<?xml version="1.0" standalone="yes"?>
<Paper uid="P87-1014">
  <Title>Functional Unification Grammar Revisited</Title>
  <Section position="5" start_page="101" end_page="101" type="concl">
    <SectionTitle>
3 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have shown how constraints on generation can be represented separately from representation of syntactic structure in FUG. Such an ability is attractive because it means that the constraint can be stated once in the grammar and can be applied to a number of different syntactic rules.</Paragraph>
    <Paragraph position="1"> In contrast, m augmented context free based generation systems, constraints must be stated locally as part of individual syntactic rules to which they apply. As a result' constraints must be duplicated. Since a main focus in language generation research has been to identify constraints on choice, the ability to represent constraints clearly and efficiently is an important one.</Paragraph>
    <Paragraph position="2"> Representing constraints separately is only useful for global constraints, of course. Some constraints in language generation are necessarily local and must be represented in FUG as they would in augmented context free based systems: as part of the syntactic structures to which they apply. Furthermore, information for some constraints may be more easily represented outside of the grammar. In such cases, using a function caLl to other components of the system, as is done in NIGEL, is more appropriate. In fact, this ability was implemented as part of a FUG in TELEGRAM \[I\]. But for global constraints for which information is available in the grammar, FUG has an advantage over other systems.</Paragraph>
    <Paragraph position="3"> Our reimplementation of FUG has demonstrated that efficiency is not as problematic as was previously believed. Our version of FUG, running in PSL on an IBM 4381, runs 9Often the surface sentences gen~ated are the same, but the syntactic structure built in producing the sentence differs.</Paragraph>
    <Paragraph position="4"> faster than Rubinoff's version of MUMBLE in Symbolics 3600 Zetalisp for the same set of input sentences.</Paragraph>
    <Paragraph position="5"> Furthermore, we have shown that we were able to achieve a slightly better speed-up over TEXT's old tactical component than Rubinoff's MUMBLE using a comparison that takes into account different machines. Given that FUG can produce sentences in time comparable to a deterministic generator, efficiency should no longer be an issue when evaluating FUG as a generation system.</Paragraph>
  </Section>
class="xml-element"></Paper>