<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2074">
  <Title>ARE: Instance Splitting Strategies for Dependency Relation-based Information Extraction</Title>
  <Section position="4" start_page="571" end_page="571" type="intro">
    <SectionTitle>
2 Related work
</SectionTitle>
    <Paragraph position="0"> There are several research directions in Information Extraction. We highlight a few directions in IE such as case frame based modeling in PALKA by Kim and Moldovan (1995) and CRYSTAL by Soderland et al. (1995); rule-based learning in Autoslog-TS by Riloff et al. (1996); and classification-based learning by Chieu et al. (2002). Although systems representing these directions have very different learning models, paraphrasing and alignment problems still have no reliable solution.</Paragraph>
    <Paragraph position="1"> Case frame based IE systems incorporate domain-dependent knowledge in the processing and learning of semantic constraints. However, concept hierarchy used in case frames is typically encoded manually and requires additional human labor for porting across domains. Moreover, the systems tend to rely on heuristics in order to match case frames. PALKA by Kim and Moldovan (1995) performs keyword-based matching of concepts, while CRYSTAL by Soderland et al. (1995) relied on additional domain-specific annotation and associated lexicon for matching.</Paragraph>
    <Paragraph position="2"> Rule-based IE models allow differentiation of rules according to their performance. Autoslog-TS by Riloff (1996) learns the context rules for extraction and ranks them according to their performance on the training corpus. Although this approach is suitable for automatic training, Xiao et al. (2004) stated that hard matching techniques tend to have low recall due to data sparseness problem. To overcome this problem, (LP)  by Ciravegna (2002) utilizes rules with high precision in order to improve the precision of rules with average recall. However,</Paragraph>
  </Section>
class="xml-element"></Paper>