<?xml version="1.0" standalone="yes"?>
<Paper uid="N04-4023">
  <Title>Feature Selection for Trainable Multilingual Broadcast News Segmentation</Title>
  <Section position="5" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
4 Results
</SectionTitle>
    <Paragraph position="0"> The overarching goal of our analysis was to identify multimedia events for each source that could be used to distinguish stories from commercials and other non-story segments in the broadcast. The results of our feature selection experiments revealed several features that were important for all seven sources we analyzed, as well as other features that were important for certain sources but not others. In this section we discuss our results.</Paragraph>
    <Paragraph position="1"> Table 2 shows the selected features for each of the broadcast sources. The first two columns show the name and type of each feature, as defined in Section 3, with start, end, and persist for durational metadata features, where relevant. The cells in the remaining columns show a &amp;quot;+&amp;quot; if the feature was automatically selected for the corresponding source; the cells are empty if the feature was not selected. The cell contains &amp;quot;n/a&amp;quot; if the feature was not available for the source; this was the case for the English-language TDT topic classification Of the hundreds of features we analyzed, only 14 were selected for at least one of the broadcast sources. The selected features varied greatly by source, with some features being used by only one or two of the seven sources.</Paragraph>
    <Paragraph position="2"> There are only three features that were selected for each of the seven sources: music segment persist, video fade start, and cue n-gram detected. Two other features, broadcaster logo detected and blue screen detected, were selected for all but one of the sources. One interesting result is that these features selected for all or most sources come from all four information sources: audio, language, image, and video.</Paragraph>
    <Paragraph position="3"> The significance of the selected features also varied by sources. For example, the blue screen detected feature was selected for all but one source; this feature thus has a much higher probability of occurring at certain points  ence of a blue screen is much more likely to occur during a commercial. For NWI it is most likely to occur at the start of a story, and for BBC it is most likely to occur outside stories and commercials. For CCTV it is equally likely in commercials and at the start of stories. For none of the sources is the blue screen likely to occur within a story.</Paragraph>
    <Paragraph position="4"> One of the most important features for all seven sources is the cue n-gram detected feature derived from the automatic speech recognition output. Interestingly, the n-grams that indicate story boundaries were extremely source-dependent, with almost no overlap in the lists of words derived across sources. Table 3 shows some examples of the highest-ranked n-grams from each of the sources (Arabic and Mandarin n-grams are shown manually translated into English).</Paragraph>
    <Paragraph position="5"> Source Top n-gram feature Aljazeera here is a report</Paragraph>
  </Section>
class="xml-element"></Paper>