<?xml version="1.0" standalone="yes"?>
<Paper uid="P82-1004">
  <Title>What's in a Semantic Network?</Title>
  <Section position="8" start_page="24" end_page="24" type="evalu">
    <SectionTitle>
6. Associations and Partitions
</SectionTitle>
    <Paragraph position="0"> Semantic networks are useful because they structure information so that it is easy to retrieve relevant facts, or facts about certain objects. Objects are represented only once in the network, and thus there is one place where one can find all relations involving that object (by following back over incoming ROLE arcs). While we need to be able to capture such an ability in our system, we should note that this is often not a very useful ability, for much of one's knowledge about an object will ,lot be attached to that object but will be acquired from the inheritance hierarchy. In a spreading activation type of framework, a considerable amount of irrelevant network will be searched before some fact high up in the type hierarchy is found. In addition, it is very seldom that one wants to be able to access all facts involving an object; it is much more likely that a subset of relations is relevant.</Paragraph>
    <Paragraph position="1"> If desired, such associative links between objects can be simulated in our system. One could find all properties of an object ol (including those by inheritance) by retrieving all bindings of x in the query 3x,r ROLE(x,r,ol).</Paragraph>
    <Paragraph position="2"> The ease of access provided by the links in a semantic network is effectively simulated simply by using a hashing scheme on the structure of all ROLE predicates.</Paragraph>
    <Paragraph position="3"> While the ability to hash on structures to find facts is crucial to an efficient implementation, the details are not central to our point here.</Paragraph>
    <Paragraph position="4"> Another important form of indexing is found in Hendrix where his partition mechanism is used to provide a focus of attention for inference processes \[Grosz, 1977\]. This is just one of the uses of partitions. Another, which we did not need, provided a facility for scoping facts within logical operators, similar to the use of parentheses in FOPC. Such a focus mechanism appears in our system as an extra argument on the main predicates (e.g., HOLDS, OCCURS, etc.).</Paragraph>
    <Paragraph position="5"> Since contexts are introduced as a new class of objects in the language, we can quantify over them and otherwise talk about them. In particular, we can organize contexts into a lattice-like structure (corresponding to Hendrix's vistas for partitions) by introducing a transitive relation SUBCONTEXT.</Paragraph>
    <Paragraph position="7"> To relate contexts to the HOLDS predicate, a proposition p holds in a context c only if it is known to hold in c explicitly, or it holds in a super context of c.</Paragraph>
    <Paragraph position="8"> (A.9) v p,t,c,c' SUBCONTEXT(c,c,)A HOt.DS(p,c') --, HOLDS(p,c), As with the SUBTYPE relation, this axiom would defy an efficient implementation if the contexts were not organized in a finite lattice structure. Of course, we need axioms similar to (A,9) for the OCCURS and IS-RF_.AL predicates.</Paragraph>
  </Section>
class="xml-element"></Paper>