<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-2021">
  <Title>Initial Study on Automatic Identification of Speaker Role in Broadcast News Speech</Title>
  <Section position="8" start_page="83" end_page="83" type="concl">
    <SectionTitle>
5 Summary and Future Work
</SectionTitle>
    <Paragraph position="0"> In this paper we have reported an initial study of speaker role identification in Mandarin broadcast news speech using the HMM and Maxent tagging approaches. We find that the conditional Maxent generally performs slightly better than the HMM, and that their combination out-performs each model alone. The HMM and the Max-ent model show differences in identifying different roles. The impact of contextual role information is also examined for the two approaches, and a significant gain is observed when contextual information is modeled. We find that the beginning and the end sentences in a speaker's turn are good cues for role identification. The overall classification performance in this study is similar to that reported in (Barzilay et al., 2000); however, the chance performance is quite different (35% in that study). It is not clear yet whether it is because of the difference across the two corpora or languages.</Paragraph>
    <Paragraph position="1"> The Maxent model provides a convenient way to incorporate various knowledge sources. We will investigate other features to improve the classification results, such as name information, acoustic or prosodic features, and speaker clustering results (considering that the same speaker typically has the same role tag). We plan to examine the effect of using speech recognition output, as well as automatic speaker segmentation and clustering results. Analysis of difference news sources may also reveal some interesting findings. Since our working hypothesis is that speaker role information is important to find structure in broadcast news, we will investigate whether and how speaker role relates to downstream language processing applications, such as summarization or question answering.</Paragraph>
  </Section>
class="xml-element"></Paper>