<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1706">
  <Title>The Effect of Rhythm on Structural Disambiguation in Chinese</Title>
  <Section position="6" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
5 Experimental Results
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
5.1 Training and Test Data
</SectionTitle>
      <Paragraph position="0"> A Chinese corpus of 200K words extracted from the People's Daily are segmented, POS-tagged and hand-labeled with content chunks in which all the trees are binary. The corpus is divided into two parts: (1) 180K for training set and (2) 20K for test set.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
5.2 Metrics and results
</SectionTitle>
      <Paragraph position="0"> We take two kinds of criteria to measure the system's performance: labeled and unlabeled.</Paragraph>
      <Paragraph position="1"> According to the labeled criterion, a recognized phrase is correct only if a phrase with the same starting position, ending position and the same label is found in the gold standard. According to the unlabeled criterion, a recognized phrase is correct as long as a phrase with the same starting position and ending position is found in the gold standard.</Paragraph>
      <Paragraph position="2">  Within each criterion, precision, recall and F-measure are given as metrics for the system's performance. Precision represents how many phrases are correct among the phrases recognized, recall represents how many phrases in the gold standard are correctly recognized, and F-measure is defined as follows:  Table 4 gives the experimental results in three different conditions: the first row gives the result of PCFG model; the second row gives the result of PCFG model integrated with rhythm feature model (RF) where only the features of simple phrases are considered; the last row gives the result of PCFG model plus RF where the rhythm features in all the phrases are considered. The results indicate that the rhythm features in both simple and complex phrases contribute to the improvement of performance over PCFG model. We see that the rhythm feature improves the labeled F-measure 6.88 percent and the unlabeled F-measure 4.24 percent over the unaugmented PCFG model.</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
5.3 Effect of rhythm feature on parsing
</SectionTitle>
      <Paragraph position="0"> The experiment shows that the rhythm feature can help the performance of a parser in Chinese.</Paragraph>
      <Paragraph position="1"> Specifically, the effects of rhythm feature on parsing are shown in two ways: (1) Help for the disambiguation of phrasal type. Table 5 shows the difference of the results between PCFG model and PCFG + RF model for the sequence &amp;quot;Guo /country Juan Qu /sacrifice&amp;quot; in the sentence &amp;quot;Gai /the Xiao /school You /have 900 Xue Zi /students Wei /for Guo /country Juan Qu /sacrifice&amp;quot; (`900 students from this school gave their lives for their country').</Paragraph>
      <Paragraph position="2"> In the sentence above, &amp;quot;Guo /country&amp;quot; is the object of preposition &amp;quot; Wei /for&amp;quot;, &amp;quot; Guo /country Juan Qu /sacrifice&amp;quot; is not a constituent. But the unaugmented PCFG model incorrectly parses it as a S(subject-predicate construction). Contrarily, according to PCFG+RF model, the type with greatest probability is the (correct) NC(nonconstituent) parse.</Paragraph>
      <Paragraph position="3"> (2) Help for pruning.</Paragraph>
      <Paragraph position="4"> Let's give an example to explain it. For the sentence &amp;quot;Jie Jue /solve Ju Min /resident Chi /eat Cai /vegetable Wen Ti /problem Shi Fen /very Kun Nan /difficult&amp;quot;('It's very difficult to solve the vegetable problem for the residents.'), the number of edges generated by the PCFG is 1236, but the number decreases to 348 after the rhythm feature is applied, thus pruning 73% of the edges. As indicated in Table 1, in the rule &amp;quot;NP -&gt; N V&amp;quot;, P(RF = [1,0] ) = 0, so &amp;quot;[Ju Min /N Chi /V]NP&amp;quot; is pruned after adding RF. Similarly, in rule &amp;quot;NP -&gt; V N&amp;quot;, P(RF = [0, 1] ) = 0.003, so &amp;quot;[Chi /V Cai /N]NP&amp;quot; is pruned since it has very low probability. With these two edges pruned, more potential edges containing them will not be generated.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>