<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-1052">
  <Title>Extracting Relations with Integrated Information Using Kernel Methods</Title>
  <Section position="4" start_page="419" end_page="420" type="relat">
    <SectionTitle>
3 Related Work
</SectionTitle>
    <Paragraph position="0"> Collins et al. (1997) and Miller et al. (2000) used statistical parsing models to extract relational facts from text, which avoided pipeline processing of data. However, their results are essentially based on the output of sentence parsing, which is a deep processing of text. So their approaches are vulnerable to errors in parsing. Collins et al. (1997) addressed a simplified task within a confined context in a target sentence.</Paragraph>
    <Paragraph position="1"> Zelenko et al. (2003) described a recursive kernel based on shallow parse trees to detect personaffiliation and organization-location relations, in which a relation example is the least common sub-tree containing two entity nodes. The kernel matches nodes starting from the roots of two sub-trees and going recursively to the leaves. For each pair of nodes, a subsequence kernel on their child nodes is invoked, which matches either contiguous or non-contiguous subsequences of node. Compared with full parsing, shallow parsing is more reliable. But this model is based solely on the out- null put of shallow parsing so it is still vulnerable to irrecoverable parsing errors. In their experiments, incorrectly parsed sentences were eliminated.</Paragraph>
    <Paragraph position="2"> Culotta and Sorensen (2004) described a slightly generalized version of this kernel based on dependency trees. Since their kernel is a recursive match from the root of a dependency tree down to the leaves where the entity nodes reside, a successful match of two relation examples requires their entity nodes to be at the same depth of the tree.</Paragraph>
    <Paragraph position="3"> This is a strong constraint on the matching of syntax so it is not surprising that the model has good precision but very low recall. In their solution a bag-of-words kernel was used to compensate for this problem. In our approach, more flexible kernels are used to capture regularization in syntax, and more levels of syntactic information are considered. null Kambhatla (2004) described a Maximum Entropy model using features from various syntactic sources, but the number of features they used is limited and the selection of features has to be a manual process.</Paragraph>
    <Paragraph position="4">  In our model, we use kernels to incorporate more syntactic information and let a Support Vector Machine decide which clue is crucial. Some of the kernels are extended to generate high order features. We think a discriminative classifier trained with all the available syntactic features should do better on the sparse data.</Paragraph>
  </Section>
class="xml-element"></Paper>