<?xml version="1.0" standalone="yes"?>
<Paper uid="P82-1004">
  <Title>What's in a Semantic Network?</Title>
  <Section position="9" start_page="24" end_page="26" type="concl">
    <SectionTitle>
7. Discussion
</SectionTitle>
    <Paragraph position="0"> We have argued that the appropriate way to design knowledge representations is to identify those inferences that one wishes to facilitate. Once these are identified, one can then design a specialized limited inference mechanism that can operate on a data base of first order  facts. In this fashion, one obtains a highly expressive representation language (namely FOPC), as well as a well-defined and extendable retriever.</Paragraph>
    <Paragraph position="1"> We have demonstrated this approach by outlining a portion of the representation used in ARGOT, the Rochester Dialogue System \[Allen, 1982\]. We are currently extending the context mechanism to handle time, belief contexts (based on a syntactic theory of belief \[Haas, 1982\]), simple hypothetical reasoning, and a representation of plans. Because the matcher is defined by a set of axioms, it is relatively simple to add new axioms that handle new features.</Paragraph>
    <Paragraph position="2"> For example, we are currently incorporating a model of temporal knowledge based on time intervals \[Allen, 1981a\]. This is done by allowing any object, event, or relation to be qualified by a time interval as follows: for any untimed concept x, and any time interval t, there is a timed concept consisting of x viewed during t which is expressed by the term (t-concept x t).</Paragraph>
    <Paragraph position="3"> This concept is of type (TIMED Tx), where Tx is the type of x. Thus we require a type hierarchy of timed concepts that mirrors the hierarchy of untimed concepts. Once this is done, we need to introduce new built-in axioms that extend the retriever. For instance, we define a predicate, DURING(a,b), that is true only if interval a is wholly contained in interval b. Now, if we want the retriever to automatically infer that if relation R holds during an interval t, then it holds in all subintervals of t, we need the following built-in axioms. First, DURING is transitive: (A.10) V a,b,c DURING(a,b) A DURING(b,c) --, DURING(a,c) Second, if P holds in interval t, it holds in all subintervals of t.</Paragraph>
    <Paragraph position="4"> (A.11) v p,t,t',c HOLDS(t-concept(p,t),c) A DURING(t' ,t) ---, HOLDS(t-concept(p,t'),c).</Paragraph>
    <Paragraph position="5"> Thus we have extended our representation to handle simple timed concepts with only a minimal amount of analysis.</Paragraph>
    <Paragraph position="6"> Unfortunately, we have not had the space to describe how to take the specification of the retriever (namely axioms (A.1) - (A.11)) and build an actual inference program out of it. A technique for building such a limited inference mechanism by moving to a meta-logic is described in \[Frisch and Allen, 1982\]. One of the more interesting consequences of this approach is that it has led to identifying various difference modes of retrieval that are necessary to support a natural language comprehension task, We have considered so far only one mode of retrieval, which we call provability mode. In this mode, the query must be shown to logically follow from the built-in axioms and the facts in the knowledge base. While this is the primary mode of interaction, others are also important. In consistency mode, the query is checked to see if it is logically consistent with the facts in the knowledge base with respect to the limited inference mechanism. While consistency in general is undecidable, with respect to the limited inference mechanism it is computationally feasible. Note that, since the retriever is defined by a set of axioms rather than a program, consistency mode is easy to define.</Paragraph>
    <Paragraph position="7"> Another important mode is compatibility mode, which is very useful for determining the referents of description. A query in compatibility mode succeeds if there is a set of equality and inequality assertions that can be assumed so that the query would succeed in provability mode. For instance, suppose someone refers to an event in which John hit someone with a hat. We would like to retrieve possible events that could be equal to this. Retrievals in compatibility mode are inherently expensive and so must be controlled using a context mechanism such as in \[Grosz, 1977\]. We are currently attempting to formalize this mode using Reiter's non-monotonic logic for default reasoning.</Paragraph>
    <Paragraph position="8"> We have implemented a version of this system in HORNE \[Allen and Frisch, 1981\], a LISP embedded logic programming language. In conjunction with this representation is a language which provides many abbreviations and facilities for system users. For instance, users can specify what context and times they are working with respect to, and then omit this information from their interactions with the system.</Paragraph>
    <Paragraph position="9"> Also, using the abbreviation conventions, the user can describe a relation and events without explicitly asserting the TYPE and ROLE assertions. Currently the system provides the inheritance hierarchy, simple equality reasoning, contexts, and temporal reasoning with the DURING hierarchy.</Paragraph>
  </Section>
class="xml-element"></Paper>