<?xml version="1.0" standalone="yes"?>
<Paper uid="W94-0301">
  <Title>Discovery and Format of Input Structures for Tactical Generation</Title>
  <Section position="12" start_page="7" end_page="7" type="concl">
    <SectionTitle>
CONCLUSIONS
</SectionTitle>
    <Paragraph position="0"> We have argued that a high-level specification of a speaker's goals in producing a discourse is not an appropriate input for a tactical generator. The latter must assume that some process has already decided what to say for whatever reasons, and requires a specification of what to say in symbols closely tied to potentially usable Surface linguistic expressions. These symbols should represent pragmatic or semantic (&amp;quot;rhetorical or non-rhetorical&amp;quot;) equivalence classes of expressions, and should thus be discovered by using as primary data speakers' equivalence judgments. We offered a discovery procedure based on paraphrase, intended to be more explicit and flexible than a procedure suggested by Ivir et al. We then compared our proposal with a later, independent suggestion due to Knott and Dale, finding that the K&amp;D procedure's ability to induce hyponymy as well as equivalence relationships was advantageous, but that its consideration of only certain syntactic classes as discourse relation expressors might be overly restrictive.</Paragraph>
    <Paragraph position="1"> We presented a number of arguments for tactical input specifications arranged as lattices rather than strictly as hierarchies.</Paragraph>
    <Paragraph position="2"> For illustration, we listed the discourse relations and rhetorical types needed to generate an attested paragraph-length discourse. We proposed a lattice composed of these as tactical generation input, and gave one example of the sort of discourse version which could be generated, assuming a suitable grammar and set of procedures.</Paragraph>
  </Section>
class="xml-element"></Paper>