<?xml version="1.0" standalone="yes"?>
<Paper uid="P00-1060">
  <Title>An Information-Theory-Based Feature Type Analysis for the Modelling of Statistical Parsing SUI Zhifang +++ , ZHAO Jun + , Dekai WU + +</Title>
  <Section position="18" start_page="121" end_page="121" type="evalu">
    <SectionTitle>
G7A Approach
</SectionTitle>
    <Paragraph position="0"> We use the following greedy algorithm to select the optimal feature type combination.</Paragraph>
    <Paragraph position="1"> In building a model, the first feature type to be selected is the feature type which has the largest predictive information quantity for the prediction of the derivation rule among all of the feature type candidates, that is,</Paragraph>
    <Paragraph position="3"> Where Ohm is the set of candidate feature types.</Paragraph>
    <Paragraph position="4"> Given that the model has selected feature type  G16 , the next feature type to be added into the model is the feature type which has the largest predictive information gain in all of the feature type candidate except  Among the feature types mentioned above, the optimal feature type combination (i.e. the feature type combination with the largest predictive information summation) which is composed of 6 feature types is, the headword of the current node (type1), the headword of the parent node (type2), the headword of the grandpa node (type3), the first word in the objective word sequence(type4), the first word in the objective word sequence which have the possibility to act as the headword of the current constitute(type5), the headword of the right brother node(type6). The cumulative predictive information summation is showed in Figure-2  type1 type2 type3 type4 type5 type6 feature type cummulative predicting information summation Figure-2: The cumulative predictive information summation of the feature type combinations</Paragraph>
  </Section>
class="xml-element"></Paper>