<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1004">
  <Title>Minimum Cut Model for Spoken Lecture Segmentation</Title>
  <Section position="9" start_page="31" end_page="31" type="concl">
    <SectionTitle>
7 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this paper we studied the impact of long-range dependencies on the accuracy of text segmentation. We modeled text segmentation as a graph-partitioning task aiming to simultaneously optimize the total similarity within each segment and dissimilarity across various segments. We showed that global analysis of lexical distribution improves the segmentation accuracy and is robust in the presence of recognition errors. Combining global analysis with advanced methods for smoothing (Ji and Zha, 2003) and weighting could further boost the segmentation performance.</Paragraph>
    <Paragraph position="1"> Our current implementation does not automatically determine the granularity of a resulting segmentation. This issue has been explored in the past (Ji and Zha, 2003; Utiyama and Isahara, 2001), and we will explore the existing strategies in our framework. We believe that the algorithm has to produce segmentations for various levels of granularity, depending on the needs of the application that employs it.</Paragraph>
    <Paragraph position="2"> Our ultimate goal is to automatically generate tables of content for lectures. We plan to investigate strategies for generating titles that will succinctly describe the content of each segment.</Paragraph>
    <Paragraph position="3"> We will explore how the interaction between the generation and segmentation components can improve the performance of such a system as a whole.</Paragraph>
  </Section>
class="xml-element"></Paper>