<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1004">
  <Title>Discriminative Hidden Markov Modeling with Long State Dependence using a kNN Ensemble</Title>
  <Section position="7" start_page="211" end_page="211" type="concl">
    <SectionTitle>
7. Conclusion
</SectionTitle>
    <Paragraph position="0"> Hidden Markov Models (HMMs) are a powerful probabilistic tool for modeling sequential data and have been applied with success to many textrelated tasks, such as shallow paring. In these cases, the observations are usually modified as multinomial distributions over a discrete dictionary and the HMM parameters are set to maximize the likelihood of the observations. This paper presents a discriminative HMM with long state dependence that allows observations to be represented as arbitrary overlapping features and defines the conditional probability of the state sequence given the observation sequence. It does so by assuming a novel mutual information independence to separate the dependence of a state given the observation sequence and the previous states. Finally, the long state dependence and the observation dependence can be effectively captured by a variable-length mutual information model and a kNN ensemble respectively.</Paragraph>
    <Paragraph position="1"> In future work, we will explore our model in other applications, such as full parsing.</Paragraph>
  </Section>
class="xml-element"></Paper>