VXI chassis.&amp;quot; If the discourse-in-focus list had included Mains-Control-Unit-29 instead of VXI-Chassis-36, the location would have been given relative to the mains-control-unit instead of the VXI-chassis, i.e., the text would have been: &amp;quot;The DC power supply is above the mains control unit.&amp;quot; Question space is quite large: the current prototype has 40 components, 7 basic questions, 6 user-tasks, and 3 user-expertise models, so there are over 5000 points in its question space even if variations in the discourse context are ignored, a A more realistically sized system would document several hundred components and probably would have additional user-task and user-expertise models as well; its question space could therefore easily contain several hundred thousand points. Many point., in question space represent queries that produce the same text (e.g., responses to Where-is-it do not depend on the user's task); even if only 10% of the points ir question space produce distinct responses, however, this still means that a realistically-sized IDAS system must be able to generate tens of thousands of different responses. The justification for using natural languag~ generation in IDAS is that it would be difficult to entel 20,000 different canned text responses for 200,000 different queries, and almost impossible to maintain this doe. umentation database as new ATE configurations wer~ announced; using NL generation from a domain knowl. edge base accompanied by explicit task, expertise, an( discourse models makes it feasible to supply appropriaU answers for this multitude of possible queries.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>