File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/88/c88-2129_metho.xml
Size: 13,171 bytes
Last Modified: 2025-10-06 14:12:14
<?xml version="1.0" standalone="yes"?> <Paper uid="C88-2129"> <Title>Understanding of Stories for Animation</Title> <Section position="3" start_page="0" end_page="622" type="metho"> <SectionTitle> 3. SDA Story Understanding Mechanism </SectionTitle> <Paragraph position="0"> In order to accurately understand an input story, SDA performs four distinct operations: \[1\] Extracting meanings of a sentence: Each sentence is parsed, its meanings extracted, and the meanings put into an independent block called world. Because our target story has a simple form, the sequence of its sentences becomes a chronological sequence.</Paragraph> <Paragraph position="1"> Each sentence in a story includes several assertions. An assertion extracted from a sentence may not be true in context with time of its succeeding sentence. Therefore, an individual worm is assigned to each sentence in order to store assertions which represent the situation inherent in a sentence. When a new world is created, it is linked to the sequence of worlds which is linked with each individual actors; the hare, the tortoise etc. Each worm is compared with its previous world, and its assertions are added/deleted/modified in the following processes, \[3\] and \[4\]. \[2\] Causality check among actions: When an action assertion is put into a world, its causal relationship to other actions is checked. If it is dependent on another action, the causality link is connected between the action and its independent action.</Paragraph> <Paragraph position="2"> \[3\] Interpolation of hidden actions between neighboring sentences: Each action of the present sentence is also checked for its continuity to actions in the previous world. If some action is hidden between the previous world and this sentence, it is identified and added into the present world.</Paragraph> <Paragraph position="3"> \[4\] Interpolation of a continuous action beyond a sentence: When there exist actions ~vhich are not mentioned in the present sentence but should continue from the previous sentence to the present sentence, they are added into the present world.</Paragraph> <Paragraph position="4"> Each world consists of two stages: present-state and post-state.</Paragraph> <Paragraph position="5"> Present-state, the upper part of a world, holds assertions which represent the state of the moment when the sentence is uttered.</Paragraph> <Paragraph position="6"> Post-state, the lower part of a world, holds assertions which represent the state just after the time when the sentence is uttered. For example, the world of (5) has the assertion &quot;the tortoise runs&quot; in its present-state, the assertion &quot;the-act-of the tortoise is run&quot; in its post-state (see Figure 1).</Paragraph> <Paragraph position="7"> the tortoise runs.</Paragraph> <Paragraph position="8"> the-act-of the tortoise is run.</Paragraph> <Paragraph position="9"> present-state post-state Figure 1 World of (5) This means that the tortoise is running during the sentence (5) and afterwards continues to run. Each post-state of each world is independently monitored by Truth Maintenance System (TMS) \[Doyle 79\]. This structure is similar to Viewpoints in ART \[Clayton 85\]. TMS works well in accomplishing the continuity chock of an action beyond a sentence.</Paragraph> <Section position="1" start_page="621" end_page="621" type="sub_section"> <SectionTitle> 4.2 Causality Check Among Actions&quot; </SectionTitle> <Paragraph position="0"> The dictionary contains action causalities of verbs. When a verb or a verb phrase is processed, the SDA parser consults the dictionary. The following is a part of the dictionary for the verb phrase, &quot;kick-up-a-eloud-of-dust&quot;: kick-up-a-clond-o f-dnst if the-act-of *actor is tun then ;;; *actor is a variable, the agent of this action If the tortoise kicks up a cloud of dust when it is running, the action of &quot;kick-up-a-cloud-of.dnst&quot; is assumed to be caused by the &quot;run&quot; ~ction. Therefore, the corresponding assertion of tbe action &quot;kick,i~-a-cloud-of-dnst&quot; is supported by tim &quot;run&quot; assertion. If the tor,oise is not running at the time. the assertion of &quot;kick-up-a-cloud-of-dust&quot; is justified as a premise, whic h means that the tortoise is kicking up a cloud of dust while standing at a point, This causality between different actions is used as a dependency directed link for TMS.