<?xml version="1.0" standalone="yes"?>
<Paper uid="C82-1001">
  <Title>A leerntng of object sti-uetures by verbalisl Norlhivo</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3. FraRe Representation
</SectionTitle>
    <Paragraph position="0"> The slot AKO means a well-known relation A-KIND-OF, ana the CLASS indicates whether the frame is gneric or instance frame. If the frame is generic, then it has two slot, GEN~ recording its lower clas~ of generic frames and \[NST recording its instance frames. The F\[G slot represents a pictorial reration, to its parent frame. This slot means that the part corresponding to this frame is a subpart of the ,frame stored in the PART and that it can be found by looking for the region designated in POS- And the facet 01R describes a relation which this ~ art has to its parent~ There are three relations concerning to the In as shown in Fig.2 and concering to the POS, many combinations of subregions are ~permitted which can be expressed with the symbols, L,C,R.and U,C,D.,: Especially the symbols o ** are used to designate the locations shown in Fig.4. The slot SH~P represent whether the part corresponding to this frame is a region(~EG) or a branch(BR A) The SU~P slot records its subparts and their locations of or reiations to this part are described in three facets as sho~n above. EspeCially when the ~HAP condition is 0~A, th~s frame has a ~BB slot ~nstead of ~ Fig 1.</Paragraph>
    <Paragraph position="1"> He pu~ it, He ~lls OJ|RO. OJl~O takes ~t. A sample story HI gives O$IRO the apple. A LEARNING OF OBJECT STRUCTURES BY VERBALISM SUOPslot and a branch structure is recorded here. An example :of branch is shown in Fig.5. The COL slot records a color of this part and a slot CONCEPT means that this frame is prepared for the conceptual consistency' of frames and not for pictorial relation. In additiontto these slots, there are several slots, WAKE, SEX, NUM and so on. These are prepaired ,to generate a sentence for stating a reason why thisl frame is required or an explanation about why discrepancies between an object and its model can he found out in its matching:process.</Paragraph>
    <Paragraph position="2"> 4. Basic:Strategy of Learning The system tries to generate,a model for the unknown object by referring to an analogical model and using a teacher's instruction, and simultaneously it augments the concept trees~of objects. At that time, the first key for a detection of analogy is assumed to be in locations of subparts of objects. When we are told that an unknown object is similar to a certain object amOng various points of view, we usually expect that many substructures having similar features will be found in the same location as the refered object* Of course,,there are many examples that resemblance in a Iocation~ is not useful hut preventscthe program from achieving a correct detection of analogy. At that case, the teacher should explicilty tell the program to ignore  this standard and to use other methods such as similarity in relations or shapes or colors of objects. ,As we usually employ these variety way to detect analogy and record these experinces into our memory, we can easily decide what method should be used to compare things. We can not say that the program has learned something untill these mechanisms recording a standards to compar things into memory are realized in the program.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5. Scenario
</SectionTitle>
    <Paragraph position="0"> T: YES. BUT OJIRO HAS JUST ONE HAIR. (1) C: A color of his body Is not white.</Paragraph>
    <Paragraph position="1"> It is yellow. Is it true ? (note~ that a color of OJIRO'S BODY is yellow) (2) T: YES.</Paragraph>
    <Paragraph position="2"> C: l have found hie body. I have found his hair. l have found his left hand. I have found his r'ight, hand.</Paragraph>
    <Paragraph position="3"> But I cannot find his mouth. (3) T: WHAT ARE YOU DOING? C: I have found a red branch to the left of his right.hand as a candidate of his month. But it is different from his model. Do you need another explanation of the location of his mouth ? (4)  T: YES.</Paragraph>
    <Paragraph position="4"> C: Do you want an explanation in:absolute coordinate ? T: NO.</Paragraph>
    <Paragraph position="5"> C: It Is to the right of his left hand in the image. Do you want another explanation in absolute coordinate? (5l T-&amp;quot; YES.</Paragraph>
    <Paragraph position="6"> C: It is in the~ center:of the middle part of his body In the imoge. (6) T: EXPLAIN THE DIFFERENCE BETWEEN THE MDDEL AND OBJECT.</Paragraph>
    <Paragraph position="7"> C: On what part do you want an expler~ation?  T: IqDUTII. (7) C: The mouth of the model is different from that of the model in the point~ of location, shade, subpart and color. Do you need more exp-lanation? (el T: YES.</Paragraph>
    <Paragraph position="8"> C: The mouth Is all the middle part of his body, and that of the object is in the center of the middle patr of hie* body. The shape of the model is e pink region, and that of the object is a red branch. The mouth of the model has a lip, but that of the obJect has not subpart. T: OJIBO'S NOUTH IS A RED LINE. (9) C: 1 have found his mouth. But I cannot find his right-eye.</Paragraph>
