<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-3146">
  <Title>TOWARDS A NEW GENERATION OF TERMINOLOGICAL RESOURCES: AN EXPERIMENT IN BUILDING A TERMINOLOGICAL KNOWLEDGE BASE</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
3 METHODOLOGICAL ISSUES
</SectionTitle>
    <Paragraph position="0"> In deciding on a preliminary methodology for our work, we naturally turned to the literature of both computational lexicology and knowledge engineering for inspiration, with little success.</Paragraph>
    <Paragraph position="1"> Even the world's largest knowledge acquisition project, CYC (Lenat and Guha 1990), provides only sparse methodological guidance (Skuce in press a). To date, our methodology has remained essentially grounded in that traditionally used by terminologists (Sager 1990), a reasonable starting point when one considers that, although terminologists have traditionally not built TKBs, conceptual analysis has always been a central part of theft work 7 nonetheless. Tenninologists are keenly aware of the importance of a certain depth of domain knowledge, and many of the conceptual analysis techniques that are advocated in the knowledge engineering literature - e.g. describing conceptual characteristics through attribute-value pairs, sketching concept networks - have been part of the terminological methodology for years.</Paragraph>
    <Paragraph position="2"> The methodology we have developed can be very superficially described as follows8: 1) After introductory reading on the domain, the principal 7 A detailed analysis of the role of conceptual analysis in terminology can be found in Meyer (in press).</Paragraph>
    <Paragraph position="3"> 8 A detailed description of the methodology can be found in Meyer et al. 1992 (in press).</Paragraph>
    <Paragraph position="4"> conceptual relations are sketched out, with the goals of establishing the boundaries of the domain and identifying the subdomains, from which the most fundamental is selected for further analysis.</Paragraph>
    <Paragraph position="5"> 2) A template of conceptual characteristics is established for the selected subdomain; it is used as a guide to the knowledge acquisition process, and inherits to lower levels of the conceptual hierarchy, where it can also be specialized. 3) Conceptual and linguistic information are entered into the system as they are acquired (mainly from the corpus). A concept is integrated into a hierarchy whenever its superconcept is known; when it is not, or when there is some doubt, the concept is labelled &amp;quot;unclassified&amp;quot; (unchtssified concepts can occur at any level in a hieraa~'hy, i.e. there can be different &amp;quot;degrees&amp;quot; of classification). 4) Intensional definitions are constructed with the help of the Characteristic Comparison Matrix. Steps 2-4 are then repeated for the next subdomain, until all sub-domains have been completed.</Paragraph>
    <Paragraph position="6"> A number of the more troublesome methodological issues with which we are currently grappling are briefly outlined below. Knowledge acquisition &amp;quot;paths&amp;quot;. Knowledge acquisition is not a journey down a straight path: there is no visible &amp;quot;goal&amp;quot;. Although we have followed traditional terminology method~ ology in adopting a subdomain-oriented, top-down approach to acquisition, it often seems desirable to deviate from the principal subdomain when one encounters related te.rms in a neighbouring subdomain or field, and to work bottom-up as well as top-down within the principal subdomain. While the subdomains we have investigated so far (the majority of the concepts belonging to the semantic class of artefacts) are dominated by generic-specific and part-whole relations, subdomains related to other semantic classes may be more amenable to analysis based on different relations, as has been pointed out, for example, in the literature on the WordNet project (Miller 1991, Fellbaum 1991).</Paragraph>
    <Paragraph position="7"> Multidimensionality. While terminologists are well aware that a given domain can be subdivided in different ways, depending on the expert's point of view, they have not traditionally attempted to account for it in any serious way, since this is difficult to do with pencil-and-paper techniques. Some problems that arise are how such &amp;quot;multidimensionality&amp;quot; affects knowledge acquisition &amp;quot;paths&amp;quot;, how the technology can better support the maintenance of conceptual clarity as the number of dimensions grows (for example, through masking facilities of the kind we have implemented), how multidimensionality can be reflected in definition construction, etc.</Paragraph>
    <Paragraph position="8"> ACRES DE COLING-92, NANTES, 23-28 Ao~r 1992 9 S 9 PROC. OF COLING-92, NANq ES, Aua. 23-28, 1992 Validation. Validation by experts and other terminologists, which has always been an important part of terminology work, is complicated in our approach by the fact that our TKB is very hypertext-like, and thus requires revision techniques that go beyond those normally applied to &amp;quot;flat&amp;quot; texts such as conventional terminology records. We need to investigate further at which points validators should be consulted, what elicitation techniques should be used at each point, how to handle inconsistencies in opinion, etc.</Paragraph>
    <Paragraph position="9"> Increased automation. To date, our research efforts are oriented towards facilitating (and not automating) the knowledge acquisition process for developing and implementing our concept of a TKB. This is consistent with the majority of knowledge acquisition projects in the world, including CYC. As the concept of a TKB becomes clearer, however, we hope that TKB and LKB researchers will collaborate in exploring possibilities for a more automated approach to acquisition.</Paragraph>
  </Section>
class="xml-element"></Paper>