<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1109">
  <Title>Discriminative Slot Detection Using Kernel Methods</Title>
  <Section position="10" start_page="21" end_page="21" type="concl">
    <SectionTitle>
8 Discussion and Further Works
</SectionTitle>
    <Paragraph position="0"> This paper describes a discriminative approach that can use syntactic clues automatically for slot filler detection. It outperformed a hand-crafted system on sparse data by considering different levels of syntactic clues. The result also shows that low level syntactic information can also come into play in finding events, thus it should not be ignored in the IE framework.</Paragraph>
    <Paragraph position="1"> For slot filler detection, several classifiers were trained to find names for each slot and there is no correlation among these classifiers. However, entity slots in events are often strongly correlated, for example the PER_IN and POST slots for management succession events. Since these classifiers take the same input and produce different results, correlation models can be used to integrate these classifiers so that the identification of slot fillers might benefit each other.</Paragraph>
    <Paragraph position="2"> It would also be interesting to experiment with the tasks that are more difficult for pattern matching, such as determining the on-the-job status property in MUC-6. Since events often span multiple sentences, another direction is to explore cross-sentence models, which is difficult for traditional approaches. For our approach it is possible to extend the kernel from one sentence to multiple sentences, taking into account the correlation between NE's in adjacent sentences.</Paragraph>
  </Section>
class="xml-element"></Paper>