<?xml version="1.0" standalone="yes"?>
<Paper uid="W00-0728">
  <Title>A Context Sensitive Maximum Likelihood Approach to Chunking</Title>
  <Section position="5" start_page="136" end_page="137" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> It was shown that using context made it possible to improve performance of maximum likelihood prediction. It was suggested that the limit of performance for this method is implicitly given by the size of the training set, as this determines the significance of larger contexts, and increases the chance of finding a matching longer context. In smaller collections, large patterns are a) likely to occur at a low frequency with few competing labels and b) likely to not exist in the test set. A larger collection will increase the number of different contexts, as well as the significance of picking the best, most frequent, prediction from a set of (identical) competitors with different labels.</Paragraph>
    <Paragraph position="1"> The presented method does not generalize beyond what is recorded in the training set as the most likely alternative. However, it is expected to * improve with the size of the training set, as this makes it feasible to use longer contexts, and * have a low computational complexity, as the  process is always limited to use a low' number of hash table look-ups (determined by the largest size of context). Training is limited to detecting the most likely outcome of each context (i.e., a sorting operation).</Paragraph>
  </Section>
class="xml-element"></Paper>