<?xml version="1.0" standalone="yes"?>
<Paper uid="W90-0122">
  <Title>The Computer Generation of Speech with Dlscoursally and Semantically Motivated Intonation</Title>
  <Section position="1" start_page="0" end_page="0" type="abstr">
    <SectionTitle>
UK
Abstract
</SectionTitle>
    <Paragraph position="0"> The paper shows how it is possible, in the framemork of a gygtemle funetlonal g:r~mmat, (SFG) approach to the semantics of aatttral language, to generate an output with intonation that is motivated semantically and discoursally.</Paragraph>
    <Paragraph position="1"> Most of ~ wurk reported has already been *accuse, fully implem~4 in GF.N~SY$ (the very large generator of the COMMUNAL Pro~eet; see Appendix 1). A major feature is that it does am flint generate a syntax tree and words, and then impose intonational coatogm on them (as is a common aggroaeh in modelling intonation); rather, it generates the various intonational feamrta diteetJy, as it is generating richly labe.llad m:rucratea (as are typical in SFG), and the associated items. ~ claim is not that the model p~ he,C/ solves all the problems of generating intonation, but that it points a way forward that makea natural links with semantics and ditu:ourse. A secondaJty perpoe~ of this paler is to demonstraW,, for one of many possible areas of NLO that could have been choc~n, that there is still much important work to be done in '~nt.e, mev geaeratkm'. I do this tn order to tefut(c) the augge~On, ~easiOnally head at recent eel that the major wodC/ io 'sentence generation' has already been done, and that the main (only'?) area of s/gnifleanee in NLG is ia ltighet level planning. In my expea'ieace the two are inteaxlependen% and we should expect significant developments at eveey level in the years to come.</Paragraph>
    <Paragraph position="2">  1. Purpose and Scope  The aspect of NoZural Language Generation (NLG) to be described here is the generation of spoken text that has intonation, where that intonation is motivated both semantically (i.e, in terms of the semantics - in a broad sense of the term to be clarified soon - of sentences) and dJscoursally (Le. in terms of vhat the discourse p!~..;.ng compotmnt specifies).&amp;quot; Any specification of intonation requires, of course, to be integrated with an adapted version of a speech synthesizer (e.g. one that draws on one of the currently available systems that attempts - inevitably with Appendix 1 for a brief overview of the project).</Paragraph>
    <Paragraph position="3"> L As will be C/lC/ar from what follow the model presented hcte owes a great deal to two people in particular:. Michael Halliday and Pad Teach. Through them, I am well awate, there is a debt to many others, too numerous to mention, who have worked in the diffvgult field of intonation in F~glish. \[ am grateful too for early encouragement in this area from Gillian Brown (whose work is drawn on also by Paul Teach), nod for the regulardeg ongoing stimulus of many good explorations of ideas in this and other arco, s with my colteagmt Gordon Tucker. But none of there should bc blamed for the inevitable crudities, infelicities and no aoubt ca'mrs in the model described here; these are mine.</Paragraph>
    <Paragraph position="4"> The approach is very different from that in MITalk (Allen 1986), which is essentially a text-to-speech system. So far as I am aware, the only generative model prior to ours that attempts to generate intonation that is motivated semanticaIly and discoursaIly is the impressive work of hard, Houghton, Pearson and their colleagues at Sussex (Houghton and hard 1987) and Houghton and Pearson 1988). Its limitation is the very small size of its syntax, lexis, semantics and working domain.</Paragraph>
    <Paragraph position="5"> We see our work in the COMMUNAL Project as being to build on their important achievement.</Paragraph>
    <Paragraph position="6"> (But see Appendix 2 for what we do not attempt.)</Paragraph>
  </Section>
class="xml-element"></Paper>