    <Paragraph position="9"> T: WHAT ARE YOU DOING? C: I have found a black region in the center of the middle part of hie:body as a candidate of his right-eye. But it is different from his model. (10) T: EXPLAIN OJIRO'S BIGeT-EYE.</Paragraph>
    <Paragraph position="10"> C: Which do you mean the model or object? T: HODEL. , C: It is a white region which is above his mouth in the model. Do you want another explanation? T: NO.</Paragraph>
    <Paragraph position="11"> T: DOES IT HAVE SUBPAIbTS? C: Yes, his right-eye has a pupil in the mode 1.</Paragraph>
    <Paragraph position="12"> T: OJIRO'S RICHT-EYE lS A BLACK ~ION. C: I have found his right-eye. I have found his left-eye. Finally, I have found oJiro. T: EXPLAIN OJIHO.</Paragraph>
    <Paragraph position="13"> C: He is a kind of bogle.</Paragraph>
    <Paragraph position="14"> T: EXPLAIN OJIRO'S STRUCTURE.</Paragraph>
    <Paragraph position="15"> C: OJiro is a yellow region and hie body has a hair, a mouth, eyes and hands.</Paragraph>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5. Example
</SectionTitle>
    <Paragraph position="0"> We would llke to explain how the dialog shown in 5. has been obtained. By (I~ the program generate OIJRO model from OBAQ model shown in Fig.3:~ and constructs a frame tree consisting from a BOGLE frame and the OBAQ frame, is stored as an instance of the BOGLE frame.</Paragraph>
    <Paragraph position="1"> But it records~in its STM thatlthere is just one hair as the teacher said so. Next, C~looks for a candidate region of OJIRO using the copied model. In.the second frame, OBAQ, TABLE, APPLE, CLOCK and OJIRO are drawn, but as the first four,objects have been appeared in the first frame, in this case C can find OJIRO by looking for new objects.</Paragraph>
    <Paragraph position="2"> But regrettably a color of the region (yellow) which seems to be OJ-IRO'S. body(J-BODY) being different from that of the model(white), this cause a complaint shown in (2) and by accepting a T's agreement C can believe its correctness and T can also think C in a right state.</Paragraph>
    <Paragraph position="3"> Consequently, C changes value of C0L in J-BODY into YELLOW.</Paragraph>
    <Paragraph position="4"> Next~ C tries a verificationiof J-HAIR which is the ~first member of Scots'r; where Scou'r={J-HAIR,J-HAND} As C can be aware of the fact that J-HAIR is a hair by its AK0 slot and that there is a note on the hair in STM, it can know that 0JIR0'S ~hair cannot be recognized only by referring~to the copied model. Since the just one alteration in the number of hairs is recordedf there,, C thinks their location to be same as tl~ model specification, end can find a line in the ((C)U) part of J-B0bY. It ends the ve~ffication of J-HAIR by storing (H1 NIL). into SUBB ~lor io place of:(L1 NIL L2 NIL L3 NIL). In a similar ~ay to this, C begins to A LEARNING OF OBJECT STRUCTURES BY VERBALISM identify J-HAND, however C can be aware of that it should 10ok for J-R-HAND and J-L-HAND, as ~here is a CONCEPT slot in J-HAND. So C succeeds in the identification*of them~because.of a perfect match in their locations, colors and substructureS.</Paragraph>
    <Paragraph position="5"> The result of this steps is reported in (3). Next,. the identification process proceeds to Sin and C starts a verification of J-MOUTH,~where SIxffi(J-MOUTH, J-EYE). As the locational constraint for this part is ((**)C), which means that it occupies ((L)C), ((C),C) and ((R),C) of J-BODY, the check is attempted whether just one candidate can be found for each of.these 3 subregions. In this case, nothing is found for ((L)C)land ((R)C) but several parts are found in ((C)C) of J-BODY. So this process is suspende d and identification of other parts (J~R-EYE and J-L-EYE) is attempted, but the same ambiguity as the above occurs and this causes the identification steps to be suspended. Consequently, for each one of these 3 parts,,their results are just same ~each other; there are 3~parts in the ((C)C)of J-BODY and theylare candiades for J-MOUTH, J-R-EYE and J-L-EYE. Then C avails of the relational constraint onilocations of them in order to clarify their correspondences as far as .possible. It infers, that J-MOUTH ,probably locates in a lower position than J-EYE, because the location*of J-MOUTH is ((**)C) and that of J-R-EYE and J-L-EYE is ((L)U) and ((R)Ut respectively (in this example note that the location of J-EYE, ((**)U) can be also available)~ And it is also.decidable if which black region corresponds to J-L(R)-EYE using, the relation between ((L)U) and ((R)Ut. By this assumption on availabilty of the relational constraints, C can discover one possible correspondence between the model and object. Then other properties are tested: But regrettably, discrepancies are found for both his mouth and eyes. The candidate for his mouth is a line segment, whereas the model says that it is a region and that it has a substructure. Similary the candidate for his left(right) eye is a black region,but its model description is that it is a white region with. a substructure.. At the present.state of program, any estimation on which is more plausible is.not realized regarding to the accordance of these properties, C simply complains about their disagreements in the order of their discovery.</Paragraph>
    <Paragraph position="6"> Therefore it at first complains of his mouth as shown in (4).</Paragraph>
