<?xml version="1.0" standalone="yes"?>
<Paper uid="I05-2018">
  <Title>Detecting the Countability of English Compound Nouns Using Web-based Models</Title>
  <Section position="5" start_page="105" end_page="106" type="evalu">
    <SectionTitle>
4 Experiments and Results
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="105" end_page="106" type="sub_section">
      <SectionTitle>
4.1 Detecting plural only compound nouns
</SectionTitle>
      <Paragraph position="0"> Plural only compound nouns that have not singular forms always occur in plural forms. The frequency of their singular forms should be zero.</Paragraph>
      <Paragraph position="1"> Considering the noise data introduced by search engine, we used model (1) and (3) in turn to detect plural noun. We detected plural only compound nouns with the following algorithm  The problem is how to decide the two thresholds. We preformed exhaustive search to adjust th 1,th 2 optimized on the training set. With 0 [?] th 1,th 2 20, all possible pair values are tried with the stepsize of 1.</Paragraph>
      <Paragraph position="2">  We use Recall and Precision to evaluate the performance with the different threshold pairs. The fundamental Recall/Precision definition is adapted to IE system evaluation. We borrowed the measures using the following definition for our evaluation. For one experiment with a certain threshold pair, A stands for the number of plural found correctly; AB stands for the total number of plural only compound nouns in training set (80 words); AC stands for the total number of compound nouns found. The Recall and Precision are defined in (4) and (5). We also introduced F-score when we need consider the Recall and Precision at the same time, and in the paper, F-score is calculated according to (6).</Paragraph>
      <Paragraph position="3"> Figure 3 shows the performance evaluated by the three measures when th 1=8 and 0 [?] th 2[?]10 with a stepsize of 1. We set th 2 to 5 for the test later, and accordingly the values of Recall, Precision and F-score are 91.25%, 82.95% and 87.40% respectively.</Paragraph>
      <Paragraph position="4">  compound nouns The algorithm detecting uncountable compound nouns is shown in Figure 4. Using model (1) and (2), we attempted to fully make use of the characteristic of uncountable compound nouns, that is the frequencies of their occurrence in the singular forms are much larger than that in the plural forms.</Paragraph>
      <Paragraph position="5">  The method to obtain the optimal threshold th 3 and th 4 is the same to 4.1. We set th 3 to 24, th 4 to 2, and the values of Recall, Precision and F-score are 88.38%, 80.27% and 84.13% respectively. null</Paragraph>
    </Section>
    <Section position="2" start_page="106" end_page="106" type="sub_section">
      <SectionTitle>
4.3 Performance on the test suite
</SectionTitle>
      <Paragraph position="0"> We evaluated our complete algorithm with the four thresholds (th 1=8,th 2=5, th 3=24, th 4=2) on the test set, and the detecting results are summarized in Table 2. There are 352 countable compound nouns in our test set, then when classify all the test words as countable, we can at least get the accuracy of 70.4%. We used it as our baseline. The accuracy on the total test date is 89.2% that significantly outperforms the baseline. For the 30 newcoined compound nouns, the detecting accuracy is 100%. This can be explained by their infrequence. Newcoined words are not prone to produce noise data than others just because they are not occurring regularly.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>