<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-2129">
  <Title>Understanding of Stories for Animation</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1. Introduction
</SectionTitle>
    <Paragraph position="0"> Recently computer animations have been widely used in many fields like conventional CAD, commercial films and movies. For these kinds of applications, high-level languages are now provided \[Reynolds 82\]\[Zeltzer 82\]. However, because these languages are programming languages and am hard to use for untrained personnel, it is desirable to develop an easy-to-use computer animation system to encourage more wide-spread use of computer animation.</Paragraph>
    <Paragraph position="1"> The authors have been developing Story Driven Animation (SDA) \[Takashima et al. 87\] which automatically generates the animation for a given story written in Japanese taken from a children's story book.</Paragraph>
    <Paragraph position="2"> SDA consists of three modules: 1) story understanding; 2) stage directing; 3) action generating. The first module reads a story written in Japanese and makes an action-based scenario. The second module receives the scenario and modifies it for stage setting. The last module generates animations on a graphics display according to * The author is a member of NEC Corporation, and is now staying at the  the precisely specified scenario given by the stage directing module. SDA differs from previous natural language processing systems, such as summarizing \[Young &amp; Hayes 85\]\[Lytinen &amp; Gershman 86\], depth understanding of stories \[Dyer 83\]deg question-answering \[Wilensky 82\]\[Harris 84\] and story gunerating\[Schank &amp; Riesbeck 81\], in that it interpolates hidden actions among sentences that must be discovered for realistic computer animations but are not explicitly described in the story itself.</Paragraph>
    <Paragraph position="3"> Since SDA can accept curtailed expressions in input stories, story writers do not have to describe all the acts explicitly to get the desired graphic animation. Consider the following sentences in &amp;quot;The Hare and The Tortoise&amp;quot; in Aesop's fables:  It is never thought that the hare would lie down while running with his face looking tn the reverse direction. If he were to do so, he would do a dive and his neck would be brokenl Any person can conjure up an accurate image of the hate's actions. Naturally lacking facts between sentences arc interpolated using the human reader's common-sense. We imagine that the hare stopped between (1) and (2), and looked forward between (3) and (4). However, if an animation producing program does not have this common-sense, it produces strange graphic animations when these sentences are not more explicitly described. The above scenario is merely one example of SDA's ability to interpolate curtailed expressions of action. In order for a story understanding program to laccurately accept input sentenees, it is imperative that such a common sense be built into the program.</Paragraph>
    <Paragraph position="4"> SDA story understanding mechanism was constructed based on action-oriented knowledge. Knowledge related to actors' actions is independent of the content of individual stories, and is common to everyone since actors' movements are constrained by physical limitations of a human body. Although this story understanding approach is superficial, it allows for an extensive domain.</Paragraph>
    <Paragraph position="5"> This kind of research which strictly infers occurrences of actions among natural language sentences has not been done yet. In this paper wo present various types of hidden actions among sentences and the inference mechanism to identify and interpolate them. The whole SDA system is roughly described in \[Takashima et al.</Paragraph>
    <Paragraph position="6"> 87\].</Paragraph>
    <Paragraph position="7"> 2. Types of Hidden Actions to be Interpolated There a~e three types of hidden actions and relations of actions to be interpolated among sentences. The above example shows a type of hidden actions to be interpolated. It can be stated as follows:  , Continuity of different actions between sentences: When an action in a sentence is not consecutive to any action in the previous ~ntence, bridging action(s) has/have to been found and added into the original text. The action &amp;quot;stop&amp;quot; between &amp;quot;run&amp;quot; at (1) and &amp;quot;look hack&amp;quot; at (2) and the action &amp;quot;look forward&amp;quot; between &amp;quot;look back&amp;quot; at (2) and &amp;quot;lie down&amp;quot; at (4) are examples. SDA interpolates discriminately; if in the same context it must  interpolate, if not in the same context, then it must not try to interpolate (e.g. &amp;quot;The frog lay down. The next morning, the frog went out.&amp;quot;).</Paragraph>
    <Paragraph position="8"> The following example includes the other two types of interpolations in it. The example is also from &amp;quot;The Hare and The Tortoise&amp;quot;. (5). The tortoise ran.</Paragraph>
    <Paragraph position="9"> (6). The tortoise ran as kicking up a cloud of dust.</Paragraph>
    <Paragraph position="10"> (7). The tort(rise sweat while running.</Paragraph>
    <Paragraph position="11"> (8). The tort(~ise stopped at the top of the mountain.</Paragraph>
    <Paragraph position="12"> The proper interpretation of these sequences may be the followings: The tortoLve starts running at (5). The tortoise runs while kicking up a cloud of dust at (6). Here, the action of &amp;quot;kick-up-a-cloud-of-dust&amp;quot; must be caused by the action of &amp;quot;run&amp;quot;. The tortoise sweats while running and kicking up a cloud oJ dust at (7), even though &amp;quot;klck-up-a-cloud-of-dust&amp;quot; is not specOfed in (7). Here, the action of &amp;quot;sweat&amp;quot; must be caused by the action &amp;quot;run&amp;quot;, for the tortoise is not breaking into a coM sweat. The tortoise stops running at (8). At this time the tortoise also stops &amp;quot;kick-up-a-cloud-of-dust&amp;quot;-ing and &amp;quot;sweat&amp;quot;-ing because these two actions were caused by &amp;quot;run&amp;quot;.</Paragraph>
    <Paragraph position="13"> The other two types of hidden actions are: * Causality among actions: When an action appears in a sentence, it must be verified whether it is independent of any other action or caused by other actions. If an action is caused by another action, it also ceases when its dependent action stops, (e.g. the relationship between &amp;quot;kick-up-a-cloud-of-dust&amp;quot; to &amp;quot;ran&amp;quot; and &amp;quot;sweat&amp;quot; to &amp;quot;run&amp;quot;). This verification is done between neighboring sentences whose agent is the same.</Paragraph>
    <Paragraph position="14"> Continuity of an action beyond a sentence: An action is assumed to continue until it is explicitly ordered to stop if this action is a continuous type. Inference of the continuity of the action &amp;quot;kick-up-a-eloud-of-dnst&amp;quot; from (6) to (7) is an example.</Paragraph>
  </Section>
class="xml-element"></Paper>