<?xml version="1.0" standalone="yes"?>
<Paper uid="C92-2101">
  <Title>LEARNING TRANSLATION TEMPLATES FROM BILINGUAL TEXT</Title>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
(1) Quality
</SectionTitle>
    <Paragraph position="0"> Basically, a conventional machine translation system performs word-for-word translation. That is, a TL sentence is created from words, each of which is a TL equivalent of a word in an SL sentence. An example-based machine translation system is, in contrast, capable of creating a more flexible translatiou whereby elements which do not have a word-for-word correspondence are transferre~l as an undivided whole. We can therefore expect improvement in traoslatioa quality.</Paragraph>
    <Paragraph position="1"> (2) Customizability With conventional machine translation systems based on grammar rules, users are not allowed to modify the grammar rules, because they are subtly related to each other and it is difficult to assess the overall effect of rule modification. But with the example-based machine translation, users can easily customize the system for their own fields, e.g. computer manuals, by providing their own translation examples. This system is particularly suitable for a field in which similar sentences are written repeatedly.</Paragraph>
  </Section>
  <Section position="6" start_page="0" end_page="0" type="metho">
    <SectionTitle>
(3) Trauspamncy
</SectionTitle>
    <Paragraph position="0"> A translation template is regarded as a transfer rule. It is easy to understand, compared to a tree-to-tree trausformation rule ill conventional nlachine translation.</Paragraph>
    <Paragraph position="1"> Translation is primarily performed by direct transfer of word string patterus. A highly transparent system can therefore be realized.</Paragraph>
  </Section>
  <Section position="7" start_page="0" end_page="0" type="metho">
    <SectionTitle>
(4) Contpumtion
</SectionTitle>
    <Paragraph position="0"> Generally speaking, example-based machine translation requires large amount of cotaputation. In the proposed architecture~ however, examples are transformed belorehand into intermediate forms by extracting useful information. The amount of required computation is therefore reduced compared to a system which uses tIanslalion examples directly.</Paragraph>
    <Paragraph position="1"> (5) Unified treaUnent of translation knowledge Various kinds of knowledge for translation are extracted and represented in a single translation template framework. For example, the template in Fig. 2 is a kind of transfer rule which bridges a structural gap between Japanese and English. Lexical selection based AcrEs DE COLING-92, NAN'rV:S, 23-28 ^ot~'r 1992 6 7 7 PROC. ol; C(JLING-92, NANrES, AUG. 23-28, 1992 on cooccurrence restriction is also implemented in the framework discussed in Section 5.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
6.2 Features of the algorithm for coupling
corresponding units
</SectionTitle>
      <Paragraph position="0"> Identifying the correspondence between units in a bilingual pair of sentences is essential for example-based machine translation. Sadler et at. have developed tools for constructing a bilingual corpus in which equivalent units are linked to each other.\[Sadlerg0\] Full automatization, however, has not yet been realized.</Paragraph>
      <Paragraph position="1"> There are three distinguishing features of the algorithm presented in Section 3. First, the algorithm was designed on the assumption that syntactic ambiguities cannot be resolved completely by the preceding sentence analysis. Syntactic ambiguities are resolved instead in the phrase coupling prece~. Second, ambiguities in correspondence between words is resolved simultaneously as phrases are coupled. Third, correspondence between phrases is determined without comparing their internal structures, because structural coincidence cannot always be expected between a pair of Japanese and English sentences, even if a dependency structure is adopted. These features result in a reliable and efficient algorithm.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
6.3 Is the translation template inflexible ?
</SectionTitle>
      <Paragraph position="0"> The translation template may not be as flexible as the matching expression proposed by Sato.\[Sato90\] However, the introduction of fragmentary templates has made it sufficieafly flexible.</Paragraph>
      <Paragraph position="1"> An obvious restriction of the template is that the word order is fixed. This is inconvenient for languages, like Japanese, in which word order is flexible. However, it is not a serious problem, as the system has a learning capability. If a corpus includes sentences which differ in word order, the system will learn a set of templates which differ in word order. A more important problem to be pursued is how to deal with omissible elements. It is not easy Io decide which phrases can be omitted from an example sentence. Translation templates which include descriptions of phrase omissibility, however, would certainly be effective.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>