<?xml version="1.0" standalone="yes"?>
<Paper uid="W94-0312">
  <Title>Generating Event Descriptions with SAGE: a Simulation and Generation Environment</Title>
  <Section position="2" start_page="0" end_page="99" type="intro">
    <SectionTitle>
1. INTRODUCTION
</SectionTitle>
    <Paragraph position="0"> In Text Generation, some of the most interesting issues lie at the interface between the conceptual model (the underlying program) and the generator. It is well recognized that one cannot produce sophisticated text from an impoverished underlying model (McKeown &amp; Swartout 1988). McDonald (1993) makes an even stronger claim: Nevertheless, the influence it \[the application program\] wields in defining the situation and the semantic model from which the generator works is so strong that it must be designed in concert with the generator if high quality results are to be achieved.</Paragraph>
    <Paragraph position="1"> In fact, some of the best results in text generation have come from efforts where the model and the generator were developed in tandem, from Davey's early work on describing tic-tac-toe games (Davey 1974) to Dale's recent work on generating recipes (Dale 1990). Dale found that in order to generate referring expressions in recipes, he had to work on the representation of the underlying objects and their state changes in order to be able to correctly generate the number of the noun phrases in examples such as &amp;quot;Grate one carrot...Add the carrots to the stew&amp;quot;. The most impressive results to date in event generation is the NAOS system (Novak 1987, Neumann 1989), which produces natural language descriptions of object movements in a street scene. It is designed to take is input from a vision system observing traffic, which captures both temporal and spatial relationships among the objects in the scene. The focus of the work has been on representing events and the relations among them and then connecting those events to case frames for expressing them in natural language.</Paragraph>
    <Paragraph position="2"> In narration, temporal and aspectual information must be available in the underlying model in order to describe events. For example, using the well recognized Reichenbachian model, three different temporal points, point of event (E), point of speech (S), and point of reference (R), are needed in order to adequately account for the English tense system, as shown in the following examples:  1. Peter drove to work. (E = R &lt; S) 2. Peter had driven to work. (E &lt; R &lt; S)  Such problems are generally treated as unified systems in linguistics within studies of semantics or the lexicon. However, in generation research, the issue is not just what distinctions there are, but at what level (model, text planner, syntactic component) should the information needed to make these distinctions be represented. Taking the temporal points in the Reichenbachian model as an example, two of the points, E and S, are facts of the model, when the event took place and the time the speaker is producing the utterance. However, the third point, the reference time, is a fact of the discourse, a choice to be made by the speaker. (1) and (2) above are distinguished by the reference time, but otherwise could describe the same event and be spoken at the same time.</Paragraph>
    <Paragraph position="3"> While most studies of events are done within the realm of linguistics, where the focus in on the expression of event descriptions, it is clear that the way events are modelled is also an essential element. Bach (1988) describes &amp;quot;how certain metaphysical assumptions are essential to an understanding of English tenses and aspects. These assumptions have to do with the way reality---or our experience--is structured.&amp;quot; From a generation perspective, there are two basic questions to be answered. First, what information is needed in order to produce the distinctions available in language, and secondly, what distinctions are facts of language (and thus should be in the generator) and which are better represented at the model level?  * . 7th International Generation Workshop * Kennebunkport, Maine * June 21-24, 1994 The problem of finding a general way to research such questions has led to the development of SAGE, a &amp;quot;Simulation and Generation Environment&amp;quot;, which provides components for both conceptual modelling and text production. In SAGE, a frame-based knowledge representation component models objects and their properties, an event-based simulator models the actions of multiple agents, and a graphics component provides models of the physical geography in the virtual world, in addition to providing a visual interface to the objects, agents, and actions. Text generation is provided by the SPOKESMAN system. SPOKESMAN is data directed in that it links to the other components both through mappings from concepts in the knowledge representation and through instances of objects and events created by the simulator.</Paragraph>
    <Paragraph position="4"> In this paper, I describe the components of SAGE and how they are integrated, focusing on the generation of event descriptions. In Section Two, I look at what information is needed to generate events through analysis of events and a review of the linguistic literature. In Section Three, I describe the architecture of SAGE and its representational levels, including where in the overall system event information is represented and in Section Four I illustrate these issues using paragraphs generated in SAGE, such as the following: Fluffy wants to catch a mouse. He is looking for her.</Paragraph>
    <Paragraph position="5"> The mouse wants to get cheese. She is leaving a mouse-house. She is going toward it.</Paragraph>
    <Paragraph position="6"> Fluffy is chasing the mouse. He is going toward her. He caught her.</Paragraph>
    <Paragraph position="7"> The mouse didn't get the cheese.</Paragraph>
    <Paragraph position="8"> The overall methodology applied in this research tooapproach the problem from two directions, as depicted in Figure 1. One direction is that from a situation modelled in some application program to the expression of some set of goals from that program in a natural language (in this case, English). The second direction is the use of text analysis to work backwards from the way something is said to what decision points led to that text, what alternative choices were not made, which decisions were constrained by the syntax or lexicon of the language, and what information is needed in the application program in order to make these decisions.</Paragraph>
    <Section position="1" start_page="99" end_page="99" type="sub_section">
      <SectionTitle>
Underlying Program
</SectionTitle>
      <Paragraph position="0"> in a particular situation with a set of goals to accomplish How to realize those goals through language</Paragraph>
    </Section>
    <Section position="2" start_page="99" end_page="99" type="sub_section">
      <SectionTitle>
Expressibility
Expressiveness
</SectionTitle>
      <Paragraph position="0"> How to account for the competence people demonstrate through their use of language  This methodology is exemplified in the work presented here in the first direction by the use of SAGE to model situations and generate text (described in Section 3 and exemplified in Section 4) and in the second direction, through the analysis of events and projection of that analysis onto the decisions of the generator (described in Section 2).</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>