<?xml version="1.0" standalone="yes"?>
<Paper uid="C04-1175">
  <Title>Combining Prediction by Partial Matching and Logistic Regression for Thai Word Segmentation</Title>
  <Section position="6" start_page="43" end_page="43" type="evalu">
    <SectionTitle>
5 Experimental Evaluation
</SectionTitle>
    <Paragraph position="0"> In the first experiment, we evaluate the proposed syllable segmentation method. The algorithm is trained with 2,200 syllables, manually segmented from a dictionary. The test data used is a text excerpt from a thesis written in Thai. The results in Table 6 show that the algorithm at order 4 yields the best result which is, from the 1,714 manually segmented syllables, the algorithm correctly identifies 1,694 (or 98.83%) of them correctly.</Paragraph>
    <Paragraph position="1"> Figure 2 shows an example of segmentation results.</Paragraph>
    <Paragraph position="2">  Next, we evaluate the proposed algorithm at order 4 against five 1,000-syllable test texts which are not part of the text used in the training. The results in Table 7 show 96.65 to 98.26% segmentation accuracy.</Paragraph>
    <Paragraph position="3">  To evaluate the syllable combination technique, we create 50 ambiguous test cases. The results show that 47 cases (94%) are segmented correctly using the technique proposed, in which 13 cases are correctly segmented in Step 1; 11 cases are correctly segmented in Step 2, and 23 cases are correctly segmented in Step 3.</Paragraph>
    <Paragraph position="4"> An evaluation of the entire process of word segmentation (i.e., from syllable segmentation to syllable combination) shows an accuracy of 97.17% by which 76.92% of those incorrect segmentation roots from incorrect syllable segmentation.</Paragraph>
    <Paragraph position="5">  Lastly, we use the same test data however with correctly identified syllables, the performance shows 99.35% accuracy. This emphasizes the importance of pre-segmenting syllables and at the same time indicates that the proposed syllable combining method is effective.</Paragraph>
  </Section>
class="xml-element"></Paper>