<?xml version="1.0" standalone="yes"?>
<Paper uid="N04-2006">
  <Title>Automatic Article Restoration</Title>
  <Section position="8" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> We applied the log-linear model on the article generation task, using features drawn from a statistical natural language parser and WordNet. The feature set was progressively enriched with information from both inside and outside the NP, semantics, and discourse context. The final feature set yielded very competitive results.</Paragraph>
    <Paragraph position="1"> We applied the same model to tackle the article restoration task, where sentences may have missing articles. On the one hand, article generation performance degraded significantly due to context extraction errors; this points to the need to adapt the tagger and parser to ungrammatical sentences. On the other hand, the articles that were already present in the sentence provided strong hints about the correct article; this points to the need for better methods for estimating the underlying confusion matrix of a sentence.</Paragraph>
  </Section>
class="xml-element"></Paper>