<?xml version="1.0" standalone="yes"?>
<Paper uid="I05-2002">
  <Title>A Hierarchical Parsing Approach with Punctuation Processing for Long Chinese Sentences</Title>
  <Section position="7" start_page="10" end_page="11" type="evalu">
    <SectionTitle>
5 Performance Evaluation
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="10" end_page="10" type="sub_section">
      <SectionTitle>
5.1 Test Sentences
</SectionTitle>
      <Paragraph position="0"> The primary aim of the HP strategy is to take use of the punctuation information to help to conquer the difficulty of parsing long sentences.</Paragraph>
      <Paragraph position="1"> Chinese sentences with over 20 words are generally regarded as long sentences. Therefore, we conduct experiments on the sentences with the length over 20 words.</Paragraph>
      <Paragraph position="2"> Firstly, 8,059 sentences were chosen randomly from TCT 973 as train set. The 3,795 PCFG rules used in our system are extracted from the train set after generalizing. Then, for other 847 sentences, whose lengths are less than 20 words are filtered and 420 sentences are finally conserved as our open test data set. Distribution of these sentences is shown in Table 1 below:</Paragraph>
    </Section>
    <Section position="2" start_page="10" end_page="11" type="sub_section">
      <SectionTitle>
5.2 Efficiency Evaluation
</SectionTitle>
      <Paragraph position="0"> In order to compare our HP approach with TP method of once-parsing algorithm, we do compared experiments using same data set in Table 1 and same grammar rules set.</Paragraph>
      <Paragraph position="1">  Running two systems on a PC (Pentium 4, 1.20GHz, 256M of RAM), their time consumptions are shown in Figure 5.</Paragraph>
      <Paragraph position="2">  In our experiment system, we set the upper limit execution time as 120 seconds per sentence, judging at the end of every algorithm cycle. When parsing time of the sentence is overtime, the system will exit without getting final result. Experiment results shown in Fig.5 prove that time efficiency of HP method is greatly superior to TP, especially when the sentence has more than 40 words. With the increasing of sentence length, it is more difficult for TP method to parse successfully.</Paragraph>
      <Paragraph position="3">  Firstly, Table 2 shows numbers of sentences failed to be parsed in two methods with the time  It is evident that HP method can largely reduce failed sentences in given time limitation. Then, except for failed sentences, only considering the successfully parsed sentences, the parsing accuracy and recall of the two methods should be compared. The standard PARSEVAL measures [9] are used to evaluate two methods. Results are shown in Table 3. From Table 3, we can see that the total parsing accuracy and recall of HP method are both almost 7% higher than those of TP method.</Paragraph>
      <Paragraph position="4"> Amounts of average crossing brackets are also reduced greatly.</Paragraph>
      <Paragraph position="5">  Analyzing data in Table 3, to different text types, the accuracy and improvement effect of TP method are slightly different. Sentences of literature text have the highest parsing accuracy and recall. Studied show that there are 97 'run-on sentences' in the 116 literature text sentences, covering 84%. The comparatively higher accuracy and recall of these sentences prove that our HP approach is effective.</Paragraph>
      <Paragraph position="6">  measures Sentences of application have lowest parsing accuracy and smallest improvement. Because comparing to other three types, sentences of this type have more long nested noun phrases or coordinate components, such as long organization names and commodity names, which will cause noun phrase combination disambiguation.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>