<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1666">
  <Title>Loss Minimization in Parse Reranking</Title>
  <Section position="8" start_page="565" end_page="565" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> This paper considers methods for the estimation of expected loss for parse reranking tasks. The proposed methods include estimation of the loss from a probabilistic model, estimation from a discriminative classifier, and learning of the loss using a specialized kernel. An empirical comparison of these approaches on parse reranking tasks is presented. Special emphasis is given to data-defined kernels for reranking, as they do not require the introduction of any additional domain knowledge not already encoded in the probabilistic model.</Paragraph>
    <Paragraph position="1"> The best approach, estimation of the loss on the basis of a discriminative classifier, achieves very significant improvements over the baseline generative probabilistic models and the discriminative classifier itself. Though the largest improvement is demonstrated in the measure which corresponds to the considered loss functional, other measures of accuracy are also improved. The proposed method achieves 90.0% F1 score on the standard Wall Street Journal parsing task when the SSN neural network is used as the probabilistic model and VP with a TOP Reranking kernel as the discriminative classifier.</Paragraph>
  </Section>
class="xml-element"></Paper>