<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-1528">
  <Title>k-NN for Local Probability Estimation in Generative Parsing Models</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Constraint Features for Training Set
Restriction
</SectionTitle>
    <Paragraph position="0"> We use the same k-NN estimation technique as Toutonava et al (2003) however we also found that restricting the number of examples in the training set used in a particular parameter estimation helped both in terms of accuracy and speed. We restricted the training sets by making use of constraint features whereby the training set is restricted to only those examples which have the same value for the constraint feature as the query instance.</Paragraph>
    <Paragraph position="1"> We carried out experiments using different sets of constraint features, some more restrictive than others. The mechanism we used is as follows: if the number of examples in the training set, retrieved using a particular set of constraint features, exceeds a certain threshold value then use a higher level of restriction i.e. one which uses more constraint features. If, using the higher level of restriction, the number of samples in the training set falls below a minimum threshold value then &amp;quot;back-off&amp;quot; to the less restricted set of training samples.</Paragraph>
  </Section>
class="xml-element"></Paper>