<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0703">
  <Title>Event Clustering on Streaming News Using Co-Reference Chains and Event Words</Title>
  <Section position="6" start_page="26" end_page="26" type="concl">
    <SectionTitle>
6 Concluding Remarks
</SectionTitle>
    <Paragraph position="0"> This paper presented an approach for event clustering on streaming news based on both co-reference chains and event words. The experimental results using event words only outperform the results using the co-reference chains only. Nevertheless, as to the combination of co-reference chains and event words in event clustering, the experimental results show that the introduction of co-reference chains can improve the performance of event clustering using event words much. To model the temporal behavior of event clustering of streaming news, a dynamic threshold setting using time decay function and spanning window size is proposed. The experimental results, using TDT's evaluation metric - say, detection cost, show that the dynamic threshold is useful. .</Paragraph>
    <Paragraph position="1"> We believe that the improvement of multi-document co-reference resolution will have great impact on temporal event clustering. In order to further improve our performance in even clustering on streaming news, there are still future works needed to be studied: (1) In order to verify the significance of the experimental results, statistical test is needed. (2) Instead of hand-tagging method, we will introduce automatic co-reference resolution tools to create large scale test corpus and conduct large scale experiments.</Paragraph>
    <Paragraph position="2"> (3) When the length of document is variable, the fixed number of representative sentences may lose many important sentences to degrade the performance of event clustering. The dynamic number of representative sentences for each document according to its length is introduced.</Paragraph>
    <Paragraph position="3"> (4) As the news stories are reported incrementally instead of being given totally in the on-line event clustering, the computation of event words is an important issue.</Paragraph>
    <Paragraph position="4"> (5) Apply the extracted sentences for each document to generate event-based short summary.</Paragraph>
  </Section>
class="xml-element"></Paper>