<?xml version="1.0" standalone="yes"?>
<Paper uid="C02-1004">
  <Title>Combining unsupervised and supervised methods for PP attachment disambiguation</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> We have shown that unsupervised approaches to PP attachment disambiguation are about as factor correct incorrect accuracy threshold noun attachment 5.47; 5.97 2400 469 83.65% 0.020 verb attachment 1219 381 76.19% 0.109 total 3619 850 80.98% decidable test cases 4469 (of 4469) coverage: 100% Table 5: Attachment accuracy for the combination of Back-off and cooccurrence values for the CZ test set (based on training over the NEGRA test set). decision level number coverage accuracy  Back-off and cooccurrence values.</Paragraph>
    <Paragraph position="1"> good as supervised approaches over small manually disambiguated training sets. If only small manually disambiguated training sets are available, the intertwined combination of unsupervised and supervised information sources leads to the best results.</Paragraph>
    <Paragraph position="2"> In another vein of this research we have demonstrated that cooccurrence frequencies obtained through WWW search engines are useful for PP attachmentdisambiguation (Volk, 2001).</Paragraph>
    <Paragraph position="3"> In the future we want to determine at which decision level such frequencies could be integrated.</Paragraph>
  </Section>
class="xml-element"></Paper>