<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1227">
  <Title>A Method of Incorporating Bigram Constraints into an LR Table and Its Effectiveness in Natural Language Processing i</Title>
  <Section position="7" start_page="14" end_page="14" type="concl">
    <SectionTitle>
6 Conclusions
</SectionTitle>
    <Paragraph position="0"> In this paper, we described a method to construct a bigram LR table, and then discussed the advantage of our method, comparing our method to the bigram and trigram language models. The principle advantage over the bigram language model is that, in using a bigram LR table, we can combine both local probabilistic connection constraints (bigram constraints) and global constraints (CFG).</Paragraph>
    <Paragraph position="1"> Our method is applicable not only to natural language processing but also speech recognition. We are currently testing our method using a large-sized grammar containing dictionary rules for speech recognition.</Paragraph>
    <Paragraph position="2"> Su et al. (Suet al., 1991) and Chiang et al. (Chiang et al., 1995) have proposed a very interesting corpus-based natural language processing method that takes account not only of lexical, syntactic, and semantic scores concurrently, but also context-sensitivity in the language model. However, their method seems to suffer from difficulty in acquiring probabilities from a given corpus.</Paragraph>
    <Paragraph position="3"> Wright (Wright, 1990) developed a method of distributing the probability of each PCFG rule to each action in an LR table. However, this method only calculates syntactic scores of parsing trees based on a context-free framework.</Paragraph>
    <Paragraph position="4"> Briscoe and Carroll (Briscoe and Carroll., 1993) attempt to incorporate probabilities into an LR table. They insist that the resultant probabilistic LR table can include probabilities with contextsensitivity. Inui et. al. (Inni et al., 1997)reported that the resultant probabilistic LR table has a defect in terms of the process used to normalize probabilities associated with each action in the LR table. * Finally, we would like to mention that Klavans and Resnik (Klavaus and Resnik, 1996) have advocated a similar approach to ours which combines symbolic and statistical constraints, CFG and bi-gram constraints.</Paragraph>
  </Section>
class="xml-element"></Paper>