<?xml version="1.0" standalone="yes"?>
<Paper uid="P02-1017">
  <Title>A Generative Constituent-Context Model for Improved Grammar Induction</Title>
  <Section position="7" start_page="543210" end_page="543210" type="concl">
    <SectionTitle>
5 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have presented a simple generative model for the unsupervised distributional induction of hierarchical linguistic structure. The system achieves the best published unsupervised parsing scores on the WSJ-10 and ATIS data sets. The induction algorithm combines the benefits of EM-based parameter search and distributional clustering methods. We have shown that this method acquires a substantial amount of correct structure, to the point that the most frequent discrepancies between the induced trees and the treebank gold standard are systematic alternate analyses, many of which are linguistically plausible. We have shown that the system is not reliant on supervised POS tag input, and demonstrated increased accuracy, speed, simplicity, and stability compared to previous systems.</Paragraph>
    <Paragraph position="1"> 9The data likelihood is not shown exactly, but rather we show the linear transformation of it calculated by the system. 10Pereira and Schabes (1992) find otherwise for PCFGs.</Paragraph>
  </Section>
class="xml-element"></Paper>