<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-3230">
  <Title>Applying Conditional Random Fields to Japanese Morphological Analysis</Title>
  <Section position="7" start_page="0" end_page="0" type="concl">
    <SectionTitle>
5 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> In this paper, we present how conditional random fields can be applied to Japanese morphological analysis in which word boundary ambiguity exists.</Paragraph>
    <Paragraph position="1"> By virtue of CRFs, 1) a number of correlated features for hierarchical tagsets can be incorporated which was not possible in HMMs, and 2) influences of label and length bias are minimized which caused errors in MEMMs. We compare results between CRFs, MEMMs and HMMs in two Japanese annotated corpora, and CRFs outperform the other approaches. Although we discuss Japanese morphological analysis, the proposed approach can be applicable to other non-segmented languages such as Chinese or Thai.</Paragraph>
    <Paragraph position="2"> There exist some phenomena which cannot be analyzed only with bi-gram features in Japanese morphological analysis. To improve accuracy, tri-gram or more general n-gram features would be useful.</Paragraph>
    <Paragraph position="3"> CRFs have capability of handling such features.</Paragraph>
    <Paragraph position="4"> However, the numbers of features and nodes in the lattice increase exponentially as longer contexts are captured. To deal with longer contexts, we need a practical feature selection which effectively trades between accuracy and efficiency. For this challenge, McCallum proposes an interesting research avenue to explore (McCallum, 2003).</Paragraph>
  </Section>
class="xml-element"></Paper>