<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-2026">
  <Title>A Domain-Specific Statistical Surface Realizer</Title>
  <Section position="11" start_page="155" end_page="155" type="concl">
    <SectionTitle>
6 Future Work
</SectionTitle>
    <Paragraph position="0"> Statistically-driven search offers a means of efficiently overgenerating sentences to express a given semantic structure. This is well-suited not only to our navigation domain, but also to other domains  The corpus was partially annotated for parse data, the full parses being automatically generated from the domain-trained language model. It was at this step that query extraction sometimes failed.</Paragraph>
    <Paragraph position="1"> with a relatively small vocabulary but variable and complex content structure. Our implementation of the idea of this paper is under development in a number of directions.</Paragraph>
    <Paragraph position="2"> A better option for robust language modeling is to use maximum entropy techniques to train a feature-based model. For instance, we can determine the probability of each child using such features as the POS, concept, and role of the parent and previous siblings. It may also be more effective to isolate linear precedence from the language model, introducing a non-trivial linearization step. Similarly, the lexicalization module can be improved on by using a more context-sensitive model.</Paragraph>
    <Paragraph position="3"> Using only a tree-based scoring function is likely to produce inferior results to one that incorporates a linear score. A weighted average of the dependency score with an n-gram model would already offer improvement. To further improve fluency, these could also be combined with a scoring function that takes longer-range dependencies into account, as well as penalizing extraneous content.</Paragraph>
  </Section>
class="xml-element"></Paper>