<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0402">
  <Title>Feature Engineering and Post-Processing for Temporal Expression Recognition Using Conditional Random Fields</Title>
  <Section position="9" start_page="15" end_page="15" type="concl">
    <SectionTitle>
7 Conclusion
</SectionTitle>
    <Paragraph position="0"> In this paper we presented different feature engineering and post-processing approaches for improving the results of timex recognition task. The first approach explores the different set of features that can be used for training a CRF-based timex recognition system. The second investigates the effect of the different tagging scheme for timex recognition task. The final approach we considered applies a list of core timexes for post-processing the output of a CRF system. Each of these approaches addresses different aspects of the overall performance. The use of a list of timexes both during training and for post-processing resulted in improved recall whereas the use of a more complex tagging scheme results in better precision. Their individual overall contribution to the recognition performances is limited or even negative whereas their combination resulted in substantial improvements over the baseline.</Paragraph>
    <Paragraph position="1"> While we exploited the special nature of timexes, we did avoid using linguistic features (POS, chunks, etc.), and we did so for portability reasons. We focused exclusively on features and techniques that can readily be applied to other named entity recognition tasks. For instance, the basic and list features can also be used in NER tasks such as PERSON, LOCATION, etc. Moreover, the way that we have used a list of core expressions for post-processing is also task-independent, and it can easily be applied for other NER tasks.</Paragraph>
  </Section>
class="xml-element"></Paper>