<?xml version="1.0" standalone="yes"?>
<Paper uid="H90-1072">
  <Title>Machine Translation Again? Yorick Wilks, Jaime Carbonen, David Farwell, Eduard Hovy and</Title>
  <Section position="2" start_page="375" end_page="378" type="abstr">
    <SectionTitle>
OPINION 1: An interlingual approach forces unneeded
</SectionTitle>
    <Paragraph position="0"> processing If a source language has, say, an expression which is three ways ambiguous and some target language has an expression which has precisely the same three-way ambiguity, unanalyzed why not simply carry the ambiguity from the source to the target and let the reader  figure it out? Why disarnbiguate needlessly? The response is, on the one hand, that a third language probably has different expression for each of the possible interpretations, so that if the same representational apparatus is to be applied to translations between the source language and a third language or from the target language and a third language, such processing is necessary in any case. On the other hand, a quick inspection of bilingual dictionaries shows that cases of complete correspondence of ambiguity across languages is extremely rare, even in closely related languages such as German and Dutch.</Paragraph>
    <Paragraph position="1"> The issue of &amp;quot;since we sometimes can get away with less processing, why risk doing unnecessary work?&amp;quot; can be compared with intelligence-gathering work, where much of the effort is routine; information often confirms expectations; and therefore much of the work is &amp;quot;unnecessary.&amp;quot; With such an attitude, all unexpected, important intelligence would often be ignored, much to the detriment of the analysts and policymakers. Ignoring meaning in translation because it need not always be interpreted, is an equally flawed philosophy. The times when deeper analysis is required can be absolutely crucial to produce meaningful, rather than misleading, translations. language particular bias would simply be defect of the approach, rather than wholly invalidating it.</Paragraph>
    <Paragraph position="2"> A standard example here would be the case of the verb &amp;quot;wear&amp;quot; in English and the problem of expressing the notion in Japanese or Chinese. It so happens that in Japanese the corresponding verb depends entirely on what is worn e.g. shoes (verb= hateiru ), coat (verb= kiteiru ), spectacles (verb= kaketeiru ) and so on (and similarly for Chinese). It is thus reasonable to say that Japanese does not have a concept of &amp;quot;wear&amp;quot; in the way English does. However, that observation is no kind of argument at all against an interlingual approach, merely one for intelligent generation. In an interlingual environment there will be at least one interlingual node (which may or may not correspond to &amp;quot;wear&amp;quot;) that links the relevant sense representations. The crucial point is that it would be the intelligent Japanese generator (since no problem arises in the Japanese to English direction) that makes the choice of output verb based simply on selection semantics (e.g. if the worn object is &amp;quot;koutoo&amp;quot; the verb is &amp;quot;kiteiru&amp;quot; and so on). Conclusion There are several aspects of the knowledge-based interlingua MT project that incur a measure of risk.</Paragraph>
    <Paragraph position="3"> Foremost among these is the distributed management risk OPINION 2: Interlingual approaches are heavily knowledgeamong the three centers. dependent and the task of working out appropriate representations is too demanding to be practical.</Paragraph>
    <Paragraph position="4"> It has been our experience that some, even if incomplete, level of knowledge representation is crucial to machine translation. The need for such knowledge is especially obvious in the translation of technical text, where translations based on a general knowledge of the world are markedly inferior to translations based on a specific knowledge of the subject domain of the translation. Large-scale know.ledge bases are being actively developed in the field, and domain models have come to be considered a standard component of many AI-related application systems. In our system, acquisition of world knowledge will be an ongoing task, and we fully intend to prove the feasibility of working with large knowledge bases in practical terms. To wit, EDR laboratories, and Fujitsu in Japan have come to the same conclusion and are actively building large knowledge-bases for interlingua-based machine translation, with initial success. Although it is clearly in the national interest to establish several sites developing MT technology in mutual cooperation, special effort must be made to address communication, establishment of standards, mutual responsibility relationships, fall-back positions and so on. We think this is an eminently manageable risk, but nonetheless a omnipresent one. We are fully cognizant of this risk, and are prepared to minimize it by establishing common procedures, open lines of communication, and accommodation where necessary.</Paragraph>
    <Paragraph position="5"> OPINION 3: interlingual approaches are based on a particular language, thus creating unnatural analyses for other languages.</Paragraph>
    <Paragraph position="6"> This is the &amp;quot;cultural imperialism&amp;quot; argument. If, however, there exists such a thing as a universal descriptive linguistic framework, then there is no reason to assume that language imperialism must be a side-effect of the interlingual approach. Our experience in the development of an interlingual representation based on a cross-linguistic comparison of parallel texts has indicated, at least, that such a language independent framework is possible. But even if no such framework exists, then at worst, such</Paragraph>
  </Section>
class="xml-element"></Paper>