<?xml version="1.0" standalone="yes"?>
<Paper uid="H01-1071">
  <Title>Towards Automatic Sign Translation</Title>
  <Section position="8" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
6. EXPERIMENTAL RESULTS
</SectionTitle>
    <Paragraph position="0"> We have evaluated the prototype system for automatic sign detection and translation. We have built a Chinese sign database with about 800 images taken from China and Singapore. We have tested the automatic detection module using 50 images randomly selected from the database.</Paragraph>
    <Paragraph position="1"> Table 1 shows the test result of automatic sign detection.</Paragraph>
    <Paragraph position="2"> Figure 5 and Figure 6 show examples of automatic sign detection with white rectangles indicating the sign regions.</Paragraph>
    <Paragraph position="3"> Figure 5 shows correct detection after layout analysis.</Paragraph>
    <Paragraph position="4"> Figure 6 illustrates a result with a false detection (Note the small detection box below and to the left of the larger detection).</Paragraph>
    <Paragraph position="5">  The text in Figure 7(a) is easily confused with the reflective background. The sign in Figure 7(b) is embedded in the  We have also tested the EBMT based method. We assume perfect sign recognition in our test. We randomly selected 50 signs from our database. We first tested the system includes a Chinese-English dictionary from the Linguistic Data Consortium, and a statistical dictionary built from the HKLC (Hong Kong Legal Code) corpus. As a result, we only got about 30% reasonable translations. We then trained with a small corpus of 670 pairs of bilingual sentences [7], The accuracy is improved from 30% to 52% on 50 test signs. Some examples of errors are illustrated below: Mis-segmentaion: Chinese with wrong segmentation: /ge zhong che liang qing rao xing/ Translation from MT: All vehicles are please wind profession Correct segmentation: Translation if segmentation is correct: All vehicles please use detour Lack-domain information: Chinese with segmentation: /qing wu dong shou/ Please don't touch it Translation from MT: Please do not get to work Domain knowledge needed to translate : &amp;quot;start to work&amp;quot; in domain such as work plan and &amp;quot;don't touch&amp;quot; in domains like tourism, exhibition etc.</Paragraph>
    <Paragraph position="6"> Proper Name: Chinese with segmentation: /bei jing tong ren yi yuan/ Beijing Tongren Hospital Translation from MT: Beijing similar humane hospital is translated to the meaning of each character because it is not identified as a proper name which then should only be represented by its pronunciation.</Paragraph>
    <Paragraph position="7"> Figure 8 illustrates error analysis of the translation module. It is interesting to note that 40% of errors come from missegmentation of words. There is a big room for improvement in proper word segmentation. In addition, we can take advantage of the contextual information provided by the OCR module to further improve the translation quality.</Paragraph>
  </Section>
class="xml-element"></Paper>