<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0610">
  <Title>Using Uneven Margins SVM and Perceptron for Information Extraction</Title>
  <Section position="6" start_page="77" end_page="77" type="concl">
    <SectionTitle>
4 Conclusions
</SectionTitle>
    <Paragraph position="0"> This paper studied the uneven margins versions of two learning algorithms SVM and Perceptron to deal with the imbalanced training data in IE. Our experiments showed that the uneven margin is helpful, in particular on small training sets. The smaller the training set is, the more bene cial the uneven margin could be. We also showed that the systems based on the uneven margins SVM and Perceptron were com- null of the SVM and Perceptron, respectively: macro averaged F1(%) on the two datasets CoNLL-2003 (development set) and Jobs. The standard deviations for the Jobs dataset show the statistical signi cances of the results. In bold are the best performance gures for each dataset and each system.</Paragraph>
    <Paragraph position="1">  parable to other state-of-the-art systems.</Paragraph>
    <Paragraph position="2"> Our SVM system obtained better results than other SVM-based systems on the CoNLL-2003 corpus and CFP corpus respectively, while being simpler than most of them. This demonstrates that our SVM system is both effective and ef cient.</Paragraph>
    <Paragraph position="3"> We also explored PAUM, a simple and fast learning algorithm for IE. The results of PAUM were somehow worse (about 0.02 overall F-measure lower) than those of the SVM on two out of three datasets. On the other hand, PAUM is much faster to train and easier to implement than SVM. It is also worth noting that PAUM outperformed some other learning algorithms. Therefore, even PAUM on its own would be a good learning algorithm for IE.</Paragraph>
    <Paragraph position="4"> Moreover, PAUM could be used in combination with other classi ers or in the more complicated framework such as the one in Carreras et al. (2003). Since many other tasks in Natural Language Processing, like IE, often lead to imbalanced classi cation problems and the SVM has been used widely in Natural Language Learning (NLL), we can expect that the uneven margins SVM and PAUM are likely to obtain good results on other NLL problems as well.</Paragraph>
  </Section>
class="xml-element"></Paper>