<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1095">
  <Title>Machine Learning of Temporal Relations</Title>
  <Section position="9" start_page="758" end_page="759" type="concl">
    <SectionTitle>
6 Conclusion
</SectionTitle>
    <Paragraph position="0"> Our research has uncovered one new finding: semantic reasoning (in this case, logical axioms for temporal closure), can be extremely valuable in addressing data sparseness. Without it, performance on this task of learning temporal relations is poor; with it, it is excellent. We showed that temporal reasoning can be used as an over-sampling method to dramatically expand the amount of training data for TLINK labeling, resulting in labeling predictive accuracy as high as 93% using an off-the-shelf Maximum Entropy classifier. Future research will investigate this effect further, as well as examine factors that enhance or mitigate this effect in different corpora. null The paper showed that ME-C performed significantly better than a series of increasingly sophisticated baselines involving expansion of rules derived from human intuitions. Our results in these comparisons confirm the lessons learned from the corpus-based revolution, namely that rules based on intuition alone are prone to incompleteness and are hard to tune without access to the distributions found in empirical data.</Paragraph>
    <Paragraph position="1"> Clearly, lexical rules have a role to play in semantic and pragmatic reasoning from language, as in the discussion of example (2) in Section 1.</Paragraph>
    <Paragraph position="2"> Such rules, when mined by robust, large corpus-based methods, as in the Google-derived VerbOcean, are clearly relevant, but too specific to apply more than a few times in the OTC corpus.</Paragraph>
    <Paragraph position="3"> It may be possible to acquire confidence weights for at least some of the intuitive rules in GTag from Google searches, so that we have a  level field for integrating confidence weights from the fairly general GTag rules and the fairly specific VerbOcean-like lexical rules. Further, the GTag and VerbOcean rules could be incorporated as features for machine learning, along with features from automatic preprocessing.</Paragraph>
    <Paragraph position="4"> We have taken pains to use freely downloadable resources like Carafe, VerbOcean, and WEKA to help others easily replicate and quickly ramp up a system. To further facilitate further research, our tools as well as labeled vectors (unclosed as well as closed) are available for others to experiment with.</Paragraph>
  </Section>
class="xml-element"></Paper>