<?xml version="1.0" standalone="yes"?>
<Paper uid="C90-2062">
  <Title>An Explanation Facility for a Grammar Writing System</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
5. Supplementary Knowledge
</SectionTitle>
    <Paragraph position="0"> The JEMAtt system explanation facility is very tightly coupled to the execution trace, and hence, to the grammar knowledge base.</Paragraph>
    <Paragraph position="1"> Advanced work with explanation systems have strongly indicated the need for supplementary knowledge to improve the explanation provided, and lately, some experts have even argued for the complete decoupling of knowledge used by the explanation system from the knowledge used by the expert system.</Paragraph>
    <Paragraph position="2"> Future work on the explanation facility of JEMAH will include supplementary knowledge to improve the quality and range of explanations provided.</Paragraph>
    <Paragraph position="3"> In describing NEOMYCIN, Clancey and Letsinger (1984, pg 380) stated that: &amp;quot;To explain diagnosis, it is useful to have psychological model of problem solving.</Paragraph>
    <Paragraph position="4"> In particular, we need to incorporate into our model the medical knowledge and strategies an expert uses for initial problem formulation.&amp;quot; In the grammar writing system context, there is a similar need for a linguistic knowledge model, as well as a parsing/generation strategy model. Normally, the linguistic knowledge refers to knowledge about the specific grammar under consideration, and does not include knowledge about linguistics in general, although this possibility should not be ruled out in future work. The pre-compilation process in the JEMAH system (see Tong, 1989) extracts information about rules' relationships, and this can be considered as a kind of meta-knowledge about the grammar. Furthermore, since this meta-knowledge is used in JEMAtt to optimize the control strategy, it seems fair to assume that it can also contribute towards explaining the parsing/generation strategy. This single concept is obviously insufficient to provide comprehensive explanation to the user, but does serve as an example of the kind of supplementary knowledge required.</Paragraph>
    <Paragraph position="5"> A second feature which is already available within the JEMAH system is that of grammar partitioning. Rules are grouped according to certain linguistic properties, and this can be used as a source of supplementary knowledge in providing explanations. A more elaborate method would be to construct a classification scheme in the form of a hierarchical tree structure, like the refinement structure in Xplain (see pg. 392 in Swartout, 1984).</Paragraph>
    <Paragraph position="6"> The variables and their values used in a .grammar obviously play a very important role an explanation; in NEOMYCIN, for example, relationships between variables and new variables are created solely for the purpose of explanation. In JEMAH, this type of meta-knowledge about variables may include, for example, information on mutual exclusion (between morphosyntactic class and word category), equivalent set of values (valencies 1 and 2) and hierarchical relationship (subcategory SUBA of category A).</Paragraph>
  </Section>
class="xml-element"></Paper>