<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-2109">
  <Title>Trimming CFG Parse Trees for Sentence Compression Using Machine Learning Approaches</Title>
  <Section position="7" start_page="856" end_page="856" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> We presented a maximum entropy model to extend the sentence compression methods described by Knight and Marcu (Knight and Marcu, 2000).</Paragraph>
    <Paragraph position="1"> Our proposals are two-fold. First, our maximum entropy model allows us to incorporate various characteristics, such as a mother node or the depth from a root node, into a probabilistic model for determining which part of an input sentence is removed. Second, our bottom-up method of matching original and compressed parse trees can match tree structures that cannot be matched using Knight and Marcu's method.</Paragraph>
    <Paragraph position="2"> The experimental results show that our maximum entropy method improved the accuracy of sentence compression as determined by three evaluation criteria: F-measures, bigram F-measures and BLEU scores. Using our bottom-up method further improved accuracy and produced short summaries that could not be produced by previous methods. However, we need to modify this model to appropriately process more complicated sentences because some sentences were not correctly summarized. Human judgments showed that the maximum entropy model with the bottom-up method provided more grammatical and more informative summaries than other methods.</Paragraph>
    <Paragraph position="3"> Though our training corpus was small, our experiments demonstrated that the data was suf cient. To improve our approaches, we can introduce more feature functions, especially more semantic or lexical features, and to deal with these features, we need a larger corpus.</Paragraph>
  </Section>
class="xml-element"></Paper>