<?xml version="1.0" standalone="yes"?>
<Paper uid="C86-1138">
  <Title>Forward message Smith CMUA and these would have the nesting structure: \[ForwardAction HeadForm FORWARD MsgObj \[MsgObjDesc HeadForm MESSAGE\] MsgRecipientObj \[MaiIAdrOesc HeadForm SMITH llost \[LocationDesc HeadForm CMUA\]\]\] \[ForwardAct ion HeadForm FORWARD MsgObj \[MsgObjDesc HeadForm MESSAGE\] CCRecipientOb.j \[MailAdrDBsc HeadForm S,M I TIt Host \[Locat.i onDesc HeadEorm CMUA\]\]\] \[ForwardAction tleadForn~ FORWARD MsgObj \[MsgObjDesc HeadForm MESSAGE MsgOriginObj \[MailAdrgesc HeadForm SMITlt Itost \[LocationDesc</Title>
  <Section position="4" start_page="591" end_page="591" type="evalu">
    <SectionTitle>
5. Summary and Future Directions
</SectionTitle>
    <Paragraph position="0"> We have explored an approach to parsing restricted-domain speech based on semantic caseframes. The approach was shown capable of dealing with the uncertainties and ambiguities of speech and the common ungrammaticalities. We argued thai a caseframe approach was better suited to these problems than more traditional.</Paragraph>
    <Paragraph position="1"> network-based approaches. This suitability was attributed to the high degree of abstraction with which caseframes represen't their linguistic information, and the corresponding flexibility in interpretation this allows. A simple implementation using this approach was described with a worked example.</Paragraph>
    <Paragraph position="2"> We envision continued development of cur system and enhancements to our approach in several directions: Our current approach relies too heavily on finding caseframe header words. While most are multi-syllable and easily recognizable at the acoustic level, many (e.g. 'send') are not.</Paragraph>
    <Paragraph position="3"> We are looking at ways to drive the recognition from the most reliably recognized words, whether they correspond to caseframe headers or not.</Paragraph>
    <Paragraph position="4"> * * Most of tile syntactic knowledge used by our current system is embedded in the code. While this makes for elficient and robust recognition, it poses obvious problems for syntactic extensibility and maintainability. We are looking at ways of separating out the syntactic knowledge, while retaining the power and flexibility inherent in specifying a restricted-domain language through caseframes, rather than (say) rewrite rules.</Paragraph>
    <Paragraph position="5"> o The nature of the interpretation performgd by the present system causes it to operate at large multiples of real.time. We are looking at methods of compiling the caseframe grammar into more efficient recognition systems, with the eventual goal of real.time operation, while retaining our current flexibility and robustness.</Paragraph>
  </Section>
class="xml-element"></Paper>