</Paragraph> </Section> <Section position="2" start_page="621" end_page="622" type="sub_section"> <SectionTitle> 4.3 Interpolation of Hidden Actions Between Sentences </SectionTitle> <Paragraph position="0"> Interpolation of bidden actions is accomplished by using goal directed search. When a sentence is processed and its assertions arc extracted, the system picks up each assertion, and then makes an inspection to determine whether the action of each assertion is continuous from the state of the previous world or not. The continuity is inspected by checking whether the pre-condition of the action is satisfied in the post-state of the previous world or not. Each verb is specified its pre-conditton and post-condition in the dictionary.</Paragraph> <Paragraph position="1"> Pre-condition is the constraint to be satisfied just before the act of a verb. Post-conditlon is the state to be achieved just after the act of a verb. For example, the dictionary indicates that in order to &quot;stop&quot;, the agent must be going on foot (pre-condition), and after the agent &quot;stop&quot;s, it must be standing (post-condition).</Paragraph> <Paragraph position="2"> If an action in the present sentence is continuous from the post-state of the previous world, it is simply put into the present world. If it is not continuous, the system searches for a sequence of actions which bridge it (goal point) and the post-state of the previous world (starting points) by referring the pre-condition/post-condition of verbs in the dictionary. This search process is similar to the execution of STRIPS \[Nilsson 80\]. If a bridging sequence of actions is found, the abridged actions in the sequence arc added into the original assertion. Then, the modified assertion is put into the world.</Paragraph> <Paragraph position="3"> The sentence (2) &quot;the hare looked back&quot; is modified to &quot;the hare stopped and looked hack&quot; in order to satisfy the pre-condition of the verb phrase &quot;looked back&quot;. The related pieces of the dictionary are shown below.</Paragraph> <Paragraph position="4"> the-state-of the agent is standing.</Paragraph> <Paragraph position="5"> post-condition: the-state-of the agent is in the reverse direction.</Paragraph> <Paragraph position="6"> stop pre-condition: tbe-act-of the agent is go-on-foot.</Paragraph> <Paragraph position="7"> post-condition: the-state-of the agent is standing.</Paragraph> </Section> <Section position="3" start_page="622" end_page="622" type="sub_section"> <SectionTitle> 4.4 Interpolation of a Continuous Action Beyond a Sentence </SectionTitle> <Paragraph position="0"> Interpolation of a continuous action beyond a sentence is accomplist~ d based on the assumption that actors' actions are assumed to continue until they are explicitly ordered to stop. This assumption is the same as the persistence problem in \[Shoham 88\].</Paragraph> <Paragraph position="1"> After a sentence ts analyzed and its meanings are stored into both the present-state and the post-state of the present worM, all the assertions in the post-state of the previous world are copied into the post-state of the present world. Then, the post-state of the present world is choked its consistency by TMS. This ebeck prevents the over-copying of continuous actions from the previous worM. TMS works acemding to the following monitoring rules: Duplication Elimination Rule: If there exist two or more same assertions in the present world, the copied one from the previous world is diminated.</Paragraph> <Paragraph position="2"> * Exclusive Action Elimination Rule: If there exist exclusive assertions, the one copied from the previous world is eliminated. The exclusion relations of actions and states are also defined in the dictionary. The exclusion relations are like the followings: exclusive(act-of ran, state-of standing) exclusive(act-of run , act-of walk ).</Paragraph> <Paragraph position="3"> TMS compares each of the assertions in a world with each (6). The tortoise ran as kicking up a elond of dust, Vtl~ tortoise'runs. \[ Itbe tortoise kick-up-a-cloud-of- I 1.0.s- ..... I f\] the-act-of the tortoise is run. \[ i the act of the tortoise's kick-up/ / L.R-cloud-&quot; .I -of-dnst. $' / (7). The tortoise sweat while running.