<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-0717">
  <Title>I- I Incorporating Knowledge in Natural Language Learning: A Case Study</Title>
  <Section position="6" start_page="125" end_page="125" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> Over several decades, research on high level inferences such as natural language understanding has emphasized programmed systems, as opposed to those that learn. However, experience in AI research over the past few decades shows that it is unlikely that hand programming or any form of knowledge engineering will generate a robust, non-brittle reasoning system in a complex domain.</Paragraph>
    <Paragraph position="1"> An approach that puts learning at the center of high level inferencing (Khardon and Roth, 1997; Valiant, 1995} should suggest ways to make progress in massive knowledge acquisition and, in particular, ways of incorporating incomplete and noisy knowledge from various information sources such as different modalities, teachers or experts, into a highly scalable learning process.</Paragraph>
    <Paragraph position="2"> The present work made preliminary steps in this direction. We have studied ways to incorporate external knowledge sources into a learning algorithm in order to improve its performance. This investigation was done within the SNOW architecture, a sparse network of threshold gates utilizing the Winnow on-line learning algorithm. The linguistic knowledge sources, noun-class datasets, were compiled for general reasons, irrespective of the task studied here. Knowledge incorporation resulted in a statistically significant performance improvement on PPA, a challenging natural language disambiguation task which has been investigated extensively. Using random noun classes, we have demonstrated that the semantic nature of the external knowledge is essential. In addition, the granularity of the data was shown to play an important role in the learning performance. A highly granular synset classification failed to improve the results.</Paragraph>
    <Paragraph position="3"> A lot of future work is to be done in order to substantiate the results presented here, study more tasks and prepare and investigate the effectiveness of other information sources.</Paragraph>
  </Section>
class="xml-element"></Paper>