<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1021">
  <Title>Modern Natural Language Interfaces to Databases: Composing Statistical Parsing with Semantic Tractability</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> This paper is the first to provide evidence that statistical parsers can support NLIs such as PRECISE.</Paragraph>
    <Paragraph position="1"> We identified the quandary associated with appropriately training a statistical parser: without special training for each database, the parser makes numerous errors, but creating a massive, labeled corpus of questions for each database is prohibitively expensive. We solved this quandary via light re-training of the parser's tagger and via PRECISE's semantic over-rides, and showed that in concert these methods enable PRECISE to rise from 61.9% accuracy to 94% accuracy on the ATIS data set. Even though PRECISE is database independent, its accuracy is comparable to the best of the database-specific ATIS NLIs developed in previous work (Table 2).</Paragraph>
  </Section>
class="xml-element"></Paper>