<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1413">
  <Title>GENERATING WARNING INSTRUCTIONS BY PLANNING ACCIDENTS AND INJURIES</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
GENERATING WARNING INSTRUCTIONS BY PLANNING ACCIDENTS AND INJURIES
</SectionTitle>
    <Paragraph position="0"/>
  </Section>
  <Section position="2" start_page="0" end_page="0" type="metho">
    <SectionTitle>
Abstract
</SectionTitle>
    <Paragraph position="0"> We present a system for the generation of natural language instructions, as are found in instruction manuals for household appliances; that is able to automatically generate safety warnings tO the user at appropriate points. Situations in which accidents and injuries to the user can occur are considered at every step in the planning of the normal operation of the device, and these &amp;quot;'injury sub-plans, are then used to instruct the user to avoM these situations.</Paragraph>
  </Section>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> We *present a system for the generation of natural language instructions, as are found in instruction manuals for household appliances * , that is able to automatically generate safety Warnings to the user at appropriate points. Situations in which accidents and injuries to the user can occur are considered at every step in the planning of the normal operation of the device, and these &amp;quot;injury sub-plans&amp;quot; are then used to instruct the user to avoid these situations. Thus, unlike other instruction generation systems, our *system tells the user what not to do as well as what to do. We will show how knowledge about a device that is assumed to already exist as part of the engineering effort, together with adequate, domain-independent knowledge about .the environment, can be used for this. We also put forth the notion that actions are performed on the materials that thedevice operates upon, that the states of these materials may change as a result of these actions, and that the goal of the system should be defined in * terms of the final states of the materials.</Paragraph>
    <Paragraph position="1"> We take the stand that a complete natural language instruction generation system for a device should have, at the top level, knowledge of the device (as suggested by Delin et al. (1993)). This is one facet of instruction generation that many NLG systems have largely ignored by instead incorporating the knowledge *of the task at their top level, i.e., the basic content of the* instructions is assumed to already exist and does not need to be planned for. In our approach, all the knowledge necessary for the planning stage of a system i s contained (possibly in a more abstract form) in the knowledge of the artifact together with the world knowledge. The kinds of knowledge that Should .be sufficient for this planning are device knowledge *(topological, kinematic, electrical, thermodynamic, and electronic) and world knowledge.</Paragraph>
    <Paragraph position="2"> The IDAS project of Reiter et al. (1992; 1995) served as a key motivation for this work. One of the primary goals of the IDAS project was to automatically generate technical documentation I Address correspondence to the second author. E-maili gh @cs.toront0.edu</Paragraph>
    <Paragraph position="4"> from a domain knowledge base containing design information (such as that produced by an advanced computer-aided design tool) using NLG techniques. IDAS turned outto be successful in demonstrating the usefulness, from a cost and benefits perspective, of applying NLG technology to partially automate the generation of documentation. If work in qualitative process theory, using functional-specifications such as those in e.g., (Iwasaki et al., 1993), can yield the device and world knowledge that are required for text pianning, then the need for cost effectiveness would be met.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="119" type="metho">
    <SectionTitle>
2 A situation calculus approach to the generation of instructions
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="119" type="sub_section">
      <SectionTitle>
2.1 Overview
</SectionTitle>
      <Paragraph position="0"> In this section we shall present some of the planning knowledge for a toaster domain, in the form of axioms in the situation calculus 2 (see (Reiter, 1991 )). This planning knowledge formally characterizes the behaviour of the artifact, and it is used to produce a basic plan of actions that both the device and user take to accomplish a given goal. The axioms together with the goal are the input to Our system.</Paragraph>
      <Paragraph position="1"> We will explain how the instructions are generated from the basic plan. This plan is then used to derive further plans for states to be avoided, and warning instructions about these Situations.</Paragraph>
      <Paragraph position="2"> We shall use the term device--environment system to refer to the device, the user, and any objects or materials used by the device.</Paragraph>
      <Paragraph position="3"> We can conceptually divide the actions that are performed in the device--environment system into user actions and non,user actions, the latter of which are actions that are carried out either by the device on its components and the materials it uses, or by some other agent. Because the majority of non-user actions are actions performed by the device, we shall only consider device actions henceforth.</Paragraph>
      <Paragraph position="4"> Natural language instructions are directed to theuser of a device, and usually they mainly describe the actions that are executed by the user.</Paragraph>
      <Paragraph position="5"> A device action may be carried out by a component of the device on another component; for example, the heating element of a toaster may carry out a heating action (i.e., a continuous, physical process) on the bread slot, which in turn may heat the inserted bread slice.</Paragraph>
      <Paragraph position="6"> Instead of using a qualitative or quantitative simulation system, such as the Device Modelling Environment (Iwasaki and Low, 1991), we have used device actions to discretely model the continuous processes, for simplicity.</Paragraph>
      <Paragraph position="7"> Table 1 shows the components of our toaster and the materials used for its operation. Table 2 shows the user actions, device actions, and fluents.</Paragraph>
      <Paragraph position="8"> 2In the situation calculus, the initial state is denoted by the constant So, and the result of performing an action a in situation s is represented by the term do(a,s). Certain properties of the world may change depending upon the situation. These are calledfluents, and they are denoted by predicate symbols which take a situation term as the last argument. Positive (negative) effect axioms describe the conditions under which performing a in situation s causes a fluent to become true (false) in do(a,s). Action precondition axioms describe the conditions under which a can be performed in s. We use these axiomatic forms in order to avoid the frame problem. Following Pinto (1994), we shall abbreviate terms of the form do(a,,(do( .... do(al,s)...)) as do(\[a1,...,a,l, s).</Paragraph>
    </Section>
    <Section position="2" start_page="119" end_page="119" type="sub_section">
      <SectionTitle>
2.2 Some axioms for the toaster system
</SectionTitle>
      <Paragraph position="0"> The following are some of the more important axioms for our toaster example (see Ansari (I 995) for the complete set). Some of them are essentially domain-independent, whereas the others relate Specifically tothe appliance. Where free variables appear in formulas, they are assumed to be universally quantified from the outside.</Paragraph>
      <Paragraph position="2"> * an agent can touch an object if it is exposed; . the agent can get burned by touching something with a temperature of at least 70degC; and * the device can cause the bread slot to pop up its contents if the temperature of&amp;quot;the bread slot reaches 200degC.</Paragraph>
      <Paragraph position="4"> These axioms state that: burned in the new state; and * if the device Causes x to pop up in state s, then x becomes exposed in the next s~te.  inserting x into y in state s results in y containing x in state do(a, s); if it is possible for the agent to get burned (by the get_burned action), then the agent might be</Paragraph>
      <Paragraph position="6"> This axiom states that an action of the user pressing the ON lever causes anything in the bread slot to become unexposed; this happens because the object in the bread slot gets &amp;quot;pushed down&amp;quot;.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>