<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0505">
  <Title>Summarising Legal Texts: Sentential Tense and Argumentative Roles</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
4 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> The work reported forms the initial stages in the development of a automatic text summarisation system for judicial transcripts from the House of Lords. We have presented an initial annotation scheme for the rhetorical structure of the domain, assigning a label indicating the argumentative role of each sentence in a portion of the corpus. A number of sophisticated linguistic tools have been described that identify tense information. Finally, correlation scores were presented illustrating the utility of this information.</Paragraph>
    <Paragraph position="1"> The next phase of the project will involve refining our annotation scheme. Once we have done this, we will create formal instructions and complete the annotation of the larger corpus. As part of the process of annotating our corpus, we will continue to examine possible indicators of the rhetorical role for a sentence.</Paragraph>
    <Paragraph position="2"> We are also interested in improving the tools we use to identify tense features. One way to do this is retraining the clause identifier. The legal language of the HOLJ domain is considerably different than the expository newspaper text from the Penn Treebank. Furthermore, the Penn Treebank is American English. Ideally, we would like to hand-annotate a portion of the legal judgments with syntactic parse information and train a clause identifier from this. However, this kind of work is very labour intensive and a more realistic approach to ensuring that the training data is slightly more representative might be to retrain the clause identifier on a corpus of British English like the British National Corpus (Burnage and Dunlop, 1992).</Paragraph>
    <Paragraph position="3"> Finally, as mentioned above, we are specifically interested in employing feature construction and selection techniques for identifying the relationship between tense features. We are also interested in employing feature mining techniques for automatically identifying cue phrases within sentences. This could be similar to (Lesh et al., 1999), where sequential features are mined from the textual context for a context-sensitive approach to spelling correction.</Paragraph>
  </Section>
class="xml-element"></Paper>