<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1315">
  <Title>An Investigation of Various Information Sources for Classifying Biological Names</Title>
  <Section position="6" start_page="5" end_page="6" type="evalu">
    <SectionTitle>
5 Results and Evaluation
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="5" end_page="6" type="sub_section">
      <SectionTitle>
5.1 F-Terms and Suffixes
</SectionTitle>
      <Paragraph position="0"> Table 1 gives the precision and recall values for the first partition for both f-terms and suffixes.</Paragraph>
      <Paragraph position="1">  As can be seen, the recall for 'Chemical' is very low as compared to the other classes. This is due to two reasons--firstly most chemical names consist of only one word and secondly we found that chemical names do not end with an indicative word.</Paragraph>
      <Paragraph position="2"> The number of f-terms and suffixes extracted by our program was considerably less for Chemicals and Protein Parts as compared to Proteins and Others. While this is consistent with the the explanation of poor recall for chemicals, it can be noted that the low number of f-terms and suffixes extracted for protein parts did not affect its recall in the same manner. As expected the precision remains high for all classes.</Paragraph>
      <Paragraph position="3"> method.</Paragraph>
      <Paragraph position="4">  The scores were simply the conditional probability of a class given a word</Paragraph>
    </Section>
    <Section position="2" start_page="6" end_page="6" type="sub_section">
      <SectionTitle>
5.2 Using Examples
</SectionTitle>
      <Paragraph position="0"> For the string matching, we tried three different set of values for the parameters #0B, #0C and k,that is (0.3, 2, 3), (0.7, 2, 1) and (0.7, 2, 5). We found that the results were marginally better for the set (0.3,2,3) and, hence, show the results for this set only. Table 1 shows the results of applying the string matching to the first partition -- all by itself and on names not classified after the suffix stage. As can be seen, the recall is higher than the previous stages but with a slight reduction in precision.</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="6" end_page="6" type="evalu">
    <SectionTitle>
5.3 Results for Context
</SectionTitle>
    <Paragraph position="0"> We ran the context classifier for different values of the parameters f, a and b but finally chose a value of 5forf because choosing a higher frequency threshold does not improve the precision but hurts the recall. Figure 1 shows the precision plotted against the recall for different choice of a and b.</Paragraph>
    <Paragraph position="1"> The values of the precision and recall on the first  partition for each individual class and the two sets of thresholds are shown in Table 1. The first set, that considers stronger evidences (since a is higher), achieves higher precision but recall is not satisfactory. Most of the word evidences chosen tended to indicate a classification of proteins and hence the higher recall for this class. Allowing weaker evidences (because a =2) means more contextual evidences were selected and hence a higher recall was obtained (particularly for protein). But precision is lowered except for Source and Others (which incidentally don't show an increase in recall).</Paragraph>
    <Section position="1" start_page="6" end_page="6" type="sub_section">
      <SectionTitle>
5.4 Overall Results
</SectionTitle>
      <Paragraph position="0"> Table 2 shows the precision and recall for all the different classes, averaging it out for the 4 different partitions. We observed very little variance between the results for the different partitions.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>