<?xml version="1.0" standalone="yes"?>
<Paper uid="J92-3001">
  <Title>Making DATR Work for Speech: Lexicon Compilation in SUNDIAL</Title>
  <Section position="7" start_page="264" end_page="264" type="evalu">
    <SectionTitle>
6. Implementation and Coverage
</SectionTitle>
    <Paragraph position="0"> DIALEX is implemented in Quintus Prolog; benchmark tests indicate that compilation time is linear in the size of the lexicon. Development of very large scale lexicons is somewhat hindered by the current lack of effective debugging tools. We have, however, succeeded in constructing lexicons that cover a broad range of syntactic phenomena in both French and English. For example, the English DATR lexicon covers all distinctive lexical forms in our corpus gathered from simulations of flight enquiry dialogues (Fraser and Gilbert 1991). Furthermore, one of the major advantages of DATR's inheritance-based approach is ease of adding new lexical entries. For example, a large number of entries for cities is required in the flight information domain. With the definition of a CITY_PROP node to specify general properties of proper nouns identifying cities, individual cities such as Paris are simple and quick to define: (30) Paris: &lt;&gt; == CITY_PROP &lt;mor root&gt; == paris &lt;sem thecity value&gt; == paris.</Paragraph>
    <Paragraph position="1"> Extending the lexicon to include new verbs, especially verbs with idiosyncratic properties like try, takes more time and effort.</Paragraph>
    <Paragraph position="2"> This paper has been mainly concerned with the definition and compilation of lexicons for understanding. In fact, SUNDIAL applications are such that a production lexicon shares a considerable portion with its recognition counterpart. To this end, DIALEX has been adapted for compilation of a generation lexicon (Youd and McGlashan 1992). This is derived from the same DATR definitions but differs from the parser lexicons in that indexing is based on semantic type and complexity, rather than the surface string, and inflection is factored away from the lexical entries.</Paragraph>
  </Section>
class="xml-element"></Paper>