</Paragraph> <Paragraph position="4"> the tortoise sweats. I copy ~tl~ wrtoise kick-up-a-cloud-of- I'~ the to ise ts \ I ~te-aet-of the tortoise is sweat. I created I the-act-of the tortoise is klck-up- I~,/ / Lg-cloud-of-du 7 ...... I/ \[ (8). The tortoise stopped at the top of the mountain. copy I the tortoise stops at the top of the I t ~ ~0untatrL :-----I |t~-state-of the tortoise is stand- I \ ling . l \, r/~-e- ................... / --~/. o~.. ~ =, .,- .... 2~o. ~. L4 ~t, ,,. I other. If two assertions cannot coexist, the one copied from the previous world is deleted.</Paragraph> <Paragraph position="5"> Figure 2 shows the worlds corresponding to sentence (6), (7) and (8). Here, the world of(6) is already troth-maintained. After the sentence (7) is analyzed and its meanings are stored into the world of (7), the two assertions In the post-state of (6) are copied into the post-state of (7). TMS then deletes the duplicated assertion, &quot;the-act-of the tortoise is run&quot;. The assertion &quot;the-act-of the tortoise is kick-up-acloud-of-dust&quot; remains in the post-state of (7). Generally an assertion in a post-state corresponds to an assertion which presents the causal action in a present-state of the same world. For example, &quot;the-act-of the tortoise is ran&quot; is corresponding to &quot;the tortoise runs&quot;. When an assertion is added into the post-state of the present world by copied from the previous world and has no correspondence in the present-state of the present world, its corresponding assertion is created and put into the present-state of the present world by the system. Therefore, in this situation the corresponding assertion, &quot;The tortoise kick-up-a-cloud-of-dusts&quot; is created and added into the present-state of (7). The present-state of world of (7) shows that the tortoise is running while sweating and kicking up a cloud of dust (present-state), and afterwards continues to mn while sweating and kicking up a cloud of dust (post-state). After the meanings of the sentence (8) are stored into the worm of (8), three assertions are copied from the world of (7) to the world of (8). Next, because &quot;act-of run&quot; and &quot;state-of standing&quot; are exclusive, the assertion &quot;the-act-of the tortoise is run&quot; is deleted by TMS according to the exclusive action elimination rule. Then, two other assertions in the post-state of (8) which were supported by the deleted assertions are subsequently eliminated according to the dependency-directed backtracking mechanism of TMS. Now, the world of (8) has only one assertion, &quot;the-state-of the tortoise is standing&quot;, in the post-state of (8), which means the tortoise is standing and stopped kicking up a cloud of dust and sweating.</Paragraph> <Paragraph position="6"> After all the story is processed and represented as chronologieal sequences of worlds, assertions in thi~ present-state of each worm are gathered and transformed into a scenario for the stage directing module.</Paragraph> </Section> </Section> <Section position="4" start_page="622" end_page="623" type="metho"> <SectionTitle> 5. System Configuration </SectionTitle> <Paragraph position="0"> Figure 3 indicates a high level view of the whole story understanding system, Each sentence is processed individually and its assertions are extracted by SENTENCE-PARSER. The sentence grammar in SENTENCE-PARSER is described using Definite Clause Grammar in Prolog \[Pereira & Warren 80\]. The assertions are then put into the MORE-MEANING-EXTRACTOR which is based on forward-reasoning. Here as many assertions as possible are extracted from the inputs. For example, &quot;If the weather is fine and it is night, then the background for the drama stage is colored in black with lots of stars&quot;, etc. The extracted assertions are put into a separate worm for each sentence. Each world is monitored by TMS.</Paragraph> <Paragraph position="1"> Path #1 between neighboring worlds indicates the interpolation of a continuous action beyond a sentence. Path #2 indicates goal directed search to discover the path of transition between different actions. The whole story understanding system is implemented using Prolog on vax111780.</Paragraph> </Section> class="xml-element"></Paper>