<?xml version="1.0" standalone="yes"?>
<Paper uid="W01-1414">
  <Title>Adding Domain Specificity to an MT system</Title>
  <Section position="6" start_page="2" end_page="2" type="concl">
    <SectionTitle>
&lt;-&gt; Transaction Log)waslearnedondifferent
</SectionTitle>
    <Paragraph position="0"> pairs of sentences during the alignment phase due to the presence of the word log introduced by the word association file.</Paragraph>
    <Paragraph position="1"> Without the addition of log at alignment time, the alignment process mapped Journal to Log, but not the more complex mapping for Journal des transactions. Compare the translations from the FE dictionary to the pairs from the word association file (where ordering represents frequency of each translation). Note that the WA list has learned the most relevant technical translation (log), which was lacking in the FE dictionary, but also the most frequent general  A similar case below (3) shows that the inclusion of the word push as a translation of emission in the word association file allows for a correct pattern in the transfer database: replication par emission push replication (3) Source: Pour configurer un serveur WINS afin d'utiliser une replication par emission, vous pouvez faire votre choix parmi plusieurs options configurables de la console WINS.</Paragraph>
    <Paragraph position="2"> Reference: To configure a WINS server to use push replication, you can choose from several WINS console configurable options.</Paragraph>
    <Paragraph position="3"> ALL translation: For configuring a WIN server to use a push replication, you can do your choice among options configurable of the WIN console.</Paragraph>
    <Paragraph position="4"> NONE translation: For configuring a WIN server to use a replication by program,you can do your choice among options configurable of the console WIN.</Paragraph>
    <Paragraph position="5">  This example is quite interesting, because the link of push to emission is helpful, even though is would be judged incorrect in a standard evaluation of the pairings themselves We have described the improvements so far as increases in domain specificity, but the effect is more wide-spread. We find that the added information allows for creation and retention of such generally better patterns as those in example (4): (4) Source: Assurez-vous qu'il y a du papier dans l'imprimante.</Paragraph>
    <Paragraph position="6"> Reference: Make sure there is paper in your printing device.</Paragraph>
    <Paragraph position="7"> ALL translation: Make sure that there is a paper in the printer.</Paragraph>
    <Paragraph position="8"> NONE translation: Provide that a paper in the printer becomes.</Paragraph>
    <Paragraph position="9"> We note the improved transfer patterns for Make sure and there is.</Paragraph>
    <Paragraph position="10"> The incidence of faulty translation patterns learned because of incorrect word-associations has been difficult to measure, but appears to be low. One instance was the learned correspondence of eteindre &lt;-&gt; off (instead of turn_off). We believe this could be avoided by more accurate preservation of information from our Logical Form representation in step one of the Moore algorithm.</Paragraph>
    <Section position="1" start_page="2" end_page="2" type="sub_section">
      <SectionTitle>
5.1 Future improvements
</SectionTitle>
      <Paragraph position="0"> The experiment presented here is the first step in our search for techniques that contribute to the quality of the translations by providing domain specific additions.</Paragraph>
      <Paragraph position="1"> We are working to find the most productive method for pruning low accuracy pairs (but still without hand-editing). We have already seen that if the data is truncated to maximize the accuracy of the word associations, the impact on the translation quality drops off, presumably because the high frequency pairs in the word association file contribute fewer unknown translations than the larger noisier file. This suggests that in the process of seeding an automatic alignment process such as ours, recall is more important than precision.</Paragraph>
      <Paragraph position="2"> .</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>