    <Paragraph position="7"> Given teacher*s instruction on a. shape of mouth, C is convinced of his decision and add a new slot SUBB in plac e of SUBP and records (H~ NIL) into it becase it has found that his mouth is not a region but a line segment. Here instead of the instructionJ(9), T can say that C should be believe the given image correct. In that case, C suppose its decision to be right and does the same thing as the above. The difference between these two cases is that the latter has a high risk in the correctness of its conclusion.</Paragraph>
    <Paragraph position="8"> Next, C complains about the discrepancies of his eyes. Note here that nothing is stated, about his left-eye oncean instruction on his right-eye is given to it, because they have the same properties concerning to both their models and object parts. In case where one of them is not same, a question is asked:about the difference by:C.</Paragraph>
    <Paragraph position="9"> T. Use of Generic Frames As mentioned in 4., OBAQ frame causes BOGLE frame to~be generated as a generic ene, and OJIRO frame ~is obtained through learning process. At present our program ~ust makes frame, trees in which OJIRO and OBAQ frame are child of BOGLE.</Paragraph>
    <Paragraph position="10"> A reason for this~is partly due to a lack of condsideration how simple pictorialidescriptions can be compiled from various types of deviations in slot values. An another reason is that there is a danger of partial rearrangements of frames trees~ In the example, we at first believe OBAQiframe to be an instance, frame but it may turn out thatiit is not an instance when other examples not mat.ched to this frame appears in:image, because there are many varietions in his shape as he can wink or move his eyes or open his mouth. After program have N. ABE and S. TSUJI experinced these example, it should ~:,~ke a general concept of OBAQ and arrange frame trees by erasing unnecessary instances about him.</Paragraph>
    <Paragraph position="11"> As amore important problem, s.rategies to discover cues for finding analogy between subparts mus. belstored in some slots of their model; that is, the locational con~ aint is a useful cue for human, animals and so on, but is not adeq, i for doors and windors of houses.</Paragraph>
    <Paragraph position="12"> Though there are some inco~ ,ete points in the construction of frame trees, program can use a po, 1on of them to identify subparts of the object to be learned. For e~ample, suppose that we,would like to teach a character Q-KO by referrin~ to OJIBO. Let suppose that Q-RO resembles to him very much except for her eyes but that they are rather similar to OBAQ's. In the course of identification of her, if OBAQ frame is not stored, program will complain about her eyes as well as in the learning of OJIRO from OBAQ. However it can use OBAQ'S eyes in the 1recognition of her eyes by tracing its AKO link and finding OBAQ frame, after a failure in the matching of her eyes to OJlRO's. Of cause, it does not do that without teacher's permission~ hut will ask for his approvals; 8. Explanation Capability &amp;quot; It is necessary for teacher to be given sufficient explanations about the level of knowledge the computer has attained. Unless the computer~can tell him what it is~looking for, what it has already found, what sort of descrepancies it suffers from, he cannot give ~ roper instructions leading the ~computer to a satisfastory state.</Paragraph>
    <Paragraph position="13"> here are many sentence generating and explanation systems, however an explanation system like this research has not been investigated in the point that our system tries to give its partner an explanation or pictorial features of objects to:be modeled by translating sentences not from the case frame, of sentences but from frames corresponding to the pictorial models. Naturallysuch an explanation is on locations, shapes, :colors and relations that models or objects have, and must be given in, the forms that the partner can easily understand what the system knows. 'For this purpose, the explanation on locations is first attempted using the referred things in the dialog, and is finally given in an absolude coordinate based on the 9 subregions if there is no reference or the reference stack becomes empty. (4),(5),(6) in the Scenario~ shows &lt;this mechanism. The next important thing is that a partner may expect a detail explanation for something, but expect just a simple one for others. Regrettably the present sate of our program cannot detect his demand like this or resolve ambiguous points of his questionl then it must ask him about his require as shown in (7). In this case, thereiare also many things to be explained, however the points are only stated by the program and the detail explanation is left to the partner as in (8). We believe this method proper because of easiness of explanations.</Paragraph>
    <Paragraph position="14"> The comparison between things are listed in above of (9) in order to clarify their~differencies. If more detail on the lip is needed, the partner can ask the system about itL On account of limited space, though we cannot state a sufficient :discussion, there are many problems to he improved on how the system should grasp partner's intention or requirments. They must be solved for giving simple explanation to the partner.</Paragraph>
  </Section>
class="xml-element"></Paper>