<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1604">
  <Title>Real-Time Stochastic Language Generation for Dialogue Systems</Title>
  <Section position="9" start_page="7" end_page="7" type="relat">
    <SectionTitle>
6 Related Work
</SectionTitle>
    <Paragraph position="0"> Stochastic work on the FERGUS system [Chen et al., 2002] uses a TAG grammar to produce a word lattice of possible realizations. The lattice is traversed to find the most likely path. The work in [Chen et al., 2002] generated sentences in 0.28 seconds for an Air-Travel Domain. This paper differs in that the input to FERGUS is a shallow syntactic tree, containing all lexemes and function words. In addition, surface syntax trees were mapped one-to-one with each template in the Air-Travel domain. There was little, if any, flexibility in the semantic input. This paper presents a result of 0.37 seconds that includes both the sentence planner, surface realizer, and a grammar that generates multiple realizations based on both syntax and semantics.</Paragraph>
    <Paragraph position="1"> Work was done on the Oxygen system [Habash, 2000] to improve the speed of the two-phased Nitrogen generator, a predecessor to HALogen. The work pre-compiled a declarative grammar into a functional program, thus removing the need to match rules during forest creation. This paper differs in that similar performance was achieved without the need for pre-compiling nor a more complex grammar syntax. This paper also described lexical movement and trickle down features not supported in Oxygen.</Paragraph>
    <Paragraph position="2"> Chambers [Chambers and Allen, 2004] used HALogen in a dialogue system and performed a human evaluation of the mixed syntax/semantic input. Their input converted their domain independent logical form into the HALogen input. This work differs in that we obviously did not use the HALogen system, but implemented a more efficient two-phased approach. The work by Chambers and Allen did not analyze runtime, perform sentence planning (not a full semantic input), nor provide results from the common String Accuracy metrics for comparison to other approaches.</Paragraph>
  </Section>
class="xml-element"></Paper>