<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1110">
  <Title>Advances in Discriminative Parsing</Title>
  <Section position="8" start_page="879" end_page="879" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> Our work has made advances in both accuracy and training speed of discriminative parsing. As far as we know, we present the first discriminative parser that surpasses a generative baseline on constituent parsing without using a generative component, and it does so with minimal linguistic cleverness. Our approach performs feature selection incrementally over an exponential feature space during training. Our experiments suggest that the learning algorithm is overfitting-resistant, as hypothesized by Ng (2004). If this is the case, it would reduce the effort required for feature engineering. An engineer can merely design a set of atomic features whose powerset contains the requisite information. Then, the learning algorithm can perform feature selection over the compound feature space, avoiding irrelevant compound features. null In future work, we shall make some standard improvements. Our parser should infer its own POS tags to improve accuracy. A shift-reduce parsing strategy will generate fewer training inferences, and might lead to shorter training times. Lastly, we plan to give the model linguistically more sophisticated features. We also hope to apply the model to other structured prediction tasks, such as syntax-driven machine translation.</Paragraph>
  </Section>
class="xml-element"></Paper>