<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0601">
  <Title>Techniques for Text Planning with XSLT</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> We have described a successful implementation of the classic NLG pipeline that uses XSLT template processing as a top-down rule-expanding planner.</Paragraph>
    <Paragraph position="1"> Implementing the necessary steps using XSLT was generally straightforward, and the ability to use offthe shelf, well-tested and well-documented tools such as Java and Xalan adds to the ease of implementation and robustness.</Paragraph>
    <Paragraph position="2"> Our implementation creates logical forms for the OpenCCG realizer; this allows the realizer to be used for those parts of the generation process to which XSLT is less well-suited. We also take advantage of the fact that the realizer uses statistical language models in its search for a surface form by generating logical forms with embedded alternatives, allowing the realizer to choose the one to use. This both adds robustness to the system and eliminates the need for backtracking within the text planner.</Paragraph>
    <Paragraph position="3"> The current implementation is fast and reliable: it correctly processes all input from the dialogue manager, and the time it takes to do so is relatively short compared to that required by the other processing and communication tasks in COMIC.</Paragraph>
    <Paragraph position="4"> The entire COMIC demonstrator will shortly be evaluated. As part of this evaluation, we plan to measure users' recall of the information that the system presents to them, where that information is generated at different levels of detail.</Paragraph>
    <Paragraph position="5"> At the moment, the logical form for each message is created in isolation. In future versions of COMIC, we plan to use ideas from centering theory to help ensure coherence by planning a coherent sequence of logical forms for a description. We will implement this in a way similar to that described by Kibble and Power (2000).</Paragraph>
    <Paragraph position="6"> We will also incorporate a model of the user's preferences into a later version of COMIC. The model will be used both to rank the options to be presented to the user, and to generate user-tailored descriptions of those options, as in FLIGHTS (Moore et al., 2004).</Paragraph>
    <Paragraph position="7"> Finally, we plan to extend the use of data-driven techniques in the realizer, and to make use of these techniques to help in choosing among alternatives in the other COMIC output modalities.</Paragraph>
  </Section>
class="xml-element"></Paper>