<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0915">
  <Title>Interpreting Communicative Goals in Constrained Domains using Generation and Interactive Negotiation</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Perspectives
</SectionTitle>
    <Paragraph position="0"> We have presented a practical approach to content analysis at the level of communicative goals, in which a strong emphasis is put on document content well-formedness. Providing the expert is willing to spend enough time, the communicative content of a document can be interactively built.</Paragraph>
    <Paragraph position="1"> The better the system performance, the less time is needed to identify the correct candidate content representation. The fact that the expert can read the corresponding normalized text (on the MDA view) can help guarantee that the whole validation process was carried out correctly.</Paragraph>
    <Paragraph position="2"> We now need to grow our grammars for Unix commands and drug leaflets, and to enrich our test corpus of annotated documents (raw text and abstract semantic structure)6 for these classes in 6Documents for the test corpus can be obtained by using order to be able to carry out evaluation. Evaluation should be performed on two aspects. First, the performance of fuzzy inverted generation could be measured, for a given normalization model and on a given source of documents, by the position and relative score of the candidate content representation corresponding to the normalized document. Second, we want to evaluate the usability of our user interface supporting interactive negotiation.</Paragraph>
    <Paragraph position="3"> An evaluation corresponding to the number of steps and the time needed to obtain the normalized version of a document would be a good indicator.</Paragraph>
    <Paragraph position="4"> Moreover, we plan to implement the possibility for the expert to add new formulations found in documents to better match communicative goals in subsequent normalizations. It will then be interesting to evaluate the impact of this kind of supervised learning on system performance and user acceptance. Our next challenge will be to investigate how our approach can be applied to documents in less-constrained domains for which normalization models cannot be entirely built a priori.</Paragraph>
  </Section>
class="xml-element"></Paper>