<?xml version="1.0" standalone="yes"?>
<Paper uid="C82-1001">
  <Title>A leerntng of object sti-uetures by verbalisl Norlhivo</Title>
  <Section position="2" start_page="0" end_page="0" type="abstr">
    <SectionTitle>
1. IntrodUction
</SectionTitle>
    <Paragraph position="0"> We have reported the story understanding system which uses both linguistic and pictoriallinformation in order to, resolve the meaning of givenisentences and images. In this research, we have believe that a correct meaning of the given sentences is obtained if the relations among noun .phrases, which correspond I to objects in the images, consistent with the relations observed among objects in the images.</Paragraph>
    <Paragraph position="1"> The Jfact that this identification : of objects and the interpretation ~:of the given sentences supplements each other simplifies both the detection of objects and disamibiguation of word sense or prepositional groups. In Spite of these effects, this forlalisn has a defect that it requires additional knowledge sources, the model of objects that will appear inpthe images. All of models of objects or actors that are supposed to appear in, the picture must be given to our system in order to achieve its purposes. But it is not easy for us to store all of such models in a computer. If a person who does not know. well about the details of this system wants to interact,with it~ he will give up to use the system, as he knows nothing !of the representation~ of models in the computer. To make matters worse, there, are quite many variations in real objects which we will. encounter in the real world. For example,we can see various type of houses. ~:In the traditional AI system, :a generic model is utilizedJ to identify such class of objects. But it is not easy for such a system to discriminate idiosyncrasy of varous objects, iFig.l shows a part~ of sample story used to experiment its story understanding capability. Even if the system is supposed to be given a generic model (for example, BOGLE) that represents both OBAQ and OJIRO, the system will not be able to discriminate them. The system needs some ,proper model for 0BAQ and OJIRO. But if a new character which has some similar points to;0BAQ and OJIR0 apperes in the story, some modificationsl to the BOGLE model are required., Thus generalization process could not be acomplised in advance, but should be achieved through experiance.</Paragraph>
    <Paragraph position="2"> When, we are asked to do some task, we are usually given informations concerning to the objects of that task and ~their processing method. In case where we encounter some~unkown objects ~n the course of the taskL we can construct aJ more generic ,model including them ~together with a,ereation of instance models for those individuals by d~manding an explanation to a person who knows well about those objects. In this~real situation, it cannot be expected that a learning process proceeds successfully like the experiment studied :by Winston, as the assumption fails of success that the samples can be arranged conveniently for the learning. We usually augment our knowledge by explicitly being taught about missing or insufficient parts of the known models.</Paragraph>
    <Paragraph position="3"> In order to realize this type of learning, there are two important problems to be solved. First is an explanation capability. Unless a 2 N. ABE and S. TSUJI capability to convey one's obscure points ta his partner:is given to the system, it is difficult for the system to obtain good instructions from its partnerL Second is a::point that fromiwhat kind of levels of knowledge state the system~:sheuld start its learning process. Should an initial state of~knowledge be given in forms of an inner representation or be explained in natural language? We select the former approach by just the following reason. Ve think it quite difficult to give a clear view to unknown object without referring to models. So we restrict a class of objects learned by our system:to the group of objects of which the system can .obtain clear views concerning to their conditions through the comparison with their similar example.</Paragraph>
    <Paragraph position="4"> But the assumption is not required that, examples should be different in only one or two points at most from the unknown object.</Paragraph>
    <Paragraph position="5"> Many, discrepancies between the object and its models are permitted to exis~ because such differences can be explained explicity in the language, by a teacher. And through a cognition of analogical or discrepant points of objects belonging to the same conceptual class, a generalization process is invoked that creates a common concept to them.</Paragraph>
    <Paragraph position="6"> ~. Description for Object The 'model description used in this paper is, the same one shown in the paper\[I\] except for the usage of the frame representation to describe~ relations among subpartsof the model. Let explan using an example. Fig.2 shows the OBAQ, who is an actor of the 'sample story shown in Fig.li To describe location of subparts of this model, its main part is enclosed by a rectangle as shown in Fig.2. Then this rectangle is devided into 9 subregions and the location of its subparts, is described in terms of these subregions. Yhen some of these subparts ihas also subparts, they:are hierarchically described in the similar way. And the relations between these subparts is represented using the frame, structures. The frame structures corresponding to:the DBAQ model is given in Fig.3 (this figure shows a hypothetical model of OJIRO obtained from the copy of OBAQ frame,)</Paragraph>
  </Section>
class="xml-element"></Paper>