<?xml version="1.0" standalone="yes"?>
<Paper uid="H93-1037">
  <Title>LINGSTAT: AN INTERACTIVE, MACHINE-AIDED TRANSLATION SYSTEM*</Title>
  <Section position="7" start_page="193" end_page="193" type="evalu">
    <SectionTitle>
5. RESULTS
</SectionTitle>
    <Paragraph position="0"> The system as described here (without the finite-state parser) was evaluated by DARPA in July 1992. The performance of two Level 2 translators was measured on a test set of 18 Japanese documents, each translator translating 9 with the aid of the system and 9 by hand.</Paragraph>
    <Paragraph position="1"> In general, the quality of translation with and without the system was found to be comparable, but the system provided a speedup of approximately 30%.</Paragraph>
    <Paragraph position="2"> Since the tested system provided no help with the analysis of the Japanese sentences, this savings was achieved by drastically reducing the time spent doing tokenization and lookup. It might appear surprising that so much time could be saved from these activities alone, but the many unusual features of Japanese described above conspire to produce a large overhead in this phase of translation compared to other languages. This result is also consistent with an analysis of how the translators allocated their time: without the system, their principal effort involved dictionary lookup, but with the system most of their time was spent analyzing sentence structure. null Productivity tests have also been conducted on the rudimentary Spanish version of the workstation. This system incorporates a Spanish de-inflector, provides word for word translation to English, and has fast access to an on-line dictionary. On a scaled down version of the DARPA test (6 documents instead of 18, including 3 by hand and 3 with the aid of the system), a fluent speaker of Italian (a language very similar to Spanish) showed no productivity gain. At the other extreme, a user with no Spanish knowledge and no recent training in any European language was about 50% faster using the system's on-line tools than with a paper dictionary.</Paragraph>
  </Section>
class="xml-element"></Paper>