<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1109">
  <Title>Discriminative Slot Detection Using Kernel Methods</Title>
  <Section position="4" start_page="2" end_page="3" type="relat">
    <SectionTitle>
2.4 Related Work
</SectionTitle>
    <Paragraph position="0"> There have been a number of SVM applications in NLP using particular levels of syntactic information. (Lodhi et al., 2002) compared a word-based string kernel and n-gram kernels at the sequence level for a text categorization task. The experimental results showed that the n-gram kernels performed quite well for the task. Although string kernels can capture common word subsequences with gaps, its geometric penalty factor may not be suitable for weighting the long distance features. (Collins et al., 2001) suggested kernels on parse trees and other structures for general NLP tasks. These kernels count small subcomponents multiple times so that in practice one has to be careful to avoid overfitting. This can be achieved by limiting the matching depth or using a penalty factor to downweight large components.</Paragraph>
    <Paragraph position="1"> (Zelenko et al., 2003) devised a kernel on shallow parse trees to detect relations between named entities, such as the person-affiliation relation between a person name and an organization name. The so-called relation kernel matches from the roots of two trees and continues recursively to the leaf nodes if the types of two nodes match.</Paragraph>
    <Paragraph position="2"> All the kernels used in these works were applied to a particular syntactic level. This paper presents an approach for information extraction that uses kernels to combine information from different levels and automatically identify which information contributes to the task. This framework can also be applied to other NLP tasks.</Paragraph>
  </Section>
  <Section position="9" start_page="21" end_page="21" type="relat">
    <SectionTitle>
7 Related Work
</SectionTitle>
    <Paragraph position="0"> (Chieu et al., 2003) reported a feature-based SVM system (ALICE) to extract MUC-4 events of  terrorist attacks. The Alice-ME system demonstrated competitive performance with rule-based systems. The features used by Alice are mainly from parsing. Comparing with ALICE, our system uses kernels on dependency graphs to replace explicit features, an approach which is fully automatic and requires no enumeration of features. The model we proposed can combine information from different syntactic levels in principled ways. In our experiments, we used both word sequence information and parsing level syntax information. The training data for ALICE contains 1700 documents, while for our system it is just 100 documents. When data is sparse, it is more difficult for an automatic system to outperform a rule-based system that incorporates general knowledges.</Paragraph>
  </Section>
class="xml-element"></Paper>